Multimodal large language models (MLLMs) have achieved remarkable progress in recent years, yet their ability to perform left–right reasoning in mirror contexts—a fundamental element of spatial cognition—remains underexplored. To address this gap, we introduce MirrorQA, a manually constructed benchmark with 5,549 samples, designed to evaluate MLLMs’ capability to distinguish left from right from a subject-centered perspective. MirrorQA is built through a three-stage pipeline (annotation, verification, and final review) to ensure high-quality labeling. Comprehensive evaluations on both open- and closed-source MLLMs show that even the best-performing models achieve only 65.40% accuracy, far below the 99.28% accuracy of humans. These results highlight substantial challenges in current MLLMs when reasoning about left and right, and point to promising directions for future research. MirrorQA and its code are publicly available at anonymous link https://github.com/stargazer-zeno/MirrorQA.
论文检索
输入标题、作者或关键词，从 9,460 篇学术成果中精准定位
Instruction-following is essential for aligning large language models (LLMs) with user intent. While recent reasoning-oriented models exhibit impressive performance on complex mathematical problems, their ability to adhere to natural language instructions remains underexplored. In this work, we introduce MathIF, a dedicated benchmark for evaluating instruction-following in mathematical reasoning tasks. Our empirical analysis reveals a consistent tension between scaling up reasoning capacity and maintaining controllability, as models that reason more effectively often struggle to comply with user directives. We find that models tuned on distilled long chains-of-thought or trained with reasoning-oriented reinforcement learning often degrade in instruction adherence, especially when generation length increases. Furthermore, we show that even simple interventions can partially recover obedience, though at the cost of reasoning performance. These findings highlight a fundamental tension in current LLM training paradigms and motivate the need for more instruction-aware reasoning models.
Post-training hybridization of large language models (LLMs) often replaces quadratic self-attention with sliding-window attention (SWA) to reduce KV cache usage and improve latency. Existing hybridization schemes are typically defined either at the layer level (e.g., interleaving) or at the head level via static rankings from local to global. Layer-level schemes ignore that local and global dependencies are routed through heads within the same layer, while static head-level rankings suffer from entanglement: a head’s local/global behavior can change after hybridization. We propose BOSCH, Black-box Binary Optimization for Short-context Head Selection, a training-free method that formulates the problem as a Large Neighborhood Search and decomposes it into three subproblems: (i) layer-importance detection via small-budget black-box probes, (ii) adaptive per-layer SWA-ratio assignment based on these sensitivities, and (iii) grouped head-level optimization within ratio buckets. Extensive experiments on 4 LLMs ranging from 1.7B to 30B parameters, across 4 SWA ratios, show that BOSCH consistently outperforms layer-level heuristics and 6 strong static head-level methods, with larger gains at higher SWA ratios. Under continual pretraining, BOSCH recover original long-context performance faster and to a higher level. Analysis of the selected heads reveals substantial turnover for BOSCH across different SWA ratios, underscoring the importance of performing head-level selection for each target ratio rather than relying on fixed locality rankings.
Recent advances in multimodal LLMs and systems that use tools for long-video QA point to the promise of reasoning over hour-long episodes. However, many methods still compress content into lossy summaries or rely on limited toolsets, weakening temporal grounding and missing fine-grained cues. We propose a multi-agent framework in which a master LLM coordinates a grounding agent to localize question-relevant segments and a vision agent to extract targeted textual observations. The master agent plans with a step limit, and is trained with reinforcement learning to encourage concise, correct, and efficient multi-agent cooperation. This design helps the master agent focus on relevant clips via grounding, complements subtitles with visual detail, and yields interpretable trajectories. On our proposed *LongTVQA* and *LongTVQA+* which are episode-level datasets aggregated from TVQA/TVQA+, our multi-agent system significantly outperforms strong non-agent baselines. Experiments also show reinforcement learning further strengthens reasoning and planning for the trained agent.
Emotion-Wheel-Guided Audio-Referred Text Representation for Multimodal Emotion Recognition in Conversation
PDF ↗Multimodal Emotion Recognition in Conversation aims to identify emotions within a dialogue with multimodal data, including audio, visual, and textual features. While existing methods have made significant improvements, there are two fundamental limitations to be addressed. From the modality fusion perspective, current approaches treat all modalities as functionally equivalent during fusion, overlooking their distinct communicative roles and information capacities, in which text conveys explicit semantic meaning while audio provides paralinguistic cues. From the emotion label perspective, many works ignore the continuous structure of emotion characterized by psychological theory and apply uniform penalties regardless of affective proximity. To address these limitations, we propose EMART, EMotion-Wheel-Guided Audio-Referred Text Representation for ERC, specifically focusing on audio and text modalities. First, we propose a modality-aware fusion strategy capturing linguistic features from text as the primary source and audio as a complementary component. Secondly, we propose an emotion-wheel-guided supervised contrastive loss to encode emotional proximity based on Russell’s circumplex model. Experimental results on IEMOCAP and MELD demonstrate outstanding performance. The code is available at: https://github.com/DILAB-HYU/EMART.git.
The evaluation of navigation instructions remains a persistent challenge in Vision-and-Language Navigation (VLN) research. Traditional reference-based metrics such as BLEU and ROUGE fail to capture the functional utility of spatial directives, specifically whether an instruction successfully guides a navigator to the intended destination. Although existing VLN agents could serve as evaluators, their reliance on high-fidelity visual simulators introduces licensing constraints and computational costs, and perception errors further confound linguistic quality assessment. This paper introduces GROKE (Graph-based Reasoning over OSM Knowledge for instruction Evaluation), a vision-free training-free hierarchical LLM-based framework for evaluating navigation instructions using OpenStreetMap data. Through systematic ablation studies, we demonstrate that structured JSON and textual formats for spatial information substantially outperform grid-based and visual graph representations. Our hierarchical architecture combines sub-instruction planning with topological graph navigation, reducing navigation error by 68.5% compared to heuristic and sampling baselines on the Map2Seq dataset. The agent’s execution success, trajectory fidelity, and decision patterns serve as proxy metrics for functional navigability given OSM-visible landmarks and topology, establishing a scalable and interpretable evaluation paradigm without visual dependencies.
Retrieval-Augmented Generation (RAG) has been proposed to mitigate hallucinations in large language models (LLMs), where generated outputs may be factually incorrect. However, existing RAG approaches predominantly rely on vector similarity for retrieval, which is prone to semantic noise and fails to ensure that generated responses fully satisfy the complex conditions specified by factual queries, often leading to incorrect answers. To address this challenge, we introduce a novel research problem, named Exact Retrieval Problem (ERP). To the best of our knowledge, this is the first problem formulation that explicitly incorporates structural information into RAG for factual questions to satisfy all query conditions. For this novel problem, we propose Structure Guided Retrieval-Augmented Generation (SG-RAG), which models the retrieval process as an embedding-based subgraph matching task, and uses the retrieved topological structures to guide the LLM to generate answers that meet all specified query conditions. To facilitate evaluation of ERP, we construct and publicly release Exact Retrieval Question Answering (ERQA), a large-scale dataset comprising 120{,}000 fact-oriented QA pairs, each involving complex conditions, spanning 20 diverse domains. The experimental results demonstrate that SG-RAG significantly outperforms strong baselines on ERQA, delivering absolute improvements from 20.68 to 50.88 points across all evaluation metrics, while maintaining reasonable computational overhead.
Multilingual language models often exhibit performance disparities across languages that can arise as early as the tokenization stage. Widely-used subword tokenization approaches favor high-resource languages, and tokenizer-free methods still yield longer sequences for scripts with a higher bytes-per-character ratio. To address these shortcomings, we propose to use the International Phonetic Alphabet (IPA) as a language-agnostic input representation for multilingual tokenizers. IPA provides a compact symbol inventory, greater cross-lingual character overlap, and a more balanced byte-per-character distribution across languages. We train matched pairs of text vs. IPA subword tokenizers across 24 languages and 14 scripts and demonstrate that IPA tokenizers consistently improve tokenization quality, especially for non-Latin scripts, and generalize more effectively to unseen languages and scripts.
MessToClean: Evidence-Grounded Structure-Preserving Reconstruction for Real-World Degraded Exam Paper Images
PDF ↗Intelligent education systems often collect exam sheets as in-the-wild photos. These photos often suffer from distortions and noise caused by handwriting and occlusions, collectively referred to as Real-World Degraded Exam Images (RDEI). Structure-preserving reconstruction is key to converting RDEI into structured assets for downstream educational applications. Existing Multimodal Large Language Models (MLLMs) often fail under RDEI, leading to disrupted structure and evidence-unsupported hallucinations. To tackle these challenges, we propose MessToClean, a backbone-agnostic, evidence-driven pipeline that treats off-the-shelf MLLMs as interchangeable components. By grounding extraction in pixel-aligned evidence and enforcing post-hoc consistency auditing on recovered structures, MessToClean mitigates unsupported hallucinations and enhances both controllability and structural fidelity in question-level reconstruction. We curate RDEI-Exam from our educational platforms and evaluate across 12 state-of-the-art MLLM backbones. Across these, MessToClean improves stem consistency by 1.01-3.18%, figure consistency by 0.50-49.16%, and refusal F1 by 1.06-10.88% across question types.
Precise spatial reasoning is fundamental to embodied intelligence, yet current Vision-Language Models (VLMs) remain bottlenecked by text-based Chain-of-Thought (CoT) that relies solely on textual reasoning trajectories, often bypassing active engagement with fine-grained visual details. To address this, we present E-ViC (Embodied Visual Chain), a framework that moves reasoning beyond text and directly into the visual domain. By formulating visual operations (e.g., zooming, marking) as executable primitives, E-ViC transforms perception from static prediction into an active verification process. Distinct from approaches relying on supervised step-wise trajectories, E-ViC is trained via an agentic reinforcement learning paradigm. This enables the model to autonomously discover optimal policies, leading to the emergence of human-like “look-and-confirm” strategies driven solely by task-level rewards. To facilitate this, we curate a comprehensive 24.4K-sample dataset covering diverse embodied tasks. By grounding reasoning in pixel-level interactions, E-ViC reframes spatial intelligence as a verifiable, tool-using capability. Extensive evaluations on external benchmarks demonstrate that our approach consistently outperforms strong VLM baselines with an average gain of 10.1%.
Africa is home to over one-third of the world’s languages, yet remains severely underrepresented in multimodal AI research. We introduce Afri-MCQA, the first Multilingual Cultural Question-Answering benchmark containing 7.5k Q A pairs across 15 African languages from 12 countries. The benchmark offers parallel text and speech modalities and was entirely created by native speakers. We find that models show poor performance across evaluated cultures, with near-zero accuracy on open-ended VQA when queried through native language or speech. To test linguistic competence, we include control experiments meant to assess this specific aspect separate from cultural knowledge, and we observe significant performance gaps between native languages and English for both text and speech. These findings underscore the pressing need for speech-first approaches, culturally grounded pretraining, and cross-lingual cultural transfer. We release Afri-MCQA to support more inclusive multimodal AI development.
TAIGR: Towards Modeling Influencer Content on Social Media via Structured, Pragmatic Inference
PDF ↗Health influencers play a growing role in shaping public beliefs, yet their content is often conveyed through conversational narratives and rhetorical strategies rather than explicit factual claims. As a result, claim-centric verification methods struggle to capture the pragmatic meaning of influencer discourse. In this paper, we propose TAIGR (Takeaway Argumentation Inference with Grounded References), a structured framework designed to analyze influencer discourse, which operates in 3 stages: (1) identifying the core influencer recommendation–takeaway; (2) constructing an argumentation graph that captures influencer justification for the takeaway; (3) performing factor graph-based probabilistic inference to validate the takeaway. We evaluate TAIGR on a content validation task over influencer video transcripts on health, showing that accurate validation requires modeling the discourse’s pragmatic and argumentative structure rather than treating transcripts as flat collections of claims.
UrbanGeoEval: A City-Scale Benchmark for Evaluating Large Language Models in Geospatial Reasoning
PDF ↗Current evaluations of geospatial reasoning in LLMs are frequently impeded by the entanglement of factual recall and spatial logic, which often obscures the models’ true capabilities in complex city-scale environments. To address this, we introduce UrbanGeoEval, a comprehensive benchmark featuring a dual-module framework designed to disentangle these competencies. The Knowledge Module assesses urban memory via scalable map-based queries, while the Reasoning Module isolates pure logical inference across 3,148 realistic tasks by providing necessary geospatial context. Unlike prior benchmarks that hand the model pre-computed spatial text, UrbanGeoEval provides raw geometry and forces the model to act as a spatial computing engine. Our evaluation methodology introduces a reliable hybrid pipeline that merges deterministic programmatic checks with an LLM-as-a-Judge, achieving expert-level evaluation accuracy. Extensive experiments on 18 widely used LLMs uncover critical insights: (1) models exhibit severe geographic biases and resolution gaps; (2) failures in complex multi-hop tasks often stem from brittle foundational spatial skills rather than high-level logic deficits. UrbanGeoEval provides a precise diagnostic tool for advancing urban geospatial intelligence in LLMs.
From Curated Data to Scalable Models: Continual Pre-training of Dense and MoE Large Language Models for Tibetan
PDF ↗Large language models (LLMs) have achieved remarkable success across a wide range of natural language processing tasks, yet their performance remains heavily biased toward high-resource languages. Tibetan, despite its cultural significance and large speaker population, is still substantially underrepresented. In this work, we present a comprehensive pipeline for advancing Tibetan language modeling through large-scale data curation and continual pre-training. We construct a 72 GB high-quality Tibetan corpus, the largest to date, and adapt Qwen2.5-7B through balanced multilingual continual pre-training with Tibetan, Chinese, and English, followed by multilingual instruction tuning. To further scale capacity efficiently, we extend the dense model to a 50B-A10B Mixture-of-Experts architecture. Due to the absence of standardized Tibetan benchmarks, we build multiple evaluation datasets via high-quality translation and human verification. Experimental results show that both dense and MoE models consistently outperform existing open-source and Tibetan-focused models of similar scale across diverse tasks. Our work advances Tibetan-centric LLM research and provides transferable insights for extending LLMs to other low-resource languages. We will release the model weights, evaluation benchmarks, and detailed data processing documentation in the follow-up.
Quantifying and Mitigating Socially Desirable Responding in LLMs: A Desirability-Matched Graded Forced-Choice Psychometric Study
PDF ↗Human self-report questionnaires are increasingly used in NLP to benchmark and audit large language models (LLMs), from persona consistency to safety and bias assessments. Yet these instruments presume honest responding; in evaluative contexts, LLMs can instead gravitate toward socially preferred answers—a form of socially desirable responding (SDR)—biasing questionnaire-derived scores and downstream conclusions. We propose a psychometric framework to quantify and mitigate SDR in questionnaire-based evaluation of LLMs. To quantify SDR, the same inventory is administered under HONEST versus FAKE-GOOD instructions, and SDR is computed as a direction-corrected standardized effect size from item response theory (IRT)-estimated latent scores. This enables comparisons across constructs and response formats, as well as against human instructed-faking benchmarks. For mitigation, we construct a graded forced-choice (GFC) Big Five inventory by selecting 30 cross-domain pairs from an item pool via constrained optimization to match desirability. Across nine instruction-following LLMs evaluated on synthetic personas with known target profiles, Likert-style questionnaires show consistently large SDR, whereas desirability-matched GFC substantially attenuates SDR while largely preserving the recovery of the intended persona profiles. These results highlight a model-dependent SDR–recovery trade-off and motivate SDR-aware reporting practices for questionnaire-based benchmarking and auditing of LLMs.
Large Reasoning Models (LRMs) have achieved remarkable success on complex tasks by generating detailed Chain-of-Thought (CoT) reasoning. However, they tend to apply a uniform, computation-intensive deep reasoning strategy to all problems, leading to unnecessary overhead on simple tasks. This significantly hinders their efficiency in real-world applications. While existing methods have improved reasoning efficiency to some extent, they still face critical challenges such as conflicting objectives, limited adaptability. To address these limitations, we propose AdaMix, an adaptive reasoning framework via decoupled optimization. To mitigate optimization conflicts, AdaMix first constructs two specialized adapters: an efficiency-oriented short adapter and an accuracy-oriented long adapter. It then incorporates a difficulty-aware routing model that assesses problem complexity to predict a reasoning intensity coefficient. This coefficient is used to dynamically interpolate a mixed adapter from the two base adapters, enabling fine-grained reasoning control. Our experiment demonstrates that our AdaMix reduces the average response length of DeepSeek-R1-Distill-Qwen-7B by 54.9% while improving accuracy by up to 4.8% on five mathematical datasets, thus indicating a favorable accuracy-efficiency trade-off.
Vision-Language-Action models have recently emerged as a powerful paradigm for general-purpose robot learning, enabling agents to map visual observations and natural-language instructions into executable robotic actions. Though popular, they are primarily trained via supervised fine-tuning or training-time reinforcement learning, requiring explicit fine-tuning phases, human interventions, or controlled data collection. Consequently, existing methods remain unsuitable for challenging simulated- or physical-world deployments, where robots must respond autonomously and flexibly to evolving environments. To address this limitation, we introduce a Test-Time Reinforcement Learning for VLAs (TT-VLA), a framework that enables on-the-fly policy adaptation during inference. TT-VLA formulates a dense reward mechanism that leverages step-by-step task-progress signals to refine action policies during test time while preserving the SFT/RL-trained priors, making it an effective supplement to current VLA models. Empirical results show that our approach enhances overall adaptability, stability, and task success in dynamic, previously unseen scenarios under simulated and real-world settings. We believe TT-VLA offers a principled step toward self-improving, deployment-ready VLAs.
Uncertainty estimation is a promising approach to detect hallucinations in large language models (LLMs). Recent approaches commonly depend on model internal states to estimate uncertainty. However, they suffer from strict assumptions on how hidden states should evolve across layers, and from information loss by solely focusing on last or mean tokens. To address these issues, we present Sequential Internal Variance Representation (SIVR), a supervised hallucination detection framework that leverages token-wise, layer-wise features derived from hidden states. SIVR adopts a more basic assumption that uncertainty manifests in the degree of dispersion or variance of internal representations across layers, rather than relying on specific assumptions, which makes the method model and task agnostic. It additionally aggregates the full sequence of per-token variance features, learning temporal patterns indicative of factual errors and thereby preventing information loss. Experimental results demonstrate SIVR consistently outperforms strong baselines. Most importantly, SIVR enjoys stronger generalisation and avoids relying on large training sets, highlighting the potential for practical deployment.
Steering methods have emerged as effective tools for guiding large language models’ behavior, yet multimodal large language models (MLLMs) lack comparable techniques due to architectural diversity and limited availability of multimodal steering vectors. Inspired by this gap, we demonstrate that steering vectors derived solely from text-only LLM backbones can effectively guide and enhance their multimodal counterparts, revealing a novel cross-modal transfer that enables reuse of existing interpretability tools. Using community-standard methods—Sparse Autoencoders (SAE), Mean Shift, and Linear Probing—we validate this transfer effect across diverse MLLM architectures and visual reasoning tasks. Text-derived steering consistently enhances multimodal performance, with Mean Shift achieving up to +7.3% improvement in spatial relationship accuracy and +3.3% in counting accuracy on CV-Bench, and exhibits strong generalization to out-of-distribution datasets, for example reaching +34.2% on CLEVR counting tasks. This reveals that textual representations alone can effectively enhance visual grounding in MLLMs, bridging the mature ecosystem of text-based steering to MLLMs with minimal additional data collection or computational overhead.
To reliably interpret the evolving context of an LLM as a reasoning trace, the underlying belief of the LLM needs to transition consistently with the progression of the context.We focus on evaluating whether the beliefs held by a model remain consistent before and after the extension of the context.Previous research on consistency evaluation typically uses datasets with ground-truth answers, which is problematic because task-solving ability acts as a confounding factor, obscuring the direct evaluation of consistency.Furthermore, evaluating cases where inconsistency stems from multiple errors poses difficulties.We propose a new evaluation method to assess the consistency of LLMs in a multiple-choice question answering format, designed so that any option chosen is correct, allowing for the evaluation of the proposed belief consistency.It also supports isolation of errors such as reasoning failures and biases.We reveal that the belief consistency does not improve solely with model size scaling,whereas continual pre-training on code and mathematics text improves it.Furthermore, models trained on code and mathematics text show a seemingly contradictory result of increased logical failures, indicating that belief consistency and superficial consistency are not necessarily directly linked.