Papers

While Multimodal Large Language Models (MLLMs) show remarkable advancements, their cross-modal capabilities introduce complex vulnerabilities that easily bypass unimodal filters. Existing benchmarks lack fine-grained intent-related annotations and rely on unidimensional metrics, hindering comprehensive robustness evaluation. To address this, we propose MME-Safety, a rigorously verified benchmark featuring a unique four-dimensional annotation schema that categorizes risk scenarios, harm severity, and modality-specific stealth levels. Furthermore, we introduce a hierarchical evaluation framework to assess fundamental response reliability, actual risk exposure, and the structural integrity of defensive behaviors. Extensive zero-shot evaluations across 17 state-of-the-art MLLMs provide a comprehensive safety profile of current multimodal systems. Our analysis systematically investigates cross-modal input configurations and uncovers safety implications associated with Chain-of-Thought (CoT) reasoning. These multifaceted findings underscore the urgent need for robust, reasoning-aware safety alignment in the multimodal landscape.

The complexity of legal language and limited accessibility to legal information pose significant challenges to justice delivery in Nepal. Traditional legal services remain inaccessible to many citizens due to language barriers, information fragmentation, and a critical shortage of legal expertise, particularly in rural areas. We present NepLEGiT (Nepali Legal Expertise through Generative and Extractive Pre-trained Transformers), a specialized small language model (SLM) designed to democratize legal knowledge and enhance legal-service delivery in Nepal. We pre-train a decoder-based GPT-2 SLM from scratch on a curated corpus of ~4 million tokens of Nepali legal text, covering constitutional law, civil and criminal codes, and administrative regulations. The model comprises ~30 million parameters in a 6-layer, 6-head, 384-dimensional transformer trained with warmup cosine-decay scheduling, gradient accumulation, and mixed-precision arithmetic. On a held-out validation split, NepLEGiT attains a cross-entropy loss of 0.5684, a perplexity of 1.8, and a next-token prediction accuracy of 82.9%. We further evaluate continual masked-language-model pre-training of mBERT and MuRIL on the same corpus; mBERT achieves a perplexity of 2.35 (eval loss 0.8565), outperforming MuRIL (perplexity 6.07, eval loss 1.8026), providing a strong encoder baseline complementary to NepLEGiT's generative orientation.

Recent multimodal large language models (MLLMs) have advanced document understanding, visual question answering, and text extraction. However, their reliability in low-resource, non-Latin settings remains uncertain. Khmer form documents present particular challenges because they contain complex script forms, mixed Khmer-English fields, and monetary values in both Cambodian Riel and US Dollars. Available resources for Khmer Document VQA are also limited. This paper presents a pilot diagnostic evaluation of open MLLMs on Khmer document images. We construct an evaluation subset from the previously introduced KH-FUNSD collection, covering invoices, receipts, quotations, and other business forms. The subset includes questions in English and Khmer, with answers retained in their original English, Khmer, mixed-script, or numeric forms. Rather than introducing a full public benchmark, this study examines the capabilities and failure modes of existing models. We evaluate representative open Qwen-VL models using direct image-based prompting and compare parser-assisted and external OCR-assisted configurations with Qwen3-VL-8B. Direct Qwen3-VL-8B outperforms smaller models, achieving 51.9% overall accuracy, although performance remains limited for Khmer-script and mixed-script answers. External OCR produces the strongest results, reaching 61.9% with Tesseract and 61.6% with PaddleOCR. Nevertheless, Khmer-script answers remain substantially more difficult than English and numeric fields. The results indicate that current MLLMs can process visually clear English and structured numeric content, but reliable native Khmer document understanding remains an open challenge.

Conversational assistants increasingly rely on persistent long-term memory to personalize responses across sessions. However, when stored user information is reintroduced into the model context, it can also influence responses in inappropriate or unrelated settings. We study two such failure modes in memory-augmented LLMs: cross-domain leakage, where memories from one life domain affect responses in another, and memory-induced sycophancy, where stored user beliefs make models more likely to agree with the user rather than respond truthfully. We apply a simple inference-time modification to how memories are presented to the model, without changing the model or the memory contents. Across seven models on PersistBench, we compare the commonly used all-in context format, where memories are injected as an unstructured list, with structured formats that partition memories by domain. This simple modification consistently reduces cross-domain leakage while preserving utility, with our strongest method reducing leakage by $8.8\%$ on average relative to the baseline.

Although large language models can translate some historical languages surprisingly well, their usefulness in digital humanities workflows is limited by the lack of reliable evaluation. We investigate whether existing automatic evaluation metrics developed for modern languages are reliable in this setting, using translation from Classical Chinese to English as a test case. We introduce a diagnostic framework based on minimal pairs capturing error types salient in scholarly use, probing both reference-based and reference-free metrics for error sensitivity and tolerance to valid variation. We find that all metrics exhibit blind spots, however MetricX24 performs best overall. Our findings highlight the need for more robust and interpretable metrics for historically and culturally distinct translation settings.

Pretrained Transformer encoders such as AraBERT, MARBERT, and CAMeLBERT have become the standard backbone for Arabic natural language understanding, but their self-attention mechanism scales quadratically with sequence length, which limits efficiency on long documents. Mamba, a selective state-space model (SSM), offers linear-time sequence modeling as a competitive alternative to attention, yet no dedicated bidirectional Mamba encoder pretrained specifically for Arabic currently exists. We introduce AraSSM, a bidirectional Mamba encoder pretrained via masked language modeling on a corpus combining Arabic Wikipedia and CulturaX text, trained end-to-end on four consumer-grade NVIDIA RTX 2080Ti GPUs (11GB) over approximately ten days. We evaluate AraSSM by fine-tuning on four established Arabic NLU benchmarks covering sentiment classification (HARD), named entity recognition (ANERcorp), extractive question answering (ARCD), and natural language inference (XNLI-ar), following the per-task evaluation protocol introduced by AraBERT, and report results as mean +/- standard deviation across three fine-tuning seeds. AraSSM matches or exceeds published base-sized Transformer baselines on sentiment classification (96.37 +/- 0.03% accuracy on HARD), is competitive on extractive QA (32.19 +/- 1.07 EM, 63.79 +/- 0.25 F1 on ARCD) and named entity recognition (81.54 +/- 0.30 entity-level F1 on ANERcorp), and trails the base-sized Transformer range on natural language inference (72.83 +/- 0.07% accuracy on XNLI-ar), despite being trained entirely from scratch on consumer hardware rather than large-scale accelerator clusters.

Agentic reinforcement learning (RL) often suffers from delayed and sparse rewards in real-world environments. A promising solution to this challenge is credit assignment, which aims to decompose trajectory-level rewards and provide more fine-grained supervision for intermediate decisions. However, existing credit assignment approaches ignore the rich process information naturally generated during environment interaction, e.g., interaction history. We argue that such information provides valuable supervision for identifying the contribution of individual actions. To this end, we propose Environmental Feedback-based Credit Assignment (EFCA), a multi-timescale credit assignment approach for long-horizon agentic RL. EFCA complements the long-term outcome signal with two environment-grounded process signals: a short-term feedback signal that captures the immediate effect of the current action and a medium-term state-history signal that identifies ineffective patterns from recent interactions. Both signals are directly extracted from environment feedback and integrated through a return reweighting mechanism. Experiments on ALFWorld and WebShop demonstrate that EFCA consistently improves both task success and task quality over strong baselines, highlighting the effectiveness of environment-grounded multi-timescale credit assignment for long-horizon agentic RL.

LLM routers promise efficiency by matching each request to the cheapest adequate model, and are increasingly applied per step inside multi-step agents. Yet agentic routers are evaluated like single-turn routers: by replaying logged trajectories and substituting another model's recorded outputs, assuming the rest of the trajectory is unaffected. We test this assumption with branching rollouts: we fork live SWE-bench agent trajectories at controlled points, rebuild the environment, continue each fork with a different model, and compare against same-model control forks that isolate sampling and replay noise. Across six paired runs (~900 rollouts), swaps exceed their matched control floors by +0.25 to +0.66 normalized edit distance (multiplicity-corrected CIs exclude zero), rewriting 61-94% of post-fork actions; 74-77% of early swaps diverge at the first post-fork action, versus 6-35% of controls, leaving only 3% of replayed states valid. Divergence decreases with fork depth in both directions. All five outcome flips we observe occur in swap arms, upgrades rescuing unsolved instances and a downgrade losing the sole solve, and zero occur across 359 control forks. Scoring these same swaps with a log-stitching replay evaluator, replay mispredicts every success-relevant outcome call and predicts patches with 0.00-0.11 similarity to reality. Auditing the noise floor, temperature-0 "determinism" is configuration-dependent: FP8-served controls diverge on over 90% of forks while AWQ-served ones remain near-identical; and under tight budgets the stronger model more often exhausts its steps without submitting. Replay-based benchmarks score the wrong world for agentic routing; we release our harness and all trajectories.

Retrieval-Augmented Generation (RAG) systems in production operate under strict service level objectives (SLOs) on tail latency and infrastructure cost. However, standard retrieval pipelines rely on fixed retrieval budgets that ignore query difficulty, over-retrieving for easy queries and under-serving hard ones, forcing operators to trade answer quality against SLO compliance. This paper proposes SAGE, a learned SLO-aware adaptive retrieval policy that dynamically selects the number of passages k per query. SAGE uses lightweight features derived from initial retrieval (e.g., score distributions, rank gaps, lexical signals) and is trained offline via imitation learning from an oracle that approximates optimal latency-quality trade-offs. At inference, it adds no LLM calls and minimal overhead. On Natural Questions, under a 5s P95 latency SLO, SAGE achieves 95% SLO compliance versus 30% for the best static baseline (k=20), reduces P95 latency by 36% and retrieval cost by 51% with only 2 percentage points Exact Match (EM) loss. A single policy trained on Natural Questions generalizes across HotpotQA, UnSeenTimeQA, and four LLM families (Llama, Qwen, Mistral, Gemma), consistently yielding +45-52 point SLO improvements without quality degradation.

Multi-agent LLM systems often fail not for lack of candidate answers, but because they have no persistent mechanism for deciding which incompatible claim should currently be trusted. Majority vote, debate, and judge-based selection choose an output without recording which claim wins, which is contested, or why a later update supersedes it. We present \term{LatticeMind}, a conflict-aware structured memory that handles contradiction at write time. It maintains explicit item status, applies cheap symbolic conflict checks, and invokes LLM reconciliation only for unresolved semantic cases. On a label-blind ConflictBank evaluation that removes source-name hints, LatticeMind reaches 0.97 accuracy versus 0.61 for the strongest aggregation baseline, with the gap significant at $p<10^{-6}$ by paired McNemar test. Ablations show that removing the checker or the reconciler costs 12 to 14 points. On four secondary planning benchmarks the picture is mixed: LatticeMind beats naive merge on three of four, but does not replace deliberation methods on tasks rewarding iterative search.

Focus particles such as "even" and "only" are central to formal semantic theories that posit structured representations over sets of alternatives. "Even" highlights unexpected or extreme alternatives, while "only" enforces exclusivity. If such scalar representations are robust and generalizable, they should give rise to consistent judgments across contexts and systems. In this work, we test whether humans and large language models (LLMs) construct stable scalar representations from sentences containing these particles. Using a dataset of approximately 100 items, participants and models were asked to make scalar judgments. Preliminary results suggest that similar outputs across humans and LLMs may arise from different underlying mechanisms.

In-context learning (ICL) can induce emergent misalignment (EM), where narrow misaligned examples alter answers to unrelated questions. Existing prompts, however, conflate harmful-text exposure with an invitation to continue assistant behavior. We hold harmful answers fixed while varying their delivery as demonstrations, evidence, assistant history, or tool output. Across ten independently sampled contexts, demonstration framing raises broad EM by $30$–$32$ percentage points on a susceptible Gemini model; the gap survives domain exclusion, semantic clustering, unseen questions, and four prompt templates. Format and length-matched controls show that harmful content is necessary but insufficient. A role times continuation factorial further reveals model-dependent provenance effects: Gemini follows both assistant and tool histories, whereas Grok largely resists tool-framed continuation. Several other frontier and open-weight models show no gap. Blinded human audits confirm every main contrast and show that the model judge underestimates active-condition failures. Thus continuation framing is a strong, model-dependent moderator of ICL-EM, not a universal consequence of harmful context.

Post-training quantization reduces the deployment cost of large language models, yet how severely a quantized model degrades is not determined by bit-width alone. We systematically study weight-only post-training quantization across bit-widths, quantization methods, model scales and downstream tasks on multiple model families. We observe that such degradation varies substantially across these factors: 4-bit quantization usually preserves performance, 2-bit often causes broad degradation, and at 3-bit, degradation becomes apparent but varies markedly with task type, quantization method and model scale. To explain this variability, we use the signal-to-noise ratio (SNR) to measure how strongly quantization perturbs full-precision representations. We trace degradation back to two linked processes: how quantization errors arise within individual modules, and how they accumulate across layers. First, a source SNR decomposition shows that newly introduced errors depend on three factors: the magnitude of the weight error, the strength of the task-specific signal, and how strongly the quantization error aligns with task-specific activations. Different factors affect these components in distinct ways. Second, a cross-layer propagation analysis shows that these errors can be attenuated, preserved, or amplified as they pass across layers, and that larger models benefit from weaker error amplification. Together, these results establish that quantization degradation is governed by how errors are introduced at the source and how they accumulate across the network.

Abstractive text summarization systems frequently generate fluent yet unfaithful summaries by fabricating or distorting relationships between entities and events. Such relation-level hallucinations undermine the reliability of generated summaries, particularly in high-stakes domains. In this work, we present a refined and grounded framework for evaluating relation hallucination in abstractive summarization. We present the empirical Relation Hallucination Index (RHI) by introducing a dependency-aware relation extraction algorithm that incorporates lemmatization-based normalization, named entity grounded subject resolution, passive agent recovery, negation-aware verb modeling, reporting verb filtering, nominal relation fallback, clausal propagation, and systematic deduplication. These enhancements improve the structural fidelity of extracted relation triples and reduce spurious matches during evaluation. In addition, we introduce a normalized formulation of RHI to ensure scale-invariant comparison between datasets and models. The revised metric decomposes hallucination into interpretable components, aggregates relation hallucination metric into a normalized relation faithfulness score. Extensive evaluation across multiple state-of-the-art summarization models demonstrates that the grounded extraction process yields more stable and discriminative hallucination measurements. The proposed framework advances automated relation-level faithfulness evaluation and supports coherence-aware, hallucination-sensitive model analysis.

While Large Language Models (LLMs) employing Chain-of-Thought (CoT) exhibit superior reasoning capabilities, the neural mechanisms distinguishing this explicit Thinking mode from direct answer generation (NoThinking mode) remain poorly understood. To deconstruct this cognitive process, we apply Top-K Sparse Autoencoders (SAEs) to the intermediate representations of DeepSeek-R1-Distill-Qwen-7B and examine the model's divergent behaviors across math-solving tasks of three distinct difficulty levels. Observationally, we identify a clear distinction in how the model functions under two reasoning modes: Thinking mode relies on sparse and high-intensity feature activations driving verbal deduction independent of problem complexity, whereas NoThinking mode exhibits an adaptive and diffuse pattern prioritizing symbolic manipulation. Causally, suppressing the three most active sparse features by Total Activation Volume reveals three principles: (i) reasoning and syntactic structure are tightly coupled, as interventions consistently degrade \LaTeX{} and boxed-solution formatting; (ii) Thinking responds to disruption with compensatory over-generation marked by increased metacognitive cues and repetitive, low-information continuations; and (iii) coherent CoT behavior depends on a fragile coordination among specialized features, yielding distinct failure modes under perturbation but a consistently impaired output structure.

Knowledge Distillation is a widely adopted technique in the training and fine-tuning of large language models (LLMs) enabling transfer of structured information and functional behavior from a large teacher model to a smaller student model while significantly reducing computational costs. However, as the use of distillation increases in both scale and complexity it raises an important question about what kind of knowledge is really transferred from the teacher model. In this work, we argue that apart from the functional knowledge, student models also learn behavioral patterns, specifically how a model represents its own identity raising concerns about output homogeneity, model biases, and accountability. To address this challenge, we introduce STEMMA, a multi-modal and multi-agent framework in which role specific agents collaboratively probe self identification behavior in different models. We also contribute a set of adversarial prompts designed manually to evaluate identity consistency in LLMs. Our results show that to an extent most models are vulnerable to inconsistencies in self-representations.

The rapid advancement of Large Language Models (LLMs) is revolutionizing AI for Games by enabling open-ended and fluid interactive storytelling. However, existing research has largely overlooked the critical challenge of maintaining long-horizon logical consistency and narrative integrity against unconstrained user interventions. To address this, we formulate this challenge as Narrative Commitment Preservation (NCP), and take interactive narrative as our testbed. We introduce NCP-Bench, a benchmark of 100 narrative environments derived from movie synopses. Each environment includes a structured narrative specification (trajectory, commitments, and initial facts) that we can automatically check throughout the interaction between the player agent and the narrator agent. Experiments across state-of-the-art LLMs reveal a substantial long-horizon consistency gap: high linguistic quality does not guarantee commitment preservation; even strong models frequently generate logically conflicting content under adversarial interventions, with the best-performing model (GPT-5.2) achieving only 42% survival rate after 20 turns and fact conflict rates ranging from 40% to 68% across models, and only isolated runs satisfying all achievement commitments within the 100-turn limit.

We extend the DS@GT ARC working-note submission to the Touché 2025 Retrieval-Augmented Debate task. The task has two subtasks: generating the next utterance in a simulated debate, and evaluating debate responses according to the Gricean maxims of Quantity, Quality, Relation, and Manner. The DS@GT ARC submission consisted of six leading LLMs from three providers through a retrieval-augmented prompting pipeline. We summarize the results from the working paper and explore whether multi-LLM evaluator agreement is a reliable proxy for official evaluation performance. The analysis shows that frontier LLM systems are strong response generators, and as evaluators they agree strongly within model families. However this consensus does not reliably track the official evaluation target, with the largest gap on the Quality maxim. The accompanying source code for this paper is located at https://github.com/dsgt-arc/touche-2025-rad and https://github.com/dsgt-arc/touche-2025-rad-analysis.

Credit scoring increasingly relies on models whose decision logic cannot be read off their parameters, in tension with supervisory expectations that adverse decisions be explainable. A common proposal closes that gap with a language model: compute feature attributions, hand them to an LLM, and let it write the rationale. We build such a system end to end and test whether the second half of the promise holds. The predictive component is a multi-scale stacking ensemble fusing four differently regularised gradient-boosting learners with a residual network through a neural meta-learner trained on out-of-fold predictions. On a public 32,581-application credit dataset it reaches test ROC-AUC 0.9539 (95% CI [0.9462, 0.9616]) and PR-AUC 0.9137, beating the best single model by Delta-AUC = 0.0143 (p = 0.016 under a conservative independence assumption). Our central finding is asymmetric. The ranking gain is real but operationally small: at the F1-optimal threshold the ensemble avoids only six additional missed defaults out of 1,422 against a tuned random forest, cutting cost-weighted loss by under 2%. The narrative layer fails in a way prompt engineering alone does not fix. In an audited case the model named three factors as risk-increasing that the supplied attributions scored as risk-reducing, omitted the dominant driver, and introduced a feature never given to it. We trace this to properties we measure rather than assume: SHAP and LIME agree on which features matter (overlap@10 = 0.80) but not on their order (tau = 0.43, p = 0.18), and the attribution sign for the model's most sensitive input is near a coin flip across applicants (modal-sign share 0.53). Calibration (ECS = 0.117) and perturbation stability (DPD = 0.078) both fall short of our own thresholds. Constrained prompting is necessary but not sufficient: grounding must be verified after generation, not assumed.

Multimodal expansion of large language models (LLMs) enables new perceptual capabilities but often compromises the language intelligence acquired during pretraining. In this work, we investigate this phenomenon from the perspective of internal adaptation dynamics and discover that neurons in pretrained LLMs exhibit heterogeneous plasticity during multimodal learning: some neurons are critical for preserving language capabilities, while others are more adaptive to multimodal knowledge. Based on this insight, we propose NeuPAT (Neuron-aware Plasticity Allocation Tuning), a lightweight and architecture-agnostic framework that allocates neuron-wise update constraints during multimodal instruction tuning. NeuPAT uses a small-scale probing stage to estimate neuron adaptation patterns and selectively protects language-sensitive neurons while promoting multimodal adaptation through more plastic neurons. Experiments across diverse LLM families demonstrate that NeuPAT recovers 94.5% of the language capability degradation caused by vanilla tuning on 11 language benchmarks while maintaining comparable multimodal performance, providing an effective approach for capability-preserving multimodal expansion.

Although multilingual approaches to figurative language identification are not new, the shift beyond language-homogeneous training data requires a clearer understanding of the contribution of translated multilingual supervision. We examine this question using 742 proverb concepts across 6,787 translated instances for seven languages. We evaluate five models including multilingual encoders and instruction-tuned LLMs through progressively increasing levels of multilingual supervision. Moreover, we introduce multidimensional annotation framework for proverbs that characterizes proverbs through four complementary figurative forms: Metaphorical, Moral/Advisory, Cause-Effect, and Culture-Specific. Our findings show that overall, adding multilingual training data beyond 50% provides only limited additional improvement, although the best supervision level varies across models and languages. Also, we show that combining diverse figurative forms yields the strongest overall performance. A notable finding is that the least frequent figurative form culture-specific exhibits the largest performance gains under multilingual supervision. Furthermore, the moral/advisory and culture-specific forms of proverb contribute more to instruct tuning LLM overall figurative identification performance. These findings motivate multilingual figurative identification to move beyond metaphor-centric taxonomies toward concept-level multidimensional frameworks that explicitly model complementary forms of figurative meanings that are context representative.

Diffusion language models (DLMs) can revise earlier predictions as context evolves, but rollback requires costly full-sequence recomputation. Conventional key-value (KV) caching assumes immutable historical states and is difficult to reconcile with such revision. We introduce Adaptive Reuse of Cached Hidden States for Efficient Rollback (Archer), a training-free caching method for rollback-capable DLMs. Archer recomputes the mutable response at every update while reusing prompt K/V within a bounded response-state neighborhood. This asymmetric boundary amortizes prompt computation without restricting rollback. Bounded reuse also delays feedback from tentative tokens while keeping response revision active. Our analysis characterizes the cache boundary, bounds state-dependent approximation error, and gives a decoder-margin condition for preserving full-refresh decisions. On code-generation benchmarks, Archer achieves the best mean performance of 33.63% with a 2.57x mean speedup, reaching 2.95x speedup on a single benchmark. On non-code benchmarks, Archer achieves up to 4.33x speedup over LLaDA-Confidence while improving average accuracy. Controlled analyses support the delayed prompt feedback mechanism. Our code is available at https://github.com/Hxnng/Archer

Classifier-free guidance (CFG) is usually kept on throughout masked diffusion language model decoding, although its benefit varies across prompts and over time. We study when CFG is actually needed by comparing, from any partial output, the probability of eventual constraint satisfaction under continued CFG and under base-only continuation. Their difference defines the remaining value of guidance. Guidance dependence is highly prompt-specific. Many prompts already succeed without CFG, while for others it provides no measurable benefit or can be harmful. For prompts that do benefit, the gain is often concentrated early. We define the commitment horizon $\astar$ as the earliest point from which switching all remaining decoding to the base model reduces final success by no more than a chosen tolerance. Under the base model, the corresponding success probability, or committor, is a martingale. To first order, CFG's per-step effect is governed by the covariance between the guidance logit direction and the successor committor. This gives a local account of when guidance can help, but it does not by itself locate the horizon. Among prompts with an observed preterminal horizon, $\astar$ is usually early and varies more within constraint families than between them. Freezing each prompt at its own cross-fitted horizon is noninferior to full CFG on all 13 subtasks at the prespecified margin, even while many tokens remain masked. This separates commitment from realization. The boundary also identifies a later region in which higher parallelism adds only a small cost in constraint success, although fluency still degrades with parallel width. For failed trajectories, reopening committed positions improves recovery in both failure modes.

Current end-to-end speech dialogue models are primarily optimized for mainstream languages and remain limited in low-resource dialect scenarios due to the scarcity of dialect speech data. Moreover, during dialect adaptation, the semantic representation space of speech dialogue models continuously evolves, while conventional speech supervision remains unchanged, leading to semantic inconsistency between hidden representations and speech targets and degrading speech stability and naturalness. To address these issues, we propose DialectS2S, an end-to-end speech dialogue model for Chinese dialects. We first develop a scalable dialect speech dialogue synthesis pipeline for efficient data construction. We further introduce a two-stage post-training strategy with self-aligned speech supervision, which aligns the semantic content of speech supervision with the evolved semantic representations of the model to improve dialect speech generation quality. Experimental results show that DialectS2S consistently outperforms existing baselines across multiple Chinese dialects in speech dialogue, achieving substantial improvements in dialect consistency, response quality, and speech intelligibility. Our work provides an efficient and scalable solution for end-to-end speech dialogue modeling in low-resource dialect scenarios. To facilitate future research and practical applications, we fully open-source the DialectS2S framework, including model checkpoints, training datasets, and fine-tuning code.

Post-Editing (PE) of Machine Translation (MT) output often involves repeating the same lexical and terminological corrections across many segments, especially in specialised and highly repetitive documents. Despite substantial work on Automatic Post-Editing (APE), most available corpora operate at the sentence level, others are synthetic, and overall not designed to study how corrections propagate in realistic Computer-Assisted Translation (CAT) workflows. This paper presents the APEX-VW (Automatic Post-Editing eXperiments on Virtual Wards) Corpus, a new open English-Spanish (EN-ES) dataset built from recent NHS virtual-ward documents and professional PE in Trados Studio, with controlled MT, terminology, and quality assurance settings. The corpus contains seven document-coherent source texts totalling 42k words, translated with four MT systems representing different paradigms and then post-edited by professional translators. Unlike prior resources such as WMT APE corpora, eSCAPE, MLQE-PE, or LangMark, the dataset preserves document order and CAT-tool context, making it suitable for research on terminology normalisation, correction propagation, and human-in-the-loop translation support. The paper describes the corpus design, data preparation, PE setup, and initial corpus statistics, and positions the resource as a benchmark for document-level APE and propagation-aware assistive tools.

Safety alignment in multilingual models is uneven: a model that reliably refuses a harmful request in English will often comply with the same request in a lower-resource language. We trace this gap mechanistically in sarvam, an Indic-multilingual mixture-of-experts reasoning model, and find it is not a failure to detect harm. Harm is encoded as an internal direction that is nearly language-invariant in mid-network (English-vs-Indic cosine ${\approx}0.9$ at $L11$), and steering that direction upstream causally controls refusal. But the detection direction is orthogonal to the change that actually writes the refusal, which is late and assembled over the course of generation rather than read off in a single forward pass. We attribute the write to a specific, localizable circuit, a mixture-of-experts writer held in check by an attention opposer and price every way of intervening on it: damping the opposer is cheap and effective, amplifying the writer is a cost wall, and surgical edits to the responsible heads do nothing. The circuit's organization, and the gradient method that exposes it, recur in a second, unrelated MoE model, while the lever's strength is architecture-specific. The result is a cost-measured map of where a multilingual safety repair can land, and what it costs

Large language models (LLMs) can generate fluent and useful responses but remain prone to hallucinations. We introduce Prompt Embedding Probes (PEP), a white-box method for answer-level hallucination detection from the hidden states of a frozen LLM. PEP extends standard linear probes by augmenting the input with a small number of learnable prompt embeddings. We evaluate PEP on TriviaQA, GSM8K, and MedQA using Qwen3 models at multiple scales. PEP improves hidden-state-based detection over standard linear probes in the main in-distribution setting. We further evaluate PEP for pre-generation prediction, cross-model transfer, and out-of-distribution generalization. PEP remains effective in the pre-generation and cross-model settings, whereas robust cross-dataset transfer remains difficult. These results show that prompt-based adaptation can strengthen hidden-state probing while keeping the backbone frozen and adding only a small number of trainable parameters.

In recent years, the term voiceprint has regained attention, particularly in technological applications and policy-making contexts, often carrying the assumption that a person's voice constitutes a stable and unique biometric trace analogous to a fingerprint. Yet this conception has been repeatedly criticized and rejected by forensic voice experts throughout the decades since its introduction. Although voices undoubtedly contain speaker-related information, this simplified conception obscures the highly dynamic and context-dependent nature of speech. This article revisits the voiceprint fallacy and reconsiders what can count as evidence of speaker identity by reviewing the historical development of voiceprint identification, evidence on human voice variability, developments in forensic voice comparison, research on human and automatic speaker recognition, and the recent challenge posed by deepfake speech to speaker identity. We point out that the voiceprint metaphor and its underlying implications are scientifically misleading because they transform a probabilistic source of speaker information into an imagined stable object of identity. We argue that speaker identity assessment does not require, and current evidence does not support, the existence of a stable and individually unique voiceprint. For speaker recognition and voice biometrics, this distinction motivates interpreting learned speaker representations with respect to the conditions under which they are trained and evaluated, and explicitly assessing their robustness to relevant sources of within-speaker variability, domain mismatch, and synthetic manipulation.

Reasoning language models increasingly use test-time compute to improve performance, but existing evaluations typically study this compute one question at a time. Yet when multiple problems share an end-to-end cost or latency constraint, models must decide how to divide limited inference compute among them. We introduce an exam-style evaluation framework for studying this setting, in which a model must distribute one shared token budget across questions with different difficulty and point values to maximize its total score. Across several open and frontier reasoning models, we find that models fail to allocate a shared budget strategically across questions of varying difficulties and values. Models behave largely as greedy sequential solvers: they prioritize questions by presentation order, front-load effort on early questions, and remain insensitive to value, with these tendencies becoming more pronounced as the number of questions grows. Explicit planning prompts spread compute more evenly but do not produce value- or difficulty-aware prioritization. The same behavioral pattern extends from mathematical to code reasoning. These findings establish global budget allocation as a distinct capability that is not captured by conventional per-question evaluation and remains a challenge for current reasoning models.

Retrieval-Augmented Generation (RAG) improves the factuality of large language models with external knowledge, yet conflicting evidence remains a fundamental challenge in dynamic and adversarial environments. Existing approaches often treat conflicts as static inconsistencies and select more reliable knowledge, overlooking that the same conflict may arise from legitimate knowledge evolution, malicious manipulation, or unresolved uncertainty. We formulate conflict origin attribution as a new problem in RAG: identifying which explanation of conflicting evidence is supported by observable context rather than simply which fact should be trusted. We propose EvoTrustRAG, a training-free framework for evolution-aware conflict attribution and evidence handling before answer generation. EvoTrustRAG represents span-grounded retrieved facts as a conflict evidence graph, evaluates grounded evolution and directional intervention hypotheses using temporal relations, support structure, and auxiliary consistency, and projects local decisions onto a globally consistent explanation of each conflict group. The attribution determines whether earlier and later states are preserved as temporal knowledge, an intervention candidate is separated from the primary context, or an unresolved conflict remains visible to the generator. Unlike provenance-based approaches focused on post-hoc analysis, EvoTrustRAG determines during inference whether conflicting evidence follows plausible knowledge evolution, exhibits intervention-like support, or cannot be reliably attributed. Experiments show that EvoTrustRAG achieves 81.4% average accuracy on benchmark-native conflict settings, improves attribution macro-F1 from 72.2% to 79.1% over the strongest baseline, and reduces the error rate under the strongest coordinated attack from 31.2% to 16.0%.