Papers

Multi-domain machine translation (MDMT) poses a unique challenge due to varying levels of linguistic complexity across domains. Inspired by human translators' ability to adapt reasoning effort based on difficulty, we propose TwT (Translation with Thought), a resource-rational framework that learns to modulate inference between intuitive and deliberate reasoning. TwT is trained in two stages: (1) supervised fine-tuning on difficulty-aware long chain-of-thought traces distilled from DeepSeek-R1 and rewritten by GPT-4o to reflect human-like reasoning economy, and (2) reinforcement learning with a hybrid reward to optimize translation quality and reasoning efficiency. Evaluated on 15 benchmarks spanning in-domain and out-of-domain settings, as well as 3 seen and 59 unseen languages, with ablations across three backbone models, TwT-7B and TwT-14B outperform much larger SOTA reasoning models in translation quality, while reducing token usage by 32–60%. These results confirm that aligning translation behavior with cognitive principles enables robust generalization, high translation quality, and efficient reasoning in MDMT.

Claims that language models homogenise are usually measured against human judgements collected for the study, which makes the human side an artifact of the design: a crowdworker given the model's instruction is running the model's prompt. We measure convergence against a human reference nobody built for the purpose – 2,523 reader mark sets across 120 web documents, produced by people highlighting for their own reasons on a platform where the overlay of others' marks is off by default. Agreement is the overlap between two size-matched sentence sets minus the overlap expected when each is resampled within its own depth-and-length bands. The null's calibration is demonstrated, not asserted: every pair involving a random baseline lands within 0.006 of zero. On the median document each party names 14 sentences of 70; two readers share 4.1 and two models 8.7. Across 18 model arms spanning 11 vendors, 3 countries and both weight regimes, the median of 153 model pairs is +0.093 against a human yardstick of +0.040, and 99 sit entirely above the human interval. Two frontier models from rival labs reach +0.203, twice what GPT-4o agrees with itself on a second call. The effect is not determinism, prompt wording, procedure, vendor or routing, and it is graded: the smallest models agree at the human level. No model agrees with readers detectably more than a reader does, and at equal depth and length no surface feature separates their choices. The multiples are procedure-dependent and the ordering is not: models are cut to their sharpest set while a reader's is a random draw from what they marked, and blunting the models alike halves the gap without closing it. Tested out of sample on four models released after this analysis, against predictions fixed beforehand, none clears the human interval. A population simulated from several models is not several populations.

Reliable evaluation of open-ended LLM outputs requires fine-grained rubrics, yet expert curation is costly and difficult to scale. Existing automated pipelines rely on strict judge unanimity and binary variance filters, which cannot distinguish measurable rubrics from informative ones. We introduce CalibratedRubric, a task-adaptive framework that combines type-specific scoring, Bayesian rubric-measurability filtering, and item response theory (IRT)-based bank assembly. CalibratedRubric estimates each rubric's measurability with a Beta–Bernoulli agreement posterior and uses a submodular information-coverage objective to construct compact rubric banks over the observed capability range. Across financial, healthcare, general, and legal benchmarks, measurability filtering improves human-gold agreement on JudgmentBench from $\kappa=0.604$ to $0.743$. IRT-based greedy selection improves cross-fitted rank fidelity over random selection across all six evaluated response blocks and requires only 49 rather than 131 rubrics to reach the target correlation on FinResearchBench decision-support tasks. Task-label perturbations further reduce system separation, confirming the practical relevance of task-adaptive scoring. These results support CalibratedRubric as an efficient, uncertainty-aware approach to open-ended LLM evaluation, with calibration gains depending on sufficient judge redundancy.

Small language models (SLMs) are attractive for agentic deployment due to low latency, reduced cost, and on-device privacy, yet they struggle with tool-use tasks where training data is scarce and noisy. Unlike larger models, SLMs cannot compensate for low-quality supervision through sheer capacity, making data quality the critical bottleneck. We present Data Turnstile, an open-source framework that takes user-defined API specifications and generates high-quality synthetic training data for function calling. Turnstile decomposes multi-turn tool-use interactions into constrained, stepwise generation with validation and error-feedback loops, providing fine-grained control over API diversity, conversation complexity, and output correctness. We demonstrate effectiveness of domain adaptation with Turnstile data on two challenging function calling benchmarks. On the BFCL single-turn benchmark, a Qwen3-0.6B fine-tuned on Turnstile data without chain-of-thought achieves 75.9% overall accuracy (versus 67.4% for the base model with thinking enabled), closing the gap with thinking-enabled Qwen3-1.7B (78.4%) and Qwen3-4B (79.9%) despite being 3$\times$ and 7$\times$ smaller respectively. On $\tau^2$-bench, a multi-turn agentic benchmark, Turnstile-trained Qwen3-1.7B achieves 31.1% pass^1 on the Telecom domain, improving 4.7$\times$ over its 6.6% base and surpassing Qwen2.5-32B-Instruct (27.4%), a model 19$\times$ larger. Turnstile-trained Qwen3-0.6B achieves 24.6%, improving 7$\times$ over its 3.5% base and approaching the 32B model (53$\times$ larger). We release Data Turnstile along with a dataset spanning 1,000+ APIs and 100K+ multi-turn interactions.

Optimizing modern recommender models still depends heavily on engineers manually iterating over architectural, objective, and training-strategy changes. While LLM-based agents can automate this trial-and-error process, allowing the LLM to both select modification directions and generate concrete hypotheses often leads to unstable search under limited experiment budgets. Inspired by the above challenge, we propose RecHarness, a Bandit-Routed Agentic Harness for automated recommender model optimization. RecHarness separates the optimization process into two steps: a bandit router selects the next modification direction according to historical validation feedback, while the LLM generates a concrete optimization hypothesis and executable code edit within the selected direction. To sustain long-horizon exploration, RecHarness uses a jump-basin mechanism to activate a structural-jump arm when local edits stagnate. Across multiple recommendation tasks, datasets, and model backbones, RecHarness achieves more stable performance improvements and uses limited trial budgets more effectively than LLM-reasoning search. During a 7-day online A/B test on a large-scale short-video advertising platform, the selected candidate improves ADVV by 2.084%, Revenue by 0.534%, and Exposure by 0.559%. Code is available at https://github.com/6lyc/RecHarness.

InMyStyle is a privacy-first, single-user system that adapts small language models to rewrite AI-edited text towards an individual user's writing style without an instruction prompt at inference. Given a user's documents, it uses multiple local helper LLMs to construct paired training examples and fine-tunes LoRA adapters on Qwen2.5 models ranging from 0.5B to 7B parameters. Length-aware generation budgets and automatic chunking support inputs of different lengths. We report a single-user case study: 219 evaluation pairs derived from 73 paragraphs of one author's scientific writing, with all adapters trained using the same rank-8, three-epoch recipe. The automatic composite score (0-1 scale) plateaus across model sizes under both greedy and sampled decoding ($Q=0.689$-$0.695$, with overlapping confidence intervals). In this setting, small models are sufficient for the measured rewriting task, and model size mainly determines efficiency trade-offs rather than a stable quality ranking. The gains favor content-preserving naturalization more than recovery of personal style, with authorship probabilities staying near the classifier's decision boundary (0.51–0.53) and stylometric improvement being near zero. As a secondary evaluation, 400 ratings from five LLM judges give InMyStyle outputs a mean perceived AI-ness score over 20% lower than their helper-generated inputs, with scores decreasing with model size in this sample. The study does not establish generalization across users.

Large language models generate computationally expensive yet semantically void reasoning on beyond-capability tasks, creating risks where plausible-sounding but incorrect derivations mislead users. We characterize this futile reasoning phenomenon through systematic analysis, revealing universal capability overreach and systematic miscalibration between capability and behavior. The dominant failure mode is specious reasoning, which outputs look superficially valid but contain subtle errors, escalating with task difficulty. To address this, we introduce CaRL (Capability-aligned Reinforcement Learning), which aligns model behavior with capability boundaries through reward shaping that incentivizes refusal over futile reasoning and hindsight refusal augmentation that converts failures into refusal supervision. Experiments demonstrate a substantial reduction in futile reasoning while preserving performance across task difficulties, effectively achieving capability-aligned behavior without sacrificing utility. \footnote{https://github.com/icip-cas/Knowing-When-to-Quit}

Long-running multi-turn interactions with chatbots and agents are now common, and a correct response often depends on remembering earlier details, tracking later revisions, identifying intended objects or referents, and withholding action when required conditions are unmet. Existing multi-turn benchmarks typically cover short exchanges and do not fully evaluate these capabilities in long multi-turn interactions, particularly in Chinese, while offering limited insight into how and why models fail. To address these limitations, we analyze real chatbot failures to identify six recurring mechanisms and use them to define six controlled evaluation modes in Hy-MultiTurn, a Chinese benchmark for deep multi-turn dialogue understanding. The six modes evaluate constraint memory, precise execution, constraint synthesis, object localization, action suppression, and reference resolution. Across the six modes, we construct 209 controlled tasks spanning 12-76 turns, with dialogue length, irrelevant-topic distraction, and colloquial phrasing adding further difficulty. Evaluation of 22 frontier model configurations shows that Hy-MultiTurn is broadly challenging, as even GPT-5.5, the strongest overall configuration, satisfies all requirements in only 41.1 percent of responses and no model performs best in all six modes.

Migrants traversing geographically distinct routes such as the Trans-Saharan and Balkan corridors often recount strikingly parallel lived experiences: police violence, smuggler exploitation, dangerous crossings, and family separation. We introduce the task of experiential intertextuality detection: automatically identifying shared experiential echoes across migration narratives without requiring annotated training data. From 108 French migration narratives spanning both corridors, we automatically generate sentence pairs and score them using annotation-free methods: lexical baselines, sentence embeddings, POS-based structural features, a migration-specific theme lexicon, context-aware narrative features, and zero-shot LLM scoring with Qwen2.5-7B and Mistral-7B under three prompting strategies. We validate all methods against 816 expert-annotated intertextuality judgments (inter-annotator Krippendorff's $\alpha = 0.27$). Our results reveal that all surface, structural, and embedding methods correlate only weakly with expert judgments ($r \leq 0.30$); Qwen2.5-7B zero-shot achieves the best single-method correlation ($r = 0.38$); few-shot examples degrade Qwen but dramatically improve Mistral; narrative position significantly predicts intertextuality, with departure-phase pairs showing the highest experiential echoes; and a supervised hybrid combining all 31 features achieves $r = 0.45$, a 21% improvement over the best individual method.

Reward models (RMs) are central to aligning large language models with human preferences via reinforcement learning. Although traditional scalar RMs enable efficient and probabilistic reward modeling, they rely on superficial cues that fail to generalize to complex or out-of-distribution (OOD) tasks. Conversely, generative RMs leverage extensive reasoning to improve robustness on challenging tasks, but their natural language-based scores lack the numerical flexibility and probabilistic interpretability that scalar RMs offer. While recent approaches combine both paradigms through off-policy multi-task learning, such parallel optimization does not guarantee that generated reasoning traces actively align with or benefit downstream scalar reward prediction. To address this mismatch, we propose LatentRM, a reward modeling framework that learns intermediate reasoning traces as discrete latent variables to explicitly maximize the likelihood of downstream scalar rewards. Through on-policy optimization of the latent reasoning space end-to-end, LatentRM tightly couples deep reasoning-based evaluation with precise scoring. Extensive validations on in-distribution and OOD datasets and RLHF show that LatentRM outperforms scalar, generative, and hybrid RMs on preference modeling and policy alignment across tasks ranging from open-ended conversation to complex reasoning.

Authorship Verification (AV) represents an important subfield of digital text forensics and addresses the fundamental question of whether two texts were written by the same author. Although the field has made substantial progress over the past two decades, several important challenges remain unresolved or underexplored. For instance, most AV research has focused on written texts, despite the fact that language is expressed not only in written but also in spoken form, such as in videos. Moreover, existing AV studies have predominantly concentrated on English, while other languages, including German, have received comparatively little attention. To address these research gaps, we apply AV to spoken language in the form of transcripts of German-language videos and examine the effectiveness of established AV methods in verifying a speaker's identity across video pairs. Our experimental evaluation, based on a total of ten AV methods applied to three self-compiled corpora comprising 300 videos from 150 speakers, shows that the best performance (up to 88% accuracy and 90% AUC) is achieved by traditional AV approaches based on simple character- and token n-gram representations. In contrast, more modern transformer-based approaches perform significantly worse on all evaluated corpora. Our results therefore suggest that traditional methods in the field of AV remain both competitive and relevant.

Full-duplex spoken dialogue systems (FDSDSs) can listen while speaking, enabling natural conversational behaviors such as smooth turn-taking, backchannel handling, and user barge-in handling. However, controlled comparisons of FDSDSs in multi-turn interaction remain challenging, and existing benchmarks offer limited joint coverage of multiple languages, dialogue settings, interaction timing, and response content. We propose M3-DuplexBench, a multi-turn, multilingual, multidomain benchmark for controlled evaluation of FDSDSs. The benchmark jointly evaluates interaction timing and response content in English and Japanese across casual conversation and multi-turn question answering. The English and Japanese question-answering data are derived from a common source corpus, enabling cross-lingual comparisons of timing and content performance. We also provide a reference-conditioned multi-turn setting as an optional diagnostic for examining sensitivity to dialogue history and trajectory mismatch. Experiments with diverse FDSDSs reveal model-specific interaction patterns, differences across languages and domains, and model-dependent effects of dialogue context on both timing and content. The benchmark is available at https://github.com/nttcslab-sp/m3-duplexbench.

Child malnutrition remains a major public health challenge in low- and middle-income countries, particularly in South Asia, where early identification of vulnerable children is critical for timely intervention and resource allocation. This study aims to evaluate the feasibility, fairness, and temporal robustness of using a pretrained large language model (LLM) in a zero-shot setting for child stunting prediction using population health survey data. Using Bangladesh Demographic and Health Survey (BDHS) data collected between 2007 and 2022, we transformed maternal, child, healthcare, and household characteristics into semantically interpretable prompt-based representations and evaluated GPT-4o-mini for zero-shot stunting prediction, comparing its performance against a random forest baseline and assessing fairness across demographic and socioeconomic groups as well as temporal robustness across survey waves. The results demonstrate that zero-shot inference using GPT-4o-mini achieved comparable balanced accuracy to the supervised baseline while exhibiting substantially higher sensitivity for identifying stunting cases, relatively consistent performance across child sex groups, and stable predictive behaviour across BDHS waves; however, important fairness disparities were observed across residence and household wealth categories, highlighting the need for further investigation before deployment of foundation models in public health prediction settings.

Training-free acceleration makes diffusion-based multimodal large language models (dMLLMs) more deployable, but it may silently change generated content. We study this serving-time consistency problem on 300 real images, comparing Fast-dLLM outputs with the same model's unaccelerated outputs. Across the mild parallelism induced in our long-form setting (1.05–1.25 committed tokens per step), confidence-threshold tuning changes decoding behavior but not baseline agreement. State-refresh ablations and an image-swap intervention instead identify stale visual and generated-text states as contributors to drift. For the tested Fast-dLLM implementation, shortening the KV-cache refresh interval yields a monotonic speed–agreement frontier and near-exact agreement at a measured 1.3x speedup. The initial diagnosis also appears with dLLM-Cache and LaViDa, although dLLM-Cache recovers agreement only after both caches are tightened, which removes its speed advantage. Independent prompts and images reproduce the threshold-insensitivity and refresh recovery. A targeted audit finds genuine content substitution in half of 50 low-agreement pairs. In a separate blinded two-annotator evaluation, the pooled accelerated-minus-baseline factual-error difference is 0.00 (95% CI [-0.17,+0.17]); this sample detects no difference but does not establish factual equivalence. Finally, none of the tested adaptive or smoothed-refresh variants beats the fixed interval at matched compute. Our contribution is a paired diagnostic and an implementation-scoped consistency control, not an accuracy or safety guarantee.

Deception detection has critical implications for legal proceedings, law enforcement, and online security. Although human judgment is limited in accuracy and scalability, Natural Language Processing (NLP) offers a data-driven alternative. We present a survey and comparative analysis of NLP-based Automatic Deception Detection (ADD) focusing on the legal domain, reviewing the evolution from feature-based machine learning to Large Language Model (LLM) approaches. We conduct a unified empirical evaluation across seven datasets (two legal, five general-domain), comparing six fine-tuned transformer models and seven LLMs under four prompting strategies. The results show strong domain sensitivity, with fine-tuned models excelling in data-rich general domains and few-shot LLMs remaining competitive in low-resource legal settings. Chain-of-Thought prompting often underperforms direct classification. These findings highlight the need for domain adaptation and interpretable systems in high-stakes legal contexts.

Deepseek's Engram, a conditional memory module, was introduced to trade-off storage versus reasoning in large language models. However, the module relies on token-level $N$-gram hashing for Engram embedding lookup, introducing a tight coupling to the tokenizer used: a model with a different tokenizer would have to train its own Engram embeddings from scratch. To improve the reusability of Engram embeddings, we propose a change to the hashing routine, enabling compatibility between Engram models using different tokenizers. Instead of modelling disjoint $N$-gram spaces, we treat $N$-gram as a method to sample potentially useful byte sequences, from all possible byte sequences across tokens. We replace the XOR-based hashing with the general polynomial hashing with a joint embedding space across $N$. This work investigates the possible trade-offs and shows that this simple substitution produces comparable performance and achieves tokenizer-agnosticism: hash equivalence for byte-equivalent token sequences.

Inline notes and collected commentaries are important forms of scholarly communication that evolved within the Confucian exegetical tradition, yet have received little computational attention. Drawing on traditional Chinese exegetics and philology, this paper formulates collected commentary compilation as an NLP task and proposes a computational framework that preserves the contextual dependency of inline notes while enabling their automatic compilation and exegetical knowledge organization. It combines two-step prompt chaining for identifying the associated main-text segments and exegetical functions of annotations with cross-source mention clustering for integrating commentary across editions, achieving a CoNLL F1 score above 97% in a case study on the Classic of Mountains. Our framework lays the foundation for the large-scale organization of historical exegetical knowledge, thereby supporting a broad range of downstream philological and NLP tasks.

Large language model (LLM) agents increasingly operate over long interaction histories, where effective reasoning requires identifying and exploiting task-relevant evidence distributed across past observations and actions. However, useful information encoded in previously computed representations is often underutilized during subsequent generation. We propose TransMem, a lightweight inference-time parametric memory module that transforms sparse historical hidden states from a frozen LLM backbone into reusable memory representations. TransMem uses a lightweight gating network to dynamically apply the latent intervention to the current hidden states, without repeatedly encoding the preceding context. To learn transferable memory utilization rather than task-specific knowledge, we introduce evidence-conditioned self-distillation. A memory-augmented student processes the full context and matches the predictive distribution of an evidence-only teacher that shares the same frozen backbone. Experiments on LoCoMo, HotpotQA, and MemoryAgentBench demonstrate consistent improvements across different model architectures and scales. TransMem yields gains of 11.58–29.25 $F_1$ on LoCoMo and 10.20–13.03 $F_1$ on HotpotQA, while improving the average MemoryAgentBench accuracy from 29.54% to 40.00%. These results establish sparse historical hidden states as an effective and efficient memory substrate for long-context LLM agents. Our code is available at https://github.com/Haodong-Lei-Ray/TransMem.

Retrieval-Augmented Generation (RAG) enhances large language models by incorporating external knowledge, but existing pipelines typically operate on plaintext data, raising significant privacy concerns. Prior work on privacy-preserving retrieval leverages cryptographic techniques such as homomorphic encryption (HE) and private information retrieval (PIR), but often relies on interactive protocols or ranking-based selection mechanisms that incur high latency and potential information leakage. In this paper, we propose a practical non-interactive encrypted retrieval framework for RAG based on threshold selection. Instead of performing expensive top-$k$ ranking under encryption, our approach selects documents whose similarity scores exceed a predefined threshold, reducing computational complexity from quadratic to linear in the corpus size. We implement this design using CKKS-based homomorphic computation, enabling fully encrypted similarity evaluation and document selection without revealing query content, intermediate scores, or selected indices. To bridge the gap between approximate encrypted computation and discrete token reconstruction, we introduce a precision-stable mask polarization method that ensures accurate recovery of selected documents. Experiments on standard retrieval benchmarks demonstrate that our approach achieves competitive retrieval effectiveness while significantly reducing latency compared to ranking-based encrypted methods. These results highlight threshold-based selection as a practical foundation for scalable and secure RAG systems.

Zero-shot image captioning (ZIC) describes images without paired image-caption supervision during captioner training, relying on text-only corpora and frozen pretrained image-text scorers. Existing retrieval-augmented methods score image-text alignment once, at retrieval, then commit the captioner's autoregressive beam under language-model probability alone, leaving the decoder without further visual grounding feedback. Progress has stalled, with no method improving on the strict-regime best since 2024. We propose Adjudicated Captioning, an inference-time multi-agent framework that restores grounding feedback at multiple checkpoints over an unchanged IFCap captioner. First, we install a stronger frozen Retrieval Encoder at the input. Second, between retrieval and decoding we insert a frozen Cross-Attention Verifier that re-ranks the top-9 retrievals to top-5. Third, at the output beam we attach a learned Reranker pairing TriFuse, a multilayer perceptron, with MemAttend, a memory-attended transformer, the pipeline's only learned components; both are trained self-supervised by Borda-consensus distillation across the three frozen scorers, using no paired image-caption labels and no reference captions. Under the inductive headline protocol, with rerankers fit on the disjoint COCO Karpathy validation beam and applied frozen to test, the framework reaches CIDEr 117.6 and SPICE 21.9 on COCO Karpathy, up from 108.0 and 20.3 for IFCap, a +9.6 CIDEr gain, and +7.7 above NES, the strongest synthetic-image-augmented method at 109.9, without retraining the captioner. A training-free fixed-fusion baseline reaches 115.8 CIDEr, so +7.8 of the +9.6 gain comes from the non-learned architectural intervention and the remaining +1.8 from the learned rerankers. The same recipe transfers off-COCO without captioner retraining: +8.1 CIDEr on Flickr30k Karpathy and +5.7 on NoCaps overall.

Recent language models achieve strong performance across a variety of tasks, but conventional adaptation applies updates uniformly across training samples regardless of their local update benefit. We propose PARALLEL, a prefrontal-aligned reinforcement inspired approach for language-model learning. Inspired by the complementary roles of goal-related and uncertainty-related control, PARALLEL represents these forms of information as separate controller signals and combines them with the current model representation. A reinforcement-inspired controller assigns sample-dependent update intensity using immediate utility-cost feedback. PARALLEL therefore learns when and how strongly to adapt to each sample, prioritizing beneficial updates while limiting unnecessary parameter changes. PARALLEL uses available updates more efficiently than selective baselines while retaining 94.1–99.2% of Full-adaptation performance. Beyond multiple-choice reasoning, experiments on XSum and CNN/DailyMail show that PARALLEL retains 96.9–98.6% of the ROUGE-1 and ROUGE-2 scores achieved by Full adaptation and 98.8–98.9% of the corresponding ROUGE-L scores. When compared at the same cumulative adaptation time or GPU energy, PARALLEL achieves higher ARC accuracy and exhibits a more stable late-stage adaptation trajectory than Full adaptation in the representative run. These results show that learning when and how strongly to update each sample supports stable and efficient post-deployment stream adaptation while avoiding unnecessary updates.

Heterogeneous Large Language Model (LLM) systems increasingly share contexts, retrieved evidence, and multi-agent dialogue histories, yet their internal key-value (KV) caches remain model-specific and cannot be reused across architectures. Consequently, each model must repeatedly prefill or store caches for the same context, limiting the scalability of multi-model reasoning and long-context generation. We propose Mixture-of-Translators (MoT), a cache translation framework that maps context KV caches from a source LLM into the cache space of a target LLM. Unlike prior approaches that depend on a single projection path or global shared latent space, MoT uses multiple translator modules with token-level routing to capture diverse cache translations. We further introduce a Context Correction Loss that aligns the replayed target trajectory with the native target trajectory, reducing residual translation error. Our analysis reveals two competing failure modes: propagation error from early injection and correction-deficit error from late injection. We evaluate heterogeneous translation among Llama, Gemma, and Qwen, spanning substantially different architectures and KV-cache spaces. MoT achieves an average accuracy of 57.6% and F1 of 0.42. These scores recover 95.0% and 120.6% of native target performance and reach 1.5x and 3.5x the strongest-baseline scores, respectively. In practical case studies, MoT enables accurate heterogeneous-agent discussion with scale-invariant KV-memory behavior as the number of agents grows, while retaining near-native generation quality in long-context cache-augmented generation. These results demonstrate scalable KV-cache reuse across heterogeneous LLMs.

Large language models often improve task performance by generating long reasoning traces, but the resulting computation is frequently wasted on redundant verification and revision. Existing probe-based early-exit approaches mainly inspect explicit self-doubt expressions, leaving many earlier termination opportunities undetected. Expanding inspection to ordinary reasoning boundaries improves coverage, but also exposes highly diverse intermediate states whose predictive information may reside in different hidden layers. We present Boundary-Expanded and Layer-Adaptive Dynamic Exit for Efficient LLM Reasoning (BLADE), a lightweight framework that dynamically terminates reasoning by estimating whether the generated prefix is sufficient for correct answering. BLADE constructs multi-granular checkpoints from sentence, self-doubt, and paragraph boundaries, and derives robust training labels through repeated answer completions. It further learns a compact subset of informative probe layers instead of relying on fixed choices or expensive representations from all layers. At inference time, calibrated predictions are combined with checkpoint-specific confirmation rules to balance responsiveness and premature-exit risk. Experiments on five benchmarks and two Qwen3 reasoning models show that BLADE preserves near-baseline accuracy while reducing generated tokens by 24.8% on Qwen3-8B and 15.8% on Qwen3-4B. Ablation studies further confirm the benefits of diverse checkpoints and automatic layer selection, demonstrating an effective approach to more efficient LLM reasoning.

Large language models (LLMs) are increasingly involved in the distribution of scarce resources, raising concerns about biased allocations based on characteristics like race and gender. Recent LLM audits have produced inconsistent results, however, finding evidence of both positive and negative discrimination towards women and ethnic minorities, even for the same models. We show that this disagreement can arise from differences in audit format and introduce FairFund-Bench, a benchmark that systematically varies key features of previous audit designs: the evaluation task (rating, ranking, or allocation), comparison context (single or multi-stimulus), and whether the audit is transparent or disguised. The benchmark comprises 600 English-language requests for financial assistance created from human-authored templates (calibrated against 1.3M real GoFundMe campaigns) across three domains, four race and two gender categories, and five causal framings of need derived from welfare deservingness theory. Across 14 models, audit format changes the direction of bias: models advantage minorities when rating claimants individually but penalize some groups when ranking them side by side. Bias magnitude, though small overall, is several times greater in disguised audits than in transparent ones, where, faced with appeals differing only in claimants' names, models overwhelmingly split funds equally. Causal framing effects, by contrast, exceed demographic effects by roughly an order of magnitude and are consistent across models and audit formats, indicating that current LLMs robustly reproduce human deservingness evaluations. The benchmark scores models on four criteria (demographic bias, deservingness alignment, cross-task consistency, and cross-context consistency), is publicly available, and can be readily adapted to other substantive domains.

Sycophancy refers to the tendency for large language models (LLMs) to match user beliefs at the cost of factual correctness, thereby undermining model reliability. Prior work on evaluating sycophancy in LLMs aims to assess whether a model's output matches an authority's claim, but cannot reveal which part of the prompt drives this sycophantic behavior. To bridge this gap, we investigate the relationship of sycophantic responses with an authority's credentials, their assertive claim, and the problem statement. We introduce the Authority Share Index (ASI), an Integrated Gradients-based token attribution method, which measures the degree to which a model's decision is driven by authority-related text. Through extensive experiments across five models and 30 test configurations, we find that sycophantic responses consistently direct more attention toward authority tokens than resistant ones. Moreover, our token attribution method reveals that for the sycophantic cases, the claim asserted by the authority receives more attention than the authority's credentials. Building on these findings, we propose attribution-guided contrastive activation steering to mitigate LLM sycophancy. Our method constructs a steering vector from high-attribution tokens of sycophantic and resistant responses, selectively pushing models toward resistance. This enables inference-time steering without retraining, lowering sycophancy from 96% to 25% in the strongest case. Together, our results show that token-level attribution can both explain what drives sycophancy and directly inform a practical intervention.

Phoneme-centric visual speech recognition reconstructs sentences from intermediate phoneme predictions, making overall recognition performance highly dependent on the robustness of the phoneme-to-text reconstruction model. Existing reconstruction approaches are commonly trained on clean phoneme sequences or synthetically corrupted inputs, leading to a mismatch between training conditions and the realistic phoneme prediction errors encountered during inference. To address this limitation, this paper proposes progressive error curriculum training (PECT). This curriculum learning framework progressively adapts a No Language Left Behind (NLLB)-based phoneme-to-text reconstruction model using synthetic phoneme perturbations, multi-domain pseudo-labels, and target-domain pseudo-labels generated by a visual speech recognizer. By gradually exposing the reconstruction model to increasingly realistic phoneme prediction errors, the proposed framework improves robustness while preserving sentence-reconstruction accuracy. Experiments on the LRS2 and LRS3 benchmarks demonstrate that PECT consistently improves reconstruction performance across multiple phoneme-based visual speech recognition frontends, including visual automatic speech recognition (V-ASR), point visual automatic speech recognition (PV-ASR), and head-pose-aware visual speech recognition (HP-VSR) variants. In particular, PECT reduces the word error rate (WER) of HP-VSR-FiLMFuse (L4) from 23.3% to 22.2% on LRS2 and reduces the WER of HP-VSR-ResFiLM from 30.3% to 29.7% on LRS3. Comprehensive ablation studies and qualitative analyses further demonstrate the effectiveness of progressively adapting the reconstruction model to realistic phoneme prediction errors. These results show that PECT provides an effective and generalizable curriculum learning strategy for phoneme-to-text reconstruction in phoneme-centric visual speech recognition.

Classical philosophical corpora pose three compounding challenges for language resources: they exist in several languages without parallel alignment, their vocabulary is remote from that of contemporary readers, and generated text over culturally sensitive material must be verifiably grounded. We present Viveka-Insight, a bilingual resource and open-source pipeline for the works of Swami Vivekananda (1863-1902): the nine-volume English Complete Works and the ten-volume Bengali Vani o Rachana, two related but non-parallel corpora of about 15 million characters. Four layers are released: (i) a structure-preserving parse (32,694 paragraphs, 168,842 sentences) with per-paragraph anchors deep-linking into the published editions; (ii) a cross-lingual concept graph of 8,362 language-agnostic concepts with 87,518 relation-typed paragraph-concept and 55,872 concept-concept edges, in which canonical English labels act as a string-equality key linking Bengali and English passages with no parallel data; (iii) a bilingual alias inventory of 60,850 surface forms (30,053 English, 30,797 Bengali); and (iv) a human-annotated set of 200 paragraph-concept edges judged by three annotators, released with all per-annotator judgments. We report known-item cross-lingual retrieval over 194 verified rendered lecture pairs (Recall@10 0.86 in both directions), a 30-question audit of citation integrity and modern-question bridging, and a human study placing concept-extraction precision at 0.60 under strict two-annotator consensus (Cohen's kappa = 0.61). The extractor's confidence weight is calibrated: restricting to weight >= 0.8 raises precision to 0.71 while retaining 98% of concept-bearing paragraphs. Precision is markedly lower in Bengali than English (0.54 vs 0.68), locating the weakness in exactly the half that cross-lingual access depends on. The design transfers to other multilingual classical corpora.

Charismatic speech is shaped by language and speaking style, yet their relative contribution in bilingual public speaking remains unclear. We analyzed spontaneous speeches of 10 politicians who address audiences in Luxembourgish and French, in highly comparable communicative contexts across languages. From 400 utterances, we extracted 41 acoustic-prosodic features linked to vocal charisma and fitted mixed-effects models to separate speaker- and language-related variance. Speaker identity accounted for most variance, whereas language explained less, but still showed systematic differences: French productions showed higher shimmer and phrase-final F0, indicative of a polite, respectful voice, while Luxembourgish productions exhibited stronger mid-frequency spectral energy, suggesting a more vocally present profile. These patterns align with the sociolinguistic roles of Luxembourgish as an informal identity language and French as a high-prestige institutional variety.

Speech rhythm helps structure persuasive speech, but most empirical work examines monolingual English. This study asks how politicians organize timing when speaking Luxembourgish and French. We analyze 400 sentences from ten politicians, annotated for segments and pauses. We compute rhythm metrics, including means, variability, and pairwise variability indices for consonants and vowels. We quantify speaker and language contributions and test within-speaker language effects with paired t-tests. Results show that consonant-based metrics retain speaker-specific signatures, whereas vowel-based metrics are largely driven by language choice. French tokens display longer and more variable vowels and vocalic intervals, while consonant timing differences are smaller. No robust language by gender interactions emerge. These findings show that language choice systematically reorganizes rhythmic timing in bilingual public speech.

Large language models (LLMs) increasingly serve users whose expectations are shaped by their cultural context, yet most cultural evaluations test what a model knows rather than how it behaves when giving open-ended recommendations, opinions, and guidance. We introduce AraBehave: 1,623 culturally grounded, open-ended Arabic prompts with 29,214 cultural-appropriateness judgments from native speakers across several Arab regions, plus a scoring model whose predictions correlate strongly with human judgments on unseen systems (Pearson r=0.74). Evaluating three Arabic-centric and three frontier LLMs, we find that cultural appropriateness is not a single capability but decomposes into two largely independent components: normative stance and grounded cultural accuracy. The best general-purpose and best Arabic-centric models score identically (3.84 vs. 3.83 of 5) yet almost never fail for the same reason: general-purpose models exhibit strong factual grounding but a culturally inappropriate normative stance, being penalized for secular framing and false balance on culturally settled matters (28–33% of their low-score rationales), while the best Arabic-centric model adopts the expected stance but is penalized for fabricated hadith and misquoted verses (29%). Stance is cheap and fragile: one sentence of cultural instruction lifts Gemini to 4.57, above every Arabic-specialized model. Conversely, a generic “answer clearly and objectively” prompt costs Allam-7B 0.68 points, while asking the same questions in English lowers scores for every model but one. Grounding instead tracks scale and Arabic alignment data, and disappears when culturally aware instruction tuning is replaced by a culture-neutral corpus. General safety benchmarks see none of this: they saturate above 89 while cultural scores span 2.71-3.84. We will release the benchmark, annotations, and the scoring model.