Language_Model
-
RetiJev: Fast and Source-Grounded Ophthalmic Decision Models
2026-10-22 08:59:47Bo Liu, Lianyu Wang, Meng Wang, Ching-Yu Cheng, Huazhu FuAbstract
Ophthalmic clinical decision support requires fast, accurate, and privacy-preserving inference, yet autoregressive large language models (LLMs) incur substantial latency and computational overhead for decisions among predefined actions. We therefore investigate a decision-oriented alternative to generative reasoning. Specifically, we first introduce RetiJev-Data, a source-grounded ophthalmology corpus comprising 78,158 clinical-decision and diagnosis questions derived from articles in PubMed Central and a MeSH vocabulary covering 531 eye diseases, and 50,074 guideline-grading questions spanning 11 grading systems, along with a 24,662-question evaluation benchmark. Based on this, we develop RetiJev, a family of 0.8B-9B decision models that directly score candidate answers without autoregressive decoding. Across all model sizes, RetiJev outperforms both a 27B LLM and the commercial Jev 1.13 decision model on guideline grading by 10.4-23.0 points. RetiJev-9B further achieves the highest accuracy on clinical decisions, improving over both baselines by 5.4-6.5 points while matching them on diagnosis. It answers a case in 158 ms on one GPU, 3.5x faster than LLM generation and 470x faster than reasoning. Models are released at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30947
PDF
-
Tracing Hallucinations in LLM Reasoning for UML Class Diagram Generation: An Empirical Taxonomy and Mitigation Study
2026-10-22 08:59:27Mojdeh Rahmanian, Foutse Khomh, Ashkan Sami, Yanchao YuAbstract
Large Language Models (LLMs) can generate UML class diagrams from natural language requirements, but hallucinations may compromise the correctness and completeness of the resulting diagrams. Existing evaluations mostly assess the final diagram, without examining the intermediate traces of reasoning-capable LLMs, which can reveal where these hallucinations arise. We develop an empirical taxonomy of eleven hallucination sub-types organised into three dimensions: Prompt Faithfulness (deviations between requirements and reasoning), Code Faithfulness (discrepancies between reasoning and generated PlantUML code), and Reasoning Quality (problems within the reasoning). The taxonomy was derived from 60 manually annotated DeepSeek-R1 traces, applied to 120 further DeepSeek-R1 traces through a dual-LLM annotation pipeline, and examined on 20 DeepSeek-V4-Pro traces. Code Faithfulness was the most frequent dimension for DeepSeek-R1 in both datasets, whereas Prompt Faithfulness dominated for DeepSeek-V4-Pro, suggesting model-dependent failure patterns. In the manually annotated traces, reasoning-trace length showed no significant monotonic association with hallucination frequency, whereas requirement complexity was negatively associated with it, particularly for Prompt Faithfulness errors. In a mitigation experiment on 13 Golden UML ModelSet cases with three Qwen3-based models, adding Prompt Faithfulness and Reasoning Quality constraints to the generation prompt raised the pooled mean F1 score by 4.0 percentage points, whereas a post-hoc Code Faithfulness review provided no consistent further benefit. Relationship modelling remained the lowest-scoring element throughout. These results suggest that localising hallucinations within the generation process can inform targeted, prompt-based mitigation. Whether these findings generalise beyond the DeepSeek and Qwen3 model families remains to be tested.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30946
PDF
-
Verbalizing Activation Differences with Natural Language Autoencoders
2026-10-22 08:54:58Andrea Blasi N\'u\~nez, Gianluca Barmina, Federico Torrielli, Thiusius Rajeeth Savarimuthu, Peter Schneider-Kamp, Lukas Galke PoechAbstract
Activation verbalizers translate the internal representations of large language models into human-readable text. The recently proposed natural language autoencoders (NLAs) learn this by reconstructing activations through a natural language bottleneck. However, it remains unclear how well these methods generalize beyond the setting they were trained for. Here, we study this question for NLAs along four closely related questions: (i) Do NLA verbalizations exhibit compositional structure with respect to their input? (ii) Do NLAs generalize to finetuned versions of the model they were trained for? (iii) Can NLAs recover finetuning objectives in model diffing? (iv) Can NLAs be specialized for model diffing? Across four models from three families, our experiments provide evidence supporting a largely positive answer to all four questions. Specifically, verbalizations of composed analogy activations mostly name the target while rarely mentioning the subtracted word, and an NLA can be reused on lightweight finetunes of its original model without retraining, with reconstruction quality degrading noticeably only for the largest model. Verbalizing activation differences between Taboo model organisms and their original models recovers the concealed secret word for most organisms, and reinforcement learning on activation differences further improves this recovery on held-out organisms. We also find that NLAs can recover the finetuning objective without the finetuned model generating any token, since the most informative token positions precede the user prompt. NLAs do not yet match state-of-the-art model diffing methods on the emergent misalignment detection task, but outperform a blackbox baseline. Our findings suggest that NLAs can meaningfully verbalize differences in activation space, rendering NLAs a promising direction for future work.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30934
PDF
-
ReflectPO: Reflection-Guided Policy Optimization for Long-Horizon LLM Agents
2026-10-22 08:54:24Rumei Li, Xunliang CaiAbstract
With external tools, large language models (LLMs) can tackle complex tasks through sustained environmental interaction. LLM-based agents are therefore increasingly used to tackle complex long-horizon tasks in real-world environments. However, these tasks involve sequential, interdependent decisions and changes in environmental states. Early execution errors affect subsequent interaction histories and environmental states, and their effects accumulate along the decision chain. To address this challenge, we propose Reflection-Guided Policy Optimization (ReflectPO), a reinforcement learning framework for autonomous error recovery in long-horizon tasks. ReflectPO comprises two core components. The first is Reflection-Guided Sampling. We use branch differences and repeated visits in interaction trajectories to construct hints for error recovery. The second is Tree-Structured Credit Assignment. We organize sampled trajectories into trees and assign action-level credit using parent--child reward differences and recovery gains. Experiments over two long-horizon tool-use benchmarks show that ReflectPO outperforms multiple existing baselines in most domains. It also achieves the best overall performance on average. Further analyses show that ReflectPO effectively improves agents' ability to recognize their own deviations, recover from erroneous states, and continue making progress toward task completion.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30933
PDF
-
Making MLLM Latents Matter: Factor-Aware Predictive Learning for Cinematic Video Generation
2026-10-22 08:43:48Songlin Yang, Zhe Wang, Xiaotong Zhao, Xi Wang, Ran Zhang, Shuai Li, Taiyi Wu, Songlian Li, Eric Liu, Alan Zhao, Anyi RaoAbstract
Cinematic video generation requires both temporal inference from limited inputs (i.e., sparse rather than frame-by-frame guidance) and factor-aware control of character, scene, action, and camera. Multimodal large language models (MLLMs) offer useful priors for both capabilities, but their representations are not automatically actionable for video generation: a Video Diffusion Transformer (DiT) may bypass them through native conditioning pathways, and using them does not ensure either factor-aware control or generation-compatible temporal prediction. We propose Factor-Aware Predictive Learning (FAPL), which establishes effective MLLM latent use as a shared prerequisite and then targets the two capabilities with complementary learning signals. First, latent-only generator grounding trains the DiT with a frozen MLLM and connector, strengthening the DiT's use of MLLM latents for generation. Second, for factor-aware control, paired videos specify a target-factor change and non-target preservation, and factor-span replacement trains the DiT to realize this relationship in its output. Third, for temporal inference, text-only and text-plus-complete-video observations share a video target, and the trained DiT is then frozen while the MLLM and connector are optimized through it. This final stage trains limited-input representations for generation while retaining complete-observation reconstruction as a complementary training mode. Extensive experiments demonstrate our consistently better performance on three cinematic video generation tasks (i.e., sparse-to-video, factor-aware editing, and multi-shot generation) over baselines and validate FAPL's contributions to MLLM feature use, temporal control, and factor coordination.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30926
PDF
-
Beyond Attack Success Rate: Cost-Aware Metric for Evaluating Iterative LLM Jailbreaks
Abstract
Attack success rate (ASR) evaluates jailbreaks against large language models without accounting for repeated calls to the victim model. Yet each additional call gives the deployed system another opportunity to detect the attack and intervene. We introduce $\mathrm{ASR}_{\mathrm{STEP}}$, which penalizes additional victim-model calls while retaining binary success judgments. Our evaluation considers GCG, PAIR with a step budget, and a single-step PAP baseline on JailBreakBench across three model families. GCG's scores decrease by 7.0-8.7 percentage points after discounting, while PAP's remain unchanged. While for PAIR attack, we see a wider discounting range between 3.4-12 percentage points, showing that the "step" quantity differs per iterative strategy. Reporting both metrics makes the discount due to repeated victim-model interaction explicit. Code is available here: this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.26589
PDF
-
Watch and Learn: Teaching Visual In-Context Learning by Removing Textual Shortcuts
2026-10-22 08:34:24Nimrod Shabtay, Asaf Yehudai, Avihu Dekel, Eli Schwartz, Raja GiryesAbstract
Vision-language models (VLMs) struggle with in context learning compared to their textual counterparts. Our preliminary study shows that VLMs rely on textual shortcuts and underutilize the visual demonstrations. We therefore hypothesize that VLMs are insufficiently trained to perform visual in-context learning (vICL). To this end, we construct MIME, a task-agnostic vICL training set without textual shortcuts. We ensure the instruction does not state the rule linking the demonstration images to their answers, so the model must infer it from the visual demonstrations. Fine-tuning on MIME improves vICL accuracy by 8.1 to 18.0 points across three VLMs, preserves single-image performance, and improves multi-image and text only in-context learning. In contrast, stating the rule in the training data sharply reduces this gain. We further propose DTAD (Demonstration Text Attention Dropout), which randomly withholds demonstration labels during training so that the VLM must infer the rule from the demonstration images. While achieving comparable vICL accuracy to fine-tuning on MIME, DTAD yields more consistent attention to the demonstration images. On TrueMICL, Qwen3-VL-8B trained on MIME matches frontier models such as Claude Opus 5, and surpasses them with DTAD. Finally, adding MIME to the instruction-tuning mix of LLaVA 1.5 13B more than doubles its vICL accuracy without hurting single-image abilities, whereas a same-size multi-image set improves it only slightly. These results suggest that vICL data should be a standard component of VLM training.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30913
PDF
-
Silicon Samples Match the Answers but Not the Respondents
2026-10-22 08:22:50Yusuke Kumagae, Masahiro KanekoAbstract
Silicon sampling generates survey respondents by having a large language model (LLM) answer a survey as a person with given demographic attributes, and the resulting synthetic populations are increasingly used in place of human ones. Existing evaluations focus on individual-answer accuracy, per-question answer distributions, and correlations across questions. None of these compares the respondents themselves, so a generator whose respondents have combinations of answers found in no human respondent can still score well if it reproduces every answer distribution. We propose the respondent coverage gap (RCG), which represents each respondent as a point defined by their complete answer profile and measures how far each human respondent is from the nearest generated one, relative to the distance between human samples. On two US social surveys, all six LLMs we evaluate leave human respondents uncovered, and under our main settings, they fall behind a baseline that samples each answer independently from the human marginals. Across the prompting and inference settings we test, every LLM has a higher RCG than a Gaussian copula fitted to the human data.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30900
PDF
-
Parallelizable Probabilistic Latent Reasoning for Diverse Reasoning Trajectories
2026-10-22 07:41:12Seokhyun Lee, Seongjun Lee, Changhee LeeAbstract
Language models relying on verbalized Chain-of-Thought (CoT) are constrained to token-by-token generation in the language space, where many tokens serve textual coherence rather than core reasoning. This entangles genuine differences in logical trajectories with superficial variations in verbalization. Moving reasoning into a compact, continuous latent space offers a promising alternative for efficiently representing and exploring diverse reasoning trajectories. However, current latent models remain largely deterministic, capturing only a single reasoning trajectory rather than modeling a distribution over multiple valid alternatives. While stochastic approaches parameterize distributions over latent reasoning steps, they frequently suffer from unstable optimization and limited empirical performance. Furthermore, their reliance on autoregressive latent prediction during training precludes effective parallelization, leading to prohibitively long training times. To overcome these bottlenecks, we propose Probabilistic Latent Reasoning Models (PLRM), a parallelizable variational framework for probabilistic latent reasoning. PLRM adopts a JEPA-style objective in which the posterior infers latent distributions in parallel from observed reasoning trajectories, while the prior learns to predict the subsequent latent distribution via teacher forcing. To model heterogeneous reasoning processes, we introduce a discrete strategy variable together with a mixture-based prior, allowing the model to preserve the Gaussian-mixture structure induced by multiple distinct reasoning trajectories rather than collapsing them into a single Gaussian. Across diverse benchmarks, PLRM achieves performance competitive with deterministic latent reasoning models on single-trajectory datasets while enabling efficient training and consistently outperforming competing stochastic baselines on multi-trajectory datasets.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29385
PDF
-
TRIAD: Benchmarking Omni-Modal Ambiguity in Multimodal Large Language Models
2026-10-22 07:15:40Zhaolu Kang, Yidi Wang, Yile Li, Siqi Zeng, Yingjie He, Richeng Xuan, Zhichao huAbstract
Ambiguity is a central feature of natural communication: a sentence, image, or sound may remain underspecified until evidence from the other channels is considered. Yet current multimodal benchmarks mostly test perception, recognition, or reasoning over already determinate inputs, leaving unclear whether omni-modal large language models can resolve ambiguity distributed across text, vision, and audio. We introduce TRIAD, a benchmark for omni-modal ambiguity resolution. Each item is a text--image--audio triplet paired with a question and a set of answer options, and is constructed so that the full triplet determines a unique answer while removing any one modality makes the set of consistent options non-unique. TRIAD contains $327$ hand-written queries in $86$ scene groups and covers $18$ ambiguity categories across the three modalities. Evaluating $14$ omni-modal systems reveals that strong item-level performance does not translate into robust scene-level disambiguation: the best model still trails humans by over $20$ points in group accuracy. Removing answer options widens the gap further, driving model group accuracy to the single digits while humans remain robust. Leave-one-out and transcript/description counterfactuals further show that current systems often underuse audio cues when text and image are present. TRIAD reframes omni-modal evaluation around disambiguation rather than recognition, and provides a taxonomy, benchmark, and diagnostic protocol for measuring this capability.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30841
PDF
-
RECAP: Looking Once Is Not Enough for Vision-Language Reasoning
2026-10-22 07:04:23Zhaolu Kang, Tailong Luo, Chenxin Li, Yingjie He, Zhenyu Yu, Fengyu Zhou, Jiachen Qian, Lei Wei, Shuang Chen, Jiachen Li, Eric Hanchen Jiang, Rongchao Zhang, Zhengtao Yao, Hoi Leong Lee, Guansu Wang, Kaiyue ZhouAbstract
Long chain-of-thought reasoning improves vision-language models (VLMs), but it also exposes a temporal grounding failure: as generation unfolds, models may drift from image evidence while reinforcement learning still provides only a final-answer reward. We study this gap through visual revisit, a temporally localized reactivation of image evidence during reasoning. Across three VLM backbones, revisit peaks are predictive of correctness and causally linked to performance: masking them degrades accuracy more than masking non-peaks, while injecting revisit-like peaks into failed rollouts partially restores correct answers. Control analyses show that this signal is not explained by raw attention strength, uncertainty, logit margin, or hidden-state norm. We propose RECAP, a GRPO-compatible training method that turns detached revisit traces into a credit-assignment signal. RECAP uses revisit-conditioned advantage estimation, dynamic discounting, and gated value prediction to propagate final rewards through visually important reasoning steps, without extra rewards, additional rollouts, or inference-time changes. Across Qwen2.5-VL-7B, InternVL3-8B, and LLaVA-OV-7B, RECAP consistently improves over GRPO and VPPO, yielding $+6.1$--$+6.5$ HallusionBench gains over the SFT base while preserving reasoning and text-only performance. Results on larger LoRA-tuned models, Chinese zero-shot benchmarks, visual perturbations, and human evaluation further demonstrate robust improvements in image-grounded reasoning.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30834
PDF
-
Loop to Look Better: Learning Adaptive Depth for Question-Guided Perception in Vision-Language Models
2026-10-22 06:53:13Wencheng Ye, Anning Hu, Hongye Fang, Tianyi Wang, Junchi YanAbstract
Visual perception should be iterative and task-directed, with representations refined in response to the question and the need for further processing. While Vision-Language Models (VLMs) increasingly support extended reasoning, their visual tokens are still typically acquired in a single, task-agnostic encoder pass. This creates an evidence-acquisition bottleneck: models may reason longer without refining the visual evidence needed to answer the question. We propose LoopLook, a lightweight encoder-side recurrent perception framework that equips pretrained VLMs with adaptive visual depth. LoopLook recurrently revisits a middle-to-late window of the vision encoder, injecting question representations at each loop to guide visual refinement. A learned stopping mechanism adaptively determines the recurrent depth, while a token-wise Mixture-of-Looks (MoL) combines representations from different loops, allowing each visual token to selectively integrate information acquired across recurrent steps. Across four backbones and eight benchmarks, LoopLook consistently improves performance over the baseline, yielding gains of up to 8.6% on MMStar and 8.1% on HallusionBench. Our analysis shows that LoopLook progressively refines visual attention and adaptively allocates depth according to task difficulty. It also better preserves task-relevant visual information under reduced input resolution, while increasing time to first token (TTFT) by only 5.1-6.5%. By bringing recurrence from reasoning back into perception, LoopLook transforms visual perception from a one-shot glimpse into an adaptive process of recurrent refinement.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30824
PDF
-
Adaptive Aerial Loco-Manipulation Policy Learning via LLM-Assisted Proactive Task Generation
2026-10-22 06:36:39Jingyu Ru, Xiangyuan Xie, Lixuan Jin, Jian An, Xingming Zhang, Binghao Li, Siyu Wang, Xiang Zhou, Lei Yuan, Hongli Xu, Yang YuAbstract
Loco-manipulation extends the workspace of conventional manipulation by combining mobility with physical interaction, while aerial loco-manipulation further expands this capability into three-dimensional space, enabling robots to reach environments inaccessible to ground-based systems. Moving to an aerial platform, however, gives rise to a demanding control problem, since the vehicle must track trajectories accurately in three dimensions while the motion of the arm and its interaction with the environment continuously shift the center of mass, alter the inertia, and impose force and moment disturbances on the body. Conventional controllers address these effects through explicit models and estimation, whereas learning-based controllers infer dynamics from experience, yet control rewards provide no direct target for the response information retained from history, and fixed task sampling can leave persistent tracking difficulties insufficiently explored. We present ALMA, an aerial loco-manipulation framework for adaptive flight control under external arm commands. ALMA employs a contextual flight policy that adapts from recent measurements and actions, using action-conditioned prediction of future measurement representations to train its context to retain response information relevant to flight control. To expand the experience available for this adaptation, an evolutionary process uses a large language model (LLM) to analyze observed control difficulties and propose new structures of arm motion for subsequent training. In simulation, ALMA reduces position root-mean-square error (RMSE) by up to 24.7% relative to the strongest baseline, while the same policy transfers zero-shot to the aerial manipulator and reduces hover position RMSE by 79.6% during periodic arm motion relative to the real-world baseline.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30810
PDF
-
Invisibilization as a Source of Probability Bias in Language Models
2026-10-22 06:03:35Christian Arnold, Ernesto Calvo, Guadalupe Gonzalez, Sebasti\'an Vallejo Vera, Tiago VenturaAbstract
Language models are aligned to avoid using protected attributes such as race, ethnicity, and gender to inform their responses. Therefore, they increasingly refuse or hedge when such attributes appear in a prompt. However, if a system must decide how much to learn from what people tell it, it is costly to refuse valuable information because of its source. In this article, we measure the cost of declining information in a two-urn game with an exact Bayesian answer. We present leading LLM models from two developers (OpenAI and Anthropic) with a prompt that requires them to decide how far to trust an unverifiable report from a participant whose description we randomize across ten configurations (6,600 sessions). We show that leading models often decline to assess the reporter, treating information as worthless at rates set by the description rather than by the evidence. At its widest, models decline a Black-described reporter 46% of the time and a white-described reporter 9%; yet, when they do assess the information, the Black-described reporter receives slightly more trust. The bias is invisible to audits that compare average trust or look only at assessed cases. It grows with reasoning effort, follows words rather than categories (models invoke race for "Black" but not for its Spanish translation negro, and never for "Indigenous"), and worsens under warning instructions a practitioner would try first. We also present a scoring option that reduces the refusal rate bias. When a model's trust is scored against revealed truth over repeated rounds, it credits identical records equally. In four of five configurations from Anthropic and OpenAI, the racial gap in trust falls below 0.02 by the last rounds.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30785
PDF
-
MixJudge: Learning a Language-Queryable Critic for Vocal Mixing
2026-10-22 06:02:59Kuan-Hsun Ho, Yuan-Chiao Cheng, Yen-Tung Yeh, Yi-Hsuan YangAbstract
Mixing requires not only applying processing but also assessing the resulting mix. Audio-language models can provide conversa- tional mixing advice, but existing evaluations do not separately test whether they can judge specific mixing issues. Since professional supervision for such judgments is scarce, we construct a curriculum of 11 vocal-mix issues at multiple severity levels and use it to train MixJudge, a language-queryable audio critic. MixJudge shows its ability to retrieve mixing issues, distinguish opposite issues, and order the severity of an issue. Moreover, its judgments transfer to menu selection: given a set of processing candidates, MixJudge often favors versions acoustically closer to a held-out professional version, demonstrating how explicit mixing judgment can support comparison and downstream mixing decisions.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30783
PDF
-
Abstract Priorities, Concrete Rewards: How Stated Preferences Relate to What People Reward in AI Responses
Abstract
AI assistants built on large language models (LLMs) are trained on human feedback and increasingly personalized to users' stated preferences. Yet how closely these preferences reflect what people reward in responses remains unclear. We compare stated and revealed preferences in the public PRISM dataset, in which 1,396 participants reported how much seven response attributes mattered to them and then rated 29,347 responses from 21 LLMs. Among responses to the same prompt, features the survey never asked about, such as response length, formatting and refusal language, were at least as strongly associated with ratings as text proxies of the surveyed attributes. Whether stated priorities matched ratings depended on how attributes were measured: participants who prioritized an attribute favoured responses from models that others perceived as strong on it, but not responses that scored higher on text proxies for it. Written feedback was more telling: participants who asked for shorter answers, even in earlier conversations, gave length less weight in their ratings, and those asking for longer answers gave it more. Their relative priorities over the surveyed attributes, however, did not predict this weight. Reasons given for choices reflected perceived attributes more than measured features. Participants also differed stably in how much they weighted several concrete features, especially length, but with a median of 6 conversations per participant, these weights were estimated poorly. The largest gain from personal information in predicting held-out choices was 0.13 percentage points, compared with 1.11 for a stronger pooled model. Stated priorities thus track perceived attributes of AI responses but say little about the concrete features that ratings reward, a distinction that matters for preference elicitation and AI personalization.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30763
PDF
-
Where Did the Light Come From? Diagnosing Cast-Shadow Reasoning in VLMs and Borrowing the Missing Shadow Prior from Image Generation Models
2026-10-22 05:11:07Ryugo Morita, Shunsuke Kitada, Stanislav Frolov, Brian Bernhard Moser, Ko Watanabe, Riku Takahashi, Jinjia Zhou, Andreas DengelAbstract
Shadows are observable consequences of latent physical causes, arising from the interaction between illumination, object geometry, and scene structure. A model that understands shadows should be able not only to read a shadow's direction, but also to infer where a shadow should fall when it is absent, and to judge whether a shadow is physically consistent with the scene. We introduce ShadowBench, built from real photographs, which poses these three questions to vision-language models (VLMs). From each scene, we build a correct-shadow, a shadow-removed, and a human-annotated wrong-direction state. Evaluating 19 open VLMs (2B to 78B) and frontier proprietary models, we find that VLMs read a visible shadow's direction well above chance and improve with scale, yet infer the absent direction only weakly and fall to chance at judging physical consistency even under a bias-free two-alternative test that removes the yes-bias of single-image naturalness questions. Only Gemini-3.1-pro exceeds chance on validity, and even then only modestly and far below humans. Chain-of-Thought (CoT) does not help. The missing ingredient is not reasoning but knowledge that image generation models (IGMs) already hold. Routing a VLM through an IGM lifts its inference well above its direct attempt and helps it flag the wrong-direction shadows it otherwise accepts. Cast-shadow direction is a single facet of illumination understanding, but the sharp gap it reveals between VLMs and IGMs suggests image-generation or reconstruction objectives as a route toward illumination-aware VLMs.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30740
PDF
-
Meaning Contracts: Local Semantic Geometry in Large Language Models
2026-10-22 04:32:30Andr\'es Daniel Godoy OrtizAbstract
This paper studies the local geometry of semantic meaning in large language models. Current representation analyses ask whether meanings are close in embedding space, or whether linear probes separate semantic classes, but they do not characterize the within-concept local geometry: what is invariant across realizations of a single semantic factor, what varies, and what changes when meaning is refined. This paper characterizes that geometry through two controlled probes evaluated on four architecturally diverse decoder-only LLMs (Gemma 4, Llama 3.2, Qwen 2.5, Pythia). A cross-lingual probe uses language as a perturbation to separate exact meaning matches, same-factor realizations, nearby concepts, and out-of-domain contrasts. A refinement probe separates specificity-bearing semantic narrowing from high-probability attributive elaboration and adversarial controls; both probes are validated by blind human judgments. Four findings recur across all models: under mean pooling, once dominant variance directions are removed, sentence pairs order by semantic relatedness (same factor, then same domain, then different domain) without overlap over a band of layers; semantic factors organize around an invariant core surrounded by structured realization-specific variation; refinement produces directed displacement into tighter neighborhoods, including against controls built on the same base noun; and sentence-level and token-local readouts expose complementary geometry (volume contraction at the sentence level, sharper anisotropic endpoints at token-local readout). We interpret these results as evidence that LLM semantics are organized as local anisotropic neighborhoods in which meaning contracts under refinement.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30715
PDF
-
ULME: Predictive Coupling for Mutually Reinforcing Embedding and Language Modeling
2026-10-22 04:23:07Yang Liu, Zhiqiang Yuan, Shu-Dong Huang, Jiancheng Lv, Liwei Yang, Qiwei Yan, Kailin Lv, Jinchao ZhangAbstract
Text embedding and text generation are two of the most useful capabilities of modern large language models, yet they are typically provided by separate specialized models (an embedder and a generator) trained with incompatible objectives and deployed side by side in retrieval-augmented generation and agent systems. Recent unified embedding-generation models such as GritLM share parameters between the two capabilities, but they either switch attention regimes between modes or optimize embedding and language modeling as parallel objectives that never supervise each other. We argue that these designs are only partially unified: the two capabilities coexist on the same weights, but do not shape each other. We propose ULME, a post-training framework that unifies embedding and language modeling on a causal decoder-only LLM through predictive coupling. ULME inserts a chain of learnable slots between the query and the answer, reads a semantic embedding and a reasoning embedding from two designated slots, and trains three objectives on a shared causal backbone: a contrastive loss anchoring the semantic slot to document representations, a language-modeling loss supervising answer generation, and a stop-gradient predictive loss that aligns the reasoning slot with a pooled summary of the answer, so that generation actively shapes what the embedding encodes. Across two model families (Qwen2.5, Qwen3) and eight scales spanning 0.5B-8B parameters, a single ULME checkpoint improves MTEB by 3.3-4.1 points over the strongest open-source baseline, tightens BRIGHT and AdvBench-IR simultaneously, and remains on par with a generation-only fine-tune on standard accuracy benchmarks (MMLU, GSM8K, MMLU-Pro, HumanEval), demonstrating that genuine unification yields mutually reinforcing rather than merely coexisting capabilities.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30705
PDF
-
Omni-UBench: Benchmarking Omni-Modal LLM-based Guardrails Across Safety, Utility, and Reliability
2026-10-22 04:20:51Junhao Wang, Yidou Chen, Li Lu, Hongwei Yao, Xiaodan Li, Mengjie Wu, Yuefeng Chen, Zhan Qin, Kui RenAbstract
Advances in Omni-modal Large Language Models (OLLMs) have enabled increasingly seamless human-AI interaction, but their unified processing of heterogeneous modalities also expands the attack surface, allowing adversaries to exploit richer input compositions to circumvent safety alignment. To counter such risks, omni-modal guardrails have been proposed to detect and filter unsafe queries across arbitrary modalities. However, their capabilities and limitations remain insufficiently understood due to the absence of a comprehensive evaluation benchmark. In this paper, we present Omni-UBench, the first comprehensive framework for omni-modal safety assessment, evaluating guardrail models along three critical dimensions, i.e., safety, utility, and reliability. Omni-UBench curates over 183k samples drawn from 35+ heterogeneous sources spanning text, image, audio, video, and their cross-modal combinations, annotated with safety labels under a unified risk taxonomy. Benchmarking under Omni-UBench on 15 representative OLLMs and omni-modal guardrails yields several key findings, including: (1) modern OLLMs exhibit strong content moderation potential even without guardrail-specific tuning; (2) existing guardrails still struggle to identify cross-modal queries that conceal harmful intent or involve sophisticated jailbreaks; (3) they further exhibit substantial safety hallucination and over-sensitivity to textual patterns, revealing a deficient understanding of cross-modal safety semantics. Our findings provide valuable insights for developing more trustworthy omni-modal guardrails in future research.
Abstract (translated)
URL
https://arxiv.org/abs/2610.31713
PDF
