LangTrend

E-commerce search requires distinguishing products that are merely related to a query from those that directly satisfy the user's shopping intent. We augment query-product pairs with structured LLM-generated query and product attributes and human-validated relevance, explanations, and centrality judgments, and evaluate these signals using a simple dual-encoder retriever and MLP re-ranker. On an augmented subset of ESCI, a human-feature oracle reaches $0.9382$ nDCG@10, while a human-free trained $Q+P$ configuration reaches $0.9258$. Synthetic approximations of the human signals reach $0.9150$ overall but provide substantial gains for difficult, low-performing queries. Ablations show that most of the oracle improvement comes from post-edited explanations and annotator comments rather than the scalar centrality feature, suggesting that LLMs are most useful for exposing and approximating structured semantic supervision rather than replacing human judgment directly.

Post-training turns a general next-token predictor into a chat model with a persistent assistant persona. If that persona is a character the model plays only on its own turns, its preferences should govern what the assistant says, not what the model predicts other speakers will say. We test this boundary and find that it does not hold: a safety-relevant preference of the assistant—for harmless over harmful tasks—shapes the model's predictions even on the user's turn, where the assistant is not the one speaking. We find that this preference is small or near-zero in pretrained base models, that it emerges through post-training, replicated across open-weight model families, grows with scale, and can be moved by narrow finetuning that never touches user turns. We claim that this is evidence that post-training does not merely install a shallow assistant persona, but instead generalises beyond just the local assistant turn, into the model's representation of the user.

Expressive speech synthesis has advanced through prosody modeling, yet generating structured poetic speech, such as haiku, remains challenging. Prior work on prosody transfer improves expressiveness, and fine-tuned poetry TTS (text-to-speech) systems capture verse intonation. However, these models do not model haiku's 5-7-5 syllable structure or line-ending pauses. We present a cascaded system, HaikuS2S, combining ASR (automatic speech recognition), LLM (large language model)-generated haiku, and TTS fine-tuning on both prose and custom haiku datasets. Our evaluation focuses on emotion similarity, speech quality, and prosody alignment. In our experiments, we see that our prosody and tonal alignment improve significantly with our fine-tuned systems, particularly the one trained on both general poetry and haiku. We also see that we maintain similar emotion similarity scores across all systems.

We present a comprehensive benchmark of Federated Learning (FL) for multilingual Automatic Speech Recognition (ASR), evaluating four Speech-LLM architectures on the Multilingual LibriSpeech dataset. We compare FedAvg and FedProx across frozen and unfrozen encoder configurations, demonstrating that optimized learning rates are critical for performance. Specifically, independently tuning the learning rates for the speech encoder, connector, and decoder yields the lowest error rates, with full three-component adaptation (LoRA for encoder and decoder, full training for the connector) producing the best FL results. We observe that FedProx efficacy is architecture-dependent, providing notable advantages in multilingual pre-trained architectures (e.g., EuroLLM over TinyLlama when keeping the encoder fixed); this indicates that LLM backbone capacity plays a key role in mediating resilience to heterogeneous data distributions. These findings offer concrete design guidance for deploying multilingual Speech-LLMs in privacy-sensitive, distributed environments.

Disaster-risk-reduction archives describe hazard events in prose that databases such as EM-DAT (Delforge et al., 2025) cannot ingest directly. We present an LLM pipeline that generates candidate georeferenced event records using a controlled hazard vocabulary and fixed schema, retaining evidence for review. Applied to 10,000 documents from PreventionWeb, the knowledge hub managed by UNDRR, it produced 3,572 records from 1,913 documents across 24 hazard types and resolved 81% of location mentions to OpenStreetMap geometries. On 171 human-positive document windows from a stratified 217-document reference set, GPT-5 achieved 86.0% pooled attribute $F_1$, versus 44.2% for the spaCy-gazetteer baseline. Evaluation pools hazard families, location strings, and event years within documents, without assessing their assignment to individual events. GPT-5.4 ranked highest among ten LLMs (86.6% $F_1$). Verbatim evidence occurrence was 72.0% for GPT-5 and 47.2% for GPT-5.4, measuring textual traceability without establishing attribute support. We report production failure modes and automated label and location-rule compliance checks. Prompts, schema, and outputs will be released for adaptation to national reporting archives.

We reconsider the theory of probabilistic formal languages generated by n-gram models and by probabilistic context-free grammars (PCFGs). The expected hierarchy of probabilistic grammars is established by proving that every probabilistic language generated by an n-gram model is also generated by some PCFG, while some probabilistic languages generated by PCFGs cannot be generated by any $n$-gram model. We introduce the notion of fully connected PCFGs, namely PCFGs in Chomsky normal form where every production rule only involving non-terminals has non-zero probability. Our main result shows that any probabilistic language generated by an $n$-gram model differs from any probabilistic language generated by a fully connected PCFG. Therefore, the class of probabilistic languages generated by $n$-gram models is not a subset of the class generated by fully connected PCFGs.

Textual-gradient methods automate prompt optimization through natural-language feedback, but their iterative updates can be unstable. We identify two sources of this instability: noisy gradients produced from already-correct examples and over-specialization to hard cases that degrades performance on simpler inputs. We introduce STEVE, a stabilization framework with two coupled mechanisms. Error-Driven Refinement generates gradients only from incorrectly handled examples, concentrating updates on informative failures. Regularized Verification treats every update as provisional and accepts it only when improvement on hard cases does not cause unacceptable regression on a preservation set. Across ten reasoning benchmarks, three evaluator/optimizer models, and established prompt-optimization baselines, STEVE reduces degradation and produces more robust prompts. Additional evaluations with gpt-5.4-mini/gpt-5.4 on symbolic reasoning, GSM8K-Platinum, and DS-1000 show that these gains persist with newer models and larger test sets. STEVE therefore provides a practical way to improve the stability and effectiveness of textual-gradient prompt optimization.

Large language models have shown strong performance across a range of legal tasks, but existing benchmarks rarely evaluate the ability to take and defend a legal position, reason under incomplete information, or synthesize multiple statutory provisions. This gap is particularly pronounced for Canadian law, which remains underrepresented in legal NLP. We introduce GRACE (Grounded Reasoning Adversarial Canadian LEgal examples), a dataset of 1,915 question-reasoning-answer instances grounded in Canadian federal legislation. GRACE covers three reasoning modes: adversarial advocacy, uncertainty, and applied reasoning. We develop a pipeline that partitions raw statutory text, generates scenario-based questions and reasoning, and filters examples through model-free citation verification and LLM-based quality auditing. As a proof of concept, we fine-tune CLeAR-4B (Canadian Legal Adversarial Reasoning), a lightweight model for grounded legal reasoning, and evaluate it against the unmodified Qwen3-4B base model in open- and closed-book settings. CLeAR-4B substantially improves agreement with teacher outputs and statutory citation behavior when the relevant act text is provided, while its grounding degrades sharply when the statute is withheld. These results suggest that GRACE can support the development of lightweight legal models that reason more effectively from supplied statutory text.

Large reasoning models (LRMs) incur high inference costs, often mitigated by efficiency techniques like quantization and pruning. However, the impact of these techniques on model adversarial robustness remains largely unexplored. This study provides the first comprehensive analysis of the interplay between efficiency, jailbreak vulnerability, and reasoning in LRMs. We find that while efficiency methods seemingly reduce the success rate of jailbreak attacks, this improvement is often superficial. It largely arises from degraded reasoning capabilities leading to "attempted but failed" malicious responses, rather than an increase in genuine alignment. Mechanistic analysis of representational drift confirms this, revealing a strict coupling between reasoning capability loss and the model's inability to maintain malicious semantic trajectories. Additionally, we identify quantization with pruning as the optimal strategy to balance efficiency and robustness. These findings clarify the distinction between true safety alignment and capability-induced failure, providing an empirical foundation for LRM deployment.

Non-verbal vocalizations (NVVs) carry important paralinguistic information but are often omitted by conventional automatic speech recognition (ASR) systems. The ISCSLP NVVSpeech Challenge requires joint transcription of lexical content and 16 NVV categories under limited and highly imbalanced supervision. We present a data-centric NVV-aware ASR pipeline based on cross-dataset label harmonization and a two-stage sampling schedule. We map heterogeneous source labels to the official taxonomy and exclude samples without a reliable mapping. Our schedule first uses square-root category sampling to moderate the long-tailed distribution and then applies uniform-category fine-tuning. On a fixed local validation split, square-root category sampling performs best among the tested single-stage settings. The final two-stage system obtains an official score of 63.86 and ranks fourth in Track 1.

Sink-aware deployment may identify important first-token attention heads before a model is quantized, then reuse that map at the edge. We test when this shortcut is safe for 4-bit NF4 weight-only post-training quantization (PTQ). Our Sink Topology Consistency (STC) metrics separate global rank preservation, top-$k$ set overlap, and layerwise sink-mass shift, and distinguish per-input sensitivity from calibration-map transfer. Across Qwen2.5-0.5B, Qwen2.5-1.5B, and Llama-3.2-1B, global bf16-to-4-bit ranks remain high at 4,096 tokens ($ρ_s \geq 0.980$), yet top-$k$ Jaccard overlap is only 0.619-0.793, corresponding to 76.5-88.5% membership retention. The global statistic also masks local failures: terminal Qwen layers shift by 6.2-7.9x their model means, whereas Llama-3.2-1B shows low, nearly uniform drift. Under a C4-to-LongBench shift, cross-domain overlap degrades more than the within-domain precision comparison for both Qwen models, but not for Llama-3.2-1B. Matched-domain 4-bit recalibration reaches 90% of a split-half stability plateau at the smallest tested $n=8$ for both Qwen models and $n=32$ for Llama-3.2-1B, though not as a sharp threshold; for the two Qwen models, updating only selected layers does not reach the full-map stability criterion. On Jetson Orin NX, the 16-sample workload takes seconds for the two models with valid on-device sink measurements. The practical message is precise: global rankings often transfer, but discrete head sets, layer-local policies, and cross-domain calibration should be revalidated after quantization.

Large language models (LLMs) drift out of date the moment their pretraining ends, yet retraining from scratch is prohibitively expensive. Continued pretraining (CPT) is the natural remedy, but it is typically evaluated through a continual learning lens that assumes disjoint data streams. This is a poor fit for time-incremental updates on web-scale crawls, where successive snapshots share substantial URL overlap by design. We study time-incremental CPT in this realistic regime: continued pretraining on FineWeb-Edu dumps drawn strictly from after each model's knowledge cutoff, evaluated across six open-weight models spanning three families (OLMo2, Llama-3.1/3.2, Gemma-3-1B) and four parameter scales (1B-3B-7B-8B). We organize our findings around four practical questions. (i) Is knowledge acquired? Yes, but heterogeneously, and without catastrophic forgetting: five of six models also improve on pre-cutoff factual recall, and the gains track pretraining saturation (driven primarily by token budget per parameter). (ii) What does it cost? Almost nothing: the macro-average across a thirteen-task suite stays within 0.01 of the base for every model. (iii) What is the recipe? Data quality dominates quantity (a curated 6B-token slice matches a broader 40B one); the optima for knowledge acquisition and general capability are separated by roughly an order of magnitude in learning rate; and LoRA at sufficient rank matches full CPT. (iv) Does it survive deployment? CPT gains transfer through SFT, while DPO's effect is family-dependent. Together, these results paint a more optimistic picture of time-incremental CPT than the prior continual learning literature suggests.

Synthetic essays can help reduce dependence on human-written data in Automated Essay Scoring (AES). However, they often lack realistic errors, limiting their ability to represent authentic human writing, particularly when the target texts are intended to resemble those produced by language learners. In this study, we present a simple approach that introduces error supervision into synthetic essay generation. Specifically, we fine-tune an LLM generator on error-annotated texts of the kind commonly used in Grammatical Error Detection (GED). To assess the utility of the proposed approach, we fine-tune and evaluate AES scorers under three data conditions: authentic essays, synthetic essays generated conventionally, and synthetic essays generated using our proposed approach. The results show that in the larger-data settings, the proposed approach outperforms the conventional synthetic baseline in 11 out of 12 dataset-metric comparisons, with performance in some cases approaching that of models trained on authentic essays. Despite these gains, performance under extremely low-resource settings remains mixed, with advantages over the conventional baseline only becoming more apparent at 200 training essays, although not consistently across datasets. Qualitative and quantitative analyses further show that the proposed approach produces learner-like errors whose distributions broadly resemble those observed in authentic essays.

Long-horizon conversational memory is especially challenging in multi-actor settings, where relevant evidence is distributed across participants and contexts and previously established information may later be revised. We introduce EGMEMORY, which formulates long-horizon multi-actor memory as a searchable state machine that separates persistent message-level evidence from an explicit active state. At write time, adaptive state resolution and an evidence-grounded propose-verify-commit protocol govern how this state evolves. At read time, adaptive evidence navigation iteratively resolves the state and supporting evidence required for a query, using conversational structure to narrow the search space and lexical-semantic relevance to rank candidates. The system operates through prompting and tool use without memory-specific policy training. EGMEMORY achieves 68.2% on GroupMemBench and 77.9% on EverMemBench, outperforming the strongest evaluated baselines by 22.7 and 21.4 percentage points, respectively. It further reaches 73.6% on the dyadic LoCoMo benchmark, demonstrating generalization beyond multi-actor conversations. We will release the codebase upon formal publication.

Temporal Information Retrieval (TIR) has been increasingly critical given the rise of Retrieval-Augmented Generation (RAG). Since temporally mismatched evidence can be highly misleading, TIR aims to retrieve documents that are both semantically and temporally relevant to a query. Two TIR paradigms have emerged - temporal retrievers and temporal re-rankers - differing in how temporal relevance is modeled. While these paradigms provide complementary strengths, our analysis reveals that each alone falls short of robust TIR: temporal retrievers provide flexible query understanding via learned representations, but often fail to explicitly account for temporal constraints; temporal re-rankers can enforce such constraints more explicitly, but often rely on predefined re-ranking rules. To address this, we propose Q-TIE, a re-ranking framework based on learned Temporal Intent Extraction (TIE). By introducing a TIE model that maps each query's temporal constraint into a unified interval representation (i.e., $\langle t_{start}, t_{end} \rangle$), Q-TIE generalizes beyond predefined rules via model-based learning while explicitly modeling temporal constraints as a separate signal - jointly achieving what each paradigm typically trades off. Experiments demonstrate that Q-TIE consistently outperforms existing TIR methods with stronger generalizability across temporal query types, and provides a lightweight yet effective add-on for temporally-aware RAG pipelines. Code: https://github.com/ssoy0701/Q-TIE.

Large language model (LLM) agents increasingly execute multi-step workflows through tool use and interaction with users and environments. However, current agent evaluations are largely English-centric, limiting our understanding of agent capabilities in multilingual settings. We introduce BabelFlow, a benchmark-general agentic workflow that adapts existing agent benchmarks to new languages by analyzing runtime dependencies, coordinating structure-preserving translation, and combining multi-layer verification with human review to preserve task and evaluation semantics. Using BabelFlow, we construct BabelArena, a task-aligned benchmark comprising 16,146 instances derived from 702 canonical tasks across four benchmark families, 13 domains, and 23 languages. Experiments with five frontier models show that no single model dominates across benchmark families and that cross-language disparities extend well beyond task success. Lower-resource languages exhibit distinct failure patterns, with larger shares of tool-use and control-flow errors rather than answer-quality errors alone, pointing to gaps in reliable task execution across the resource levels of these languages. On the same tasks, agents in low-resource languages also consume substantially more tokens than in English (up to roughly twice the input) without proportional increases in interaction length, and language consistency degrades further on tasks requiring structured output, where switches are directed overwhelmingly toward English. We believe BabelArena provides a foundation for advancing research on reliable and efficient multilingual agents.

Financial language models can transform unstructured firm-specific news into structured decision signals, but financial AI research lacks an integrated deployment framework for evaluating whether those signals remain useful in financial decision systems. Computer science research has developed strong methods for time-series forecasting, text classification, multimodal stock prediction, graph-based market modeling, and machine-learning operations, yet these streams do not provide a domain-specific protocol that jointly tests financial language-model outputs under event-time observability, probability calibration, execution timing, transaction costs, liquidity constraints, capacity limits, operational diagnostics, and statistical inference. We introduce MFAST, a Market-Friction-Aware Sentiment-to-Trading framework that converts timestamped financial text into auditable, reproducible, and market-feasible trading decisions. The application is news-based trading, where firm-specific text must be linked to securities before portfolio decisions can be evaluated. The framework links Refinitiv News Analytics to Center for Research in Security Prices (CRSP) equity data, restricts the primary out-of-sample evaluation to post-release news outside disclosed foundation-model data-freshness periods, and adds a public replication arm using open financial text and public price data. Results show that decoder-only language models outperform encoder baselines and dictionary sentiment in classification, calibration, return prediction, and net portfolio performance, while operational diagnostics reveal trade-offs among accuracy, latency, memory, throughput, and inference cost. The paper shows that credible evaluation of financial language models requires an end-to-end engineering approach combining language understanding, temporal discipline, market-friction-aware deployment, and reproducible validation.

Acoustic representation learning is crucial for speech processing, yet low-resource languages (LRLs) face severe data scarcity, limiting the effectiveness of traditional and self-supervised methods. As a promising alternative, in this work, we propose to enhance acoustic representation trough a cross-modal transfer knowledge approach, based on heterogeneous graph neural networks (HGNNs), where acoustic and linguistic entities are modeled as distinct node types within a unified graph. Through message-passing mechanisms, linguistic nodes explicitly transfer knowledge to acoustic nodes, enabling structured and interpretable cross-modal information flow. To highlight this knowledge transfer and its benefits, we measured standard clustering metrics as an intrinsic evaluation of acoustic representation, and to emphasize applicability, we performed isolated-word recognition tasks using an English benchmark and a Cameroonian language dataset in low resources settings . Results demonstrate that acoustic representations consistently benefit from linguistic knowledge propagated through the graph. To our knowledge, this is the first demonstration of explicit cross-modal knowledge transfer for acoustic representation learning using HGNNs, highlighting a promising direction for speech representation in low-resource settings.

Federated learning lets multiple parties train a shared model without pooling their data, but a client with far less data than the others can end up poorly served even when the group's average accuracy looks fine. We ask whether the choice of pretrained backbone affects this under LoRA fine-tuning, and whether per-word perplexity on the target text predicts which backbone helps the worst-off client before federated training starts. We ran 313 experiments across three text-classification datasets and three similarly sized backbones (RoBERTa, BERTweet, PubMedBERT), each compared against a task-specific baseline on identical data splits. Lower-perplexity backbones consistently produced larger gains for the worst-performing client, with a rank correlation of -0.87 across nine dataset-backbone pairs; a backbone held out of the analysis confirmed the pattern. Personalization with Ditto recovered only 4-12% of the gap between training alone and full federation, and removing aggregation entirely erased the benefit. A client's update also showed no sign of conflicting with the group's update; the two are close to orthogonal, ruling out one proposed explanation for this failure. Practically: measure perplexity on a sample of task text before choosing a backbone, and do not rely on personalization to protect a data-poor client. We release our code, predictions, and full results for others to test.

The visual modality, i.e., images, plays a key role in multi-modal entity alignment (MMEA). Existing approaches often directly fuse the image with other modalities to align different entities. Although simple, such strategies overlook the potential noise in the images and their semantic misalignment with corresponding entities, resulting in suboptimal fusion and degraded performance. Addressing this, we propose a novel Reliability-Aware framework for MMEA (RA-MMEA), which assesses visual reliability and adaptively improves unreliable visual representations for robust entity alignment. The core lies in two modules, including dependency-aware visual reliability prediction (DA-VRP) and stability-regularized visual embedding generation (SR-VEG). The former aims to estimate the reliability of an image by leveraging multi-modal dependency within the entity, while the latter focuses on producing alternative visual representation conditioned on semantics encoded in textual modalities for multi-modal fusion. Compared to current methods, RA-MMEA enables more reliable visual representations for modality fusion, thereby improving performance. In extensive experiments, RA-MMEA achieves state-of-the-art results, verifying the importance of reliable visual modality for entity alignment and the effectiveness of RA-MMEA. The code and results will be released.

Long-context language models interface with external knowledge through raw natural language. In retrieval-augmented systems, this creates a persistent index-payload schism: dense vectors enable searchable routing, but models must re-ingest lengthy text payloads for reasoning at O(N^2) attention cost. Existing compression methods further produce private states tied to specific architectures. We introduce Machine-Interpretable Information (MII), the first agent-to-agent (A2A) document-to-state protocol. A dual-timescale state-space Writer compiles documents into a canonical, fixed-bandwidth state (56 tokens), and a lightweight Translator maps it into any frozen Reader's embedding space, reducing query-time cost to O(K). The resulting .mii artifact unifies Retrieval (searchable geometry), Reasoning (global memory), and Reconstruction (grounded details) in a single transferable medium. We demonstrate strong cross-model interoperability across heterogeneous LLMs (e.g., Llama, Qwen, Mistral) – despite the Writer using a legacy GPT-2 vocabulary, forcing genuine semantic translation rather than token-level memorization. Mechanistic probes reveal modular latent structure: entity representations can be causally traced and zero-shot transplanted between unrelated document states while remaining decodable. To address lexical reconstruction under fixed bandwidth, we propose Residual-MII, a cache hierarchy combining compiled global memory with sparse local evidence. On HotpotQA (7,405 queries), Residual-MII exceeds full-context Exact Match at approximately 7% of the attention FLOPs, suggesting a paradigm shift toward compiled, transferable neural document formats.

FORM is a domain-specific symbolic manipulation language widely used in particle physics for processing the very large algebraic expressions arising from multi-loop Feynman diagram calculations. Despite its central role in precision theoretical physics, no artificial-intelligence tooling exists, to our knowledge, for assisting physicists in writing FORM code. We show that contemporary large language models (LLMs), including frontier models with hundreds of billions of parameters, achieve a zero-percent execution pass rate on our instruction-following and tutorial-style FORM tasks without documentation in a single attempt, establishing FORM as a genuine zero-shot language for LLMs at the time of writing. We then present a verification-driven data generation pipeline that uses the FORM binary itself as an execution oracle to produce and validate a corpus of 4,633 training examples spanning deterministic computations, open-ended programs, tutorial code, and knowledge question-answer pairs. Fine-tuning a compact open-weights model (Qwen3-8B) with quantized low-rank adaptation (QLoRA) yields a specialist that, evaluated on four complementary benchmarks (840 tasks, single attempt each), decisively outperforms frontier models with up to 756B parameters in execution rate and in strict, FORM-verified output matching on the larger benchmarks, and remains statistically indistinguishable from them on the smaller, harder ones. General reasoning and coding capabilities are preserved within 2.6 percentage points.

Long-form audio performance is often summarized by context length and aggregate accuracy, obscuring how language, evidence, and task jointly shape difficulty. We introduce MuLA-Bench: 5,038 open-ended questions over 1,769 in-the-wild recordings totaling 1,377.9 hours, covering 16 languages and eight domains. A balanced Language x Domain semantic track supports controlled comparisons, while a complementary acoustic track preserves naturally occurring non-speech evidence. Evidence-grounded generation, shortcut checks, and language-expert review provide auditable questions without translating a shared source set or injecting target sounds. We evaluate ten audio-language models and conduct pooled diagnostics on a fixed eight-model cohort. Language rankings change across domains and tasks; acoustic-semantic performance gaps vary with the requested operation; and temporal errors can persist after the correct event is identified. Long-range retrieval is comparatively strong, while precise clock alignment and factual grounding of natural acoustic events remain fragile. MuLA-Bench thus exposes conditional failure patterns that a single long-context score does not capture.

While test-time scaling enhances Large Language Model (LLM) agents in long-horizon software engineering (SWE), sparse binary rewards (Pass/Fail) create a severe credit assignment crisis and waste failed exploratory trajectories. Current trajectory optimization and scaling methods are costly and structurally limited, relying on heuristic state reuse without causal diagnosis or delayed scalar scoring without actionable online guidance. We propose FLARE (Full-Lifecycle Alignment and Reward Engine), a novel dense supervision paradigm driven by a lightweight Generative Reward Model (GRM). First, RADAR, an offline causal-aware diagnostic framework, extracts high-fidelity, hindsight-free supervision through causal-chain backtracking to distill a GRM providing real-time, step-level risk feedback. Second, FLARE uses this GRM to continuously optimize the agent across its entire lifecycle. During inference, FLARE acts as an Active Scaffold, autonomously intercepting high-risk generation steps for localized breakpoint re-execution, drastically reducing compute overhead. During post-training, the GRM's structured signals serve as process-supervised reranking scores for Supervised Fine-Tuning (SFT) and step-level dense rewards for Reinforcement Learning (RL), mitigating policy collapse in sparse environments. Extensive evaluations show that FLARE establishes a new Pareto frontier across the agent lifecycle: FLARE (N=1) outperforms Global Rollout (N=5) with a 5x reduction in token consumption. Extending FLARE to training overcomes the sparse reward problem in long-horizon interactive tasks, delivering relative performance gains of 19.13% in SFT through process-aware data curation and a consistent 9.19% improvement in RL.

A RAG system can hand you the right answer and cite a source it did not actually use. Models output these unfaithful citations via post-rationalization: they write the answer first and then attach a citation to whatever passage looks close enough. Search agents are now trained with reinforcement learning from verifiable rewards (RLVR), which pays them for getting the answer right. We asked whether that training also teaches them to cite honestly. Improving an existing methodology with a required control, we compared an instruction-tuned model against three RLVR agents trained from it, on four question-answering datasets, using only free-tier Kaggle GPUs. Post-rationalization is everywhere: on Wikipedia-based questions roughly one citation in seven is unfaithful. RLVR does not fix it. The agents post-rationalize at their base model's rate, and one lands slightly worse. Rewarding correct answers buys nothing in citation faithfulness, so faithfulness has to be trained and measured on its own terms.

Time series forecasting underpins critical decision-making across diverse domains. While large language models (LLMs) offer promising reasoning capabilities, existing LLM-based time series forecasting approaches either reduce them to numerical predictors that bypass their strengths, or allow direct forecast generation that destabilizes predictions in non-stationary settings. We introduce CTRL, a framework that decouples semantic reasoning from quantitative prediction. A frozen backbone generates base forecasts, while specialized LLM agents function as controllers that analyze backbone prediction errors through decomposed trend, seasonal, and irregular components, grounding reasoning in interpretable temporal structure. Each agent outputs compact control signals that a lightweight residual decoder translates into forecast corrections. CTRL incorporates label-free test-time adaptation that detects distribution shift from input statistics alone and readapts control signals with only 3-24 LLM calls via caching. CTRL is explicitly designed to improve robustness under non-stationary temporal dynamics and distribution shift, while remaining competitive on highly stationary time series where adaptive correction provides limited additional benefit.

Standard positional encodings represent position as a one-dimensional reading-order coordinate, but reading order alone does not determine hierarchical textual structure. We use a hierarchical rotary positional encoding (hRoPE) that represents paragraph, sentence, and token indices as separate channels, hold the token sequence fixed, intervene on the paragraph coordinate p1, and measure cross-paragraph attention with a token-distance-exact estimator. Attention is compressed relative to a token-distance-matched baseline in every corpus, but compression alone is not diagnostic of true structure: an architecturally identical channel with density-matched random labels is compressed too, more shallowly. What distinguishes real structure is the depth of compression, which is greater and corpus-dependent while the control's is not. Comparing eight corpus-only quantities across three constructs (lexical persistence, paragraph length, embedding-based coherence), none fully reproduces the cross-corpus ordering of depth, though embedding-based coherence comes closest. Compression depth, not its location, is the reproducible signature of genuine paragraph structure in our setting.

Dialectal Arabic machine translation (MT) remains challenging despite recent progress in Arabic language technologies, particularly because effective translation requires modeling not only semantic content but also dialectal variation, conversational context, speaker and addressee characteristics, and sociolinguistic appropriateness. Moreover, conventional MT metrics provide limited insight into the linguistic errors produced by dialectal systems. We present the AlexandriaX 2026 Shared Task on Dialectal Arabic MT, which addresses these challenges through three complementary subtasks: (1) context-aware English-to-Dialectal Arabic dialogue translation across 13 Arabic varieties, (2) cross-dialect Arabic translation in the financial domain covering six Arabic dialects, and (3) span-level MT error detection and classification using linguistically motivated error categories across five Arabic varieties. The shared task attracted 38 registrations for Subtask 1, 33 for Subtask 2, and 35 for Subtask 3. Twelve unique teams submitted their system description papers, all of which we accepted for publication. The best system on Subtask 1 achieved 30.42 spBLEU in the constrained track and 33.49 spBLEU in the unconstrained track. On Subtask 2, the top system obtained 28.40 spBLEU. On Subtask 3, the best system achieved an overall score of 49.82, outperforming the 24.29 baseline. Taken together, the results of the leading systems across the three subtasks highlight the benefits of explicit dialect modeling, context-aware generation, retrieval and reranking, and specialized approaches to interpretable MT error analysis. All the resources of AlexandriaX 2026 shared task are publicly available, including data, baselines, and evaluation code on our project page: https://alexandriax.dlnlp.ai.

Long-context models read a novel the way a person reads a printout: one token after another, in narrative order, with the whole history competing for a fixed budget of attention. A detective does not work that way. They sort what happened when, and they keep a map of who relates to whom, so a clue from chapter one can meet a question asked at the end of the book. We test whether a frozen knowledge graph can give a small local model that same freedom. Thirty detective novels and 234 multiple-choice questions are answered by one fixed qwen3.5:9b reader under nine conditions: five graph routes, a recent-window baseline, whole-book compression, ordinary vector retrieval, and a question-only control. The strongest graph route reaches 53.85% (126/234) against 46.15% for the recent window, 51.28% for compression, 51.71% for vector retrieval and 40.17% for question-only. On the subset that no model can answer without the book, the graph route reaches 42.86%. None of the fifteen graph-baseline contrasts survives Holm correction, so we present the result as exploratory evidence about a design. Two structural findings survive scrutiny better than the headline number: annotated evidence concentrates in the topological core of these graphs (2.35x enrichment, pooled), and the two graph-building pipelines differ so much in annotation coverage (16% versus 73% of clue paragraphs) that pooled accuracy alone would hide which bottleneck is being measured.

Educational foundation models must solve problems, understand curriculum structure, diagnose learner difficulties, and provide appropriate instructional support. Existing educational language models often focus on either problem solving or tutoring, with training mixtures organized by source or task rather than capability. We present OmniEdu, an open family of foundation models for K-12 learning and teaching. Its instruction-tuning corpus combines over 100 educational resources and general instruction sources, organized around four capabilities: subject competence, curriculum grounding, diagnostic reasoning, and pedagogical action and scaffolding. Our pipeline integrates deterministic cleaning, semantic auditing and rewriting, task-specific quality scoring, token-budgeted diversity selection, and pedagogical instruction assignment. It yields 69,999 examples and 15.96M supervised response tokens, including 60,951 education-specific examples. We fine-tune 4B, 9B, and 27B models and evaluate curriculum grounding, K-12 problem solving, and pedagogical tutoring, alongside general capability. Education-oriented tuning consistently improves all three educational benchmark groups across model scales. OmniEdu-27B achieves 63.12% EM and 76.69% F1 on K12-Bench, 85.89% on MathFish, 86.95% on EDUMATH, and 78.74% in MathTutorBench's Scaffold setting. It also achieves the highest Teaching average on LongTutor among the evaluated models, at 3.02. These results demonstrate the value of curated, capability-balanced supervision for adapting general language models to educational tasks spanning problem solving, curriculum understanding, and instructional support.

An automated fact-checking system returns a label: the claim is true, or it is false. In many such systems the verdict remains the primary output. What is generally missing is a record of which document settled the question, of what would have had to be different for the verdict to change, or of whether the same claim, reworded, would have been judged the same way. We call the missing piece a warrant: a separate statement of what was guaranteed and on what grounds. Formal methods produce evidence of this kind, and regulation is beginning to ask for it, since the Digital Services Act and the AI Act both call for auditable evidence about how systems behave. Surveys of automated fact-checking are usually organised by pipeline stage, and treat logic as one technique among many. We organise the field by what is being formalised instead, which gives five levels: the claim, the reasoning, the system doing the checking, the ecosystem the claim spreads through, and the regulatory obligation. Sorting 121 works into those levels, two patterns stand out. Most of the relevant formal machinery already exists, but it was built for other domains and has rarely been applied here, and the gap is widest for verifying the checking system itself. Several stages of the routine professional fact-checkers follow also have no stated correctness criterion, and two of them, writing a claim in checkable form and correcting a verdict already published, are not formally specified in any work we coded. We close with open problems, each with a suggested first step.

Speech-text space alignment is a multimodal representation learning method consisting to map different speech and text into a shared representation space, leading to enrichment of the representation of each modality. Proposed architectures, such as SAMU-XLSR, typically follow a student/teacher framework, with the goal of fine-tuning an audio encoder to produce representations that closely match those of the text. In this way a speech representation is semantically enriched. However, such systems generally require large amounts of training data and considerable computational resource, making them difficult to apply to low resources languages under frugal constraints. The present work proposes a data-efficient space alignment method based on Heterogeneous Graph Neural Networks and link prediction. The core idea is to leverage message passing to explicitly transfer information from the text modality to the speech modality, thereby reducing the need for large training datasets and intrinsically enriching the acoustic representations, all in a more interpretable manner. Although thoroughly explored for high-resource languages, word-level tasks in speech remain relevant for certain low-resource languages. Therefore, we conducted experiments on speech-text alignment at the word level using the TIMIT (English) dataset and Yemba (a Cameroonian language). Our approach yields results comparable to those of SAMU-XLSR, a state-of-the-art method, and even surpasses it for the Yemba language in the task of word retrieval, while using far fewer resources, demonstrating its power, frugality, and efficiency.

Language models are appealing tools for research on the past. But to trust the evidence a model provides, researchers need to know whether its responses fit the period represented. Validation is challenging, because this is not a task living people ordinarily perform, and because many questions have multiple correct answers. We use historical texts to develop a benchmark for a model's representation of English-language contexts 1831-1930, relying on pairwise comparisons to multiple ground truths and strong distractors to score the hardest questions in an appropriately graduated way. We find that generative tasks are harder than discriminative ones; in fact, reasoning models can typically discern the weakness of their own generated answers. While models pretrained exclusively on historical text lead the pack when evaluated by answer likelihood, they cannot compete with commercial models in free generation. None of the models we tested represent historical contexts in a fully persuasive way yet, but progress toward that goal is evident.

Evaluating user experience (UX) on live websites through user testing is expensive, subjective, and difficult to scale. LLM agents offer a promising route to automating UX testing by simulating realistic user behaviour. However, existing simulation approaches typically lack multimodality and require time-consuming manual review to extract actionable insights. We formalise UX improvement recommendation from simulation data as a structured natural language generation and ranking problem, and establish an evaluation protocol using expert annotation and LLM-as-a-Judge. We present AMUSER, a multimodal framework which simulates user behaviour and automatically generates prioritised UX improvement recommendations from resulting data. We evaluate AMUSER on commercial websites and show that its recommendations substantially outperform those from text-only simulation (NDCG@3 = 0.758 versus 0.359) at an 89% lower simulation cost. Our results suggest an asymmetric role of multimodality: visual access during simulation improves recommendations through richer traces, while providing visual inputs during recommendation generation can modestly degrade quality. We also discuss practical deployment lessons from applying AMUSER to commercial websites.

Non-technical stakeholders frequently cannot write the SQL needed to extract insights from operational databases. We built and evaluated a Text-to-SQL agent that closes this gap end to end: a six-node LangGraph StateGraph checks question relevance, fetches the live schema, generates PostgreSQL, validates it with a dry run, retries on failure, executes the query, and narrates the result set in plain English. The agent uses prompt engineering only; no model was fine-tuned. We evaluated it on the Chicago Crime dataset (approximately 8.5 million records, 22 attributes) against a hand-built benchmark of 100 natural language questions with ground-truth SQL, stratified into 30 Easy, 40 Medium and 30 Hard items. Comparing two prompt revisions of the same agent, the revised system (V2) reached a Valid SQL Rate of 93% (from 87%), an Execution Accuracy of 60% under a hybrid relational equivalence metric (from 47%; 19% from 12% under strict JSON matching), and a mean Synthesis Quality of 4.34 out of 5 (from 3.91). The single largest driver was removing a LIMIT 10 instruction from the system prompt, which had been truncating multi-row answers. Error analysis attributes the residual failures to relevance-checker false rejections, ambiguous question semantics, and free-tier API rate limits rather than to the language generation step. We report no comparison against an external baseline system or a public benchmark; the study is a single-model engineering evaluation.

Multilingual code language models transfer skills across programming languages (PLs), but whether any PL occupies a privileged pivot position remains contested: geometric analyses point to C-family languages and Go, while behavioral evidence highlights Python. We revisit this question under controls for representational anisotropy and length variation across PLs, two confounds that compromise prior cosine-based analyses. Across three code models on multilingual competitive-programming data, we study three views of cross-PL organization: pairwise PL geometry, PL-English alignment, and pivoted retrieval through candidate PL representation spaces. The results are relation-dependent. Code-code geometry reveals structured language regions but no universal center; code-English alignment favors high-level scripting languages; and pivoted retrieval favors different intermediate spaces for code-to-code and English-to-code transfer. These findings suggest that Python's special role is better understood as English-facing affinity than as universal geometric centrality.

The distinction between the spirit and letter of the law is a central issue across research and everyday life, and a growing concern for building safe, intelligent machines. What is this distinction based on, and how can we develop machines that follow the intention behind a rule? We used targeted adaptation that made large language models prioritize the spirit or letter of the law. With minimal modifications, our method significantly changed LLM behavior across diverse measures, novel vignettes, real-world scenarios, and influential legal cases. An analysis of model internals revealed a low-dimensional space with three interpretable dimensions matching a formal pre-specified framework for the geometry of legal concepts. These findings show how legal thought in LLMs may be organized and directed.

Cross-lingual information retrieval (CLIR) is increasingly important in multi-national industries, where critical technical evidence may exist in a different language than the query. However, existing benchmarks do not adequately capture domain-specific cross-lingual retrieval or the retrieval-depth and recoverability failures that aggregate recall hides. In this work, we benchmark CLIR in the chemical domain, with a focus on patent data. We construct a multilingual dataset from Google Patents and the European Patent Office (EPO) data, spanning five languages (covering major Eastern and Western languages) and reflecting the diversity and complexity of real-world industrial documentation. Using this dataset, we systematically evaluate eight state-of-the-art embedding models for cross-lingual retrieval. Our results show a substantial performance gap between monolingual and cross-lingual settings: for the best-performing model, Recall@10 drops from 0.72 to 0.53 in cross-lingual setting. Retrieval depth also degrades significantly, with relevant documents ranked lower across languages in cross-lingual scenarios. Furthermore, some multilingual embedding models that perform strongly in monolingual settings exhibit sharp declines when queries and documents are in different languages, providing practical insights for model selection in cross-lingual use cases. These findings highlight critical limitations of current approaches and emphasize the need for more robust cross-lingual retrieval methods in domain-specific settings. Our benchmark provides actionable insights for model selection and establishes a controlled diagnostic evaluation framework for CLIR over industrial technical text. Data and code are publicly available at https://github.com/MohammadKhodadad/Multi-Lingual-QAC.

The IEEE BigData Cup benchmark combines three prediction problems with different output structures: ordinal suicide-risk classification, multi-label psychosocial factor detection, and extraction of supporting phrases. We introduce heterogeneous readout decomposition (HRD), which separates semantic verification from output realization. A locally deployed Qwen3.8-27B model, adapted with task-specific QLoRA adapters, produces both answer-token margins and layer-63 answer states for card-conditioned queries. HRD compares four latent scores for ordinal risk, retains token margins for most factors while routing seven labels through one shared latent probe, and constructs evidence sets from verbatim span candidates with calibrated, risk-conditional constraints. On two held-out user-grouped confirmation folds, the latent risk readout improves weighted F1 from 0.8237 to 0.8372 and macro F1 from 0.7965 to 0.8185. Selective factor routing improves macro F1 by 0.0105 and tail-label macro F1 by 0.0189; in contrast, global latent replacement and independent label-specific probes fail. The constrained evidence decoder raises pooled phrase F1 from 0.7488 to 0.7609 in row-level out-of-fold evaluation. Lenormand's best public result is 0.8052 on Subtask 1 and 0.6636 on Subtask 2, giving a composite score of 0.7627. These results identify the answer-token boundary, rather than semantic representation alone, as a measurable source of error in this benchmark.

Online discourse about urban issues - walkability, cycling infrastructure, public transit, housing density, and street safety - is voluminous but unstructured, and existing city-evaluation tools capture none of it. We present a pipeline and conversational system that combines geographic entity resolution, topic modeling, sentiment analysis, and Retrieval-Augmented Generation (RAG) over 22,788 chunks of YouTube transcripts and comments spanning 309 North American cities. Beyond the system itself, our contribution is a set of measurements about what happens when standard NLP components meet short, informal, geographically ambiguous text. A Twitter-tuned RoBERTa classifier outperforms a VADER lexicon baseline by 12.6 macro-F1 points (0.589 vs. 0.464; McNemar p = 0.0001), but both models collapse on the neutral class, which dominates urbanist comment traffic; annotators disagree on the same class (Cohen's kappa = 0.53). Dense retrieval beats a TF-IDF baseline at every cutoff (P@5 0.790 vs. 0.560), and video-level relevance proxies understate chunk-level precision by a wide margin (0.660 vs. 0.94 under human rating). For groundedness evaluation, we find BERTScore unusable when a multi-sentence generated summary is compared against a single short comment - scores are nearly flat regardless of relevance - and show that ROUGE-1-based groundedness is a paraphrase-driven lower bound rather than a hallucination rate. These findings generalize beyond the urbanist domain to any RAG system built over short user-generated documents.

Bangla, the seventh most spoken language in the world, exhibits significant regional dialectal diversity, with dialects such as Barishal, Chattogram, Sylhet, Noakhali, and Mymensingh differing in lexical, morphological, and syntactic characteristics. These variations pose substantial challenges for Named Entity Recognition (NER), limiting the generalization of models trained on Standard Bangla or a single regional dialect. This paper presents a cross-dialect Bangla NER framework using the publicly available ANCHOLIK-NER dataset, comprising 17,405 annotated sentences and 101,817 tokens across five major Bangla regional dialects. A Leave-One-Dialect-Out Cross-Validation (LODOCV) strategy is adopted, training models on four dialects and evaluating on the remaining unseen dialect. Eight pretrained transformer-based models, including BanglaBERT, MuRIL, XLM-RoBERTa, and Multilingual-E5, are evaluated under identical experimental settings. Multilingual-E5 Large achieves the highest F1-score in every fold, peaking at 97.26% on Mymensingh and reaching its lowest, 82.38%, on Chattogram, the most challenging target dialect. To improve interpretability, Local Interpretable Model-agnostic Explanations (LIME) are applied to word-level predictions, revealing that the model's decisions are driven primarily by the surface form of the target entity word itself rather than by surrounding sentence context. These findings establish a benchmark for cross-dialect Bangla NER and demonstrate the effectiveness of transformer-based transfer learning for low-resource regional dialects, while highlighting the surface-form dependence of current models as a direction for future work.

Large language models (LLMs) increasingly mediate human decisions and communication, yet their behavioural regularities remain difficult to characterize systematically. We develop a cross-linguistic psychometric profiling framework and evaluate nine LLMs using seven psychological instruments, with five repeated administrations per model and language in Chinese and English. Items unresolved after a prespecified retry procedure are retained as NA. Joint analysis of scored and NA responses captures response tendencies and boundaries of self-report applicability. LLMs exhibit structured, model-specific profiles despite a shared alignment-shaped pattern of higher prosocial and self-regulatory responses and lower dominance, disengagement and harmful-intent endorsement. NA responses are structured rather than uniformly distributed, indicating where outputs are treated as inapplicable, refused or cannot be mapped to valid response options. Language condition and provider origin are associated with profile configuration and answerability, whereas repeated administrations show high reproducibility and permit recovery of model identity. Human-reference and prompt-robustness analyses further indicate that these signatures are context dependent. Joint analysis of psychometric profiling and answerability offers a framework for quantifying deployment-level behavioural signatures.

LLM-based machine translation evaluation can closely match human judgments, but in practice it remains largely diagnostic, with the signals rarely translating into direct quality improvements under real production constraints. We propose a two-stage, evaluator-guided automatic post-editing framework that turns MQM-style evaluation into targeted repairs: a retrieval-augmented LLM evaluator outputs structured, span-level MQM diagnoses under an explicit edit contract, and a separate LLM post-editor applies minimal edits restricted to those diagnoses. This separation improves controllability and reduces paraphrastic drift compared to one-stage "judge-and-refine" baselines. In a systematic study involving seven LLMs spanning three model providers and seven languages, our best configuration consistently improves both COMET-22 and COMETKiwi scores over one-stage post-edit methods, while the evaluator's error spans and severities show strong agreement with human MQM annotations and human editor preferences.

Triage in the emergency department (ED) is a sequential decision process that unfolds turn by turn. Existing evaluations of large language models (LLMs) for triage use completed retrospective records and report performance close to that of physicians. We implement a methodology for evaluating LLMs on sequential triage, the task of predicting a triage acuity label from a growing prefix of a nurse-patient conversation. We evaluate six LLMs at five sequential checkpoints on two corpora: 425 LLM-generated (SIMULATED) and 50 physician-authored (CLINICIAN) conversations, both labelled under the Emergency Severity Index (ESI). Every model, measured by quadratic weighted kappa (QWK), degrades from moderate-to-substantial agreement on completed records to fair-to-moderate agreement at every sequential checkpoint. Controlled perturbations show that the label at every checkpoint is anchored on the chief complaint exchanges, and prompting interventions fail to lift this plateau. Models extract clinically relevant content from later turns, yet the surprisal of the true label rises across the checkpoints. So the model fails to integrate the evidence. Three expert clinicians on the same conversations reach a QWK of 0.887-0.929, while the best model reaches 0.295. Predictions concentrate at ESI-2 and ESI-3, and models agree with each other more than with the ground truth, so ensembling worsens the failure. Deploying LLMs for ED triage based on offline benchmarks alone misses this sequential failure.

We introduce QuranicMMLU, a benchmark for evaluating generative AI on Quranic Arabic across multiple dimensions of linguistic complexity. Existing Quranic benchmarks center on general question answering and semantic retrieval, without probing specific linguistic competencies or stratifying by cognitive demand and verse difficulty. We construct a five-pillar Quranic taxonomy spanning Phonology, Morphology, Syntax, Semantics, and Pragmatics, with 31 leaves covering phenomena from tajwīd and root-and-pattern morphology to occasions of revelation and inter-surah coherence. For each leaf we generate questions stratified by Bloom's cognitive level and verse perplexity, then have LLM as a judge to independently answer and score every item and route the annotations to manual review. The resulting dataset comprises 980 human-reviewed questions, each issued in both open-ended and multiple-choice form. We benchmark 12 systems on these items and find that the Islamic-specialized model leads, yet every system scores higher on multiple-choice accuracy (average 84%) than open-ended answer quality (average 60%): the two rankings agree closely (Kendall's τ=0.73), but multiple-choice scoring hides failures that surface only once answer choices are removed. QuranicMMLU thus offers a rigorous, linguistically grounded framework for evaluating Arabic NLP in the Quranic domain.

As Large Language Model (LLM) APIs become increasingly integrated into privacy-sensitive workflows, ensuring inference-time privacy without compromising task utility remains a major challenge. Existing approaches preserve most of the original semantic content to maintain downstream performance, but this also leaves exploitable cues for reconstructing the original text. This work investigates semantic decoupling, which replaces original semantics with alternative content while preserving the structure needed for LLM reasoning. Based on this idea, we propose CROSS-MAP, a bidirectional framework that maps private inputs into a different semantic domain before inference and recovers the corresponding outputs afterward. Local models are trained with multi-objective optimization to maximize semantic divergence in the mapping stage while minimizing semantic inconsistency in the recovery stage. Experiments show that CROSS-MAP reduces reconstruction success across multiple attack settings while outperforming existing baselines in utility.

Effective LLM tutoring depends on correctly identifying the specific error in a student's reasoning before generating feedback. We study this problem in propositional-logic proof tutoring, where student actions can be checked against formal inference rules. We introduce a verifier-grounded architecture that separates diagnosis from language generation. Using 600 balanced student actions, we compare a zero-shot LLM detector, a fine-tuned detector, and a symbolic verifier. Each diagnosis is processed by shared rationale and feedback agents, isolating the effect of the initial diagnosis. The zero-shot detector achieves a macro-F1 of 0.191; fine-tuning raises this to 0.709 but retains systematic errors between structurally related classes. Rationales generally preserve the diagnosis supplied to them, showing that an incorrect diagnosis can be faithfully propagated through the pipeline. Feedback can likewise remain faithful to its rationale, non-revealing, and pedagogically appropriate while addressing the wrong error. Verifier-grounded feedback achieves the highest diagnostic correctness, and expert ratings largely uneven with the automatic feedback evaluations. These findings show that apparent feedback quality can conceal upstream diagnostic errors and that faithfulness must be evaluated separately from correctness. Our code is publicly available

Summarization ships in countless production systems, making model selection a routine decision that depends on measuring summary quality. Existing metrics struggle to support this: ROUGE captures only surface overlap, while LLM-as-judge scores saturate to near-identical values that fail to rank models effectively. We observe this saturation across three public datasets, two proprietary datasets, and multilingual settings. Motivated by this, we introduce Semantic Scaffold, an evaluation framework that extracts a hierarchical representation of facts, questions, and entity attributes from a source text, labeling each as a main point or supporting detail, and reusing this structure as a fixed reference for scoring summaries. From this representation, we derive three diagnostic metrics: Fact Preservation Score (FPS), Question Preservation Score (QPS), and Entity Preservation Score (EPS), designed to reward the preservation of essential information while penalizing detail overload, and position them as interpretable diagnostics that remain informative where holistic axes collapse. Finally, we analyze four recurring failure modes of ROUGE and LLM-as-judge scores, demonstrating that scaffold-based evaluation remains informative where conventional metrics collapse.

We present Apollo Restore, a 24-billion-parameter large language model for restoring lacunae—physical gaps—in fragmentary Ancient Greek texts. Fine-tuned from Mistral Small with a fill-in-the-middle objective, Apollo Restore reconstructs missing spans without requiring oracle knowledge of their length. To our knowledge, it is the first large-scale decoder model for historical Greek, and the first for any ancient Mediterranean language. Evaluated as in prior work, on short gaps of up to ten characters, Apollo Restore places the correct restoration among its top twenty candidates for 80.6%/54.6%/61.0% of documentary-papyrus, literary-papyrus, and stone-inscription lacunae, exceeding the strongest published models by $1.6\times$/$2.6\times$/$1.4\times$. Prior evaluation protocols, however, inflate scores through a bias toward trivially short gaps; under a length-balanced metric Apollo Restore's advantage over the strongest published models grows to $2.3\times$/$3.5\times$/$1.6\times$ and degrades gracefully, even given incorrect length hints. In a blind study, 20 expert papyrologists, epigraphists, and philologists strongly preferred Apollo Restore to the strongest baseline and judged its performance at least as good as human restorations in 77% of cases. Apollo Restore also improves the published reading of P.Herc. 1667—a papyrus roll carbonised in the eruption of Vesuvius in 79 CE and digitally unrolled and edited after Apollo Restore's training data was compiled. Apollo Restore is an output of the Decoding Antiquity initiative to build specialized LLMs for historical languages and manuscripts, led by the Austrian Academy of Sciences.

Chinese social media has generated a vast and continually evolving lexicon of internet buzzwords whose meanings are often non-literal and deeply rooted in local cultural and pragmatic contexts. Existing research has primarily focused on interpreting these buzzwords within Chinese, leaving largely unexplored whether LLMs can transfer such culturally grounded knowledge across languages and accurately convey the intended meanings in English. This cross-lingual capability is also critical for safety, as harmful expressions may obscure their offensive content through culture-specific homophony, euphemism, irony, or coded language. In this paper, we investigate the ability of advanced LLMs to understand Chinese internet buzzwords across languages. To this end, we introduce CIBuzzBench, the first benchmark for cross-lingual Chinese-to-English understanding of Chinese internet buzzwords. CIBuzzBench comprises 3,001 Chinese internet buzzwords annotated with English meaning explanations, English equivalents, category labels, and harmfulness labels. Based on these annotations, we design three evaluation tasks: Meaning Explanation, Cross-lingual Equivalent Matching, and Culturally Grounded Harmfulness Detection. We evaluate representative state-of-the-art proprietary and Chinese LLMs under both English- and Chinese-prompting settings. Our results show that LLMs continue to struggle with the cross-lingual understanding of Chinese internet buzzwords, particularly in fine-grained non-literal interpretation, robust equivalent matching under option perturbations, and calibrated harmfulness detection. These findings highlight the persistent challenges posed by culturally grounded language phenomena for multilingual LLMs and safety-oriented evaluation. The dataset and code are available at https://github.com/SuperYFan/CIBuzzBench.

Most work in computational ethics treats annotator disagreement on moral content as noise to be voted away, collapsed into majority vote or the more permissive any-annotator rule the moment a single annotator flags an item. We argue this uncertainty should instead be modeled and learned from. We introduce Moral Entropy, a Bayesian framework that keeps a full posterior over the true label and decomposes its entropy into aleatoric uncertainty (irreducible disagreement about the moral content) and epistemic uncertainty (from insufficient or noisy annotation) – and lets any heuristic consensus rule be audited against a calibrated ground truth via entropy methods such as cross-entropy/KL, Brier score, and expected calibration error. Across three corpora and fifteen discourse domains, auditing the standard aggregation rules against this posterior reveals bias that no current pipeline reports: the any-annotator rule disagrees with the calibrated posterior on roughly 30% of items – pooled, almost entirely false positives, though the errors invert at the foundation level (19.9%/38.9% mean FPR/FNR on MFTC) – while the stricter majority and two-vote rules miss 63-83% of true positives.

Understanding contextual causality is critical for large language models (LLMs), as it enables them to accurately identify causal relations in specific situations and support more reliable decision-making. Despite its significance, a systematic exploration of contextual causality with LLMs is still lacking. To fill this gap, we present a comprehensive survey on this topic. In this survey, we first propose a taxonomy of contextual causality, consisting of semantic, intervention, and counterfactual causality, and characterize each category by its core causal question, required model capabilities, representative tasks, and practical uses in causality analysis. We then analyze existing studies and discuss their key limitations. Finally, we examine the gaps between current benchmarks and real-world needs and outline promising directions for future research. Our goal is to clarify the research landscape of contextual causality with LLMs, emphasize its importance, and highlight promising future directions.

Bilingual language models (LMs) offer a controlled setting for studying how training conditions shape second-language (L2) behaviour, but prior work typically varies exposure structure, scale, and architecture at once, making it difficult to attribute effects to any single factor. We introduce Beetle, a controlled language model pretraining framework in which tokeniser, target language, training budget, and exposure structure are each independently manipulable, enabling systematic and comparable experimentation of training conditions. Using Beetle, we train and release 285 bilingual and 45 monolingual open-source LMs with rich checkpoints across a range of exposure schedules, data scales and first languages (L1s) to study multilingual pretraining and computational modelling of bilingualism and second language learning. Evaluating models on human bilingual and second language reading-time prediction and grammaticality judgement tasks, we find that staged and temporally structured curricula consistently improve alignment with language learner reading time compared to balanced bilingual training, with the largest gains at smaller data scales and for typologically closer language pairs. The Beetle models are well suited tools to help move computational psycholinguistics beyond its prevailing monolingual, English-centric focus toward models of human bilingual processing, to study cross-lingual learning dynamics, while supporting community-based development of controlled model families.

LLMs are increasingly used in social simulations for socially interactive agents and robots, offering more flexibility than rule-based systems. However, even though they mimic human behaviour very well, there is a persistent alienness to them. This work investigates whether personality-aware fine-tuning can reduce this gap by improving the consistency and controllability of personality-conditioned dialogue generation compared with instruction prompting alone. We fine-tune two small open-weight LLMs, Qwen2.5-7B-Instruct and Ministral-8B-Instruct, using a corpus that combines personality-labelled social media posts and dialogues to create a personality-based dialogue engine for social simulation. The resulting models are evaluated across multiple social interaction scenarios using three independent LLM judges, which assess personality fidelity and provide evidence-based behavioral interpretations. We additionally quantify inter-rater agreement and lexical characteristics of the generated dialogue. Results indicate that fine-tuned models are not better at role-playing different personalities than their respective baseline models. However, low inter-rater agreement limits the confidence with which these results can be interpreted. Concerning the quality of generated texts, fine-tuned models are mostly comparable to the baselines, with fine-tuning improving the linguistic diversity of the Qwen models. While the results appear generally usable and the baseline models offer the best overall performance, future studies should place greater emphasis on the quality and domain alignment of training data for accurate personality role-playing.

Most multilingual dysarthria-severity systems either train on a single aetiology-language pair or pool heterogeneous aetiologies into one label space. We test that pooling assumption with four matched HuBERT-base contrastive embedding models under a shared backbone, training recipe, corpus registry and held-out evaluation: one mixed-aetiology baseline and three aetiology-specific models for cerebral palsy (CP), Parkinson's disease (PD) and amyotrophic lateral sclerosis (ALS). Training combines clinically labelled speech with ordinal pseudo-labels from a training-free phonological profiling method [1], [2]. On speaker-disjoint, leakage-filtered held-out subsets, the per-aetiology models outperform the mixed baseline across all three target aetiologies: CP (macro F1 0.829 vs 0.676, +22.6 % relative), PD (0.715 vs 0.511, +40.0 %) and ALS (0.788 vs 0.596, +32.3 %). On CP, adding 144 SAP and 44 CDSD pseudo-labelled speakers lifts macro F1 from 0.786 to 0.829 over a clinical-only CP model (+4.3 percentage points). Training data span three to seven languages per aetiology. We position this as a controlled comparison of label-space design choices and discuss pseudo-label calibration, split hygiene, and confidence-thresholded deployment as important limitations for future work.

We present the Spoken Wikipedia Presentation Corpus, an extension of the Spoken Wikipedia Corpora featuring LLM-generated slide decks for multimodal ASR. Slides are created from LLM-segmented sections using a hybrid pipeline that combines LLM-based content planning with rule-based design decisions. For each section, an LLM generates a slide title, bullet points, a takeaway message, and a visual description that is used to create an illustration. Rule-based matching then selects layouts, themes, and styles to produce the final slides. A vision LLM extracts slide text as Markdown. We evaluate multiple ASR and spoken language models (SLMs). The best model achieves an average micro-WER of 10.23% and an average micro-CER of 6.48% on audio-only inputs. English yields the lowest error rates, followed by German and Dutch, while performance declines across lower-resource languages. Although audio-only baselines are strong, multimodal zero-shot prompting of omni models remains challenging. The aligned slide, text, and audio data show a strong potential to improve recognition through cross-modal context.

Chain-of-thought (CoT) can sound plausible yet be unfaithful to the model's underlying reasoning. Most prior work probes CoT faithfulness through input–output behavior or input attributions, leaving internal computation largely underexplored. We instead cast faithfulness as internal concept grounding: Does a large language model's (LLM) CoT reasoning engage the same internal concepts that support the LLM's direct prediction, and do the shared concepts causally drive its answer? Encoding a prediction pass and a CoT pass with a single shared sparse autoencoder (SAE), a reliable approximator of the latent concepts LLMs use, makes their internal concepts directly comparable. We introduce three correlational metrics of concept-level alignment and a causal metric, $Δp$, which ablates the shared concepts and measures the drop in answer probability. Across five LLMs and four datasets, concept alignment is generally high, as indicated by the correlational metrics; yet these only identify which concepts are shared, not how much they causally contribute. $Δp$ fills this gap: causal faithfulness varies substantially with model depth, peaking at mid-to-late layers rather than the final ones, and model scale reshapes the layer-wise profile. Moreover, causally important shared concepts are not always verbalized in the CoT. These dissociations suggest that faithfulness cannot be reliably assessed from surface-level or representational correspondence alone; assessing it requires causal tests of whether the internal concepts underlying a CoT actually drive the model's prediction.

Word Error Rate (WER), the most commonly used metric for Automatic Speech Recognition (ASR), treats every lexical deviation from the reference as equally costly, regardless of whether it changes meaning. This raises the question: does WER actually track how humans judge ASR transcript quality? We introduce HATS-en, an English dataset for human-centered ASR evaluation. Using this dataset, we benchmark lexical metrics against several configurations of BERTScore and SemDist, varying the language model, layer, and pooling strategy. We find that WER agrees least with human judgment among all metrics tested, that the best-performing SemDist configurations achieve the highest overall agreement, ahead of CER and BERTScore, and that no single model is best across settings. CER, despite its simplicity and low cost, remains remarkably close to these best configurations. In line with prior recommendations, our results support shifting ASR evaluation toward CER both for English and for morphosyllabic writing systems as it is a more interpretable and low-cost metric for what evaluation should actually capture, and using SemDist as a complementary evaluation.

In-context learning using Large Language Models (LLMs) offers a compelling path to training-free post-OCR correction, yet its effectiveness for Devanagari script remains entirely unexplored. We present the first systematic evaluation of LLMs (3B-32B) for post-OCR correction in Hindi and Marathi, comparing three in-context example retrieval strategies: domain-random selection, dense semantic retrieval, and our proposed CharBM25, which retrieves examples by character n-gram BM25 similarity over OCR inputs to target shared error patterns with the test sentence. Across a 20,000-sentence benchmark spanning five news domains, retrieval strategy is the decisive factor in correction quality: CharBM25 outperforms domain-random selection by 2.8-4.0pp absolute WER on Hindi and 2.9-3.8pp on Marathi, using character trigrams, which consistently outperform bigrams and unigrams. Scale dominates performance: Gemma-3-27B achieves WER reductions of 55.0% for Hindi and 33.3% for Marathi under CharBM25-5. Few-shot gains are capacity-gated: models below 8B do not reliably improve over the OCR baseline, and on Marathi the smallest models (3B) degrade more sentences than they improve. Marathi is persistently harder to correct than Hindi across all scales, reflecting its greater morphological complexity. These findings establish CharBM25 as an effective, GPU-free retrieval strategy that matches or exceeds dense retrieval at negligible computational cost, and show that combining it with a general-purpose LLM of 12B+ parameters delivers reliable, training-free Devanagari post-OCR correction without task-specific fine-tuning. Dataset: https://huggingface.co/datasets/AbhishekBhandari/Devanagari-OCR-ICL-Benchmark

Knowledge distillation (KD) aims to compress high-performance teacher LLMs into lightweight students. However, distilled students often exhibit substantial performance degradation in out-of-distribution (OOD) settings, a critical gap that remains underexplored. We identify two compounding mechanisms causing OOD performance degradation: (1) data spuriousness: students can learn spurious correlations in the distillation dataset over genuine causal relationships; and (2) teacher capability: standard KD treats all samples uniformly, ignoring whether the teacher is guided by causal features or misled by spurious shortcuts on a given sample. To address these challenges, we propose Invariance-Weighted Distillation (IWD), a theoretically grounded framework that dynamically reweights training samples using an estimate of the teacher's causal reliance derived from prediction invariance across multiple synthetic environments. IWD perturbs spurious cues while preserving core semantics, assigning higher distillation weights to samples whose teacher predictions remain invariant, indicating greater reliance on causal rather than spurious features. We theoretically show that IWD reduces the student's Spurious-to-Causal (S2C) gradient ratio compared to standard uniformly weighted KD, driving the student toward more invariant representations. Experiments on four NLP benchmarks (MNLI, SQuAD-v2, CoNLL-2003 NER, and SST-2) across two model families (DeBERTa-v3 and Qwen-2.5) demonstrate that IWD consistently outperforms strong KD baselines on OOD evaluations while maintaining competitive in-distribution (ID) performance. Specifically, IWD achieves the highest accuracy in 15 out of 16 OOD benchmarks and improves average OOD performance over standard KD by 4.34 percentage points on NLI and 14.94 percentage points on QA.

Conversational speech depends on dialogue context and the listener's immediately preceding behavior. We propose ReACT-TTS, a two-stage framework that uses a one-second pre-response listener facial sequence to plan the next utterance's emotion and prosody before speech realization. On a strict dyadic MELD protocol, Temporal conditioning yields higher mean macro-F1 and VAD concordance than Text-only across ten seeds, while accuracy remains essentially unchanged. Ablations show that temporal modeling performs best among the visual variants and that an explicit early-to-late difference is unnecessary; correct listener reactions also outperform cyclic mismatches on average. In a contextual-appropriateness study with 20 speech researchers, 76% of judgments prefer Temporal, 9% Text-only, and 15% report no preference. We further connect the predicted response style to a Grad-TTS backbone for end-to-end speech realization. Overall, the results support pre-response listener dynamics as complementary cues for conversational response planning. The source code is available at https://github.com/CYJ1/ReACT-TTS_public.

LLM-based AI systems answer political questions for hundreds of millions of people. Current audits measure what they say to an average user, but their behavior is dynamic. I argue that their political behavior is a set of policies over whom to answer, what to say, and whether to engage at all, conditional on the topic and what the system knows about the user. I call these policies the system's speech regime, which is how a developer settles the tradeoff between answering, accommodating the user, and refusing, each of which carries a cost that varies by topic. I derive a typology of five regimes from two dimensions, engagement and stance. I test six AI systems (OpenAI, Anthropic, xAI, Google, Mistral, DeepSeek) in a preregistered experiment of 7,500 multi-turn conversations that randomly assign the user's political identity across five topics: abortion, Catalan independence, climate change, Nazism, and a zero-stakes control (pineapple on pizza). Two LLM judges from different developers score every answer, validated against human coding, and refusal is treated as an outcome rather than missing data. Every system accommodates the user on the control topic, showing that political restraint is a policy. On contested topics the systems fall into different regimes: on abortion, GPT engages and mirrors every user, Gemma refuses everyone, Claude answers strongly conservative users 35 percent of the time and almost no one else, and Grok accommodates conservatives only. On settled topics such as climate change and Nazism, five systems hold firm for every user. The systems also infer the user's overall ideology, so accommodation can spill over to topics not yet discussed. A comparison of two Grok releases shows the regime changing between versions in a way current audits miss. Speech regimes matter for alignment research and for polarization, political knowledge, and the quality of democracy.

We propose a framework to analyse how strongly different linguistic relations are linearly encoded in language model embedding spaces. We formalise linear encoding via a constrained linear approximation over related and unrelated word pairs and apply this to an extended BATS dataset covering inflectional, derivational, lexicographic, and encyclopedic relations in GloVe, RoBERTa, and ModernBERT. Our experiments show near-perfect linear encodings for inflectional and derivational relations, but substantially higher errors for lexicographic and encyclopedic relations, especially for one-to-many and many-to-many associations. We also find that RoBERTa and ModernBERT generally encode relations more linearly than GloVe. These results indicate that our framework can reveal which relational structures are most linearly accessible in embeddings, offering a compact tool for probing and comparing relational geometry across models.

Gender bias remains a persistent concern in machine translation (MT), affecting both generated translations and their automatic evaluation. When a source text leaves a person's gender unspecified, translations may realize that person using masculine or feminine forms, and both MT systems and evaluation metrics may exhibit systematic preferences between these alternatives despite the source providing no basis for such a distinction. We study this behavior in the WMT 2026 Automated Translation Quality Evaluation Systems Shared Task using an occupation-balanced subset of GAMBIT+. We consider seven English-source language pairs, six from the original dataset, targeting Arabic, Czech, Greek, Icelandic, Russian, and Ukrainian, and extend the original resource with German. The subset contains 1,308 masculine/feminine translation pairs per target language, with three examples for each of the 436 ISCO-08 occupational groups. We evaluate shared-task submissions and baselines for score prediction and error annotation, examining the direction, magnitude, and frequency of gender-related differences. We find an overall tendency for masculine translations to receive higher scores, as well as differences per occupation following stereotypical gender representations, although the strength and consistency of this preference vary considerably across evaluators and languages. Our results show that gender bias remains present in MT evaluation, but that capturing its extent requires looking beyond a single aggregate measure to complementary dimensions of evaluator behavior.

Long transcripts are costly inputs for downstream NLP systems and often contain irrelevant context. We study query-conditioned topic localization: predicting the sentence span in a transcript that best addresses a topic-title query. To improve span localization, we reuse ASR encoder states as sentence-level representations and fuse them with textual embeddings. This lets lightweight span locators exploit speech information without running a separate audio encoder. Experiments on two public datasets show consistent gains over text-only baselines, especially under strict boundary-matching criteria. Cross-dataset experiments further indicate that the benefits are strongest for structured or semi-structured speech, while gains on spontaneous speech are limited and mixed.

Vision-language models (VLMs) rely on storing and transferring appropriate information across their sub-components. Verifying that the VLMs exhibit desired behaviours, while avoiding harmful ones, is central to their reliable deployment. Yet, methods that identify VLM behaviours remain scarce. We present DiaVLo, a diagnostic framework that leverages human curation and VLMs' generation capabilities to construct specifications of desired and observed VLM behaviours, surfacing potential misalignments. Beyond this, DiaVLo also provides causal estimates to identify the most influential concepts steering VLM behaviours. We evaluate DiaVLo on several open-source VLMs under both classification and generation conditions. Our experiments show that DiaVLo produces behaviour labels that correlate with model performance and provide context for measured performance. DiaVLo surfaced behaviours that are clearly aligned and misaligned, alongside patterns in how VLMs perceive, organise, and prioritise concepts.

Occupational accident narratives contain valuable information about work situations, unfavourable conditions, accident events, and their consequences. Automatically structuring these narratives can facilitate large-scale accident analysis and support occupational risk prevention. However, the terminology and writing styles used to describe accidents vary considerably across sectors and organisations, raising questions about the ability of automated coding systems to generalize beyond their training domain. In this paper, we evaluate the cross-sector generalization of accident-process role classification in French occupational accident narratives. We construct an expert-annotated corpus in which factual units are classified into four roles: work situation (A0), explicitly reported unfavourable condition (A1), accident event or deviation (B), and reported consequence (C). The role classifiers are developed and selected exclusively on 42,244 factual units extracted from 6,040 construction-sector narratives and are then evaluated on unseen corpora from the metallurgy and chemistry–plastics sectors, as well as on an independently collected company corpus, without retraining or target-domain tuning of the role classifier. We compare frozen pretrained representations with task-specific fine-tuning and supervised representation-learning strategies. The results show that task-specific adaptation consistently improves cross-domain transfer over frozen representations. Across repeated training runs, the three leading task-adapted strategies achieved average balanced accuracies between 85.6% and 85.8% across the three target corpora. These findings support the development of transferable assisted-coding systems capable of consistently structuring heterogeneous occupational accident narratives for expert review and cross-sector prevention analysis.

Probabilistic Graphical Models (PGMs), especially Bayesian Networks (BNs), expose directed structure and probabilistic parameters, making them natural symbolic targets for neurosymbolic AI. Yet training text-to-parameterized-BN systems requires paired text-to-BN resources unavailable at scale. We introduce PRISM-BN, a controlled corpus of 5054 BN-grounded descriptions paired with discrete reference BNs containing variables, states, directed edges, root priors, and full multi-parent CPDs across five domains. The instances are derived from 50 Wikipedia-seeded backbones, and their probabilities are internally constructed benchmark targets rather than externally validated causal estimates. PRISM-BN is built with PRISM, a marginal-first pipeline that elicits marginal and local joint distributions, analytically recovers normalized CPDs, and constructs locally reparameterized subgraphs. We define a benchmark with semantic node and state alignment, conditional structural scoring, and strict full-CPD evaluation. Across six LLM extractors, Node F1 ranges from 0.56 to 0.83, conditional Edge F1 from 0.90 to 0.97, and CPD-KL from 1.11 to 3.14. Conditional state and edge recovery remain consistently strong, whereas strict full-CPD agreement remains challenging. These trends persist with independently generated GPT-5.5 references, and a human pilot corroborates structural recoverability and similar probabilistic interpretations. PRISM-BN supports separate evaluation of structural recovery and probabilistic parameter estimation.

Fairness research on language models involves measuring bias, applying mitigation methods, and examining the evidence on which an evaluation rests. Existing tools offer complementary functionality through different interfaces, so combining them requires reconciling model interfaces, evidence formats, access constraints, and result types before applicability can be checked or methods compared. We introduce \textbf{FairLMs}, a Python library that connects these activities through explicit declarations of model capabilities and input requirements. It provides 33 intrinsic and extrinsic metrics, 14 mitigation components spanning four intervention categories, 14 dataset and scoring-instrument diagnostics, adapters for the three Transformer architectures and supported hosted completion APIs, and benchmark loaders. Declarations are checked before execution and results carry the configuration under which they were obtained, so that compatible components can be combined, methods compared under a common protocol, and workflows extended to new models and datasets. The source code is available at: https://github.com/FairLMs/FairLMs.

In recent years, there has been a surge of interest in Cultural NLP, with substantial efforts to create globally inclusive NLP systems. The rapid growth of literature in this field makes it difficult to track trends in methods and data resources. To address this, we analyze over 375 papers to answer three complementary questions: (1) What Cultural Capabilities (CCs) are being targeted in NLP systems? (2) How are cultural data resources being created? and (3) What methods are being used to improve the CCs of those systems? We discuss trends observed across the three questions, and identify relevant research gaps. To facilitate further research in this field, we release our full list of analyzed papers in the form of an interactive web interface, which includes a feature to allow researchers to add their work; we hope this facilitates future research and proves to be a valuable resource for the Cultural NLP community.

Air operations rely on complex rules, established procedures, and time-critical analysis under limited connectivity and strict security constraints. In such environments, analysts must combine written doctrine with images, often without access to external computing resources. This paper studies offline large language models as decision support tools, deployed in isolated and restricted environments to give analysts access to doctrinal knowledge that remains traceable to its original sources through natural language interaction. We describe a modular retrieval-augmented architecture suitable for operation without Internet connectivity, supporting both text and image input from technical manuals. As a first step toward evaluating this architecture, we report a pilot study with four image analysts of the Brazilian Air Force, combining (i) a doctrinal knowledge assessment based on their electronic-target identification doctrine, comparing human and proposed system performance on the same test, and (ii) a measurement of the cognitive workload involved in manually producing a reconnaissance target report (Relatório de Missão de Reconhecimento - REMIR) without AI assistance. The results show a demanding manual task, especially in terms of mental demand (6.0/7) and effort (5.0/7), while the proposed system matches the human score (8/10) and completes the assessment in 7.1 minutes (compared to a human average of 26.5 minutes), establishing a baseline for future AI-assisted evaluation. Finally, we describe a future evaluation protocol to systematically compare manual and AI-assisted workflows.

Recent work applies human psychometric questionnaires to large language models to elicit moral and value profiles, but it is not clear whether these instruments measure anything stable in models or whether the resulting profiles can be moved toward a target human population. We administer the Norwegian Moral Foundations Questionnaire (MFQ-30) to six open-weight LLMs and compare their foundation profiles to a sample of N = 1,282 Norwegian respondents. We test two steering interventions, prompt-level persona steering and activation-level ActAdd. Half the models engage with the questionnaire under our attention check. The other half default to flat or central-tendency outputs that look near-human on average without tracking item content. A neutral Nordic-respondent persona, written without any distributional information from the human sample, brings the engaging models 44-77% closer to the Norwegian mean in Mahalanobis $d^2$. One-pair ActAdd at a fixed mid-layer flattens the foundation profile rather than steering individual foundations. For at least one model the same persona that shifts the profile also induces engagement that was absent at baseline, a concrete instance of the cognitive phantoms that Peereboom et al. (2025) warn about.

Conversational AI systems produce fluent, socially appropriate responses, yet whether they participate in cooperative communication or merely simulate its surface forms remains unclear - a question central to how these systems are evaluated, trusted, and designed. This study investigates how morality, politeness, and alignment - three dimensions central to cooperative dialogue - function in human-AI interaction compared to human-human conversation. We analyze 15,881 human-ChatGPT and 10,784 human-human multi-turn dialogues, using mixed-effects models to identify which features predict turn-to-turn alignment. We observe a consistent dissociation: AI produces the surface features of cooperative communication without the underlying social architecture. Moral output appears preconfigured rather than negotiated; warmth is generated without face sensitivity; linguistic convergence declines persistently. Most strikingly, the cooperative mechanisms themselves reverse direction: hedging and softening associated with greater accommodation between humans are associated with reduced alignment when produced by AI, and purity framing associated with human divergence coincides with users converging toward the AI. Agency - giving users room to shape the exchange - is the most consistent predictor of alignment across both interaction types, while lower moral assertiveness in more recent models is not accompanied by better cooperation. Together these patterns suggest that AI reproduces the surface of cooperation without the mutual adaptation that grounds it between humans - and, more surprisingly, that mechanisms sustaining human accommodation can run in reverse with AI, suggesting a turn-level view may be insufficient for interaction-level success.

Relexicalization is a pivotal technique in clinical NLP, as it facilitates robust masking of sensitive information while synthesizing datasets that retain high-fidelity, real-world characteristics. However, preserving structural integrity, relational coherence, and temporal consistency during transformation remains a significant challenge. Existing approaches frequently rely on independent entity replacement, which results in clinical inconsistencies across longitudinal records. This reduces the value of such relexicalized datasets for downstream scientific analysis. To address these limitations, we introduce G-RELIC (Graph Based Contextual Relexicalization with Improved Consistency) which combines the power of LLMs with graphs. G-RELIC implements a graph-based mapping mechanism which optimizes for one-to-one correspondence between original and surrogate entities. It also introduces a deterministic temporal repositioning algorithm to preserve temporal consistency. Empirical evaluations on diverse, real-world clinical datasets validate that G-RELIC significantly outperforms state-of-the-art baselines. G-RELIC yields a 30.4 percentage point improvement in relational integrity (62.1% to 92.5%) and 45.9 percentage point improvement in temporal coherence (46% to 91.9%) without compromising on the recognized privacy benchmarks for clinical datasets. This maximizes the analytical utility of relexicalized datasets while minimizing re-identification risk.

Large Language Models (LLMs) are increasingly deployed in applications that must weigh clashing moral values, yet even strong models exhibit hidden biases and brittle instruction-following across languages. We introduce a 12,000-instance dataset of two-option dilemmas covering pairwise three value conflicts: Honesty vs. Justice, Justice vs. Autonomy, and Autonomy vs. Honesty, along with their translations into Hindi, Arabic, Spanish, and Chinese, to probe cross-lingual behavior. Benchmarking on GPT-5-mini reveals that it consistently favors Honesty over Autonomy across all five languages when no policy is given. The Llama-3.2-1/3B models exhibit strong first-option bias; however, both plain fine-tuning and Direct Preference Optimization fine-tuning effectively remove this bias, increasing accuracy to greater than 98%. In order to decouple the effect of learning correlations in the dataset from abstract values, we propose a task vector transfer based experiment where after computing the task vectors for a direction of value preference we orthogonalize it with respect to the general instruction following vector. Our experiment shows that this method is effective in isolating the direction of the specific value preference that can successfully be used to conduct task arithmetic to obtain a model with the opposite stance.

FarmerChat is Digital Green's farm advisory service for smallholder farmers. When something looks wrong with a crop, the farmer takes a photograph and sends it, and that photograph is the whole question: no symptom described, no crop named, often no text at all. The service has to determine whether the picture can be used, what crop it shows, and what is wrong with it, from images taken on cheap phones in a field, in poor light and with a moving camera. The system doing this today cannot be adjusted. It has no adjustable thresholds for photograph rejection, crops and problems cannot be added, and there is no confidence cut-off to set. We study about 1.16 million photographs sent to FarmerChat from Ethiopia, India, Kenya and Nigeria. The production quality gate rejected 46.8% of the images it judged, over a quarter of those reaching diagnosis returned no crop name, and 35.8% of the labelled problems filed under "disease" are pests, identifiable without the crop. We therefore split the work into three stages: a quality gate (M0), a crop detector (M1), and a disease or pest detector (M2). Route A fills all three with one fine-tuned vision-language model (Qwen3-VL-4B) answering in a single call. Route B fills each with a small specialist model (DaViT, YOLO26). We replace our production GPT-4o quality gate with a small MobileNetV3 gate at 86.9% F1 in 12 ms. On one test set scored the same way for every system, a hierarchical DaViT-Base achieves 95.41% crop accuracy against 91.46% for the production baseline. It also leads on diagnosis and never declines to answer, while every language model in the comparison leaves a large share of rows with no diagnosis. The fine-tuned model retains two capabilities the specialists do not have: one call for all three stages, and a request for a better photograph when the image cannot support an answer.

Introduction. Open collaborative knowledge graphs such as Wikidata increasingly ground agentic artificial intelligence, information retrieval, and language modeling systems, making systematic auditing of their demographic representation and overall equity a research imperative. Methods. Herein, we present an open-source auditing platform that ingests over 10 million statement bindings representing over 6 million humans on Wikidata via QLever, and evaluates representation of gender, sexual orientation, geography, birthplace urbanicity, ethnicity, multilingual coverage of labels, descriptions, and aliases, occupation, and select intersectional pairs of these entities. It does so by making use of Chi-square goodness-of-fit tests, 95% Wilson-score confidence intervals, and disparity ratios, in light of Rubin's missingness taxonomy. Results. Women accounted for 28.71% (CI +/-0.04) of all humans in Wikidata with a stated gender. 38.26% of humans had a citizenship statement, with Western Europe and North America (WENA) representing approximately 53% of such statements. Among 1.8 million birthplaces that could be classified, rural birthplaces were observed in 2.48% of cases (in comparison to 27.4% global baseline). Fewer than 1.2% of entities carried an ethnicity statement, and non-English Wikidata descriptions covered 18.2% of items. Discussion. Our findings reveal significant missingness across the evaluated axes. Ethnicity and sexual orientation were the most critically under-documented (missing statements) while rural birthplaces and non-WENA citizenship were the most underrepresented.

Neural morphological generation systems often achieve high aggregate accuracy on benchmark datasets, yet such performance can conceal systematic errors clustered in rare morphological subclasses. We present an orthography-aware diagnosis of Japanese past-tense verb inflection, treating hiragana not merely as a transcriptional medium but as a representational system that encodes morphophonological structure. Using two character-level Transformer architectures evaluated across five random seeds, we show that although both systems exceed 97% aggregate accuracy, a single structurally specific irregular subtype, verbs whose stems end in /e/ and require gemination before the past-tense suffix and make up fewer than 1% of the data, accounts for a disproportionate 30-43% share of residual errors and contributes roughly 34-48x its prevalence to total errors. We then move from diagnosis to causal isolation: controlled ablation experiments show that removing this subtype alone produces larger accuracy gains than removing all irregular verbs combined. These findings indicate that error concentration in neural morphological learning is not driven by irregularity per se, but by the interaction between extreme low-frequency morphological patterns and specific orthographic processes. We argue that morphological evaluation should incorporate fine-grained subclass analysis, and discuss implications for data-efficient, developmentally plausible language model pretraining.

Good tutoring adapts to the individual: it tracks what a learner knows, notices why they go wrong, and asks the next question that will help most. Most deployed tutoring tools instead serve fixed item banks and treat a wrong answer as a single bit of signal. We present CoLearn, an interactive, agentic tutor that supports an iterative tutoring loop: the learner practises, and the system builds an evidence-grounded memory of the learner's mastery and misconceptions. This memory is updated as evidence accumulates and is used to generate the next personalised question. CoLearn has three components: (i) a persistent learner-state memory that updates per-topic mastery with a soft-evidence variant of Bayesian Knowledge Tracing, where a large language model acts as a continuous observation function; (ii) adaptive question generation that targets the learner's weakest topic and recurring misconceptions; and (iii) an evidence view that makes personalisation visible and testable through live progress visualisation and blind A/B comparison. In blind A/B evaluation, questions conditioned on this memory are preferred over non-personalised ones 68-69% of the time, and in persona simulations with hidden ground-truth mastery the agent's belief converges toward the learner's true mastery.

The Viterbi algorithm has been previously used to perform forced alignment of audio to text to mine training data from online resources. However, many existing implementations have quadratic time and space complexity, scaling poorly to long input sequences. We propose two optimizations to address this issue. First, we apply the Hirschberg algorithm to perform the alignment in place using linear memory. Second, we model the alignment between speech and text as a constrained random walk, allowing us to prune the search space with arbitrary confidence while accounting for transcription errors. The Hirschberg optimization reduces memory usage from 140 GB to 5 MB for three-hour inputs while producing identical alignments in one-third the time of torchaudio when both run on a CPU. We achieve an additional 2x speedup with pruning on inputs longer than 20 minutes while preserving alignment accuracy in more than 98% of tested cases.

Modern automatic speech recognition (ASR) systems trained on extremely large datasets can produce transcripts with numbers written in Arabic numerals. This creates a need for fair comparison with models that output verbatim texts and proper processing of reference transcripts. Popular approaches often reduce text normalization to lowercase and remove punctuation, with no additional normalization applied to languages other than English. In this work, we analyze the impact of normalization of numerical expressions in the evaluation of ASR systems in various languages, using Polish as an example of a highly inflective language. We perform experiments on VoxPopuli and The Polish Parliamentary speech datasets and estimate word error rate (WER) differences for different text normalization approaches. We show that the difference due to the lack of number normalization in WER may be substantial - more than 2 percentage points, and often higher than the differences between systems in popular multilingual benchmarks.

Debate adjudication requires tracking how arguments develop through interaction, yet existing datasets rarely combine fine-grained debate transcripts with professional judgments collected during real competitions under a shared rubric. We introduce a dataset and benchmark for evaluating large language models' understanding of competitive Chinese-language debate at the match, stage, and speaker levels. We organized 182 matches and recruited 120 professional judges, with each match independently adjudicated by three judges using a predefined rubric. After excluding matches with incomplete records, the dataset contains 148 matches, 2,698 stages, and 20,542 exchange units, with manually verified transcripts and segmentation. It preserves original stage scores, match votes, best-debater ballots, and adjudication rationales. We define three tasks: winner-tendency prediction, stage-score prediction, and best-debater prediction. Zero-shot evaluation of multiple large language models yields a highest winner-prediction accuracy of 66.2%, a highest Pearson correlation of 0.250 between model stage scores and mean human ratings, and a highest best-debater prediction accuracy of 56.8%. The dataset and benchmark provide a testbed for studying large language models' understanding of interactive argumentation and their agreement with professional judges.