Long egocentric videos pose a central challenge for multimodal language models: the visual stream must be compressed into a memory representation that can be searched efficiently while still preserving the evidence needed for question answering. Building on an online textual-memory framework in which first-person videos are converted into temporally grounded captions, this work systematically studies the role of retrieval in episodic-memory video question answering. We evaluate how different retrieval units, query formulations, context budgets, and reranking strategies affect multiple-choice answering on QAEgo4D-Closed, isolating retrieval as the main variable while keeping the underlying generated memory fixed. Our best configuration, Parent, matches fine-grained caption subchunks, then provides their full parent captions to the reasoner. It achieves 56.4 % accuracy, matching strong full-memory and retrieval baselines while exposing only a small fraction of the textual memory at question time. More importantly, our analysis shows that better temporal coverage does not necessarily translate into better answers: increasing the retrieval budget improves temporal Hit@K and Recall@K, but can reduce accuracy by introducing visually plausible distractors and uncertain caption evidence. These findings suggest that retrieval-augmented egocentric memory should not be optimized only for evidence recall, but also for the reliability, specificity, and answer-discriminative value of the retrieved context. The study motivates more structured episodic memories that move beyond raw caption retrieval toward evidence-aware representations for long-form egocentric reasoning.
https://arxiv.org/abs/2610.31284
Training-free personalization adapts a frozen vision-language model (VLM) to recognize user-defined concepts from only a few exemplars by retrieving concept entries at inference time. While recent retriever-based methods enrich each concept with descriptive evidence, they largely treat concepts in isolation, so the retrieved information can be plausible yet non-discriminative, especially when the personalized concept set contains highly confusable, fine-grained categories. We argue that personalization is inherently relational: correct recognition depends on the specific distinctions between a concept and its nearest alternatives. To this end, we propose ReCoG, a Relational Concept Graph for retrieval-augmented personalization. ReCoG represents concepts as nodes and augments them with directed edges that explicitly capture discriminative cues between concept pairs, elicited by a frozen VLM during database construction. At inference, ReCoG retrieves a shortlist and resolves ambiguity via explicit pairwise comparisons conditioned on the corresponding relational edges, selecting the concept that consistently wins against competing candidates. We also introduce FINGER benchmark to test personalized recognition on confusable concept sets. Across FINGER, ReCoG substantially outperforms prior methods, with the largest gains in the most confusable regimes. ReCoG also achieves consistently best performance on established personalization benchmarks including captioning and VQA, confirming that relational evidence is broadly beneficial beyond fine-grained settings. Project Page: this https URL
https://arxiv.org/abs/2610.31282
Clinical Question Answering (QA) underpins Clinical Intelligent Decision Support Systems, but progress is severely limited for low-resource languages such as Indonesian by the absence of public domain-specific data. We introduce IndoNutriQA, an expert-curated Indonesian clinical QA benchmark of 911 question-answer pairs built from real clinical-nutrition reports. Rather than a large-scale training corpus, it is positioned as a high-quality evaluation resource: every pair is grounded in a clinically relevant context, enriched with structured metadata (anthropometric assessment, medical diagnosis, document provenance, chunk segmentation). Eight standardised question types following the Nutrition Care Process (NCP), including the Problem, Etiology, Sign and Symptom (PES) diagnosis statement, are answered for every report, which was written by a graduate-level dietetic intern, by one of three nutritionists. Answers are free-form clinical prose rather than extractive spans, and an expert validation study finds 6-10% of reference answers contestable. The dataset supports generative, extractive and retrieval-augmented QA as well as zero- and few-shot LLM evaluation. We report fine-tuned baselines (IndoBERT variants, mT5, RAG) with 5-fold cross-validation, learning curves and bootstrap 95% confidence intervals, and benchmark eight contemporary LLMs zero- and five-shot. We further calibrate our LLM-as-judge rubric against three expert raters (one RD, two nutritionists), finding it human-comparable on factuality and completeness but unreliable on clinical safety, and a human error analysis shows it misses unsupported content that the raters identify. IndoNutriQA is released to foster research in Indonesian clinical NLP and nutritional decision support.
https://arxiv.org/abs/2610.31256
Some questions about an image need commonsense or factual knowledge that its pixels do not contain. Large Language Models (LLMs) can supply this knowledge to Visual Question Answering (VQA) systems, but the depth at which their guidance should enter a vision-language model remains unclear. We prompt the 70-billion-parameter Llama~3.3 model to describe, for each question, the visual concepts to focus on, the distractors to ignore and the external facts needed. We inject this guidance into a frozen Vision Transformer (ViT) paired with an instruction-tuned Text-to-Text Transfer Transformer (T5) in two ways: decision-level late fusion and attention-matrix injection. On the Outside Knowledge VQA (OK-VQA) benchmark, and contrary to the common view that deeper is better, the shallow route wins: late fusion reaches 39.4% exact match, +10.0 points above an unguided baseline, while attention injection reaches only 31.8% under data-scarce fine-tuning. An LLM paraphrasing pipeline also enlarges the training set more than tenfold. The results suggest that integration depth should be matched to the data regime.
https://arxiv.org/abs/2610.17079
Text-to-music alignment is commonly evaluated by the global similarity between text and audio embeddings, which collapses all described attributes into a single score and obscures which ones are missing. We instead introduce MQAScore, which verifies each atomic musical concept along five clip-level dimensions (genre, instrument, mood/theme, vocal, and usage) by the probability that a large audio language model (LALM) answers affirmatively to a binary question. To assess evaluator reliability, we construct verification labels from presence/absence annotations and LLM-guided concept substitutions, and propose the Sensitivity-Invariance Ratio, which measures how selectively an evaluator responds to meaning-changing substitutions. Experiments show that concept-level scoring improves both question-answering and similarity-based evaluators, and MQAScore can be a more reliable and interpretable fine-grained complement to global similarity for evaluating text-to-music alignment.
https://arxiv.org/abs/2610.31083
Large language models (LLMs) deployed in mental-health contexts can be evaluated for safety failures, but the resulting judgments are often difficult to inspect or explain. Common approaches to improving safety rely on prompt engineering, LLM-as-a-Judge, and Constitutional AI. In this work, we describe a clinician-led alternative in which two clinicians define, finalize, and adjudicate the four safety checks of a rule base, which is formally verified for internal consistency and applied as a guardrail over LLM outputs. Here and in the title, clinician-led refers to this rule base, not to the fact extractor or the uncertainty default, which are ours. Every verdict of the resulting verification layer is recomputable from a named rule and sixteen facts. Against an LLM judge matched in model, revision, quantization, rule text, and decoding but without the layer's computed 988/911 fact, the largest disagreement class, the satisfied-rule class, is structured: the Llama-3.3-70B judge, on its primary prompt, demands a crisis resource from 55 of the 65 replies (84.6%) that already contain one, while the same model, asked as a fact extractor, reports a recognized resource present on 54 of the 55. The verification layer never raises that demand, by construction. We report with equal prominence where this judge is stronger: fewer false crisis alarms on the one reference-scored check. We also report the measured cost (a specification revision raised block-class outcomes on held-out boundary prompts of an over-refusal benchmark from 6.0% to 16.0%, chiefly by changing fact extraction), the clinicians' decision to keep our uncertainty default, and the analysis of the result. The contribution is not a safer verdict. It is a guardrail whose disagreements with the judge localize to named rules and facts and whose measured costs trace to named facts and the uncertainty default.
https://arxiv.org/abs/2610.30690
Existing datasets on cultural knowledge have been typically created by human annotators familiar with each culture. However, the process requires considerable human effort, presenting a significant challenge to scalability. To address this challenge, we introduce WikiOriginQA, a QA dataset automatically constructed by objectively representing the origins of knowledge through the creation dates of different Wikipedia language editions. Experimental results with the created dataset confirm that LLM performance tends to be biased toward Anglophone cultures. Furthermore, the results suggest the importance of knowledge propagation into the Anglosphere. These findings demonstrate the validity of our automated approach, enabling cultural bias evaluations with superior scalability across diverse languages without extensive manual annotation.
https://arxiv.org/abs/2610.30677
Reinforcement learning from verifiable rewards (RLVR) has produced large reasoning gains in language models, and verifiable video benchmarks make it applicable to causal-temporal video question answering. We study what RLVR teaches video-language models about time. We fine-tune four open models (Qwen3-VL-8B/4B, Qwen2.5-VL-7B, Gemma-3-12B) with group relative policy optimization under three data recipes: verified (synthetic CLEVRER questions with exact answer and event-order rewards), unverified (self-supervised pretext tasks over 43,751 real web videos), and a 1:1 mixture, plus a verified+real arm that adds 4,000 verifiable questions on real video. Each cell is evaluated in-domain and on out-of-domain real video (a NExT-QA temporal stress set and an MVBench subset), with frames in order, shuffled, and absent. (1) Verified training yields large in-domain gains that shrink as base competence grows (+14 to +19 points on weaker models; +6 on the strongest). (2) Much of the gain is non-visual: accuracy with no frames rises nearly as much as with frames. (3) Verified-only training can severely degrade out-of-domain accuracy with no sign during training: Qwen3-VL-8B loses 26.7 and 25.2 points on the two real-video sets, while the mixture never significantly degrades a model trained on it. Adding real verified questions removes that loss (-2.3 points, within noise of base) and keeps a +9.3 in-domain gain, so the cause is narrow synthetic-only data, not verification. (4) No recipe induces temporal-order grounding: across 41 evaluations the ordered-versus-shuffled gap is indistinguishable from zero in 39 and marginal in two, despite an event-order reward. Verifiable rewards improve benchmark accuracy without temporal understanding. Report no-frame controls, and mix in real video to guard against out-of-domain degradation.
https://arxiv.org/abs/2610.03792
LLM-based chat agents are increasingly used for complex, multi-turn, multilingual medical triage in low- and middle-income countries (LMICs), yet evaluation methods focus on single-turn, English-only benchmarks that primarily test the chatbots' knowledge. In practice, users are often confused, overwhelmed, or emotionally vulnerable, giving unclear, misspelled single-word answers or long messages in regional dialects, and the context of a conversation can shift rapidly. We propose SIM-UT (Simulation of User Behaviour for Triage), a multi-turn, simulation-based evaluation method that identifies ecologically valid, contextually relevant failure modes of LLM-based triage chat agents. Using SIM-UT, we evaluated three real-world WhatsApp-based health chatbots: a respiratory screening programme in India, a hearing health support programme in South Africa, and a diabetes and lifestyle coaching agent in Kenya. We compared a medically fine-tuned open model (MedGemma) against general-purpose frontier models (ChatGPT and Claude) across 5,000+ multilingual simulated conversation turns grounded in the personas, constraints, and communication patterns of the real user populations, over 2-4 iterative rounds. A key failure mode emerged in both model families: de-escalation risk. The agents initially identified medical emergencies correctly, but when the user minimised the symptoms without introducing new clinical information, they de-escalated, exhibiting clinically dangerous sycophancy, for example agreeing that vaginal bleeding in pregnancy was "not a problem" or that decreased fetal movements at 32 weeks "can be normal". We recommend user simulation-based behavioural evaluation, grounded in real user context and covering adversarial and edge cases, as standard practice for triage chatbots in production, so that safeguards can be added before deployment or in subsequent iterations.
https://arxiv.org/abs/2610.30412
Dynamic representations based on 4D Gaussian Splatting (4DGS) have shown strong capabilities in reconstruction and rendering, yet current methods remain limited in supporting complex natural-language referring in dynamic scenes. Existing 4D Gaussian methods are primarily designed for rendering-oriented dynamic modeling, while current semantic extensions often rely on scene-specific semantic optimization or are evaluated on relatively simple query settings. In this paper, we study Referring 4D Gaussian Splatting (R4DGS), a task for grounding temporally varying, multi-target or reasoning-intensive, and zero-target or distractor expressions in dynamic 4D scenes. To support this task, we introduce R4D-Bench-QA, a benchmark with structured query annotations. We further present ReferGaussian, a unified framework built upon 4DGS that converts low-level dynamic Gaussians into persistent, language-addressable entities for training-free referring inference. ReferGaussian organizes mask-supported Gaussian entities in an EntityBank, uses Multimodal Large Language Models (MLLMs) to decompose complex queries into trackable object phrases and structured referring constraints, and performs constraint-aware grounding without scene-specific semantic optimization. Experiments on R4D-Bench-QA and the 4D LangSplat protocol demonstrate improved temporal target selection and competitive spatiotemporal grounding without scene-specific semantic training. Project resources are available at this https URL.
https://arxiv.org/abs/2610.30326
Today's multimodal assistants are typically prompted to interpret the visual evidence presented to them. To offload the cognitive burden of remembering across long horizons, wearable assistants should instead proactively acquire and maintain useful memory as the user's surroundings evolve. Doing so on the user's behalf requires identifying what matters, understanding the scene in real time, and initiating memory maintenance in the background while the user stays focused on their activities. We introduce ProactEgoMem, a benchmark that evaluates proactive memory in streaming egocentric video. Three tasks ground the evaluation in ambient environmental needs: Information Capture (IC) preserves transient yet dense visual information, State Tracking (ST) follows changing object states, and Location Tracking (LT) records object movements. The benchmark defines its protocol at two levels of proactiveness: the Notification Level asks for a single timely notification that satisfies the user's intent, while the deeper Intervention Level asks the agent to maintain memory on the user's behalf through an active start-to-stop tool lifecycle. Parallel ProactEgoMem-Daily and ProactEgoMem-Sci tracks contain 300/100 clips covering daily and scientific use cases, paired with 1000/500 questions in ProactEgoMem-QA, which measures the quality of the proactively acquired memory. Across proprietary, open-source, video-language, and streaming baselines, even the strongest models fail to capture memory both timely and accurately. ProactEgoMem provides a rigorous evaluation foundation for proactive multimodal assistance and for video agents that co-exist with their users.
https://arxiv.org/abs/2610.30089
Retrieval-augmented generation (RAG) frequently integrates knowledge graphs (KGs) for complex multi-hop question answering. However, their performance is increasingly bottlenecked by the knowledge structuring and reasoning processes. (i) Severe semantic loss: Distilling continuous texts into discrete symbols strips away contextual nuances and narrative coherence. (ii) Pronounced modality mismatch: Fragmented graph structures clash with the continuous texts LLMs are pre-trained on, while rewriting them into natural language introduces generative noise and omits information. To address these limitations, we propose ThreadRAG, a semantic-topology-based RAG framework for context-preserving knowledge structuring and modality-native evidence reasoning. ThreadRAG introduces a Semantic Topology Mapper (STM), which organizes retrieved documents into context-intact Semantic Nodes and constructs a Semantic Topology Graph (STG) through narrative, semantic, and lexical edges. It further incorporates Energy-Driven Pathfinding (EDP), which identifies query-relevant subgraphs via energy diffusion and deterministically linearizes them into coherent textual evidence paths. Experiments on three multi-hop QA benchmarks demonstrate that ThreadRAG consistently outperforms strong baselines. Notably, it achieves a 14.4% EM improvement over text-centric methods on HotpotQA and surpasses the best baseline by 5.8% EM on 2WikiMultihop, while reducing token cost to 23.3% of the strongest baseline.
https://arxiv.org/abs/2610.29933
This paper describes our submission to the WMT 2026 Multilingual Instruction Shared Task. We adapt Qwen3-8B with LoRA using organiser-provided data, EuroAlpaca, a multilingual instruction dataset created through task-preserving localisation, and CrossEuroAlpaca, a cross-lingual augmentation that assigns different source and target languages to instructions, contexts and responses. Our system targets open-ended question answering (QA-OEG) in ten shared-task languages represented in the auxiliary data. Under matched decoding, adaptation improves the mean Gemma-4 judge score on target-language QA-OEG by 7.8% relative to the base model. On an additional evaluation using the Aya Evaluation Suite across all shared-task languages, it yields relative gains of 41.5% in macro-averaged ROUGE-L and 2.9% in BERTScore-F1, with larger improvements in target languages. However, these gains are task- and language-specific. Gemma-4 scores improve for QA-OEG and summarisation in the target languages but decline for context-grounded QA there and across all three tasks in the non-target languages.
https://arxiv.org/abs/2610.29926
Over the past decade, substantial efforts have been devoted to developing machine learning (ML) methods for extracting materials data and, where possible, materials knowledge directly from the scientific literature. This review traces the evolution of these methods from heuristic, rule-based pipelines through natural language processing (NLP) models and transformer-based encoders to, lastly, large language models (LLMs) and agentic systems capable of extracting materials information across multiple modalities, including text, tables, and figures. Beside construction of large, queryable materials databases and knowledge graphs that support downstream materials discovery, we discuss how these advances have enabled new capabilities, such as, generic question answering across materials science, AI co-scientists and chatbots for research assistance, and closed-loop optimization and discovery when integrated with autonomous laboratories. We critically assess the datasets generated using these approaches, their extraction accuracy and computational costs, and the respective strengths and limitations of different methodologies, and even provide practical guidelines for developing more reliable and scalable materials data extraction pipelines. Finally, we identify key open challenges, including the extraction of full materials structure-property-processing-performance tetrahedron at the sample level, the scarcity of reliable benchmark datasets for systematic evaluation, and the resolution of non-standard materials terminology across different document modalities.
https://arxiv.org/abs/2610.29895
LLM-generated records increasingly populate production information systems where they are treated as data assets subject to the same quality expectations as records from operational databases. A single such record can simultaneously satisfy and violate a quality rule: a grounded fact adjacent to a hallucination, a useful answer adjacent to a stereotype. Classical data cleaning cannot resolve this because the two judgments are incompatible over the same span; a single LLM judge cannot either, since it shares the training biases that produced the conflict. Existing interventions such as prompt engineering, chain-of-thought, and RLHF reduce surface error rates but provide no audit trail and cannot distinguish genuine remediation from blanket suppression. The question is whether a deterministic check over structured inter-agent records can catch the defects that referential checks alone miss. We present BiasGuard-D: two adversarial roles drawn from distinct model families exchange typed claim documents, a three-model quorum adjudicates by majority vote, and a combined referential-plus-semantic invariant validates the verdict. The semantic component fires on verdicts referential checks miss; removing the iterative loop drops invariant satisfaction sharply. The deliberation layer transfers from stereotype to hallucination remediation on HaluEval QA with only role-framing changes, raising right-answer recovery by 49 percentage points; the triage gate required a cross-family second auditor to maintain recall on objective hallucination. Every claim, vote, and instruction is stored as a typed record giving operators a provenance trail auditable under emerging AI regulations.
https://arxiv.org/abs/2610.29669
The deployment of small and quantized language models in resource-constrained settings raises an important fairness question: when a compressed model produces fewer biased responses, does this reflect fairer behavior or greater abstention? We study this problem in Question Answering (QA) using the BBQ benchmark across multiple model sizes and quantization settings from Qwen3, Gemma-3, and SmolLM2. We also examine whether the observed trends remain stable across alternative prompt formulations and when the ``Unknown'' option is removed in a forced-choice setting. Our results show that larger models achieved higher QA accuracy, but bias showed no simple scaling trend and quantization effects varied across configurations. Notably, 20 of the 26 configurations with fewer biased responses also show increased over-abstention, indicating that lower observed bias can partly reflect greater abstention. These findings show that lower observed bias can partly reflect greater abstention, and that fairness evaluation should consider biased responses, abstention, prompt formulation, and answer-option design jointly.
https://arxiv.org/abs/2610.29657
Digital image transmission typically optimizes source coding, channel coding, and modulation separately, limiting how source representations adapt to channel impairments. We propose SemQAM, a joint semantic-channel coding and modulation scheme for generative image transmission converting source to modulated symbols, such as standard QAM symbols. Channel impairments create two challenges: they can corrupt the latent representation at the receiver, and this corruption can compound through autoregressive coding, degrading reconstruction quality. First, SemQAM represents latent features with vector quantization (VQ), maps each VQ index to a QAM symbol, and updates the codeword corresponding to the detected symbol so that channel impairments directly affect codebook learning. Moreover, SemQAM uses the conditional expectation of received codewords to approximate the receiver reconstruction and reduce error propagation during autoregressive coding. SemQAM supports multiple SNRs and CBRs in a unified model, while also supporting conventional source coding through SemQAM-C. Experiments show that (i) under the same LDPC and QAM settings, SemQAM-C outperforms OneDC in LPIPS and DISTS with substantially lower complexity, showing that the unified model retains competitive capability for source coding. (ii) Compared with this separated counterpart, SemQAM reduces channel uses by 18.57% at matched LPIPS and outperforms the compared JSCC methods, demonstrating the benefit of joint coding.
https://arxiv.org/abs/2610.29643
Multi-modal Large Language Models (MLLMs) have seen broad adoption, but their biases are still far less examined than those of standard LLMs, creating possible social harms when deployed in real-world settings. Prior efforts to extend bias analyses from LLMs into the multi-modal space either cover only a few bias categories or do not ensure that the groups being compared are clearly addressed by visual cues in the images. To overcome these limitations, we introduce Melange, a wide-ranging synthetic visual benchmark covering seven categories for assessing social biases in MLLMs. It is grounded in the widely used language-only BBQ dataset (Bias Benchmark for Question Answering) from BigBench, carefully extending its constructed contexts via controlled generation of visual scenarios so the target attributes appear visually and clearly. To maximize reliability, our generated images pass through two stages of validation: automated screening using visual question answering (VQA) and subsequent manual inspection by human annotators. By systematically evaluating 23 open-source MLLMs across three matched evaluation settings---text-only, image-only, and image-with-context---Melange isolates the fairness effect of the input modality. Switching from text to vision reduces fairness by 0.291 on average, a gap that varies by model family rather than by scale alone. A perception control confirms that this gap is not an artifact of visual comprehension: models identify the depicted attributes at 90--95\% accuracy yet still fail to abstain. When models do commit to a biased answer, polarity-aware stereotype alignment shows they choose the human-stereotypical response roughly three-quarters of the time.
https://arxiv.org/abs/2610.29537
Visual grounding (VG) in visual question answering (VQA) is a fundamental capability of Vision-Language Models (VLMs), enabling the detection of the most relevant region in an image to answer the question. This reflects the model's capability of multimodal comprehension, which is crucial when interacting with visually impaired people (VIP). Indeed, visual grounding, when applied to assistive scenarios, is not merely a technical feature, but rather a trust and safety requirement. In this work, we study visual grounding in VLMs for visually impaired people. To this end, we assess open-source VLMs on visual grounding using the VizWiz-VQA-Grounding dataset and challenge, which contains real-life images taken by VIP. We propose VG-SHAP, which extends an existing explainability framework to provide visual grounding explanations, allowing the evaluation of black-box VLMs. Our results on the VizWiz-VQA-Grounding challenge show that despite generating correct answers, VLMs struggle to ground in the proper region in the image. However, these results are constrained by the quality of the segmentation model of our explainability framework and further research is necessary.
https://arxiv.org/abs/2610.33078
Quantization-aware training (QAT) aims to recover reasoning capabilities in low-bit large language models (LLMs). Existing recipes typically supervise the quantized student with the original full-precision model on fixed response sequences. However, the resulting low-bit quantized models still struggle on long-horizon reasoning tasks and often fail to terminate within a token budget that suffices for the full-precision model. We identify two challenges in closing this gap. First, fixed-response supervision leaves a mismatch between training prefixes and those encountered during student generation, and this trajectory divergence grows as responses lengthen. Second, on its own prefixes, the student underpredicts some next-token candidates and overpredicts others, and this distributional misalignment becomes more frequent at high-entropy prefixes. We present OnQ, a low-bit recovery pipeline that combines on-policy distillation and candidate-wise Dynamic KL. Starting from block-wise calibration, the quantized student generates responses and receives full-precision teacher supervision on its own prefixes. Dynamic KL weights forward and reverse KL by each candidate's teacher-student probability ratio to balance correction of under- and overprediction. Across Qwen3-1.7B, 4B, and 8B at 2- and 3-bit precision, OnQ achieves the highest nine-benchmark average, with the largest gains on long-horizon reasoning benchmarks. At 2 bits, it outperforms every baseline on every benchmark at every model size, improving the Qwen3-8B average by 10.26 points over the strongest baseline. It also produces shorter responses with fewer truncations and repetitive errors.
https://arxiv.org/abs/2610.29378