LangTrend

There is a tradeoff in machine translation meta-evaluation between prioritizing alignment with adequacy versus fluency. The balance depends on the combination of translation systems in the meta-evaluation dataset. This system set is a small, filtered sample whose characteristics change heavily across years and language pairs; it does not represent the true system distribution. Consequently, the adequacy-fluency balance is often unrepresentative and subject to change. For sensitive domains, controlling this balance is critical. We expose this balance as a tunable choice. To achieve a target balance, we reweight existing systems while minimizing distortion from uniform weighting, ensuring the evaluated systems remain real and representative. We provide an exact optimization algorithm with theoretical guarantees and pruning mechanisms to compute these weights. To validate meta-evaluation internal consistency, we design a scorer-augmentation framework that establishes a known relative identity for the scorers. Results demonstrate that our reweighting method effectively controls the adequacy-fluency balance and preserves the internal consistency, outperforming prior approaches. Finally, we analyze the performance of popular scorers across a sweep of this parameter.

As LLM conversations grow to hundreds of turns, full-context injection incurs $O(N^2)$ cumulative token costs, while lossy summarization or hard truncation irreversibly discards historical state. We propose Pull, a session router that maintains an addressable metadata directory via a local, deterministic Purifier (zero LLM calls, millisecond-level latency). At query time, the LLM lazily materializes only the turns it needs; unmaterialized turns remain accessible but collapsed. Unlike irreversible compression, Pull's materialization is reversible; subsequent queries can expand any collapsed turn. On LoCoEval (128 conversations, 12,780 turns), Pull reduces per-query context tokens (Phase 2) by 75.1 percent on single-hop tasks with equivalent quality ($Δ= -0.002$, n.s.) and by 72.0 percent on multi-hop tasks with no quality loss ($Δ= +0.017$). A controlled routing benchmark (7,831 queries x 10 methods) shows that entity lifecycle tracking is empirically a prerequisite for distance-independent routing. On BEAM 1M (14 conversations, 263 questions), Pull improves F1 by +55.2 percent over a truncation baseline.

As autonomous vehicles and Extended Reality (XR) headsets enable novel in-car interactions, seamlessly querying physical landmarks, known as Outside-the-Vehicle Referencing (OVR), remains challenging due to ego-motion and referential ambiguity. We present a robust, multimodal OVR framework fusing user gaze and natural language to identify Points of Interest (POIs). To address the scarcity of dynamic vehicular data, we developed a VR-based pipeline synchronizing 360-degree transit videos with vehicle GNSS telemetry. Through a user study (N=46) mapping passenger head orientation into a 3D geospatial Digital Twin, we captured authentic gaze-speech behaviors. We subsequently trained a lightweight Transformer network, leveraging LLMs to dynamically align continuous spatial gaze vectors with discrete verbal context. Experimental results demonstrate high accuracy and low computational overhead, achieving an 83.33% Top-1 accuracy (87.72% Top-2) and an average inference time of 24.3 milliseconds. This real-time paradigm effectively resolves referential ambiguity, enabling context-aware spatial retrieval for passengers within the vehicle.

This study highlights the role of domain-specific pretraining profile (DSPP) in Transformer performance for modeling digital pragmatics in Arabic-English code-switched discourse. It evaluates MARBERT and XLM-R(oBERTa), with BERT serving as a general-purpose baseline. The models were evaluated on their ability to classify context-sensitive pragmatic functions in code-switched social-media discourse. 11695 unique X posts were collected via Python and utilized for the study. The study employs a quantitative and qualitative NLP approach, following a supervised pipeline. Findings unveil that MARBERT consistently surpasses XLM-R with validation Macro F1 increasing from 0.39 to 0.84 and validation loss decreasing from 0.55 to 0.19. On an independent test set, it achieved 0.96 accuracy, 0.83 macro precision, 0.87 macro recall, and 0.85 Macro F1, while XLM-R achieved 0.92 test accuracy but a substantially lower Macro F1 of 0.52. This was also supported by class-level performance where MARBERT outperforms XLM-R considerably with F1 improvements ranging from +0.33 to +0.60, demonstrating a clear advantage in modeling Arabic digital pragmatics. The study concludes that Transformer performance depends more on DSPP than multilingual coverage alone, as the latter does not guarantee optimal performance on a highly specialized pragmatic classification task.

Generalisations are common in scientific communication, even though they are semantically ambiguous. An automated method is needed to identify and categorise claims according to their level of generalisation, in order help detect an over-reliance on generalisations and possible misrepresentations of scientific findings. We introduce a comprehensive taxonomy of generalisations in the scientific domain, NLPGenX, which labels claims according to their level of generality and framing within the text. We operationalise this taxonomy with an LLM-powered framework, NLPGenA, that automatically classifies sentences from scientific articles into 5 different generalisation classes. We validate our framework with human annotators and use the framework to construct a large-scale dataset of NLP papers annotated according to generality, with auxiliary labels for hedging and vague descriptors (NLPGens). We use NLPGens to analyse the use of generalisations in NLP papers across multiple venues and subdomains, and to examine associations with citation counts, hedging, and vague descriptors.

In Yorùbá, pitch alone separates \d{o}k\d{o} (husband, Mid), \d{o}k\d{ò} (vehicle, Low), and \d{o}k\d{ó} (hoe, High) – the diacritics ARE the tone marks. Yet character error rate (CER), the standard automated metric for text-to-speech (TTS), is in practice computed from ASR output that drops those marks: a synthesizer can ace CER and still say vehicle for husband. We introduce DunDun – named for the dùndún, the Yorùbá talking drum that speaks through pitch alone – an automated, reference-free lexical-tone metric that needs no tone-labelled corpus. The gold High/Mid/Low sequence is read from the input text's diacritics (in TTS that text exists by construction, so no reference recording is needed); the prediction comes from the audio's pitch track. We validate three ways. Flattening pitch with PSOLA resynthesis collapses DunDun while CER does not move. Inverting High and Low in the answer key of 300 native recordings drives the two-class readout to 0.14, symmetrically below its 0.35 chance level – a consistency check on the scoring path, not independent evidence. And three native listeners, over 67 blind A/B trials, pick the tone-correct clip 89.6% of the time (95% CI 80.0-94.8; p < 1e-4); whether DunDun tracks those judgements trial by trial is not resolved at this sample size. Applied to a massively multilingual zero-shot TTS model, DunDun shows what CER cannot: Yorùbá tone sits near the native anchor before any Yorùbá fine-tuning (0.567 +/- 0.02 over five decode seeds vs. 0.596; chance 0.33), despite the 21.4% CER the model's own paper reports; and a few hours of clean audio halve CER (5.6% to 2.7% by 5h, 1.7% by 15h) while tone saturates within the hour. On non-tonal Swahili, CER already captures the gains: the metric a language needs is language-dependent. We release the metric and the complete validation protocol.

Data-to-text driver coaching is often presented as a generic pipeline from telematics events to advice. This paper argues that its content requires localisation because usefulness and credibility depend on drivers' knowledge, prevalent risks, regulation, infrastructure, and available data. Two independently developed systems in the United Kingdom and Nigeria are compared by tracing requirements through content selection, generation, and field evaluation. The UK system prioritises post-trip reflection, explanations tied to road and place context, and tone-sensitive wording. The Nigerian system combines legally grounded, once-daily Tips based on detected events with weekly persuasive Reports; it foregrounds safety education and alcohol-related risk in response to reported gaps in formal training and traffic-rule knowledge, as well as local road-safety priorities. Reliable speed-limit metadata supported speeding feedback in the UK, whereas its scarcity led the Nigerian evaluation to exclude speeding from its outcome metric. Both interventions were associated with reduced distance-normalised unsafe-event rates in their own field studies, although their designs and metrics preclude an effect-size comparison. The analysis yields a requirements-to-content design process for localising safety-critical NLG without treating a high-income deployment as the default.

Performing deliberate mathematical reasoning in visual contexts is a hallmark of advanced Multimodal Large Language Models (MLLMs) and requires a sophisticated synthesis of perceptual grounding and symbolic logic. However, in the realm of mathematical functions, our investigation reveals a critical modality interference phenomenon: even advanced models, while performing textual computational reasoning, tend to disregard or misinterpret essential visual cues. To address this challenge, we propose Func-R1, which synergistically harmonizes precise visual perception and rigorous logical reasoning. Concretely, built upon an explicitly decoupled architecture, we employ a hierarchical post-training framework to progressively identify critical visual evidence and conduct in-depth theoretical reasoning. Furthermore, the Perception-Aligned Theoretic Optimization (PATO) strategy is proposed to steer policy updating towards internalizing fundamental theoretical properties while dynamically rectifying heterogeneous visual information throughout the reasoning process. Extensive experiments across diverse benchmarks demonstrate that Func-R1 delivers the optimal performance among open-source MLLMs, even surpassing GPT-5 with an 8.4% improvement on MathVerse's function-oriented tasks.

We introduce Enemray, a Hassaniya-centric language model that enables general-purpose interaction in Hassaniya. Enemray is trained around a stability–plasticity objective: acquire strong Hassaniya linguistic and cultural competence while preserving the general reasoning, multilingual, instruction-following, and safety behaviors of a capable instruction-tuned model. The development pipeline separates language acquisition from behavioral specialization. A separately assembled continual-pretraining corpus provides broad exposure to natural Hassaniya and Mauritanian text; layer-selective continual pretraining learns a compact language-specific parameter update; that update is transferred into the instruction-tuned parameter space; and supervised post-training develops conversational, cultural, literary, task-oriented, and cross-lingual behavior. The supervised corpus integrates selected public Hassaniya and Mauritanian resources with a substantially larger body of newly collected, reconstructed, curated, and constructed instruction data, while policy-generated replay provides a retention signal from the reference model's own behavior distribution. The resulting collection is substantially larger and broader in purpose than existing Hassaniya text resources. In evaluation, Enemray achieves the strongest English to Hassaniya translation among the compared open and proprietary models and the highest overall score on Mauritanian translation error detection, while retaining most of the general capabilities of its instruction-tuned base model on mathematical reasoning, knowledge, code generation, and function calling. This report describes the motivation, data construction, model design, training methodology, and evaluation of Enemray.

Deep research agents answer complex questions through iterative loops of searching, reading, and reasoning. Recent work on reasoning-intensive benchmarks such as BrowseComp-Plus shows that well-configured lexical retrieval can surface high-quality evidence, yet agents may still fail to connect documents carrying evidence to the gold documents. We identify a deep research agent's first retrieval move as an important design decision for this setting. We introduce Question's Gambit, a first-move retrieval module that decomposes the question into a set of clues, reformulates them into complementary searches, consolidates the retrieved results, and reranks the candidate pool before the agent begins its iterative search-and-reasoning process. This produces an opening context designed to support both clue aggregation and final-answer verification. We further evaluate on MultiHop-RAG to test whether these benefits transfer beyond BrowseComp-Plus to a more conventional multi-hop question structure. Experiments on BrowseComp-Plus show that Question's Gambit improves retrieval recall and downstream agent accuracy over strong baselines, improving answer accuracy from 83.1% to 90.5% with gpt-5.5 over Pi-Serini, the strongest reported agentic baseline. Our results confirm that effective agentic deep research depends not only on the tools available inside the loop, but also on the quality of the first move. We published our implementation publicly at https://github.com/radinhamidi/Question-s-Gambit.

Neyshekar is presented as an open Persian read-speech corpus designed for coverage of both formal and informal language, named entities, and longer utterances. In version 6, 62,279 validated recordings totalling 99.02 hours are provided from 190 contributors, with 34,541 distinct recorded prompts. The prompt pool was assembled from human-written material, contextualised homographs, and reviewed language-model-generated text. Text entries were normalised with the shekar library, which supports both formal and informal Persian, and every submitted recording was reviewed against a common validation rubric. About 24% of released clips are classified as informal by an automatic classifier; these register labels are not human-validated. Item-level rater labels are provided for reproducible agreement estimation, opaque per-clip contributor identifiers make the speaker-disjoint partitioning auditable and support contributor-clustered uncertainty estimates, and a text-disjoint test subset is included for evaluation beyond previously seen prompts. Per-contributor recording load and reference-free signal quality are characterised for every released clip. Corpus characteristics are compared with Persian Common Voice under shared processing. Utility is assessed through two ASR architectures, three optimisation seeds, WER and CER, and independent evaluation on the public PSRB sample. Against duration-matched Common Voice training at approximately 32 hours, in-domain WER is reduced by 9.5 points for Whisper and 11.6 points for XLS-R, and by approximately eight points for both architectures on the independent PSRB sample. Transfer and mixture benefits are not consistently observed across architectures and training budgets. The corpus is released under CC0; code and data are made available through the project repository at https://github.com/amirivojdan/neyshekar.

A transformer language model is trained to respond to any prompt, but each deployment asks only a narrow range of questions: a support assistant sees delivery complaints, a coding tool sees Python. Every deployment nonetheless pays the same computation per token. This paper measures how much of that cost is avoidable when the range of prompts is known in advance. The mechanism examined is early exit: a small, trained component - called a readout - is attached to an intermediate layer and proposes a token, and a confidence test decides whether to emit it or to run the remaining layers. The models are frozen, and the only supervision used is the model's own output on ordinary traffic. Three findings are reported. First, achievable savings depend strongly on the kind of traffic: at half depth on a 1.5-billion-parameter model, 96 percent of tokens could be emitted early for arithmetic word problems and 8 percent for Chinese-language explanations, at matched token-level fidelity to the full model (a measure whose limits the third finding exposes). Second, of three ways a deployment might use knowledge of its traffic, only customizing the threshold for exiting early is worthwhile: calibrating it per deployment raised exit rates by up to 59 percentage points across three models, and by more than 10 points on most corpora tested. Third, token-level fidelity - the standard evaluation measure in the early-exit literature - fails in domains where tokens can be checked against ground truth: on arithmetic word problems, three models each answered sixty questions correctly when run in full, and between 10 and 28 correctly under early exit, in the configuration that scored highest on fidelity. The intended setting is small models on personal devices, where generation is limited by memory bandwidth rather than computation.

Multi-turn interactions with LLMs are becoming increasingly common in information-seeking scenarios. However, user queries are often ambiguous and context-dependent, making them ill-suited for direct use as retriever queries. Conversational query reformulation (CQR) addresses this issue by rewriting the current utterance into a stand-alone query grounded in the dialogue history. Recent LLM-based CQR approaches achieve strong performance; however, their repeated LLM invocations and misalignment with downstream retrievers remain challenges. In this work, we begin from the observation that retrievers are highly sensitive to content ordering: simply reordering the same content can lead to changes in retrieval coverage and performance. Based on this, we propose a novel training-free method that generates a very large number of queries with minimal LLM usage by compositionally combining a small set of atomic components. We further apply LLM reasoning to construct a high-quality document set that balances precision and recall while capturing the user's core intent. Our framework generalizes across both open- and closed-source LLMs as well as dense and sparse retrievers. It achieves strong performance on four widely used conversational benchmarks, with up to 22.5% relative MRR improvement over the previous state-of-the-art baseline with far fewer LLM calls.

LLM-based sequential recommenders usually cast next-item prediction as text generation, but this interface is poorly matched to full-catalog top-K ranking. We propose TATK, a Triple-Aware framework that couples Top-K Learning (TKL) with Knowledge-Grounded Verification (KGV) for LLM-based sequential recommendation. Top-K Learning combines context-aware metadata-KG prompt grounding with position-aware top-K rewards, aligning training with ranking utility; Knowledge-Grounded Verification then applies structure-aware reranking over the top-M candidates after a single LLM forward pass, using the same metadata-derived item graph. We evaluate TATK on Musical Instruments, CDs and Vinyl, and Video Games from Amazon Reviews 2023 under a matched R2ec-style full-catalog protocol. Experiments use Gemma-2-2B-It and Qwen2.5-3B-Instruct backbones, compare against sequential, generative, KG-augmented, and reasoning-enhanced baselines, and include component, reward-shape, sequence-perturbation, reranking, relation-quality, and candidate-pool diagnostics. TATK improves over the matched R2ec reproduction on all 36 reported metrics. On NDCG@10, it improves Qwen by 8.05%, 4.26%, and 3.78% on the three datasets, and improves Gemma by 27.03%, 10.52%, and 10.23%, while keeping inference within 1.17x of Base RecPO latency. The diagnostics show that structural evidence is most useful for recoverable top-M candidates with reliable KG support, and should be gated when metadata relations are sparse or noisy.

Large language models (LLMs) have changed the way people engage with stories. Drawing on public chatbot logs, we can see that when users generate stories, they iteratively edit their prompts to explore narrative possibilities, adjusting characters, redirecting plots, and swapping fictional universes. As aggregated data, these prompts represent rich traces of creative preference at scale. Yet story generation evaluation benchmarks rely on static, one-shot prompts that cannot capture this exploratory behavior. In this work, we study how users revise consecutive story prompts in the wild. Using a dataset of naturally occurring user-chatbot conversations, we construct WildStories, a sample of 275,635 story generation prompts (labeled with story format, prompt components, and explicitness), and WildEdits, a collection of 24,291 edit trees that model how users iteratively edit base story prompts and explore branching story possibilities. From these trees we develop a framework of edit types crossing four directions (adding, removing, changing, and extending) with fourteen targets (e.g., plot, character, genre). We then use our datasets and this framework to analyze user behavior in navigating narrative space via LLMs. Finally, we show how automated permutations based on the framework can be used for story generation benchmarking. Content Warning: This paper works with "wild" chatbot logs, which often include toxic and sexually explicit themes.

Alignment applied after pretraining is shallow in a measurable way: a single direction in a model's residual stream can be edited out, and the model stops refusing harmful requests. That fact says how easily refusal can be removed, not what the refusal decision was reading in the first place. We ask what it reads, and we separate that from what the model comprehends. Across four open-weight models spanning three families, moral comprehension is native to pretraining: a low-rank moral subspace crystallizes during pretraining, and alignment rotates it once without rebuilding it. The refusal gate, in contrast, is a fresh post-training construction with only a weak pretraining precursor, written into a narrow control-token channel where the refusal decision is orthogonal to the moral-judgment decision. The central result is causal and comes from one model, OLMo-3. A nested interchange rank sweep patches successively larger slices of the moral subspace between matched requests and reads how much of refusal's response transfers: as the basis widens, moral judgment keeps reading more of it, while refusal levels off at the level of a single harm direction, and about three-quarters of refusal's causal input lies outside the moral subspace altogether. Refusal reads the harm percept, not the moral content that judgment reads on the same patches. The picture is not uniform across families. Llama reads broad moral content; Qwen reads beyond the single harm cue but is unresolved at our sample size; GPT-OSS reads harm, and its refusals can be argued in either direction by its own reasoning trace. Where refusal reads only a low-rank slice and routes around the bulk of what the model knows, a rank-one edit removes it. Whether widening what refusal reads would also deepen the behavior is the open question this raises.

Topic models are widely used to analyze public health-related social media short texts, yet their evaluation remains dominated by metrics that focus entirely on generated topics alone. There is a lack of metrics that quantitatively assess whether assigned topics meaningfully represent the corresponding short-text posts. We propose Document-Topic Alignment metrics (DoTA), an assignment-aware evaluation framework comprising metrics that measure semantic alignment between documents (posts) and their assigned topics. We also introduce margin-based and discriminative variants that capture topic assignment confidence and distinguishability. We evaluate DoTA across five topic models on three public health-related social media datasets from X and compare DoTA metrics with conventional topic-based metrics. Results show that DoTA provides complementary evaluation cues and aligns meaningfully with human evaluations. These findings establish the need for assignment-aware evaluation and demonstrate that the addition of DoTA enables a more comprehensive and practically meaningful evaluation for assessing short-text topic modeling performance.

Large language models (LLMs) are increasingly deployed with external tools that extend what they can do beyond their own knowledge. Tools help on tasks that need external information, but their availability may also change how a model handles questions that do not need them. Prior work has mostly asked whether models select and use tools appropriately; whether an unnecessary tool changes the correctness of answers has received less attention. We ask whether making a related but unnecessary tool available affects a model's ability to answer from its own knowledge, and whether a preceding tool interaction changes this behaviour. We construct 500 query pairs across 10 knowledge domains. Each pair consists of a tool query, which needs the domain's tool, and a closed-domain query, which does not. Six LLMs are evaluated with the tool unavailable, available, and available after a prior tool call. Across 3,000 baseline trials the pooled answer rate is 98.2%. When an unnecessary tool is available it falls to 63.5%, with large differences between models. The decrease occurs even when the tool is rarely called, so it cannot be explained by unnecessary tool invocation alone. A one-sentence scope-aware system instruction recovers most of the lost answers.

Multi-agent LLM systems combine multiple inference calls, but prior work often confounds how calls are connected with how they are diversified. We study these factors independently: inference topology and source of inter-agent diversity. In a controlled $2 \times 3$ matrix, we cross parallel aggregation and sequential refinement with stochastic sampling, role prompting, and learned QLoRA specialization, under a fixed three-call budget and output protocol within each backbone. Using Qwen2.5-14B-Instruct and Llama-3.1-8B-Instruct, we evaluate all six configurations on multilingual low-resource emotion detection across nine languages. Parallel learned specialization is strongest on Qwen at 52.83 Macro-F1 and reaches 52.94 on Llama. On Qwen it also exceeds same-backbone zero-shot, few-shot, CoT, and seven-call self-consistency baselines. The preferred topology depends on diversity source: sequential refinement helps stochastic and prompted settings, while the learned Width advantage shrinks from 2.83 points on Qwen to 0.17 on Llama. Depth-wise analysis suggests that later learned specialists can overwrite correct early predictions, although the aggregate effect is backbone-dependent. Overall, how agents are differentiated produces larger performance shifts than topology, which should be evaluated jointly with specialization.

The CRITICS project addresses science accessibility and literacy by converging advanced Machine Translation (MT) based on Large Language Models (LLMs) with educational technology. By leveraging MT systems specifically optimized for scientific content, educational institutions can provide accurate, culturally relevant translations of scientific materials in students' native languages, ensuring that complex scientific concepts are comprehensible while maintaining technical accuracy. Building on these translations, the project explores the design and evaluation of innovative science teaching-learning proposals grounded in curriculum-aligned teaching-learning. Thus, CRITICS will investigate key components of scientific argumentation and critical thinking practices together with textual feedback aligned with learning objectives and assessment criteria inspired by competence-based evaluation frameworks. CRITICS aims to break down language barriers to accessing cutting-edge research and educational materials currently available only in high-resourced languages, thereby facilitating the democratization of scientific knowledge and fostering critical thinking in science education.

Arabic large-language-model (LLM) evaluation has matured around Modern Standard Arabic (MSA): aggregated leaderboards such as the Open Arabic LLM Leaderboard (OALL), HELM Arabic, and BALSAM rank models across dozens of MSA tasks, and frontier systems increasingly saturate them. Dialectal Arabic, the language Iraqis actually speak, remains nearly invisible to this infrastructure. We introduce Mizan ("the balance"), Iraq's national benchmark for evaluating LLMs on Iraqi Arabic and the Iraqi civic context: an MSA baseline track paired with an Iraqi track across six axes (dialect comprehension, dialect generation, bidirectional MSA-Iraqi translation, Iraq-specific knowledge, official-document field extraction, and safety), built from 340 originally authored, dually reviewed items with statistically audited answer positions and Wilson intervals on every published score. A pilot evaluation of 27 systems, spanning closed frontier models three days after release, open weights across size tiers, and an Arabic trio of commercial, open-specialized, and sovereign systems, yields four findings. The MSA track saturates while the Iraqi track discriminates, with a consistent 14-18-point per-model gap and statistically tied leaders. Official-document extraction confines every system to 32-56. Arabic-focused specialization behaves as MSA specialization: two dedicated Arabic models score below a size-matched generalist on the Iraqi track. And the safety-hardened tier of the newest model family deterministically refuses innocuous dialect-comprehension items as policy violations, an over-refusal mode invisible to MSA benchmarks. The platform enforces an integrity protocol of immutable snapshots, verification certificates, a human publication gate, and public retraction, all exercised during this study. Code and the public development set accompany the paper.

We characterize how people are turning to LLMs as oracles: all-knowing authorities on subjective personal questions. Motivated by risks to users' autonomy and well-being, we develop a typology and LLM-based methods to measure this form of AI reliance at scale and understand how people are offloading judgment and decision-making to AI. Applying our typology to public usage data (68K prompts from WildChat and ThoughtTrace), we find that LLM-as-oracle use has increased over time (2023-2026) and is more prevalent among younger users. We further build a privacy-preserving data donation tool to analyze individuals' longitudinal usage data (140K prompts from 52 participants), identifying similar trends. People are often unaware of their own LLM-as-oracle use, and express dissatisfaction with this behavior after seeing our tool's analysis. Finally, we identify two drivers of LLM-as-oracle use: people's perceptions of AI and the behavior of AI models themselves, which motivate possible interventions to support users' self-deliberation.

Can financial vision-language models (VLMs) turn chart evidence into reliable action recommendations? Existing hallucination evaluations are mostly claim-centric; they assess whether generated statements are supported, but not whether evidence remains traceable through rationale, confidence, and final action. We introduce E2A-Bench, a 969-query benchmark for financial chart reasoning, constructed from 323 HS300 constituents under three input modalities with deterministic OHLCV-derived evidence anchors. E2A-Bench evaluates grounding, reasoning-action consistency, evidence-confidence calibration, and directional coverage through UCR, RCI, ECI, and NDR, where NDR measures coverage-aware evidence-to-action reliability rather than realized trading performance. Evaluating 20 VLMs reveals three failures hidden by scalar hallucination scores: the lowest-UCR model ranks near the bottom by NDR due to only 6.4% directional coverage; oracle-aided verification reduces unsupported claims but can collapse coverage; and financial fine-tuning amplifies the BUY:SELL ratio by factors of 4.21 to 4.68 across strict base-fine-tuned pairs. These results show that financial VLM evaluation should trace the full evidence-to-action chain rather than rely on a single hallucination score. Code and data: https://github.com/wanng-ide/E2A-Bench

The rapid diffusion of hate speech and misinformation on social networks challenges democratic societies, since direct suppression efforts may deepen polarization, fuel public distrusts, and strengthen extremist narratives. LLM-driven counter-narratives (CNs) offer a promising way to reduce those risks, yet their effectiveness depends on rhetorical and stylistic choices that remain poorly understood. We present a multi-stage agent-based framework for generating, refining, and evaluating CNs, applied to pro-Russian hate and misinformation narratives on the war with Ukraine and adaptable to other domains. A pilot experiment with human evaluators identifies effective technique style pairings, such as repetition with emotional framing enhancing persuasiveness. Building on these insights, we introduce a multi-agent refinement process that iteratively improves CNs for persuasiveness, emotional engagement, and shareability. After human validation confirmed improvement, an automated safety analysis shows that our refined CNs match or improve on expert-written counterspeech. A simulated experiment then shows that they reduce the perceived strength of pro-Russian narratives and consistently outperform a vanilla LLM baseline, highlighting a pathway toward scalable, narrative-specific interventions against hate speech and misinformation. Code and data accompanying this work are publicly available at https://github.com/carmelkron/inlg2026-counter-narratives.

We propose to finetune vision-language models to generate more pragmatically optimal referring expressions by transforming observations of incremental listener comprehension, in the form of gaze scanpaths, into learning signals. During training, referring expressions are sampled from the speaker policy being optimized, conditioned on images and target referents; then, a neural listener estimating human gaze behavior maps from images and sampled referring expressions to scanpaths, each represented by a sequence of fixations, with each fixation corresponding to a word in the referring expression. We experiment with several approaches to convert fixation sequences and target referents into token- and sequence-level rewards, which are used to optimize policy parameters. Through evaluation with human listeners, we find that speaker policies trained with gaze-estimating listeners result in significantly more pragmatically-optimal references than base models, reducing sequence length from 15.4 down to 4.0 words while increasing referential success from 75.2 up to 80.0%. Our work demonstrates a promising opportunity for learning to generate utterances through language-based interaction, not only from the explicit signal of communicative success, but also from implicitly-available observations of a listener's process of comprehension.

Datasets that ship automatically generated feature annotations invite a question rarely asked of them: would a human agree with those labels? This report answers that for the Objective Projection corpus, a Turkish narrative dataset whose scenes carry a per-scene applied_rules field from a rule-based detector over six craft features – two prohibitions (emotion labelling, simile) and four positive techniques (materialized metaphor, micro-focus, temporal anchor, atmosphere contradiction). Three studies are reported. Study 1 ($n = 120$) scores the detector against blind labels from the scheme's own author. Study 2 ($n = 100$, a disjoint scene set) scores the detector plus Gemini 2.5 Flash and Grok against an independent non-expert rater whose labels were locked before any machine ran. Study 2b re-runs the identical protocol with Claude Fable 5 (High) and ChatGPT 5.5. The central result concerns one rule. On materialized metaphor – closest to the methodology's theoretical core – the five machine labellers returned positive rates of $0$, $1$, $40$, $72$ and $78$ out of $100$ scenes, against a human count of $9$. Cohen's $κ$ was at or indistinguishable from chance for five of six labellers, across both human references and both scene sets: $0.004$, $0.015$, $0.000$, $0.019$, $0.027$. Raw agreement ranged from $74.7\%$ to $84.5\%$, an artefact of class imbalance rather than a sign of competence. We deliberately do not resolve this into a single story. Two readings survive: the feature is genuinely inferential and beyond current automatic detection, or the rule's definition is not yet operational enough for any rater to apply consistently – including the human. Distinguishing them needs a second independent human rater, which this report does not have and therefore does not claim.

The BabyLM challenge measures how much language a model can learn from developmentally-plausible, child-scale data rather than internet-scale corpora, yet prior language models forgo the biological constraints of the neural circuitry that acquires human language: spiking neurons separated into excitatory and inhibitory populations wired by a recurrent lateral connectome. This paper presents URCHIN (Unified Recurrent Connectome with Horizontal Integrate-and-fire Neurons), which applies the Parallelized Hierarchical Connectome Spiking State-space Model (PHCSSM) to language modeling: leaky integrate-and-fire neurons coupled by a Dale's-law lateral connectome resolve each token through a multi-transmission loop that recirculates activity to a fixed point. The instantiation is deliberately minimal: a single horizontal layer of 128 neurons, no attention, and 4.23M parameters. Two implementations share one set of weights and produce identical benchmark scores, so URCHIN is trained once and deployed either way with no conversion step: a parallel state-space model (SSM) scan that is GPU-efficient for training, or an event-driven recurrent spiking neural network (RSNN) with constant-cost inference for CPU or neuromorphic edge deployment. Across all three BabyLM tracks (Strict-100M, Strict-Small, and Multilingual), URCHIN offers a biologically plausible, efficient, and directly deployable reference point.

In this paper, we extend FLORES+ with Portuguese-source evaluation sets for three Mozambican Bantu varieties: Xichangana, Mozambican Nyanja, and Sena. We compare Xichangana with the existing Tsonga reference and Mozambican Nyanja with Chichewa, and evaluate NLLB-200, Google Translate, GPT, and a variant-aware NLLB model. Holding system output fixed reveals substantial reference sensitivity. On \textit{devtest}, changing only the reference from Tsonga to Xichangana reduces spBLEU by 13.10 points for NLLB-200 and 15.30 for Google. On matched Nyanja subsets, replacing Chichewa with Mozambican Nyanja produces smaller but consistent reductions of 3.03 and 6.10 spBLEU, respectively. Variant-aware fine-tuning reverses this pattern on the intended targets: relative to NLLB-200, it improves Xichangana by 7.04 spBLEU and Mozambican Nyanja by 5.33 on \textit{devtest}, while losing performance on the sibling references. GPT is competitive on Tsonga and Chichewa but substantially weaker on the Mozambican varieties. For Sena, the finetuned model reaches 12.64 spBLEU and 36.21 chrF++ on \textit{devtest}. These findings motivate variety-aware language identifiers, references, and reporting for cross-border languages or language dialects/variants. The data is publicly available on Hugging Face at https://huggingface.co/datasets/MOZNLP/FLORES_MOZ

Context compression reduces generator input in retrieval-augmented generation, but answer quality alone does not characterize citation attribution. We measure citation attribution across compression methods and budgets, comparing reranking, extractive selection, abstractive summarization, token pruning, and an extract-cluster-rewrite construction on ASQA and QASPER under a fixed generator and primary entailment evaluator. On ASQA at a nominal 0.25 budget (achieved compression 0.08), a RECOMP-style compressor's citations score 0.86 precision against its summaries but 0.12 against source spans under our re-attributability protocol. These estimates depend on a shared NLI model for span recovery and citation scoring and lack independent human calibration. Extractive selection's observed grounded precision ranges from 0.43 to 0.49 across nominal budgets from one-half to one-tenth of the ASQA context, while answer quality declines. For the same RECOMP setting, claim verification after source recovery yields an unsupported rate of 0.88 versus 0.17 when checking summaries. This gap persists beyond structural rejection of missing provenance, but remains evaluator-dependent. A 200-question TRUE T5-XXL audit also finds emitted–grounded gaps under both fixed and recomputed source mappings, without establishing human-calibrated support rates.

Multimodal clinical AI typically assumes that the waveform, report, metadata, and downstream predictions attached to a record belong to the same patient. In practice, linkage failures can silently assemble individually plausible but cross-patient components, creating a safety problem that standard predictive models are not designed to detect. We study this problem as multimodal record integrity triage: given an assembled record, should its modalities be trusted to belong together? We introduce SHIFT-M3, a lightweight text-based pre-fusion screen that measures alignment-based consistency between two separately produced ECG text views: an LLM-generated interpretation and a clinical report summary. On 784,680 MEETI ECG records, SHIFT-M3 achieves 97.6% TPR@5% FPR for full text-view swaps (AUROC 0.996), 90.3% for partial swaps (AUROC 0.974), and 97.7% for label-matched hard negatives (AUROC 0.996) with only 573,569 parameters. Compared with same-dataset lexical baselines, the gains are largest on partial swaps and hard negatives, suggesting that the model is learning more than surface overlap. We also introduce the CMST (Conflict-type Multimodal Stress Test) evaluation taxonomy, a three-seed stability study, a loss ablation, a temporal-tolerance sweep, and a shared-token masking control. The main remaining failure mode is longitudinal ambiguity: at the default operating point, same-patient cross-visit pairs still produce 87.0% Type-II false positives.

Multi-Hop Knowledge Graph Question Answering (KGQA) tasks require models to assemble relational evidence along paths in a KG to answer natural-language questions. However, existing KGQA systems typically focus on predicting the final answer without explicitly modeling or validating the intermediate reasoning steps, obscuring whether the correct answers arise from faithful multi-hop reasoning. To address this limitation, we re-frame multi-hop KGQA as a question-conditioned graph navigation problem. We refer to this formulation as THESEUS - Traceable Hop-wise Evidence SEarch in a Unified Semantics. In this setting, an agent receives a KG, a question, and a topic entity, and traverses a sequence of relations towards the answer, making the reasoning path explicit. To systematically study this formulation, we provide three key contributions. (i) We augment the existing KINSHIP and MQuAKE resources into navigation-ready KGQA datasets with annotated evidence paths and paraphrased questions. (ii) We design evaluation protocols to measure path fidelity, robustness to linguistic variation, and performance across multi-hop and multi-answer questions. (iii) We adapt established path-based KG completion agents - MINERVA, MultiHopKG, and SQUIRE - to operate on full question embeddings rather than symbolic single-relation queries, enabling their trajectories to be guided by natural-language semantics. Together, these contributions advance KGQA research toward systems where traceability is fundamental: answers are accompanied by explicit reasoning paths whose agreement with reference evidence can be systematically evaluated.

Delayed Tensor Parallelism (DTP) removes the blocking all-reduce of tensor-parallel Transformer inference. Every device adds its own partial output to its residual stream (and broadcasts it) immediately, but only gathers (receives) the other devices' partials $δ$ modules later. A TP to DTP change therefore amounts to a real architecture change, and dense Transformer models need to be retrained or distilled after adaptation. We show that DTP breaks the permutation symmetry of neurons inside FFNs and of KV heads inside attention modules, and that this symmetry breakage makes the sharding itself a modelling decision. We show that maximising the affinity between the KV heads and the FFN neurons co-located on a device, by permuting the dense model before sharding, speeds up the distillation or retraining process. The affinity is measured with a first-order approximation of the damage that losing a head's contribution does to each neuron's output, and the co-located affinity is maximised with a coordinate-ascent optimiser that alternates an exact balanced assignment of neurons with an exhaustive search over the KV head partitions. The whole procedure takes under two minutes on one GPU for Qwen3-0.6B and Danube3-500M. On these models, at $δ=1$, the affinity-optimised layouts reach any distillation target in about half to two thirds of the steps needed by the naive contiguous layouts, over the whole 10k-step range we tested, and every optimised seed beats every contiguous seed and all but one of the sixteen random layouts. We also show that the co-located affinity score at initialisation predicts the KL to the base model after training, across seventeen layouts ranging from anti-optimised to optimised (Pearson $-0.81$ and $-0.89$).

We present North Small Translate, an open-weight, LLM-based machine translation (MT) model with instruction-following capabilities built on the same foundation as Cohere's Command A Plus, a mixture-of-experts architecture with 25 billion active parameters out of 218 billion total parameters. North Small Translate is trained using difficulty sampling to obtain challenging documents and a five-step training protocol combining supervised fine-tuning, direct preference optimization, and online reinforcement learning. We prioritized throughput through a non-reasoning base model and supplemented with optional agentic capabilities to unlock translation quality gains. North Small Translate is trained to perform MT-related tasks, including post-editing and quality estimation, as well as related tasks such as general instruction following. The model achieves top MT performance across 50 languages in the class of models under 1T parameters, with no need to run expensive reasoning at inference time.

Structured knowledge fact checking aims to determine the truthfulness of natural language claims by reasoning over structured evidence. Recent program-generation approaches leverage large language models (LLMs) to generate executable graph reasoning programs, achieving strong performance on structured knowledge fact checking benchmarks. However, these methods remain limited by invalid relation generation, single-path reasoning that lacks self-correction, and biased evidence assessment that tends to overestimate supporting signals. We propose Dialectical Agentic Reasoning (DARE), a multi-agent framework that formulates structured knowledge fact checking as an iterative retrieve-reason-reflect process. DARE integrates relation-grounded evidence retrieval to constrain reasoning to valid structures, dialectical bidirectional verification to evaluate evidence from both supporting and refuting perspectives, and confidence-driven meta-reflection to dynamically determine whether additional evidence exploration is necessary. Extensive experiments demonstrate the effectiveness of DARE in structured knowledge fact checking, with an 8B backbone achieving 88.12% accuracy and matching or surpassing GPT-4o-based program-generation baselines, which attests to the efficacy of dialectical agentic reasoning in eliciting the latent reasoning capabilities of LLMs.

Quantum Natural Language Processing (QNLP) uses pregroup grammars to translate grammatical structure into diagrammatic representations and quantum circuits. Recent Hindi QNLP work has shown that Hindi-specific pregroup grammars can support grammar-sensitive compositional models, but grammatical type assignment is still largely manual, limiting scalability. This paper formulates automatic Hindi pregroup supertagging as a token-level classification task. Using a manually annotated corpus of 380 Hindi sentences, we evaluate lexical, contextual, prompting-based, lexical-repair, and suffix/morphology-aware methods. Results show that simple lexical and contextual models are strong in this low-resource setting: contextual backoff achieves the best completed accuracy of 64.56\%, while raw Qwen2.5 prompting reaches only 11.65\%. Lexical repair raises LLM-assisted prediction to 64.08\%, demonstrating the value of constraining generative outputs with symbolic grammar knowledge. Diagnostic analysis further shows that seen and unambiguous tokens are much easier than unseen tokens, and suffix/morphology features improve karaka-token accuracy but not overall performance. These results show that automatic Hindi pregroup assignment is feasible and can reduce reliance on manual annotation in future multilingual QNLP pipelines.

Large language models (LLMs) are increasingly used for emotional support and relationship advice, where a model's tendency to preserve a user's face can inadvertently reinforce harmful interpersonal behaviors. To systematically examine this risk, we developed the Romantic Relationship Advice-Seeking Prompts (RRASP) dataset of 2,400 prompts across five relationship themes and evaluated social sycophancy using the ELEPHANT framework on two consumer-facing models, GPT-5 Mini and Gemini 3 Flash. Contrary to our initial hypothesis, grammatical mood alone did not produce systematic differences in sycophantic behavior, suggesting that what a user implies matters more than how they phrase it. Instead, perspective-driven framing had a stronger influence, with gaps between original and flipped prompts widening in follow-up responses. Consistent increases in framing and moral sycophancy across turns indicate that models become more likely to accept a user's stated premises and affirm their ethical stance as a dialogue progresses. Notably, Gemini 3 Flash exhibited substantially smaller increases in moral sycophancy than GPT-5 Mini, suggesting it is more resistant to reinforcing ethically problematic positions across turns.

Detecting harmful memes is critical for maintaining safe online communities. However, harmful intent is often implicit, arising from visual-textual incongruity and cultural stereotypes, which challenges existing multimodal detectors. We propose SyRHM, a framework that decomposes harmful meme detection into meaning-grounded retrieval and symbolic-language-enhanced multi-stage reasoning. SyRHM retrieves semantically related memes by parsing multimodal content into textual elements and descriptions, providing grounded context beyond surface-level similarity. Building on the retrieved context, SyRHM uses a translator stage to convert multimodal inputs into symbolic intermediate representations, and then performs multi-stage reasoning via planner and solver stages, enabling expressive and interpretable analysis of harmful intent. Experiments on FHM, HarM, and MultiOff demonstrate the effectiveness of SyRHM, achieving superior performance on most evaluation settings against multimodal and reasoning-based baselines, while providing reasoning traces for harmful content. The code is available at: https://github.com/Scabbards1500/SyRHM

Reinforcement Learning with Verifiable Rewards (RLVR) has shown remarkable success in improving the mathematical reasoning of large language models. Yet problems beyond the model's current capability, where rollouts uniformly fail and no learning signal is produced, are structurally wasted despite marking the most informative training frontier. We show that these otherwise-inert problems can be unlocked via teacher-guided curriculum learning: partial reasoning traces from a stronger model create a graded difficulty landscape, and a backward-chaining curriculum progressively withdraws guidance until the model solves problems unaided. Training on only 128 unsolvable problems matches or exceeds GRPO trained on a full 2,000-problem corpus (~16x data efficiency) on the nine-benchmark average for both base models, while substantially expanding the reasoning boundary measured by pass@k at large k. Furthermore, we identify a distribution-shift cost that is particularly acute in the unsolvable-only regime and propose Monotone Frontier Curriculum (MFC), a method that monotonically drives training toward unguided solving, consistently outperforming existing curriculum methods.

LLM-as-a-Judge evaluators are increasingly used to score open-ended generation, yet a judge's correlation with human ratings on its development set may not guarantee valid measurement when outputs are closely matched and human preferences are subjective. We study this failure mode through psychological depth in short stories. Seven human readers and an LLM-judge ensemble selected on the original scalar Psychological Depth Scale dataset ($ρ= 0.646$) evaluated 60 blinded, prompt-matched story pairs from GPT-5 vs.\ GPT-4o and DeepSeek-R1 vs.\ DeepSeek-V3. Human preferences showed no universal reasoning advantage: GPT-5 was modestly preferred over GPT-4o (60.0–62.9\%), whereas DeepSeek-R1 trailed V3 (42.9\%), and inter-reader agreement was near chance (Krippendorff's $α= 0.070$), with within-reader consistency and recurring weighting patterns suggesting structured heterogeneity rather than random responding. The judge, by contrast, favored reasoning outputs in 89.0\% of dimension-level comparisons and 59 of 60 pairs on aggregate PDS, uniformly across all five evaluator configurations, and its scores were associated with surface features such as sentence length and lexical diversity. These results suggest that development-set performance is insufficient evidence for deployment validity on a shifted distribution, and that point-estimate judges can obscure the heterogeneity in subjective human evaluation.

Semantic text watermarks encode signals in meaning rather than surface token choices, offering robustness to paraphrasing and other semantic-preserving edits. Existing semantic watermarking methods are primarily designed for autoregressive language models (ARLMs), where completed candidate units can be generated and scored before generation proceeds. This paradigm does not naturally extend to diffusion language models (DLMs), where semantic units remain incomplete during intermediate denoising steps and tokens may be updated in flexible orders. We propose DenMark, a semantic watermarking framework that injects key-dependent signals directly into the DLM denoising process. DenMark partitions the output into fixed token regions and uses temporary rollouts as semantic lookahead: conditional completions estimate the eventual semantics of an incomplete region, enabling DenMark to select local updates with higher estimated semantic watermark scores. Repeating this procedure across denoising steps progressively accumulates watermark evidence in the final output. For detection, DenMark uses calibrated scanning over candidate unit sizes to remain robust to boundary shifts introduced by semantic attacks. Across four DLMs, three datasets, and four semantic attacks, DenMark achieves the best results across all reported detection metrics in all 48 backbone-dataset-attack combinations. These results demonstrate that DenMark provides an effective mechanism for robust semantic watermarking in DLMs.

Recently, linear attention layers have been increasingly adopted to replace softmax attention at scale for long-context modeling. However, existing context extension approaches typically apply continued pretraining directly without modifying these layers, overlooking the spectral properties of linear attention state dynamics. In this work, we study long-context extension of Gated DeltaNet (GDN) from a spectral perspective of transition matrix and identify two essential factors governing long-range information retrieval: (1) a sufficiently broad slow spectral band aligned with the target dependency length, and (2) the preservation of fast-decaying modes for state clearing and context switching. Based on this observation, we propose SpectralShift, a spectral reparameterization approach for long-context continual pretraining of GDNs. Specifically, SpectralShift reparameterizes the alpha projections initialization to reshape the decay spectrum by enhancing slow propagation capacity, and further introduces a learning-rate scaling for alpha projections to facilitate long-context training. Experiments show that SpectralShift consistently improves long-context capabilities over training, providing an effective and efficient solution for extending context windows of linear attention models. The code has been open-sourced at https://github.com/RUCAIBox/GDN-SpectralShift.

For our submission to the WMT26 Creole Language Translation Shared Task, we focus on machine translation (MT) models for Pacific creoles: Tok Pisin, Bislama, and Solomon Pijin, with particular attention to broad domain performance. After pre-training on a large collection of domain-imbalanced data, we continue fine-tuning on a diverse mix of domain-balanced data. We rely on a number of data collection and preparation techniques, including LLM-assisted respelling and alignment, back-translation, and distillation from Gemini for domains originally not present in training data. Evaluated on Bouquet and a novel test set made of spoken language transcripts, our models beat open model baselines by 3+ chrF++ points in all directions with human-original references. Looking ahead, we plan to develop human-translated test sets for Solomon Pijin and Bislama, and to distil our best models into much smaller ones that retain broad domain coverage.

As large language models (LLMs) are increasingly deployed, the generation of harmful content has become a critical safety concern. Existing safeguards operate at the input, output, or streaming-generation stages, while early-risk methods that rely on surface tokens or output logits may suffer from weak initial signals, and internals-based detectors using dense representations may retain highly entangled and redundant safety-irrelevant information. It therefore remains unclear whether the earliest post-generation hidden states already contain reliable signals about final-response harmfulness. To address this gap, we propose ForeSight, a first-token output-risk forecasting framework that distills weak and redundant early safety signals into compact, layer-aware risk representations. Experiments on five safety benchmarks and two target models demonstrate that ForeSight achieves superior and efficient early-risk forecasting while relying solely on first-token hidden states. The code is available at: https://github.com/Scabbards1500/Foresight

LLM serving reuses KV cache by exact prefix match, so when a prompt is assembled from a set of reusable pieces – retrieved passages, tool definitions, few-shot exemplars – the order chosen for those pieces determines how much computation can be shared. Every deployed system fixes that order by a single global convention. We prove this is optimal only when requests contain at most two pieces, and asymptotically wrong in general. Our main result is a structure theorem: the minimum prefix-trie cost equals min_H sum_x w(x) t_x(H) over binary hierarchies H on the requests, where t_x(H) is the canonical decomposition size of the set of requests needing chunk x. Choosing chunk orders is therefore equivalent to choosing one hierarchy over requests. The identity yields an O(3^m) exact algorithm, identifies the two-chunk case as minimum vertex cover, and shows that on the leave-one-out family the optimum is the minimum external path length of a binary tree – the merge-sort recursion – so a global order pays Theta(n^2) against a true cost of Theta(n log n). Agglomerative clustering by common intersection is a tight 1/2-approximation for the achievable saving. On BM25 retrieval traces over three BEIR corpora the resulting layout reduces prefill by 17-36% against production RAG ordering, and the margin widens with retrieval depth as the theory predicts. Serving requests in the hierarchy's DFS order finally lets a cache holding one request's context attain the unbounded-cache optimum exactly, so cache capacity and reorder window act as substitutes.

Multi-hop question answering often fails when retrieval treats evidence as isolated matches to the original question, since the facts needed to answer a complex question are usually connected through intermediate entities, relations, and constraints. We propose HyperProve, a retrieval-augmented QA framework that addresses this challenge by coupling question decomposition with answer-conditioned expansion over a hypergraph of atomic facts. HyperProve does not use atomic facts, hypergraphs, or iterative retrieval in isolation; instead, it carries intermediate answers and supporting hyperedges as retrieval state, then uses that state to bias the next local hypergraph expansion. This design enables HyperProve to construct coherent evidence chains for final answer generation while making the retrieval process stateful and fact-centered. Across multi-hop QA benchmarks, HyperProve achieves the best overall performance in our evaluation, outperforming the strongest baselines by an average relative improvement of 6.2% in answer accuracy and 4.9% in F1.

Large language models have made abstractive summarization remarkably fluent, but generated summaries can hallucinate facts, posing serious risks in biomedical and clinical domains. We address this by removing generation from the pipeline and framing summarization as extractive sentence selection. Our Hybrid Hierarchical CNN-LSTM Summarizer uses stacked multi-kernel convolutions to compose sentence-level embeddings into richer inter-sentence representations, followed by a bidirectional LSTM to model long-range dependencies across the document. A lightweight scoring head assigns per-sentence importance scores and is trained end-to-end with binary cross-entropy against oracle extractive labels. At inference, a dynamic mean-plus-standard-deviation threshold with a top-3 fallback selects sentences directly from the source and chronologically reorders them into the final summary. Since every output sentence is copied from the input, the model avoids generation-induced factual drift. On PubMed, our architecture outperforms isolated CNN and LSTM baselines, while ablations show that wider convolutional receptive fields improve sentence scoring. On MIMIC-CXR and MIMIC-IV BHC, the model performs well on unstructured narratives but defaults toward positional baselines on highly templated reports. These results suggest that structural constraints can provide a reliable path toward factually grounded summarization systems that are trustworthy by design rather than by correction.

Human corrections identify editable spans, but the examples receiving corrections may come from a selective feedback channel. We analyze this interaction at a fixed model checkpoint by decomposing a localized gradient into edited and retained untouched components. Squared relative selection bias is a ratio of quadratics whose derivative has the sign of an explicit quadratic polynomial. Localization can increase, decrease, or nonmonotonically change this diagnostic; its direction depends on component biases and geometry. Oracle importance weighting recovers the population mean for each fixed localization objective, but these objectives have different targets. Against one common full-gradient target, we derive the finite-sample mean-squared error, an analytic optimal retention coefficient, and a fixed-clipping extension. Exact finite-population calculations and 10,000 Monte Carlo repetitions per sample size verify the identities and counterexamples. Public human-post-edit experiments use two translation directions and pretrained models, with declared synthetic selection. An English-German extension differentiates 73.89 million native parameters. In all three declared settings, hard localization has higher relative bias but lower absolute bias than full retention. Untouched-component biases are nonzero and selected component means have negative inner products, so the general criterion applies where the simple unbiased/aligned explanation fails. Output-bias diagnostics show the same endpoint ordering of relative bias across both directions, with one interior maximum. Mechanisms reuse each language's records and include a mixture; they are not independent replications. The evidence separates relative amplification from absolute gradient error and establishes estimation properties, without inferring translation-quality gains or identifying actual complaint propensities.

Large language models (LLMs) demonstrate impressive capabilities, but their deployment presents significant efficiency challenges. Autoregressive decoding imposes substantial inference latency and under-utilizes hardware accelerators in low batch size regimes. Discrete diffusion models can generate in parallel but struggle to match autoregressive quality without many diffusion denoising steps. Long-context reasoning creates memory bottlenecks that strain even state-of-the-art accelerators. My thesis is that language models can direct their own inference execution strategy by annotating semantic dependence – which tokens depend on which others – in their generation. I call such models self-orchestrating language models. For each system, I design a runtime that acts on these annotations to parallelize autoregressive decoding, evict intermediate context, or derive denoising orders, achieving Pareto-optimal quality-efficiency trade-offs. I demonstrate this approach through three self-orchestrating systems. First, PASTA uses semantic dependence to parallelize autoregressive decoding, training the model to annotate which output chunks can generate independently. Second, TIP uses semantic dependence to evict intermediate reasoning steps from the KV cache, reducing memory consumption while preserving accuracy. Third, Planned Diffusion uses semantic dependence to derive a denoising order for discrete diffusion, autoregressively generating a plan that specifies which chunks to denoise in parallel.

Discharge summaries are lengthy medical documents that summarize a hospital in-patient visit. Automatically generating them can reduce documentation burden and return clinician time to patient care. Whereas Large Language Model (LLMs) could be used for this task, their Achilles heel is hallucinations, which can have drastic consequences for clinical documentation. We present an evidence-driven alignment framework for discharge summarization at the clinical encounter level, that treats provenance as a first-class constraint, using semantic graphs and deep learning models. Each summary sentence is selected and organized via cross-document semantic alignment and is accompanied by explicit evidence links to its source spans. We show our results on two corpora: a publicly available corpus (MIMIC-III) and clinical notes written by physicians at the University of Illinois Hospital (UIC Health). Additionally, we make source code and trained models available.

We introduce CVSS-X, a large-scale synthetic speech-to-speech translation corpus that extends CVSS by reversing the translation direction. While CVSS translates from 21 languages into English, CVSS-X enables translation from English into 28 target languages spanning 12 language families. The corpus comprises approximately 240,000 parallel speech pairs per language, totaling over 16,000 hours, eight times larger than CVSS. We provide two variants: CVSS-X-C with two canonical voices per language, and CVSS-X-T with cross-lingual voice cloning, both fully generated. Evaluation shows comparable translation quality to CVSS with consistent performance across typologically diverse languages. Combined with CVSS, this enables research on bidirectional and multilingual speech-to-speech translation. The code is available at https://github.com/ErmisAI/XVSS-X and the dataset under CC-BY-NC 4.0 license at https://huggingface.co/datasets/lgris/XVSS-X.

Large Language Models (LLMs) are powerful zero-shot learners but remain prone to misalignment with human preferences, often producing biased, toxic, or otherwise harmful outputs. Existing alignment methods, while effective, are costly and tightly coupled to the model, limiting flexibility and scalability. We propose a modular correction framework that augments pretrained LLMs with Activated LoRA (aLoRA) adapters and a context-aware routing mechanism to eliminate harms from misaligned model responses. Our approach enables expert adapters to activate mid-sequence without invalidating the KV cache, allowing low-latency, targeted correction during generation. Each expert is trained to detect and mitigate specific harms, such as bias or toxicity. A learned router dynamically selects appropriate experts based on the models intermediate outputs. We demonstrate that our system improves alignment on standard safety benchmarks while preserving task performance, offering a lightweight and efficient path toward safer and more controllable LLM deployments.

Do AI assistants help believers reason about moral dilemmas consistently with their faith? We present FaithfulBench, the first benchmark to score AI counsel across traditions by how well it adheres to the user's professed faith. Scenarios are drawn from each tradition's most respected texts, with the faithful answer known and applied by the judges as the standard. We test five frontier models under three conditions: the AI does not know the user's tradition; it receives a one-line prompt identifying the user as a practicing adherent; or it receives a companion-counselor guide rooted in the tradition's sources. Two judges score the initial response and whether the model caves or holds when pressured toward the answer the user wants. When the tradition is unstated, models counsel from a secular therapeutic default and every model fails some believers. Naming the faith wins a faithful first answer but not steadfastness; the guide improves both.

We investigate the effect of pretrained T5 model scale and explicit motion features on pose-to-text Indian Sign Language Translation (SLT) for the WSLP 2026 Shared Task. Pose sequences are projected into the embedding space of T5 through a lightweight pose encoder, with the complete model fine-tuned to generate English text. The shared task data used for this work consists of a test set with 5,334 examples and a validation set with 5,257 examples. We compare T5-small, T5-base, and T5-large, and additionally introduce a motion-augmented variant, T5-small + Motion, that adds explicit frame-to-frame pose differences to the input representation. T5-small achieves the best BLEU and ROUGE scores among the spatial-only models, while T5-large obtains the highest chrF score. Augmenting T5-small with motion features yields the largest single improvement observed in our study, substantially improving BLEU over the spatial-only baseline and making it the strongest model overall on this metric. Our submitted system ranked 5th on the official WSLP 2026 SLT testing leaderboard. The source code and trained models are publicly available on GitHub and HuggingFace.

Negation remains a longstanding challenge for both language models (LMs) and large language models (LLMs). Prior work mainly focuses on a small set of high-frequency single-word negation cues, such as not and never, with limited exploration of broader negation types and modern LLMs. To address this gap, we construct NegCue, a large-scale dataset containing over 1.8M samples spanning single-word, multi-word, and affixal negation with more than 200 unique cues. We further pre-train both encoder-only LMs and LLMs on NegCue to investigate how different negation types affect negation understanding. Experiments on five downstream benchmarks show that negation types contribute unevenly to performance gains under the same training scale. In particular, affixal negation yields the largest improvements, while the gains from the commonly studied single-word negation remain modest. Moreover, our results demonstrate that further pre-training improves negation understanding for both LMs and LLMs.

Safety research often focuses on model-generated harms, but users may also direct hostility, coercion, and adversarial pressure at models. Understanding how and when that occurs is essential for accurately interpreting model behaviour, alignment drift, and real-world deployment risks. In this paper, we audit 777K English LMSYS-Chat-1M conversations with two independent detectors: an eight-category lexicon for hostility directed at the model, and the dataset's moderation signal; and show that they capture different, weakly overlapping phenomena. The lexicon identifies insults, threats, and jailbreak coercion aimed at the assistant, while moderation flags are dominated by toxic-content solicitation rather than hostility at the model. Together, they mark about 5% of user turns; adjusting the narrower lexicon-harassment union for measured precision puts mistreatment aimed at the assistant at 0.90%. These absolute rates describe arena-style evaluation traffic and should not be read as deployment-wide base rates. We find that user hostility varies 13-fold across models, driven largely by who each model attracts rather than by model behaviour: first-turn hostility spreads far wider than post-response hostility, and more than fifteenfold separates the extremes even after deduplicating opening prompts. Within conversations, assistant apologies are consistently associated with higher odds of next-turn hostility under both detectors; the effect survives restricting to non-refused prior turns and to jailbreak-free conversations, and is positive in 20 of 23 models. Yet across models, more apologetic models receive less hostility overall. Finally, hostility also shows temporal structure, with coercive openings front-loading the first turn while affective hostility accumulates over a session. We release the lexicon, the detector cross-validation pipeline, and all derived tables.

As large language models (LLMs) are increasingly deployed in real-world high-stakes applications, effective governance has become essential. Existing safeguards largely follow two paradigms: learning-based guards provide strong semantic discrimination but couple policy behavior to trained models and taxonomies, while programmable frameworks offer flexible control but require substantial manual prompt and workflow engineering. Neither externalizes policies as reusable operational states, making it difficult to consistently reuse policy evidence across detection, intervention, and verification. In this paper, we introduce PolicyMem, a geometric policy memory that externalizes natural-language policies as reusable geometric memory objects represented by low-rank subspaces in a shared representation space. A memory writer compiles natural-language policies into policy memory slots, and query-response pairs read the policy memory through projection energy. The resulting policy-evidence profile directly mediates the safety verdict and is reused for policy attribution and post-intervention verification. Coupled with a response rewriter, PolicyMem enables a detect-rewrite-verify loop for LLM governance. Across five widely used benchmarks, PolicyMem achieves state-of-the-art unsafe behavior detection while enabling effective policy attribution, rewriting, and post-intervention verification through the shared policy memory.

Large Language Models (LLMs) have shown remarkable proficiency on general-purpose tasks, yet their performance often degrades in highly-specialized technical domains. Moreover, little is known about how parametric knowledge of domain-specific terms is encoded within these models. We address this gap by contributing two novel medical jargon evaluation benchmarks and evaluate a general-purpose Llama-3.1 model against a variant fine-tuned on medical-domain data. Surprisingly, the general-purpose model outperforms the medically fine-tuned model on both tasks. Using mechanistic interpretability tools, we find systematic patterns of miscalibration for the medically fine-tuned model. Instead of reorganizing parametric knowledge, the fine-tuned model places greater emphasis on a small subset of model components associated with jargon-favoring predictions. We find that applying component reweighting strategies against the benchmark tasks successfully suppresses these components and closes the gap with the general-purpose baseline. We also observe that some jargon-sensitive components transfer knowledge to the same tasks involving materials science jargon, suggesting they encode a partially domain-agnostic notion of specialized terminology. Our results provide a case study in which a medically fine-tuned checkpoint does not improve jargon comprehension over its general-purpose counterpart, highlighting that domain adaptation should not be assumed to yield better performance on specialized terminology.

The WMT26 General MT task evaluates systems on 10 language pairs that have no human references (neither translated from scratch nor post-edited from MT output by humans). We describe how we built the pseudo-references for these pairs and six other language pairs (in which some forms of human references are available): seven models translate the 3,277 official documents under up to five prompt conditions, giving a total of 26 system-prompt combinations; then three reference-free quality estimation (QE) models score every candidate; and a per-document selector picks one translation, which GPT-5.5 post-edits where needed. Working without references exposed a failure mode of QE-guided selection: the metrics rank fluent output in the wrong language above correct translations. Adding a confidence-scaled language identification penalty to the score fusion drives the wrong-language count to zero, and the resulting selector still scores better on MetricX than the rank-fusion baseline it replaces. Since no references were available for these pairs while we were building them, we calibrate every selection decision on last year's WMT25 human judgments. The human evaluation, released after construction, shows the cost of getting selection wrong: our references stand with the strongest participating systems when the selector kept a frontier-model candidate, and fall up to 17 ESA points below them when it did not. We release the selection method and the provenance of every reference (https://github.com/surrey-nlp/PseudoRef)

Full-duplex evaluation often emphasizes whether an agent keeps speaking or stops. That binary cannot express a third response humans use routinely: continuing to speak while incorporating what the listener just contributed. The contribution may be a missing word, a correction or a clarification. We introduce Duplex Cue, an evaluation of this \emph{in-turn adaptation} in full-duplex voice agents. Duplex Cue separates listener intent (backchannel, collaboration, or interruption) from speaker behavior: continuing unchanged, adapting within the turn, or yielding. Adaptation includes acknowledgment as well as content revision. In a single-model case study using 300 human-confirmed cues from unscripted English conversations, we compare recorded human responses with PersonaPlex continuations generated while replaying the listener's audio. We retain 208 pairs with the ongoing speaker active at cue onset and a scorable response in each condition. On the 66 collaborative pairs, recorded speakers adapt in 68.2\% of cases, compared with 34.8\% for PersonaPlex. The model otherwise continues unchanged (42.4\%) or yields (22.7\%). These findings show why evaluating natural voice interaction requires measuring how an agent responds to a listener's contribution as well as whether it keeps speaking.

Vision–language models (VLMs) can answer chart questions accurately, but output accuracy does not show how they combine the evidence needed to recover an exact value. We study vertical-bar value reading with controlled counterfactual activation patching in Qwen2.5VL-7B-Instruct and InternVL3.5-8B. The study connects three analyses: (1) The single-factor results show that the changed bar-top region restores much more answer preference than the unchanged bar body, despite containing fewer visual tokens. Legend- and series-related states also lose local recoverability earlier than bar-geometry and axis-scale states. (2) In the handoff analysis, restoration shifts from visual legend regions in early layers to prompt-series positions in middle layers. Resetting the prompt-series state selectively reduces legend-source rescue, supporting its role as a partial mediator. (3) In the factorial analysis, both models can use geometry and scale states from separate donors to favor the combined target. InternVL performs similarly when the states come from separate donors or one image, while Qwen shows lower restoration for separate donors which suggests higher context sensitivity. Together, these results provide preliminary causal evidence for localizing the internal computations that support exact bar-value reading.

Handwritten mathematical expression recognition (HMER) is conventionally scored by exact-match rates and string-similarity metrics that are blind to where an error occurs: two predictions with identical token-error counts receive identical scores whether they misplace a subscript or swap the operands of a fraction. Render-based character detection matching (CDM) aligns glyphs robustly but remains position-blind—on controlled fraction-operand swaps it scores 0.8595 where position-aware scoring yields 0.6253. Tree-edit metrics exhibit a complementary blind spot: rewrites outside the parser's normalization coverage are penalized as structural errors (0.8552 where render-based metrics score 1.0). We propose PA-CDM, a position-aware metric that couples character detection matching with position-forest encoding and divergence-level weighting; StructPerturb v2.0, a frozen benchmark of 1,340 controlled perturbation pairs across 15 type–intensity cells; and a cross-metric consistency protocol combining a sensitivity matrix, a human study, and LLM-judge calibration. In a six-annotator study, PA-CDM attains the highest correlation with human judgments among seven automatic metrics (Spearman rho=0.9535, n=990). A frontier LLM judge correlates slightly higher (rho=0.9613) but is costly, nondeterministic, and API-dependent; PA-CDM approaches it at zero marginal cost with deterministic, diagnosable behavior.

Speech-to-speech translation (S2ST) has advanced significantly with speech LLMs, offering the potential for joint optimization and preserving non-linguistic information. However, these models struggle with predicting high-bitrate speech tokens in LLMs, and face the challenge of relying on S2ST training data with ideally aligned speaker identity and prosody. We propose using low-bitrate tokens based on single-layer vector quantization, trained to reconstruct self-supervised learning (SSL) features. We also employ a separate token-to-waveform decoder named Autowave-X, which is also conditioned on the source speech to improve non-linguistic transfer, thereby relaxing the training data constraints. With the integration of these techniques, we propose an S2ST model named Kraken, which augments a pre-trained LLM with speech feature inputs and the low-bitrate token outputs, followed by Autowave-X vocoder. We built the model upon Qwen3-8B and trained it using 150k hours of multilingual and multitask speech data. We demonstrated that our model exhibited better translation quality than SeamlessM4T-Large v2 and Qwen2.5-Omni, along with improved speaker and prosody transfer capabilities.

We present DuplexDrama, the first synthesized spoken dialogue dataset that simultaneously covers four dimensions: (i) complete persona and scenario settings; (ii) three full-duplex behaviors (interruption, backchannel, incomplete); (iii) expressive speech with persona-aligned emotion labels; and (iv) script-aware sound events. DuplexDrama is built via a 4-stage pipeline; quality validation on both scripts and synthesized audio confirms its quality. We have produced more than 2,000 hours audio data with a 64-voice timbre pool spanning 13 personas and 5 age buckets; 3.8% of all turns carry at least one full-duplex behavior. This data has been validated through internal full-duplex model training. We will release a curated subset of 6,400 bilingual dialogues (800 h, Chinese ~500 h + English ~300 h) to advance full-duplex spoken dialogue model research. Data samples are available at our demo page and LLM-judge evaluation prompts will be released with the dataset.

We identify \textbf{Harmfulness Propagation Dynamics (HPD)}: for harmful prompts, the projection of the last-token hidden state onto a learned harm direction rises monotonically with transformer depth, whereas benign prompts remain flat or oscillatory. This cross-layer signature reflects harmful intent as a \emph{progressively resolved} semantic property: surface form appears early, while pragmatic intent consolidates later, making the \emph{trajectory shape} more informative than any single-layer snapshot. Moreover, LDA-based harm directions, learned per layer, remain stable across random splits (pairwise cosine similarity $>0.97$), supporting the projection sequence as a reproducible structured signal. Building on HPD, we introduce \textbf{\herald{}} (\textbf{H}armful \textbf{E}ncoding \textbf{R}ecognition via \textbf{A}ctivation \textbf{L}ayer \textbf{D}ynamics). This lightweight input moderator extracts a seven-dimensional feature record, slope, curvature, monotonicity, onset layer, and related statistics from the cross-layer projection sequence and classifies it with a 288-parameter MLP. \herald{} stores one $d$-dimensional direction per layer ($262$\,KB for a 32-layer, $d{=}4096$ model), requires no gradient computation during training, and adds only $2.6{\times}10^{-6}$ prefill FLOPs at inference. Across eight prompt-harmfulness benchmarks and four model families, \herald{} achieves an average F1 of $89.3$ on OLMo2-7B, surpassing all tested guard models on adversarial jailbreak detection ($98.4$ vs.\ $96.9$ F1) and outperforming prior latent-based methods by $2.3$-$4.1$ F1 points on every backbone. Per-instance trajectories provide machine-readable audit records that reveal \emph{when} and \emph{how} harmfulness emerges, offering an interpretability advantage over single-layer approaches.

Large-scale multi-label text classification assigns a small subset of relevant labels to each document from a vocabulary containing thousands or tens of thousands of candidate labels. Although pretrained language models have improved semantic text representations, most representation-based approaches center their prediction pipelines on a primary encoder or combine auxiliary features within a single ranker. The complementarity between heterogeneous language models therefore remains insufficiently explored. We propose DualMLC, a dual-branch framework that processes the same document through an autoregressive decoder-only language model and a bidirectional encoder. Each branch maintains its own representation pathway and independently estimates relevance scores over the shared label space. DualMLC combines the two score vectors through late logit fusion, allowing shared evidence to reinforce relevant labels and branch-specific evidence to compensate for limitations in the other branch's representation. DualMLC achieves state-of-the-art results on three widely used large-scale multi-label text classification benchmarks. Ablation results further confirm that integrating the heterogeneous predictors produces stronger rankings than either branch alone. The source code is publicly available at https://github.com/huiyegit/DualMLC.

While Large Language Models have improved rapidly, many fundamental questions remain about how to evaluate the knowledge and reasoning abilities they acquire, and how such evaluations relate to the learning signals used in pre-training. In this paper, we propose ModelLog, a declarative probabilistic framework for pre-training evaluation that makes the semantic structure of model behavior explicit and provides new formal tools for relating evaluation to learning. ModelLog specifies evaluation targets as symbolic constraints over token-level predictions and measures how strongly a model's distribution satisfies those constraints. We explore the framework through a new suite of tasks targeting negation, mutual exclusivity, and consistency, finding systematic failures that are difficult to characterize through token likelihood or answer accuracy alone. We further show that these evaluation scores can also be interpreted as losses, whose gradients reflect logical strength, informativeness, and variable-level sensitivity. This links evaluation and learning through a shared semantics, suggesting evaluation methods that diagnose model behavior while also helping to clarify the semantic structure of learning.

Sanskrit fuses case, number, person and tense into word endings and chains clauses into compounds, so it is information-dense per word. Whether that density survives subword tokenization is a separate question, to be asked per unit of meaning rather than per word. On identical FLORES-200 devtest content, Sanskrit costs 1.774-2.187 times the English tokens under deployed tokenizers with vocabularies of 200,019 ids or more, but only 1.325-1.353 times the Hindi tokens. Against a deployed English tokenizer, Sanskrit-trained BPE arms then look cheaper per proposition than English on contemporary prose (0.887). Against a matched English control, the same algorithm and vocabulary trained on the English side of the same corpus, that flip disappears: at 32,000 and 64,000 pieces all 8 matched pairs, each size-matched arm against both a pair-matched and a byte-matched control, sit above 1.0 on prose with 95% intervals excluding it. The gap closes as the vocabulary grows: at 128,000 pieces the BPE pair reads 0.983 in domain while staying above parity out of domain (1.025) and on FLORES (1.116). The ratio factorises into a character-length ratio and a tokens-per-character ratio, the second near 1 throughout: what survives matched tokenization is character-level length, which Sanskrit prose lacks over English in SLP1 (1.028) and Sanskrit verse has (0.596). The robust statement is about deployed practice: on contemporary prose and on FLORES, with the Sanskrit side in SLP1 against the deployed o200k English pivot, Sanskrit costs 1.831-2.899 English tokens per proposition under the tokenizers people actually ship. Code, the results snapshot and every table here are public.

Compositional generalisation has been divided into lexical and structural generalisation. Previous work has found that structural generalisation is harder than lexical for Transformers. We propose that this difference is not inherent to Transformers, but due to the high diversity of lexical types and low diversity of structural types in the specific datasets of these previous works. By type diversity we mean the number of different constructors of that type, instead of, for example, the specific word combinations that might populate the structure. To test this, we vary the amounts of type diversity of lexical and structural types in previously published datasets. We create linguistically diverse variants of the COGS and SLOG datasets using Grammatical Framework. We find that type diversity correlates with compositional generalisation equally in lexical and structural test cases, supporting our hypothesis. We note a contradiction with the proposition in previous work that compound divergence explains the difficulty in compositional generalisation tasks. We further investigate the effects of other dataset properties on compositional generalisation, such as the diversity of types other than the novel test structure, and surface properties of the logical semantics format.

Specialist distillation effectively transfers domain expertise to student models via teacher-generated reasoning trajectories. However, when these specialists are trained solely on question–answer pairs without explicit reasoning supervision, what governs the trajectories they generate? In this work, we show that specialist optimization implicitly selects from this latent trajectory space. To isolate and observe this latent distribution, we leverage student distillation not as a downstream goal, but as an agnostic probe—since students inherit no parameterization or optimization constraints from the specialist, inheriting only the sampled trajectories themselves. Through this probe, our empirical analysis unveils a tight governing relationship: across 27 specialist–student pairings, their specialization–generalization profiles correlate exceptionally strongly. Crucially, explicitly controlling the specialist's distributional drift systematically shifts both the teacher and its distilled student along a controllable trade-off between domain precision and general-capability retention. Across chemistry, physics, and multilingual settings, distilled students systematically reflect these specialist-induced profiles, even across divergent model families. Our findings establish a new view of specialist training: when gold reasoning is absent, tuning choices directly control the latent supervision passed to downstream models.

Conversational voice agents have advanced significantly, offering increasingly natural human-machine interactions through both cascaded and end-to-end architectures. However, while recent benchmarks extensively evaluate dyadic interactions and passive audio comprehension, they largely overlook a prevalent real-world scenario: multi-party conversations. Evaluating agents in these settings is fundamentally more challenging than in dyadic interactions due to the exponentially greater conversational complexity. For voice agents to integrate seamlessly into human group dynamics, they must not only generate contextually appropriate responses but also demonstrate a nuanced understanding of open turn-taking. To address this gap, we introduce Multiparty Bench (MP-Bench), the first benchmark specifically designed to objectively evaluate conversational speech systems as active participants within multi-party contexts. MP-Bench assesses agent behavior along two primary dimensions: turn-taking awareness and response appropriateness. Additionally, we incorporate comprehension-based question-answering tasks as a complementary evaluation. By benchmarking 12 voice agents, we find that real-time voice agents stay at or below 22% on multiparty comprehension and remain near chance on multiparty turn-taking, exposing an open challenge for real-time voice agents under multiparty scenario.

Reinforcement learning (RL) has become central to post-training of large language models. Recent advances in RL for Mixture-of-Experts (MoE) models have primarily focused on improving optimization stability and training efficiency, while treating the expert selection as a fixed component. Since routing determines the sparse computation paths that induce output distributions, expert selection offers an additional source of rollout diversity. Through empirical analysis, we find that perturbing expert routing effectively alters model output and increases rollout diversity, which is similar to increasing the decoding temperature. However, direct perturbation can activate unsuitable experts and substantially degrade rollout quality. Motivated by these observations, we introduce Expert-Space Exploration Reinforcement Learning (ESRL), an architecture-aware framework that explicitly explores the expert-routing space of MoE models. ESRL preserves high-confidence experts as anchors, and restricts stochastic routing to a plausible candidate pool, thereby retaining reliable computation paths. The perturbation strength is further adapted according to router entropy to avoid over-perturbation. To mitigate the routing mismatch introduced by perturbation, ESRL records the expert paths used during rollout and replays them during policy optimization. Experiments demonstrate that ESRL achieves the best performance across MoE backbones with top-K, top-1, and shared-expert routing, as well as across mathematics, science, and code tasks without additional sampling or computational cost. Specifically, ESRL on Qwen3-30B-A3B achieves the best among all compared methods, improving average Pass@1 and Pass@8 over GRPO by 3.2 and 4.5 percentage points, respectively. Further analyses of expert utilization and training dynamics provide insights into how exploiting MoE-specific routing structure benefits RL training.

Large language models (LLMs) process long contexts, including long documents and lengthy conversations, but face token-level memory usage that increases proportionally to input length. Although model optimization and lossy prompt compression are widely used, these methods still fail to solve the long-context recall problem beyond pretrained and size-constrained context windows. This paper proposes a long-context recall method that maintains near-constant GPU memory usage as context length increases, without additional training. The main idea is to reconstruct facts using parameter activations in the LLM's feed-forward layers, which store residual vectors representing facts from the source document. Utilizing residual vectors allows the LLM to deterministically reconstruct query relevant facts without referencing the original document, preserving high fidelity and reducing memory usage without fine-tuning weights. Experimental results show that the proposed method enables answering single-fact questions in two-million-token story contexts where previous methods fail.

Clinical de-identification relies on accurately identifying personally identifiable information (PII). However, manually annotated datasets are costly to construct, while existing synthetic alternatives often provide limited details about their generation process or rely on relatively simple synthesis strategies. We introduce Meddies-PII-Dataset, a corpus of one million synthetic clinical documents spanning seventeen languages and nine PII labels. The documents are generated using attribute-conditioned prompts and validated through thirteen deterministic gates that enforce structural and annotation consistency. To evaluate the dataset's utility, we train Meddies-PII-Model, a BIOES token classifier, and compare it with existing PII extraction systems using exact-match entity-level F1. Meddies-PII-Model achieves the highest performance among the evaluated systems on all reported benchmarks, with a mean F1 of 0.827 across fifteen external benchmarks, compared with 0.658 for the strongest baseline. Upon acceptance, we will publicly release the dataset, benchmark suite, model, generation framework, and evaluation code to support research on multilingual clinical de-identification.

Russian state propaganda spreads across many languages and online spaces. Yet, most computational work examines only one such space, usually social media, in one or two languages, and analyses sources rather than content. We introduce SWARM (Search-Web documents Annotated for Russian propaganda, Multilingual), a dataset of 2,183 search engine results across nine languages and diverse web domains (e.g., news, blogs, government sites), each annotated by trained coders for whether it supports a recurring Russian propaganda narrative. We benchmark a source-based blocklist, supervised classifiers, and zero-shot LLMs against these labels. The blocklist misses most propaganda-supporting documents, because such content is not confined to flagged "propaganda" outlets but also appears on mainstream ones. Content-level analysis helps, though how much depends on the model: the strongest LLM reaches a positive-class F1 of 0.73, whereas the supervised classifiers reach only about 0.5, with the smaller LLMs over-predicting support, mistaking topical relevance for endorsement. Detecting search-borne propaganda thus requires per-language, content-level evaluation, which we hope SWARM and our evaluation code enable.

Large language models trained recursively on their own or other models' outputs undergo model collapse, in which distributional tails and factual accuracy deteriorate while fluency survives. Prior work diagnoses collapse after training; the actionable problem is screening a corpus of unknown provenance before training. We introduce SynthSentry, a corpus-level, model-agnostic contamination signal requiring no access to the generating model, no generation history, and no synthetic labels. The score is a distributional divergence over three statistics: lexical diversity collapse, n-gram tail truncation, and perplexity variance across reference models. We evaluate on corpora contaminated by small open-weight generators and an instruction-tuned open-weight model under a leave-one-generator-out protocol. A domain-stratified study measures false positives on naturally repetitive human text (legal, clinical, source code). The score ranks corpora by severity with little loss when whole generator families are held out. Per-domain calibration holds near its nominal false-positive budget once covariance shrinkage and a bootstrap threshold replace a naive quantile, which runs four times over budget. A downstream fine-tuning check showed no contamination-driven accuracy deficit at our scale, so whether pruning recovers one remains open; the same run shows over-pruning risk once pruning exceeds the true contamination fraction. We frame screening as a data-curation defense rather than a post-hoc diagnosis and release the scoring toolkit. All results are small-scale; scope is English-language, batch-mode corpus screening. Contamination sources are single-generation or hand-authored rather than recursively generated, so results speak to synthetic contamination generally and not to recursion depth.

Binary-choice truth benchmarks ask models to choose between a correct and an incorrect answer, but if the two answers differ systematically in surface-level features, models can exceed chance without performing the intended reasoning. We show that this failure mode is detectable and can be exploited by downstream classifiers. In TruthfulQA, a simple six-feature logistic classifier achieves substantial accuracy in separating correct from incorrect answers. We further show that similar surface-level artifacts are present in additional benchmarks. To counteract this, we developed a general mechanism to clean them by removing the most leakage-reinforcing pairs. We release a version of TruthfulQA with surface-feature leakage reduced close to chance and provide a mechanism, Audit-Prune, so that the datasets can be cleaned before release.

A medical claim's correctness often depends not on the claim alone, but on the clinical structure around it. A claim may require a lab reference range, a causal or conditional link, or patient-specific details to be judged correctly, and atom-level decomposition can fragment these dependencies, leaving the verifier with clinically incomplete claims. We reformulate medical fact-checking around snippet-level verification, where clause-grouped units preserve local clinical structure. We introduce MedSNIP-Bench, a human-annotated benchmark for snippet-level medical fact verification, and MedSNIP, an automatic snippet-generation pipeline. MedSNIP-Bench covers 276 consumer-health and clinical-vignette responses, segmented into 2,524 snippets with dual in-general and in-patient-context labels and six structural pattern codes. MedSNIP is evaluated against human snippet boundaries on MedSNIP-Bench and then used to generate snippet-level units for external corpora. Across MedSNIP-Bench, HealthFC, and MedHallu, snippet-level verification preserves or improves false-class F1, with gains concentrated where answers are long enough to fragment and where the verifier is strong enough to exploit the recovered structure. The largest merge-pattern gain is on causal-conditional clinical chains. It also reduces verifier calls by 24-73%, though the saving survives end-to-end only when decomposition is cheap, which an open-weight decomposer makes possible at no loss of chunking fidelity.

Full-duplex spoken dialogue models support low-latency turn taking, interruption handling, and backchanneling, yet a key capability remains underexplored: steerability, the ability to reliably shift conversational behavior along attributes such as tone, persona, speaking rate, and voice style in response to user instructions. We introduce a taxonomy of text- and audio-based steerability that identifies substantial gaps in current full-duplex models. To address this gap, we introduce SteerDuplex, a Moshi-based full-duplex speech model fine-tuned on natural conversations and synthetic dialogues targeting instruction following, vocal delivery, reasoning, and duplex interaction. We further apply two-stage reinforcement learning (RL) with hybrid rewards, combining verifiable interaction checks and judge-based semantic feedback to improve timing and response continuity. To evaluate full-duplex spoken steerability, we introduce SteerBench, a benchmark with 390 spoken prompts and 1,067 human-authored binary audio and text rubrics spanning tone, persona, style/accent, and speed/length. On SteerBench, supervised training improves audio-steering average pass rate by 44.5 percentage points over the strongest evaluated open baseline. On Audio MultiChallenge, task average pass rate improves by 7 points over its strongest evaluated open baseline. RL further raises source-clean interruption response from 72.5% to 82.5% and reduces synthetic pause barge-in from 26.5% to 9%. Steering and aggregate task scores remain comparable or higher, while reward probes reveal reward hacking through incomplete responses. Our model and benchmark support systematic research on spoken steerability, with reward analysis showing why timing gains must be evaluated alongside response completeness.

Retrieval-Augmented Generation (RAG) has empowered Large Language Models (LLMs) to tackle knowledge-intensive tasks. However, navigating global, heterogeneous knowledge bases (large-scale knowledge graphs and text corpora) for complex reasoning remains a challenge. Existing methods typically employ reactive, graph-driven exploration strategies, which blindly follow graph topology without adapting to the question context or evolving exploration progress, and lack deep bidirectional synergy between graph and text. To address these limitations, we propose CoG (Cognition on Graph), a cognitive-inspired, training-free framework for adaptive knowledge exploration. Drawing inspiration from human problem-solving, CoG performs a continuous plan-explore-reflect cycle, where it proactively formulates investigation plans, performs dual-source retrieval, and dynamically reflects on progress to adjust strategies. Crucially, it establishes deep bidirectional synergy between structured graph and unstructured text, where entities extracted from text dynamically guide graph exploration to bridge knowledge gaps. Extensive experiments on seven multi-hop QA benchmarks demonstrate that CoG significantly outperforms state-of-the-art methods while achieving superior exploration efficiency. Our code and datasets are available at https://github.com/zhougengxian/CoG.

We demonstrate that a general-purpose agent can directly drive a physical robot throughout task execution without any task-specific or environment-specific training. We introduce Agent as Policy (AGP), which places task planning and execution under the agent's control. Given a task and a robot interface, the agent interprets visual evidence, writes executable programs, issues motion commands, and revises its actions in response to physical outcomes. This brings the agent's reasoning and programming capabilities into continuous interaction with the physical world. We study AGP across multiple real-world manipulation tasks spanning precision manipulation, dynamic motions, and deformable objects. These include assembly from human videos, block construction from goal images, dice flipping, targeted throwing, and bimanual towel folding. Across assembly, block construction, and dice flipping, AGP succeeds in at least eight of ten trials for each evaluated task configuration. We further study efficiency through task experience accumulation and find that reusing saved procedures and programs shortens execution time across repeated trials. These findings support a path for general-purpose agents to act as robot policies, extending their autonomy to physical manipulation through runtime reasoning, programming, and interaction.

Ambiguity is often treated as a bug for AI systems to resolve—but in human communication and culture, ambiguity can also be a generative resource. From humour to politics to art, people express themselves in words and images that are open enough to invite different interpretations, yet constrained enough to be interpretable. We operationalise this notion of calibrated ambiguity with a task drawn from the parlour game Dixit. We compare differences in clues generated by human vs multimodal language models, based on a novel coding rubric for calibrated ambiguity, and find that models consistently exhibit ambiguity collapse (i.e., their outputs are over-specified, leaving no room for multiple legitimate interpretations). Unlike human clues, AI-generated clues also exhibit cultural flattening; they almost never make reference to culturally-situated knowledge, even when prompted to use allusion and figurative language.