Papers

Agentic AI is rapidly proliferating across diverse real-world domains such as software engineering, yet public trust has not kept pace. The central reason is that responsibility, despite being widely discussed, remains a subjective and unenforced concept, as no current agentic framework produces the quantifiable, traceable, and interventionable provenance needed to assign it when harm emerges from compositions no single party designed. We position that what is missing is not better benchmark-level evaluation but $\textbf{explicit provenance}$ across the full agentic lifecycle, which is the only viable basis for making responsibility computable and actionable. We advance this agenda along four axes: establishing $\textit{why}$ such provenance is a structural necessity by identifying responsibility gaps across sociotechnical dimensions, formalizing $\textit{what}$ it must encode through a causal attribution function and responsibility tensor, discussing $\textit{how}$ it can be made computable across four lifecycle layers with preliminary experiments showing that provenance is estimable and interveneable online before irreversible harm accumulates, and examining $\textit{who}$ bears responsibility through a concrete agentic incident. Explicit provenance is not a discretionary refinement but the necessary condition for responsible agentic AI, and no stakeholder across its ecosystem can afford to treat it as optional.

Multimodal LLMs are evolving from vision-language to tri-modality that see, hear, and read, yet pipelines and benchmarks remain English-centric and compute-heavy. The tutorial offers an overview of this emerging research area for multilingual multimodality across text, speech, and vision under limited data/compute budgets, synthesizing foundations, recent multilingual models (PALO, Maya), speech-text LLMs. We cover low-cost data creation/curation; adapter stacks for tri-modal alignment; culture-aware evaluation beyond English and hands on resources for fine-tuning a compact multilingual VLM and wiring a speech->text->LLM pipeline. The content will be delivered as an interactive half-day tutorial, designed for researchers and practitioners working on multilingual, multimodal AI in low-resource language settings.

Brain tumor diagnosis is largely dependent on Magnetic Resonance Imaging (MRI) evaluation, which requires radiologists to synthesize thousands of images across multiple 3D sequences and longitudinal studies. This process requires advanced neuro-radiology training, poses substantial cognitive load, and is highly time-consuming. Despite increasing demands in radiology, this expertise is difficult to scale, straining the current health systems. Vision-Language Models (VLMs) provide an opportunity to reduce this burden through a semi-automated, interactive interpretation of complex brain MRIs. However, they are currently underutilized in neuro-oncology due to a lack of specialized benchmarks for evaluating them. We introduce a clinically relevant visual question answering (VQA) benchmark – the UCSF-PDGM-VQA dataset – consisting of 2,387 QA pairs from 473 glioma-related MRI studies in the public UCSF-PDGM dataset. We further establish a performance baseline for six state-of-the-art vision-language models (VLMs) and one large language model on this dataset. We find that current models are incapable of effectively processing multi-sequence, 3-dimensional MRI scans, thus resulting in a suppression of visual features and over-reliance on language priors, causing modality collapse. These findings underscore a critical deficiency in current model reliability and safety within clinical settings, necessitating the development of robust, domain-specific VLMs.

Existing deception datasets label completed outputs as honest or deceptive, treating deception as a property of the final response rather than a function of the model's reasoning trace. This obscures a more fundamental question: when does a language model become committed to deception? We introduce counterfactual localization: for each sentence prefix in a reasoning trace, we fix the prefix, resample continuations, and estimate the probability of a deceptive outcome. To scale this, we construct five environments (spanning strategic bluffing, maze guidance, financial advice, used-car sales, and offer negotiation) in which deception is never prompted but emerges from strategic incentives and labels follow mechanically from environment state rather than subjective human judgment. The resulting corpus localizes $\sim$1.46M sentences across four reasoning models, drawn from over 94.1M sampled continuations, 91.5B generated tokens, and over 100K scenarios. Sentence-level human evaluation confirms that detected commitment points correspond to interpretable shifts in decision state. Using this resource, we show that lexical cues for commitment prediction transfer poorly across environments, whereas attention-based transition features generalize out of distribution, suggesting that deceptive commitment is reflected in reusable changes in reasoning dynamics rather than surface form. We further identify compact attention-head sets (under 10% of heads) that, selected on one environment, causally suppress deceptive commitment across held-out environments. We release the corpus as a substrate for studying deception, and more broadly commitment, in language-model reasoning.

Production LLM deployments increasingly maintain heterogeneous model pools spanning order-of-magnitude cost differences. Existing routers make binary strong-vs-weak decisions and couple learned parameters to specific model identities, requiring retraining whenever the catalog changes. We present HyDRA (Hybrid Dynamic Routing Architecture), a framework that predicts fine-grained, multi-dimensional capability requirements per query and matches them against configuration-defined model profiles via shortfall matching. A ModernBERT encoder with K=4 independent sigmoid heads scores each query along reasoning, code generation, debugging, and tool use; a shortfall-matching algorithm then selects the cheapest model whose capabilities meet the predicted requirements. The deployed predictor runs at 86 ms median CPU inference latency in production, and is fully decoupled from the model catalog – adding or removing models requires only a configuration change, with zero retraining. On SWE-Bench Verified (5-model pool: GPT-5.4-mini, Claude Haiku 4.5, GPT-5.3 Codex, Claude Sonnet 4.6, GPT-5.4), HyDRA's tunable shortfall threshold spans three regimes: peak-quality exceeds the always-strong Claude Sonnet 4.6 baseline (75.4% vs. 74.2% resolution) at 12.9% cost savings; iso-quality matches Sonnet at 54.1% cost savings, a 6x improvement over our prior in-house binary router at 9.1%; aggressive pushes savings to 72.5% for a 3.2-point quality trade. Results generalize across LiveCodeBench, BigCodeBench, and tau-bench. HyDRA is deployed to all users in GitHub Copilot's VS Code Chat auto-mode and – to our knowledge for the first time in the LLM routing literature – demonstrates language-invariant routing across CJK, European, and other script families.

Retrieval-Augmented Generation (RAG) is widely employed to mitigate risks such as hallucinations and knowledge obsolescence in medical question answering, yet its predominantly single-round, static retrieval paradigm misaligns with the multi-stage process of clinical reasoning. This compressed workflow induces two structural deficiencies: question-to-query translation often lacks clinically grounded semantic interpretation, and retrieval lacks iterative sufficiency feedback, making it difficult to form reliable evidence chains. We argue that both issues stem from a deeper cause: overloading a single reasoning chain with heterogeneous tasks of interpretation, exploration, and adjudication. The remedy is to reconstruct the workflow via task decoupling and dynamic multi-round exploration. To this end, we propose SEMA-RAG, a Self-Evolving Multi-Agent RAG framework for medical question answering, which assigns these roles to three specialist agents: the Interpreter Agent for clinical schema interpretation, the Explorer Agent for sufficiency-driven self-evolving retrieval, and the Arbiter Agent for evidence adjudication and answer selection. Across five benchmarks and five LLM backbones, SEMA-RAG improves the strongest baseline by +6.46 accuracy points on average, measured per backbone.

Distilling vision-language models into faster hybrid architectures, such as 3:1 Mamba-2/attention mixes, is now standard practice for making inference efficient. Aggregate benchmarks suggest that this works but they hide selective failures. When we distill Qwen3-VL-8B-Instruct into a 3:1 Mamba-2/attention hybrid, student model stays within 2 points of the teacher across visual reasoning benchmarks like MMStar, MMBench, and MMMU-Pro, while dropping 13 points on optical-character-recognition and document tasks. The student can still understand the scene but loses the fine-grained text needed to answer. We localize much of the failure to a specific kind of position. In a high-resolution image, most patches are sky, wall, or smooth texture, while a small fraction carries text, edges, object boundaries, or other local details. In a token-level diagnostic, the top 10% highest-density patches have 3.6$\times$ larger residual drift than the bottom 10% lowest-density patches and 3.5$\times$ larger teacher-masking answer contribution. Uniform weighting devotes many loss terms to low-information background patches, whereas sparse answer-bearing patches receive no special protection. The required intervention is minimal: we replace uniform residual alignment with density-weighted residual alignment, using patch self-dissimilarity as a training-free proxy for position importance. We call this HEED. Compared with normal end-to-end distillation, HEED increases performance by 8.7 points on OCRBench v2 and 5.13 points on a 10-benchmark average. The gain is realized on different teacher models and hybrid architectures. After standard post-training, the student reaches teacher-level performance on the 10-benchmark average with a 4.12$\times$ throughput and a 68% memory saving at 128k context, with no additional parameters and no inference-time cost.

Recent advances in large language models (LLMs) have shown that Chain-of-Thought (CoT) reasoning can substantially improve performance on complex reasoning tasks. At the same time, In-Context Learning (ICL) has become an important mechanism for adapting LLMs to new tasks without updating model parameters, using only examples provided in the prompt. However, standard ICL often struggles on tasks that require multi-step reasoning, because the demonstrations usually contain only input-output pairs and lack explicit intermediate reasoning steps. This paper introduces an Automatic Chain-of-Thought (Auto-CoT) framework to improve ICL by automatically constructing reasoning-enhanced demonstrations. Auto-CoT generates reasoning chains for input-output examples, augments the prompt context with structured intermediate explanations, and removes irrelevant or low-quality demonstrations through a systematic selection process. By incorporating high-quality reasoning examples into the ICL prompt, Auto-CoT guides the model toward more reliable reasoning and improves prediction accuracy. Experiments across multiple reasoning tasks demonstrate that the proposed framework improves ICL performance by providing explicit intermediate reasoning guidance.

Interpreting, editing and compressing language models often involve assessing which parts of their internal representations matter for prediction. Many of these assessments rely on distance, such as reconstruction error or the length of an intervention. For next-token prediction, however, the logit differences that matter most are those among the few most plausible tokens. These differences can be small. We therefore test whether changes to the final hidden state that move the logits less also change the prediction less. Specifically, changes along the smallest singular directions of the unembedding, which maps this state to the logits, move the logits less than random changes of the same length. In open-weight language models across several families, their prediction change and their change in the next-token loss are nevertheless on average several times those of random changes. Across all singular directions, the prediction change is out of proportion to the logit change only at the smallest. These directions carry a large share of the difference between the model's leading candidate tokens. Consequently, distance-based methods undervalue these directions, and rankings by singular value place them last, although they matter disproportionately for prediction. More broadly, our study highlights that comparing internal representations by distance can miss differences that change what the model predicts.

LLMs are increasingly used as “digital consumers” to simulate public opinion, pre-test marketing decisions, and anticipate audience response. However, existing evaluations rarely ask whether a model can reconstruct the concrete reaction patterns that real consumers surface in public discourse. We introduce ConsumerSimBench, a benchmark built from 1,553 real Chinese social-media topics and 23,122 atomic, rule-audited criteria spanning four reaction families. Rather than scoring open-ended generations with a holistic preference judge, ConsumerSimBench decomposes each task into auditable yes-no decisions over concrete reaction points, raising three-judge agreement from 65.8% to 92.1% with 98.4% agreement between pointwise judge decisions and human-majority labels. Across 13 frontier generators, the strongest model, Gemini-3.1-Pro, covers only 47.8% of real reaction criteria, while GPT-5.2 and Claude-4.6 trail far behind despite their strength on technical benchmarks. The failures reveal a sharp gap between technical-benchmark performance and socially grounded consumer intuition. A direct structured reasoning prompt decreases coverage, while a generate–reflect multi-agent pipeline improves MiMo-V2.5-Pro from 32.9% to 37.6% on a subset. ConsumerSimBench reframes consumer simulation as a forecasting problem over real public-discourse reactions, showing that frontier LLMs remain far from reliably predicting what consumers will actually care about in high-context Chinese consumer discourse.

Existing LLM-driven knowledge graph (KG) construction methods predominantly employ stateless batch processing pipelines, exhibiting structural deficiencies in cross-chunk semantic relation capture, entity disambiguation, and construction process interpretability. These limitations undermine KG quality, retrieval precision, and deployment trust in high-stakes domains. We propose RAGA (Reading And Graph-building Agent), an LLM-based autonomous KG construction and retrieval fusion framework. RAGA provides an atomic toolset supporting full KG lifecycle CRUD operations and embeds a Read-Search-Verify-Construct cognitive constraint into a ReAct tool loop. A KG-vector synchronization mechanism enables hybrid symbolic-vector retrieval, while evidence-anchored verification links every knowledge entry to its source text for auditable provenance. Preliminary experiments on a subset of the QASPER scientific QA dataset indicate that RAGA's fusion retrieval outperforms zero-shot baselines, with KG integration providing measurable gains in both answer and evidence quality. The framework design and experimental baseline serve as a reference for agent-driven autonomous KG construction.

Autonomous AI coding agents are becoming a core tool for ML practitioners in industry and research alike. Despite this growing adoption, no standardized benchmark exists to evaluate their ability to design, implement, and train models from scratch across diverse domains. We introduce **1GC-7RC** (*Single Graphic Card: Seven Research Challenges*), a benchmark comprising seven ML tasks spanning language modeling, image classification, semantic segmentation, graph learning, tabular prediction, time-series forecasting, and text classification. Each task provides a locked data-preparation and evaluation script together with a baseline training script; the agent may only modify the training code, has no access to pretrained weights (with one controlled exception for semantic segmentation), no internet access, and must complete each task within a task-specific wall-clock budget (40-120 minutes) on a single GPU. We evaluate seven coding agents: five proprietary (Claude Code with Sonnet 4.6, Opus 4.6, and Opus 4.7; Codex CLI with GPT 5.5; and OpenCode with Qwen 3.6+) and two open-source (OpenCode with Kimi K2.5, Kimi K2.6). Across 5 runs per agent-task pair, we report substantial performance differences that reveal varying levels of implicit ML knowledge, planning ability, and time-budget management. The benchmark, harness, and all evaluation artifacts are publicly available on GitHub at https://github.com/Strolchii/1GC-7RC-Benchmark to facilitate reproducible comparison of future agents. Because our benchmark design is modular, the benchmark can be extended to new tasks and domains, adapted to different GPU budgets, and used to study multi-agent settings, making it a flexible platform for future research on autonomous research agents.

We present Agentic AI Translate, an agentic translator prototype that operationalises the thesis of Yamada (forthcoming) – that the metalanguage of Translation Studies has become an instruction code for generative AI. The system replaces the dominant text-in / text-out paradigm of machine translation with a four-stage agentic cycle (Identify -> Prompt -> Generate -> Verify), preceded by an interactive specification phase in which the user composes – through model-assisted dialogue – a structured translation brief grounded in skopos theory, register, audience, and genre conventions. The verification stage adopts the GEMBA-MQM error-span protocol (Kocmi & Federmann, 2023) for evidence-grounded scoring, and document-level coherence is preserved through a DelTA-lite memory of proper nouns and a running bilingual summary, after Wang et al. (2025). We describe the philosophical motivation, the architectural commitments, the four reference-material categories the system consumes, and the principal design tensions the architecture makes explicit. Empirical validation is left for future work; the contribution here is conceptual and architectural – an executable embodiment of the position that translation in the GenAI era is communication design, not text conversion.

Reinforcement learning (RL) has demonstrated potential for enhancing reasoning in large language models (LLMs). However, effective RL training, which requires medium-difficulty training samples, faces two fundamental challenges: Effective Data Scarcity and Dynamic Difficulty Shifts, where medium-difficulty samples are scarce and become trivial as models improve. Existing methods mitigate this scarcity to some extent by generating training samples. However, these approaches suffer from anchor-free generation, ignoring co-evolution, and difficulty mismatch. To address these issues, we propose D$^2$Evo, a Dual Difficulty-aware self-Evolution RL framework. In each iteration, our method mines medium-difficulty anchors based on the current Solver's capability, trains the Questioner to generate diverse questions at appropriate difficulty levels, and jointly optimizes both components to enable progressive reasoning gains. Extensive experiments demonstrate that D$^2$Evo outperforms existing methods on mathematical reasoning benchmarks with fewer than 2K real mathematical samples, and exhibits strong generalization on general reasoning benchmarks.

Large language models (LLMs) hallucinate with confidence: their outputs can be fluent, authoritative, and simply wrong. In medical, legal, and scientific applications this failure causes direct harm, and detecting it from internal model states offers a path to safer deployment. A growing body of work reports that this problem is increasingly tractable, with recent methods achieving high detection performance on widely used benchmarks. We show, however, that much of this apparent progress does not survive scrutiny. Four of the six corpora embed the ground-truth answer directly in the input prompt. A naïve text-similarity baseline we call TxTemb exploits this to achieve near-perfect detection scores without any access to model internals. To measure what genuine detection capability remains once these artifacts are controlled, we conduct a large-scale evaluation spanning twenty-two detection methods, twelve open-source models spanning six architectural families, and six corpora. We further introduce DRIFT, a supervised probe over inter-layer hidden-state transitions, as a point of comparison for live-generation detection. Our findings suggest that the field's reported progress on hallucination detection is substantially explained by benchmark construction artifacts in widely used corpora, and that the majority of established baselines perform near chance under controlled conditions; the consistent exceptions are SAPLMA and DRIFT, both supervised probes on upper-layer hidden states.

Algorithmic feeds have become primary environments for encountering information online, yet while they shape what people see, less is known about how sustained feed exposure shapes how people write. Drawing on Cultivation Theory, we examine whether algorithmic feeds function as online environments that leave measurable traces in users' language. We leverage a large-scale longitudinal dataset of 235M posts by 4M users on Bluesky, and conduct a quasi-experimental study matching an initial pool of 368,513 users exposed to one of three feeds – News, Science, and Blacksky – with a pool of 2,001,915 active control users who did not engage with any of these feeds. We examine linguistic evolution across three dimensions: lexico-semantics, psycholinguistics, and topics. We find that users exposed to these feeds show significantly greater stylistic accommodation, semantic alignment, and register formalization than matched controls. These effects vary markedly by feed identity – Blacksky produces the deepest psycholinguistic restructuring, with significant shifts in cognitive processing, affective expression, and pronoun use, while News and Science effects are largely confined to register and topical focus. Regression models reveal that reposting is the most consistent predictor of linguistic convergence across all feeds, whereas posting and bookmarking show feed-dependent effects, with effects differing more than fourfold across feeds. Our work extends Cultivation Theory beyond belief formation to linguistic behavior, demonstrating that feeds function as persistent linguistic environments that gradually shape what and how users write online. Our work has implications for studying algorithmic influence, online identity formation, and the design and governance of feed-based platforms that mediate online interactions.

We introduce IsalProgram (Instruction Set and Language for Programming), a novel assembly-like programming language with three distinctive theoretical properties: (1) it is a regular language in the sense of formal language theory, meaning its programs are accepted by a finite automaton; (2) every finite string over the instruction alphabet is a syntactically valid program; and (3) it makes no explicit use of memory addresses or variable names, absolute or relative. Programs are finite sequences of tokens drawn from a fixed instruction set, and are executed on a virtual machine whose sole data structure is a circular doubly linked list (CDLL) navigated by three data pointers, with control flow governed by two code pointers. We give a complete formal definition of the language and its virtual machine, prove its regularity, and demonstrate its expressive power. We further discuss IsalProgram's potential advantages as a target language for neural program synthesis, the amenability of its program space to metric-based exploration via the Levenshtein edit distance, and directions for analyzing computability and complexity within this framework.

Large language models (LLMs) have achieved remarkable progress in natural language generation, but remain susceptible to hallucination. In response to growing concerns about hallucinations, several benchmarks have been developed, primarily in English and Chinese. However, Arabic remains underrepresented, with limited benchmarks for LLMs hallucination due to scarce annotated resources and the language's morphological complexity. Consequently, existing benchmarks do not adequately reflect the linguistic, cultural, and reasoning characteristics of Arabic. To address this gap, we introduce HalluScore, a structured Arabic question answering benchmark designed to evaluate hallucination behavior in LLMs across different levels of reasoning difficulty, various knowledge domains, historical timelines, and culturally grounded Arabic scenarios. It contains 827 carefully curated questions for evaluating, detecting, and mitigating hallucination in LLMs. The dataset was constructed through a structured pipeline involving quality assurance, filtering for clarity and factual validity, and model-driven selection to retain questions that consistently trigger hallucinations. Each question is linked to verified ground-truth evidence, answer explanations, and multi-label annotations. Using the HalluScore benchmark, we conduct a comprehensive empirical analysis of hallucination patterns across 17 Arabic, multilingual, and reasoning LLMs. Moreover, we provide high-quality human annotations identifying hallucinated, non-hallucinated, and partially hallucinated responses of all evaluated LLMs. These results suggest that hallucination in Arabic LLMs extends beyond factual inaccuracies, encompassing challenges related to cultural understanding, linguistic reasoning, and logical consistency. We release HalluScore to support future research on improving the reliability and cultural competence of LLMs in Arabic.

Can large language models reliably express a human-like personality, or are they merely mimicking surface cues without a stable underlying profile? To investigate this, we induce personality in LLMs by fine-tuning them on the long-form essays, where each essay is associated with a target Big Five personality profile. We then evaluate the stability and fidelity of the induced personality using the IPIP-NEO questionnaire. Specifically, we ask: (i) does post-training (SFT, DPO, ORPO) stabilize questionnaire scores under prompt rephrasings, and (ii) can it induce target Big Five profiles from unguided essays? Our results demonstrate that fine-tuning consistently reduces variance in questionnaire responses across five models, directly mitigating the evaluation fragility reported in pre-trained models. However, this newfound stability reveals a more fundamental limitation: accuracy on the full five-dimensional profile remains near chance, even when single-trait scores improve. This indicates that unguided essays lack the cues needed for faithful personality expression. We therefore argue for scenario-grounded datasets or interactive elicitation that accumulates test-aligned evidence over time.

Item difficulty must often be estimated before test administration, when no responses are yet available for calibration. While most response-free difficulty modelling approaches derive item-text features by hand for a separate statistical model, we fine-tune a transformer end-to-end on the wording, avoiding the theory-based feature design and the preprocessing that discards information. We address reading-comprehension multiple-choice items, whose difficulty depends on inferential demands spanning passage, question, and options, yet the simplest model sees one undifferentiated sequence and is trained on difficulty alone. We introduce and investigate two extensions to the joint-encoding baseline: a component-wise variant, which encodes the wording parts separately, and a multi-task variant, which adds an auxiliary task of question answering. We compare the methods across three training-set sizes sampled from a corpus of nearly 30,000 items whose labels approximate response-based Rasch difficulty. At the smallest training size, both extensions improve on the baseline, the multi-task variant across every metric, and component-wise encoding in rank ordering. Further research may ground the auxiliary supervision in observed responses and extend the approach to other item types.

Additional test-time compute can give LLM agents access to more past experience, yet expanding the context or adding rollouts does not necessarily yield greater agent capability. We call this challenge test-time compute-to-capability conversion and propose SkillTTA, which retrieves task-relevant training trajectories and synthesizes a temporary skill conditioned on the visible target context for a solver with fixed parameters. To pursue a higher performance ceiling, SkillTTA further uses meta prompt optimization (MPO) to adapt the policy that writes these skills. MPO evaluates candidate prompts on paired tasks and emphasizes informative transitions. It also confines updates to benchmark-specific atomic slots, reducing the variance caused by observing each edit only indirectly through skill synthesis and solver rollout. Across ALFWorld, SpreadsheetBench, BigCodeBench, and WebShop, SkillTTA outperforms state-of-the-art reuse and optimization baselines. It attains a higher performance ceiling at lower compute cost than baseline reuse and sampling strategies.

We present our submission to the LLM track of the 2026 Computational Models of Reference, Anaphora and Coreference (CRAC 2026) shared task. With an average CoNLL F1 score of 74.32 on the official test set, our system ranked first in the LLM track, and third overall. Our system is based on the Gemma-3-27b model, fine-tuned using a two-stage strategy with a multilingual base adapter followed by dataset-specific adapters. We represent mention spans by their headword using an XML-inspired format with local reindexing and annotate documents iteratively. These design choices proved effective across languages, document lengths, and annotation guidelines.

While AI-generated hallucinations pose considerable risks, the underlying cognitive mechanisms by which humans can successfully recognize or be misled by these hallucinations remain unclear. To address this problem, this paper explores humans' neural dynamics to characterize how the brain processes hallucinated content. We record EEG signals from 27 participants while they are performing a verification task to judge the correctness of image descriptions generated by a multi-modal large language model (MLLM). Based on an averaged event-related potential (ERP) study, we reveal that multiple cognitive processes, e.g., semantic integration, inferential processing, memory retrieval, and cognitive load, exhibit distinct patterns when humans process hallucinated versus non-hallucinated content. Notably, neural responses to hallucinations that were misjudged versus correctly judged by human participants showed significant differences. This indicates that misjudged AI-generated hallucinations failed to trigger the standard neurocognitive fact verification pathway.

Diffusion Large Language Models (DLLMs) promise fast parallel generation, yet open-source DLLMs still face a severe quality-speed trade-off: accelerating decoding by revealing multiple tokens often causes substantial quality degradation. We attribute this dilemma to a train-inference mismatch amplified by irreversible decoding. While training reconstructs tokens from randomly corrupted states, efficient inference requires an adaptive denoising order, where easier tokens are revealed earlier and context-dependent ones are deferred. This view motivates two complementary methods: an inference-time method that makes parallel decoding revokable, and a training-time extension that distills the reliable order exposed by this revokable process. Accordingly, we first propose Wide-In, Narrow-Out (WINO), a training-free decoding algorithm that enables revokable parallel generation. WINO aggressively drafts multiple tokens, verifies generated tokens with enriched global context, and re-masks unreliable ones for later refinement. Building on this discovered order, we further introduce WINO+, which injects the verified denoising trajectories produced by WINO into model parameters, aligning training with efficient inference. Experiments on LLaDA and MMaDA show that WINO improves both quality and efficiency, while WINO+ further strengthens this progression. On GSM8K, WINO improves accuracy from 73.24% to 75.82% with a 6.10x step reduction, and WINO+ further achieves 76.58% with a 6.83x reduction. On Flickr30K, WINO+ reaches a 16.22x step reduction with improved CIDEr. These results demonstrate that DLLMs can serve as their own efficiency teachers by first discovering reliable denoising orders through revokable decoding and then learning to follow them for faster generation. Code is available at https://github.com/Feng-Hong/WINO-DLLM/tree/WINO-plus.

Large Reasoning Models (LRMs) generate chain-of-thought traces whose length tracks human reaction times across cognitive tasks, but recent debate questions whether this alignment reflects genuine computational structure or surface verbosity. We test whether the alignment varies with inference-time reasoning effort. Across GPT-OSS-20B and GPT-OSS-120B, three effort levels, and six reasoning tasks, within-task and cross-task alignment remain invariant: Bayes Factors lean toward the null, and mean alignment is numerically near-identical across conditions. A manipulation check reveals that the effort parameter sets an upper budget on generation rather than driving real-time allocation, suggesting that the allocation policy is crystallized at training time. Arithmetic complexity contrasts further show that token allocation tracks fine-grained, format-dependent human difficulty patterns, with model scale improving the match. Cognitive cost alignment between LRMs and humans appears to be a training-time achievement, robust to inference-time perturbations, supporting a compiled rather than online account of LRM problem-solving.

Long-context inference in large language models is bottlenecked by the quadratic cost of full attention. Existing efficient alternatives often rely either on native sparse training or on heuristic token eviction, creating an undesirable trade-off among efficiency, training cost, and accuracy. In this work, we show that full-attention LLMs are already intrinsically sparse and can be transformed into highly sparse models with only minimal adaptation. Our approach is built on three observations: (1) only a small subset of attention heads truly requires full long-context processing; (2) long-range retrieval is governed primarily by a low-dimensional subspace, allowing relevant tokens to be retrieved efficiently with a 16-dimensional indexer; and (3) the useful token budget is strongly query-dependent, making dynamic top-$p$ selection more suitable than fixed top-$k$ sparsification. Based on these insights, we propose RTPurbo, which retains the full KV cache only for retrieval heads and introduces a lightweight token indexer for sparse attention. By exploiting the model's intrinsic sparsity, RTPurbo achieves sparsification with only a few hundred training steps. Experiments on long-context benchmarks and reasoning tasks show that RTPurbo preserves near-lossless accuracy while delivering substantial efficiency gains, including up to a 9.36$\times$ prefill speedup at 1M context and about a 2.01$\times$ decode speedup. These results suggest that strong sparse inference can be obtained from standard full-attention training without expensive native sparse pretraining.

Contextual Automatic Speech Recognition (ASR) faces challenges with large-scale keyword dictionaries, as excessive irrelevant candidates introduce noise that degrades accuracy. To address this, dynamic filtering typically uses a base ASR model to generate preliminary hypotheses, followed by semantic text retrievers to fetch a concise subset of relevant keywords. However, this approach frequently fails in Chinese ASR. Base models often produce homophonic or near-homophonic errors that preserve the phonetic cues of the target keywords but severely distort their semantic meaning, rendering standard semantic retrievers ineffective. To resolve this, we propose a filtering framework that jointly integrates Semantic, Pinyin, and Glyph features (JSPG). Pinyin effectively retrieves targets based on phonetic similarity, while glyph provides complementary structural cues to filter out numerous irrelevant homophones inherent in Chinese. To bridge the gap between character-level pinyin/glyph metrics and sequence-level filtering, we introduce an extended Smith-Waterman algorithm that computes similarity scores between the N-best hypothesis sequences and keywords. Experiments on the Aishell-1 and RWCS-NER datasets demonstrate that JSPG significantly outperforms single-feature baselines. Furthermore, downstream contextual ASR models guided by JSPG achieve substantial improvements in keyword recognition accuracy.

End-to-end LLM trading agents have moved quickly from research curiosity to a small ecosystem of named systems, including FinCon, FinMem, TradingAgents, FinAgent, QuantAgent, and FLAG-Trader. Several of these report headline Sharpe ratios that would be material if read at face value on a deployment desk, and associated benchmarks such as FinBen report trading-task Sharpe statistics in the same range. The gap between architecture research and deployment claim has been crossed too freely on both sides of the academia–industry divide. We take a position on that gap: reported alpha from end-to-end LLM trading agents should not be treated as deployment evidence. Before such returns can support claims of deployable trading capability, they must survive structural validity tests for temporal integrity, real-world frictions, counterfactual robustness, predictive calibration, numerical execution, and multi-agent disaggregation. Current public evidence cannot yet distinguish robust predictive ability from temporal contamination, unmodeled frictions, short-window Sharpe uncertainty, narrative fitting, and parametric priors. The problem is not only evaluative but structural. Language confidence is not tradable probability, narrative reasoning is not numerical execution, and model priors may become undisclosed implicit factor exposures. We contribute a minimum reporting protocol suite, P1–P6, with tiered applicability by claim strength, and a conservative modular alternative that uses LLMs as auditable information interfaces upstream of independent calibration, risk, and execution modules. Code and reproduction harness: \url{https://github.com/hj1650782738/Trading}.

Comprehensive situational awareness is essential for autonomous vehicles operating in safety-critical environments, as it enables the identification and mitigation of potential risks. Although recent Multimodal Large Language Models (MLLMs) have shown promise on general vision-language tasks, our findings indicate that zero-shot MLLMs still underperform compared to domain-specific methods in fine-grained, spatially grounded risk assessment. To address this gap, we propose DriveSafe, a framework for risk-aware scene understanding that leverages structured natural language descriptions. Specifically, our method first generates spatially grounded captions enriched with multimodal context, including motion, spatial, and depth cues. These captions are then used for downstream risk assessment, explicitly identifying hazardous objects, their locations, and the unsafe behaviors they imply, followed by actionable safety suggestions. To further improve performance, we employ caption-risk pairings to fine-tune a lightweight adapter module, efficiently injecting domain-specific knowledge into the base LLM. By conditioning risk assessment on explicit language-based scene representations, DriveSafe achieves significant gains over both zero-shot MLLMs and prior domain-specific baselines. Exhaustive experiments on the DRAMA benchmark demonstrate state-of-the-art performance, while ablation studies validate the effectiveness of our key design choices. Project page: https://cvit.iiit.ac.in/ research/projects/cvit-projects/drivesafe

Low-resource deployment constraints have made model quantization essential for deploying neural networks while preserving performance. Meanwhile, model merging has become an increasingly practical low-resource strategy for integrating multiple task- or domain-specialized experts into a single model without joint training or multi-model serving. Together, quantization and model merging enable an efficient low-resource deployment pipeline by integrating multiple experts into one low-bit model. We formulate this setting as Post-Merge Quantization (PMQ). We show that directly applying post-training quantization (PTQ) to a merged model is unreliable because two distinct deviations are coupled: the quantization deviation introduced by low-bit reconstruction and the expert-relative merging deviation inherited from model merging. To mitigate these deviations, we propose E-PMQ, an expert-guided PMQ framework that uses source expert weights to provide expert- guided output targets during layer-wise calibration, together with merged-weight anchoring to stabilize the calibration and preserve the integrated behavior of the merged model. On CLIP-ViT-B/32 eight-task merging, E-PMQ improves 4-bit GPTQ from 65.0% to 73.6% under Task Arithmetic and from 69.1% to 74.8% under TIES-Merging. On harder settings, E-PMQ improves GPTQ from 34.8% to 76.7% on 20-task CLIP-ViT-L/14 and from 78.26% to 83.34% on FLAN-T5- base GLUE. These results demonstrate that E-PMQ enables effective post-merge quantization and low-bit deployment.