Overview

Multimodal reasoning studies how models perceive information across modalities, connect that evidence to knowledge, and use it to answer questions, plan actions, or explain decisions. Common approaches include instruction tuning and chain-of-thought training, reinforcement learning with verifiable rewards, tool-using agents, and structured methods such as graphs or symbolic solvers. Benchmarks such as MathVista and MMMU test visual problem solving, while video, document, scientific, and medical tasks probe reasoning across time and specialized evidence. A central challenge is ensuring that correct answers actually depend on relevant multimodal evidence rather than language shortcuts; grounding, cross-modal integration, and reliable generalization remain difficult.

Latest papers

531 papers29 papers in the last four weeks, down 44% on the four weeks before. 0.2% of all new papers.

Nov 18, 2026

Benchmarking Reasoning Reliability,Prior_Evidence Contrast,and Spectral Discrimination in Remote-Sensing Geological Image Interpretation

Reliable lithology interpretation requires accurate geological recognition and reasoning grounded in visual evidence. Although Multimodal Large Language Models show promise in remote sensing, their reliability in geological interpretation remains unclear. We introduce HyGeoBench, a diagnostic benchmark of 10,000 questions based on hyperspectral imagery across three dimensions: reasoning reliability, prior_evidence contrast, and spectral discrimination. It covers ten task types, with both multiple-choice and open-ended response formats. Integrating hyperspectral satellite imagery with geological annotations and verified evidence, HyGeoBench examines reasoning errors, conflicts between geological priors and visual evidence, and cross-band inconsistencies. Evaluations reveal limitations in fine-grained recognition and evidence-grounded interpretation. To address these challenges, we propose HyGeoAgent, a tool-augmented framework integrating region observation, text knowledge retrieval, same-band image retrieval, and native PCA analysis through adaptive tool use. Experiments demonstrate consistent improvements over the corresponding base models, improving geological image interpretation across diagnostic dimensions.
Nov 17, 2026

PhysScope: Agentic Multimodal Physics Reasoning with Grounded Visual Memory

Vision-language models (VLMs) can correctly interpret physics diagrams yet apply constraints outside the states or conditions in which they hold, causing premise errors to propagate through subsequent calculations. We introduce PhysScope, an inference-time visual harness that supports self-correction of these scope mismatches without updating model weights. PhysScope enables the model to link physical scopes directly to its visual observations and explicitly constructs derived views to support reasoning. The harness imposes explicit constraints on state distinctions, conditions, and physical scopes during model reasoning, and checks agent-proposed revisions against sources and scopes. Accepted revisions invalidate registered dependents, thereby maintaining scope-aware working memory. On PhysAlign, PhysScope achieves higher solving scores than direct solving across all six evaluated VLMs. With Gemini-3.8-flash, it outperforms Canvas-CoT by 9.34 percentage points and substantially improves solving accuracy on the out-of-distribution (OOD) PhysUniBench benchmark. Ablation studies support explicit scope management and dependency maintenance as design principles for multimodal physics reasoning agents.
Nov 17, 2026

Synergizing Evidence Retrieval and Utilization for Self-Improving Multimodal Scientific Document Reasoning

Scientific document reasoning requires an agent to acquire and integrate evidence distributed across text, figures, and tables. Retrieval can identify relevant sources while missing complementary support, and answer-level feedback alone does not reveal whether an error stems from evidence acquisition or evidence use. We present an agentic Retriever--Perceiver--Reasoner framework for multimodal scientific document reasoning. The Retriever sequentially selects source-aligned evidence under a context budget. Given the selected evidence and question, the Perceiver constructs a question-conditioned latent ontology, an implicit organization of multimodal evidence in a shared continuous space. The Reasoner combines this latent guidance with the original sources to produce grounded answers. We train the three components through diagnostic-guided co-evolution: alternating updates optimize them with a shared answering reward, while Retrieval Potential Feedback (RPF) and Latent Efficacy Feedback (LEF) identify questions that need better retrieval or latent guidance and direct targeted resampling. Across four benchmarks and seven evaluation settings, our framework obtains average scores of 69.9 and 75.0 with 7B and 8B Reasoner backbones, exceeding the strongest baseline by 4.0 and 9.1 points, respectively.
Nov 17, 2026

PACD: Provenance-Aware Consensus Decoding for Mitigating Cross-Modal Hallucinations in Omnimodal Large Language Models

Omnimodal large language models (Omni-LLMs) integrate audio, video, and text for multimodal reasoning, yet remain vulnerable to cross-modal hallucinations, where one modality improperly influences predictions about another. Existing training-free decoding methods mitigate these errors through modality perturbation or relevance-based weighting. However, due to interference from irrelevant modality cues during fusion and difficulty detecting shared errors, these methods may suppress correct unimodal predictions and allow erroneous consensus to persist. To address these limitations, we propose Provenance-Aware Consensus Decoding (PACD), a training-free framework for mitigating cross-modal hallucinations. PACD comprises two complementary paths based on whether the branches reach consensus. For non-consensus cases, Relevance-Guided Candidate Arbitration identifies task-relevant unimodal evidence and selectively arbitrates between predictions using alternative candidates. For consensus cases, Provenance-Based Consensus Recovery leverages the outranking principle to evaluate the consistency and strength of layer-wise modality evidence for alternative answers, selectively resolving erroneous consensus under confidence constraints. Extensive experiments on four Omni-LLMs and five benchmarks across two tasks demonstrate the superiority of PACD. Compared to existing methods, PACD achieves up to 27.6% relative improvement over standard decoding across CMM, AVHBench, and SVHalluc for cross-modal hallucination mitigation, and an average relative improvement of 7.7% over best baseline across OmniVideoBench and FutureOmni for general audio-visual reasoning.
Nov 17, 2026

PURE: A Diagnostic Benchmark for Multi-Granularity Multimodal Identity Matching

Identity matching asks whether two records denote the same identity, yet the answer depends on the target granularity. Two records may match at the base level but not at the variant level: they share an underlying design while differing in attributes such as color or size. Final matching accuracy does not separately characterize the capabilities involved. To study this problem, we use e-commerce as a challenging real-world domain. We introduce PURE, a diagnostic benchmark for multi-granularity multimodal identity matching covering four capability dimensions: Perception, Understanding, Reasoning, and Evidence-grounded Judgment. The benchmark comprises 1,457 multiple-choice questions across 14 tasks, derived from real-world matching pairs and verified by two domain experts. Across ten recent MLLMs, overall scores range from 47.1% to 86.0%. Understanding receives the highest mean score and Judgment the lowest, with weaknesses in visual difference localization and variant-level rule judgment. Modality ablations show that models benefit from multimodal evidence but can incorrectly identify conflicts between matching records. In direct matching, most models recognize matching pairs more reliably than non-matching pairs at the base level, whereas the reverse holds at the variant level. PURE therefore reveals what aggregate matching accuracy hides: not only whether models judge a pair correctly, but where they struggle to interpret evidence and apply matching criteria.
Nov 12, 2026

Asymmetric Multimodal Learning with Decomposed JEPAs

Cross-modal alignment is the standard objective for building multimodal representations. We show that alignment as a sole objective is lossy wherever modalities are informationally asymmetric---where neither can express what the other carries. Visual question answering is an obvious case: a question is a selective relational operator probing a dense visual field whose fine-grained features no query recovers, so what alignment preserves is not what the task requires. We prove an Alignment Redundancy Bound: alignment-only Dirichlet energy minimization retains at most the redundant information atom \Red(Z;X1,X2)\Red(Z; X_1, X_2), annihilating private modality-specific structure \Unq(Z;Xm)\Unq(Z; X_m) and cross-modal synergy \Syn(Z;X1,X2)\Syn(Z; X_1, X_2). We introduce Decomposed JEPA (D-JEPA), which factorizes latents into an aligned shared core and private modality-specific fibers. The shared core retains most of the structure alignment targets; the private fibers recover what alignment discards, via contextual masked prediction and asymmetric cross-subspace independence. Under information asymmetry, alignment-only training undergoes catastrophic semantic collapse on compositional reasoning benchmarks, surviving only on binary queries. D-JEPA averts this collapse, substantially improving downstream accuracy over both alignment-only and alignment-with-reconstruction baselines, and restoring a positive multimodal synergy gap in place of the baseline's inverted gap. Information-theoretic audits confirm that D-JEPA recovers most of the mutual information available from a frozen backbone, and causal fiber-swap interventions demonstrate that the recovered information is causally localized in the private fibers.
Nov 11, 2026

A Unified Interpretable Semantic Framework for Multi-Modal Fiber Sensing Under Data Scarcity

We propose a unified interpretable semantic framework for multi-modal fiber sensing under data scarcity. The framework maps heterogeneous measurements from state of-polarization (SOP) monitoring, distributed acoustic sensing (DAS), and optical time-domain reflectometry (OTDR) into a shared set of physics-grounded semantic descriptors capturing impulsiveness, persistence, structural characteristics, and energy concentration. Inference is performed through reasoning over semantic class profiles, optionally assisted by retrieval of semantically similar examples to resolve ambiguity and support adaptation without model retraining. Experimental validation on real SOP, DAS, and OTDR datasets demonstrates robust performance across sensing modalities, achieving accuracies of 95.0%, 91.1%, and 94.6%, respectively, while providing transparent and physically interpretable decisions. The proposed framework establishes a unified and data efficient approach for optical fiber sensing in evolving and data scarce deployment environments.
Nov 10, 2026

A Cross-Scenario Survey and Analysis of Self-Evolving Multimodal Agents

Recent work shows that multimodal agents increasingly self-improve without recurring human annotation. However, each self-improvement mechanism is designed and evaluated within a single scenario, leaving its preconditions unverified under cross-scenario transfer. This survey investigates the structural source of asymmetric mechanism transfer across four scenarios: multimodal reasoning, multimodal generation, GUI and Web navigation, and embodied manipulation. We first formalize the self-evolving multimodal agent by four criteria and a differentia of self-reference, and adopt the three-stage self-evolving loop of experience generation, feedback acquisition, and update consolidation as our analytical framework to position the scenarios along ceilings on rollout count and feedback fidelity. Comparing self-evolving with fixed loops within each scenario separates the limitations that the scenario imposes from those inherent in the self-evolving design. We then recast each mechanism as a trade-off that raises a scarce factor by presupposing a task-structural resource, such as output equivalence or checkable outcomes, and judge it portable, conditional, blocked, or redundant across scenarios. We attribute the asymmetry to divergent combinations of scarce factors and resources across opposite transfer directions, and identify the matched-budget measurements that could overturn these still untested verdicts.
Nov 9, 2026

Layer Roles Matter: Mitigating Modality Imbalance in Multimodal Emotion Reasoning via Adaptive Layer Profiling

Multimodal Emotion Reasoning (MER) requires perceiving and integrating emotion cues from multiple modalities. Recent advances increasingly leverage Multimodal Large Language Models (MLLMs) to enhance reasoning capacity. However, due to the inherent heterogeneity in conveying emotions across modalities, these models often suffer from modality imbalance and confront flawed emotion understanding. In this paper, we conduct an in-depth analysis on MER models and uncover two key empirical insights. (i) In MER, modality imbalance manifests two distinct failure modes: unimodal over-reliance and insufficient cross-modal utilization, which call for different intervention strategies. (ii) MLLM layers exhibit implicit functional hierarchies in multimodal emotion processing, playing separable roles in stabilizing unimodal representations and integrating cross-modal information. Motivated by these observations, we propose a Layer-Aware Modality Rebalancing (LAMR) framework by explicitly aligning intervention strategies with layer roles. Specifically, we first introduce two quantized metrics for adaptive layer role profiling. In each layer, Modality Entropy (ME) measures the dispersion of modality contributions, and Cross-modal Intensity (CMI) captures the strength of cross-modal interactions. Then, we respectively enhance modality equilibrium in stabilization layers and cross-modal propagation in integration layers. Extensive experiments on two comprehensive MER benchmarks demonstrate that LAMR effectively mitigates modality imbalance and improves emotion reasoning across diverse MLLM backbones, without additional training.
Nov 8, 2026

DISCOVER: Online Variance-Guided Data Discovery for Budgeted Multimodal GRPO

Recent advances in reinforcement learning have made verifiable rewards a central post-training paradigm for improving multimodal reasoning. However, methods such as GRPO remain data- and compute-inefficient, spending rollout and annotation budget on prompts that provide little learning signal. We propose DISCOVER, an online data-discovery framework for budgeted multimodal GRPO that enables vision-language models to dynamically identify prompts that are most useful for the current policy. DISCOVER first organizes the multimodal sample pool into semantic clusters and queries a small set of density-diverse representative samples to obtain initial reward observations. During training, it estimates the utility of unannotated samples by transferring sparse reward feedback from queried samples. Since sample utility changes as the policy evolves, DISCOVER dynamically updates these estimates with recency-weighted rollout observations. It then selects only high-utility prompts for annotation using a policy-conditioned score that accounts for reward variance and difficulty. Once the discovery budget is exhausted, DISCOVER continues with utility-based replay. Experiments on ViRL39K and VLAA-Thinking with Qwen3-VL-2B show that DISCOVER substantially improves multimodal mathematical reasoning under a strict 10% unique-sample budget, outperforming budget-matched random selection, hard-example sampling, full-pool GRPO under the same training-step budget, and online selection baselines. Ablations further validate the importance of semantic coverage, online reward transfer, and stable discovery cadence, establishing DISCOVER as a practical framework for data-efficient multimodal RL post-training.
Nov 7, 2026

EndoSCOP-V: A Multi-Turn Video Understanding Evaluation Framework for Multimodal Models in Endoscopy Reporting and Clinical Reasoning

Endoscopic diagnosis unfolds over native video, where a single positive finding during traversal of the gastrointestinal tract sets off a multi-minute chain of reasoning over location, morphology, severity and intervention. Existing endoscopy multimodal benchmarks test only static images or single-disease videos, which miss the temporal and clinical reasoning that real-world reporting demands. We introduce EndoSCOP-V, a video-grounded multi-turn evaluation framework for structured endoscopy reporting and clinical reasoning. It contains 1,307 real-world endoscopy and colonoscopy videos (63.2 h; 97% 1080p) averaging about 174 s. Its 20,274 question-answer turns are instantiated deterministically from clinician-filled structured endoscopy-reporting EHRs, so no per-case annotation is synthetic or LLM-rephrased. They come from 1,033 clinician-reviewed templates, about 18 per disease, that span a clinically aligned 9-task x 30-sub-task reporting taxonomy over 54 World Endoscopy Organization (WEO) MST 3.0 diseases. 22% of procedures carry co-existing pathologies, up to six per case. A Markovian context-pruned oracle-forcing protocol evaluates the dialogue turn by turn, injecting the gold answer before each subsequent turn so that per-turn reasoning is isolated from cascading dialogue errors. A panel of 15 board-certified endoscopists sets a 78% strict-accuracy baseline on 1,500 questions. Across 14 open-weight LMMs, the best 27B-class model reaches 59.8% strict accuracy, falling to 53.4% once format-trivial disease-rejection items are excluded. On a between-subjects spot-check, the best of 3 frontier proprietary models trails the clinician panel by 23 pp on strict NDR accuracy. Sub-task analysis reveals consistent gaps in quantitative measurement, multi-variable grading and intervention reasoning on long-horizon video. Data: Hugging Face (EndoSCOP-V). Code: GitHub (abhi25t/EndoSCOP-V).
Nov 7, 2026

Mental Imagery That Matters: Curing Latent Collapse in Multimodal Reasoning

Mental simulation, the ability to reason by constructing and updating internal visual states, is central to flexible human cognition, but remains elusive in current multimodal large language models. Latent multimodal reasoning offers a promising route toward this capability by interleaving text with continuous visual latent blocks, allowing models to maintain internal visual workspaces without decoding them into pixels. In this paper, we ask whether these latent visual states actually matter. Through a direct perturbation analysis across representative latent-reasoning MLLMs, we find that they often do not: dropping, replacing, or shuffling generated visual latents has only minor effects on the model's own final-answer probability, while equivalent perturbations to the linguistic reasoning trace are substantially more damaging. We call this failure mode Mental Simulation Collapse (MSC): the model appears to perform interleaved visual reasoning, but routes most answer-relevant computation through text. We further diagnose that MSC follows naturally from existing training objectives: visual alignment can make generated latents resemble helper-image embeddings, but it does not require the decoder to rely on them; subsequent text-only supervision further leaves the semantic role of the latent pathway unidentifiable. To address this, we introduce AnchorSim (Anchored Mental Simulation), a contrastive distillation framework that trains models to make their generated visual latents consequential. Across multiple visual reasoning benchmarks, AnchorSim improves performance, while diagnostic re-evaluation shows that visual perturbations now meaningfully affect predictions. Our results expose a fundamental failure mode of latent multimodal reasoning and demonstrate that mental simulation must be anchored to model behavior in order to matter.
Nov 6, 2026

Seeing vs. Counting: Complexity Twins Show Where Multimodal LLMs Fail on Diagrams

When a multimodal LLM miscounts in a diagram, does it fail at seeing or at counting? Existing visual graph and puzzle benchmarks vary visual density and algorithmic hardness together, confounding the two. We pair counting questions about the same drawing that sit in different counting-complexity classes, such as its spanning trees (a determinant) and its simple paths (#P-complete), and use a transcription probe and a text-only control to separate reading from computing. We test Gemini, Gemma and Qwen3-VL models; for the Gemini models, the answer depends on how the model computes. In chain of thought, most of them read small drawings correctly and then fail at counting, at a rate set far more by the size of the answer than by its complexity class (for the primary model, the odds of an exact answer fall 2.38-fold per decade of the answer). Given code and an exact transcription, the primary model counts correctly even on 10x10 grids, and reading becomes the bottleneck: a single view misses horizontal edges, and a second reading of a rotated copy with crop checks reads 86/96 large grids exactly, against 57/96 when the upright reading is simply repeated.
Nov 6, 2026

AIM: Adaptive Interaction in Multi-Agent Debate for Multimodal LLM Inference

Multi-Agent Debate (MAD) improves reasoning on multimodal tasks, but its accuracy and token efficiency depend on how agents interact during debate. A largely overlooked design choice is which interaction modalities agents use to exchange evidence. Existing methods rely on fixed text- or graph-based interaction, which fails to convey fine-grained multimodal evidence and thus degrades accuracy while inflating token cost due to verbose descriptions. To address this, we propose Adaptive Interaction in Multi-Agent Debate (AIM), an adaptive routing framework that dynamically selects the optimal modality combination for inter-agent interaction in multimodal MAD. Beyond text and graph, AIM introduces task-dependent modality interaction, where agents exchange targeted regions of interest from the task-specific multimodal input, such as image regions, audio segments, or spatio-temporal video clips. To determine the best combination for each instance, AIM first generates a structured baseline response and extracts interpretable routing features that capture which evidence the instance requires, how confident the baseline agent is, and how much new evidence each modality would add beyond the baseline response. A lightweight router then uses these routing features to select a modality combination that avoids both under-fusion (i.e., agents miss decisive evidence) and over-fusion (i.e., redundant modalities inflate token cost) risks. Across eight benchmarks spanning three multimodal task families, AIM achieves the highest accuracy. Combining modalities improves accuracy by up to 6.4% over the best single-modality baseline, and adapting the combination per instance adds up to 4.2% over the most accurate fixed combination of each benchmark with up to 53.4% fewer tokens.
Nov 4, 2026

Trick-or-Treat: Adversarial Evaluation of Multimodal Medical Reasoning with Synthetic Images

We used medical image synthesis to create a challenging multimodal evaluation framework that exposes hallucination behaviors in frontier models. Evaluating multimodal reasoning has historically been difficult because existing medical datasets rarely pair different image modalities, such as radiographs and external clinical photographs. To address this gap, we introduce Trick-or-Treat, an adversarial multimodal medical benchmark that pairs real chest radiographs with synthetic clinical photographs to challenge models' cross-modal reasoning. For each of a wide range of conditions, we created two pairs of images, where the pairs share a single chest radiograph but contain different external photographs. In matched pairs, the chest radiograph and photograph both contain findings suggesting the same disease (e.g., varicella pneumonia suggested by lung opacities and crusted skin lesions seen in chickenpox). In mismatched pairs, the photograph depicts an alternative clinical finding that would not lead to the same unifying diagnosis (e.g. replacing crusted skin lesions with ordinary acne). Synthetic photograph generation allows clinical findings to be varied while the person's features and photographic setting stay consistent, reducing confounding. In an investigation of three frontier models, we found that models hallucinated clinically significant photograph findings in 45-57.5% of mismatched pairs, up from 25-37.5% for the same photographs shown alone. When faced with mismatched pairs, models tended to ignore mismatches and instead hallucinated findings that better matched the radiograph (e.g. misreading acne as crusted skin lesions to justify diagnosing chickenpox). Trick-or-Treat both demonstrates concerning error pattern in frontier models' medical reasoning and provides a proof of concept for the use of synthetic images in adversarial multimodal evaluation.
Nov 4, 2026

A Systematic Benchmark for Evaluating Multimodal Personalized Reasoning of Vision-Language Models

People often make decisions by holistically reasoning over heterogeneous forms of personal information. In this paper, we study the relatively underexplored capability of multimodal personalized reasoning in multimodal large language models (MLLMs). To this end, we introduce MPRBench, a novel and comprehensive benchmark specifically designed to evaluate personalized reasoning across a wide range of tasks and real-world challenges. MPRBench consists of 12 sub-tasks and additionally supports interpretable analysis of representative error cases. Through extensive experiments, we find that current personalized MLLMs still struggle with personalized reasoning due to several key challenges, including inaccurate recognition of personalized concepts and sensitivity to irrelevant personal information. We hope that MPRBench will stimulate further research on personalized reasoning in MLLMs. Code and data are available at https://github.com/kkdh15/MPRBench
Nov 4, 2026

Reasoning-Aware Cross-Modal Association Unlearning in Multimodal Large Reasoning Models

Multimodal large reasoning models (MLRMs) integrate visual inputs with textual and parametric knowledge, allowing sensitive information to be accessed through cross-modal associations during reasoning. This raises privacy concerns when a visual input should no longer provide access to such information. Machine unlearning can remove this influence without retraining the model from scratch. However, multimodal unlearning methods mainly target samples, concepts, or knowledge content, without separating the knowledge from the cross-modal association through which it is accessed. In reasoning models, target information may disappear from the final answer while remaining in intermediate reasoning, or be replaced by incorrect substitute knowledge. We therefore study cross-modal association unlearning, where the forget target is a designated visual-to-knowledge association: the model should suppress target-bearing behavior under the designated visual context while preserving access to the same knowledge in permitted contexts. We propose Visual-Keyed Counterfactual Rebinding (VKCR), a parameter-efficient framework for cross-modal association editing. VKCR first localizes the edit to a compact association-conditioned parameter subspace. It then reduces support for the designated target at task outputs and counterfactually rebinds the affected reasoning behavior by discouraging target-bearing or unsupported continuations while favoring valid non-sensitive alternatives. We further introduce CLEAR-R and MLLMU-R to evaluate target leakage, hallucinated reassociation, reasoning preservation, and modality-specific utility. Experiments show that VKCR substantially reduces image-conditioned target leakage and hallucinated reassociation, including across heterogeneous task formulations, while preserving non-sensitive visual reasoning and permitted text-only knowledge.
Nov 3, 2026

VisSearch-Bench: Evaluating Long Horizon Visual Search in Multimodal Agents

Multimodal agents use web and image tools to answer questions beyond their internal knowledge. Some search tasks use images mainly to identify an initial entity, then rely on text retrieval. These tasks offer limited insight into how agents connect information across images, objects, and search steps. We therefore focus on visual context management. This involves organizing and using visual information and linking it to other findings during search. Visually intensive tasks require agents to combine object appearance, local details, or spatial relations with external information. We introduce VisSearch-Bench, with 204 human-written tasks across 7 main domains and 25 subdomains. It covers Set-wise Visual Search, Sequential Visual Search, Hybrid Wide-Deep Search, Competitive Visual Disambiguation, and Visual-Structure-Guided Search. Each task provides a reference answer, traceable evidence, and scoring rules. Accuracy measures complete task success, while Process Score measures the correctness and completeness of individual results and subquestion answers. Across 16 multimodal models, the best system fully solves only 26.47% of tasks, even with search tools. These results highlight the challenge of linking visual evidence with search results to produce complete answers.
Oct 30, 2026

Multimodal In-Context Learning with Explicit Rule Induction

A central challenge in multimodal in-context learning (MICL) arises from tasks that are ill-defined without context and require uncovering the underlying rules from demonstrations. Existing methods predominantly treat MICL as direct input-output mapping, overlooking the importance of explicitly inducing and reasoning over such rules, limiting both performance and interpretability. We propose ERIM, Explicit Rule-Induction for Multimodal In-Context Learning, a unified framework that enhances MICL by explicitly inducing the latent rules governing demonstration samples. Leveraging the reasoning ability of multi-modal large language model (MLLM), ERIM constructs rule-centric MICL datasets, introduces explicit supervision for rule induction, and fine-tunes MLLM to perform explicit rule-induction via a two-stage training strategy that combines mixed cold-start supervised fine-tuning with reinforcement learning. Extensive experiments show consistent performance improvements across diverse MICL settings and demonstrate that correctly inducing underlying rules directly leads to stronger and more stable performance.
Oct 29, 2026

Diagnosing the Reference Gap: What Visual Scaffolds Can and Cannot Fix in Multimodal Reasoning

Multimodal large language models can recognize visual content yet struggle to reason reliably over it across multiple steps. In counting, a model may see all objects but lose track of which ones it has counted. We argue that such failures, commonly attributed to perception, reveal a reference gap: the failure to bind language to the relevant visual entities, regions, or locations. Motivated by the theory of visual routines, we test whether external references such as numbering can narrow this gap. We introduce ReGap-Bench, a paired benchmark of five visual reasoning tasks in which the scene, question, and answer are fixed and only the reference varies. Results show that task-matched references improve accuracy by up to 25%, while mismatched references provide little benefit or even hurt performance. These findings identify the reference gap as a bottleneck distinct from visual perception and show that constructing task-appropriate visual references can serve as a training-free visual skill for improving multimodal reasoning.
Oct 29, 2026

Grounding Multimodal Reasoning with Evidence-Ablated Negatives

Reinforcement learning with verifiable rewards (RLVR) has become a powerful tool for improving multimodal reasoning. However, standard outcome supervision suffers from credit assignment ambiguity, as a model may receive positive rewards by exploiting language priors or dataset shortcuts rather than grounding its reasoning in the requisite visual evidence. This makes high-quality negative responses critical, since effective negatives should remain plausible under the current policy while exposing the shortcut that should be suppressed. In this paper, we propose EVAN, a multimodal reinforcement learning framework utilizing EVidence-Ablated Negatives to train LMMs against their own language-prior failures. EVAN first identifies highly visually dependent examples using a scaled image information gain metric and preserves data diversity with a mixed-ranking strategy. It then constructs evidence-ablated images by masking the most discriminative image patches, eliciting plausible but visually ungrounded responses from the current policy. During policy optimization, these negative responses are incorporated as bounded auxiliary signals under the clean image, penalizing shortcut learning while maintaining RL stability. Experiments across challenging multimodal reasoning benchmarks demonstrate that EVAN mitigates modality imbalance and improves over strong RLVR baselines. Our code is available at https://github.com/DawnLIU35/EVAN.git.
Oct 28, 2026

Leaving, Staying, or Living There? Duration Commonsense Reasoning with Human-LLM Co-annotations

Gardening in the backyard implies living there, while checking into an Airbnb more likely means staying for a few days. Such duration judgments are important for understanding human mobility, travel behavior, and social context. However, duration is rarely stated explicitly in social media posts and must instead be inferred from noisy textual, visual, and contextual cues. In this paper, we investigate whether vision-language models possess such duration commonsense. We create a new corpus of geotagged tweets co-annotated by humans and LLMs to serve as a testbed for this task. We find that state-of-the-art models struggle with duration inference, even when guided by multi-turn Chain-of-Thought prompting. Error analyses further reveal that models often fail to integrate critical contextual information, or they overly rely on prominent visual details while overlooking subtle hashtags and semantic nuances. These findings underscore the significant challenges remaining in context-aware multimodal understanding.
Oct 28, 2026

Alpha-TCM: Auditable Knowledge Integration for Multimodal TCM Inference

Knowledge-guided multimodal systems need a way to measure how supplied knowledge affects predictions, including when that knowledge is wrong. Alpha-TCM addresses this problem for traditional Chinese medicine (TCM) through an implemented workflow that separates measurements, supplied knowledge, predictions, and reviewer decisions. Observations and contextual knowledge enter separately, and a replaceable prior allows controlled interventions with fixed modality encoders. Bayesian committee accounting removes the reference prior already contained in each encoder before inserting contextual knowledge. A fully simulated confirmation study uses ten paired seeds and four training sizes. Knowledge context reduces prediction error relative to knowledge-free fusion at 600 and 1,200 training cases; equal-information learned context is competitive at 600 and has lower error at larger sizes. Controlled corruption identifies overconfident and displaced priors as the most damaging challenges. Case-level conflict detection remains weak, and training-time prior regularisation shows no consistent benefit. Functional tests verify the connected workflow separately from predictive performance. The simulation demonstrates how Alpha-TCM measures knowledge dependence, separates it from gains due to additional context, and exposes harmful prior assumptions before clinical evaluation.
Oct 27, 2026

Beyond the Readout: A Latent Numerical Frame Shared Across Modality and Operation

When a multimodal model computes 17+35 from text or 91-39 from a rendered image, it produces the same answer, 52. Does it use one internal concept of number for both, or separate machinery for each path? Similarity metrics such as RSA or CKA cannot settle this question. They give the same score to identical and to rotated directions, and two representations can look alike only because both prepare the same output digits. We therefore use causal interventions. For four conditions (text or image, addition or subtraction), we learn a small subspace of Gemma4-12B-it's residual stream in which swapping coordinates swaps the model's answer. The subspaces answer three questions. First, do the conditions write numbers along the same directions? They do not. Reusing a subspace from text in an image on the same operation recovers only 53.2% of its causal effect. A single rotation and rescaling maps one subspace onto the other and raises recovery to 91.4%, so the conditions share one geometry expressed in different coordinates. Second, does this shared geometry only reflect the digit the model is about to print? To test this, we split each subspace into 9 directions that the output layer reads to choose the next digit and 13 directions that it cannot see. The 13 hidden directions still carry the number, yet changing them barely alters the digit being written. Third, what are these hidden directions for? A model must emit a multi-digit answer one digit at a time, so it can compute the full result early and keep it in reserve. When two numbers share their first digit, swapping the number in the hidden directions leaves the digit being written intact and redirects the next digit in 77.6% of cases. The model thus forms an abstract numerical concept upstream and unfolds it across time: multimodal representations share meaning through relational geometry across distinct coordinate frames.
Oct 26, 2026

Hierarchical Multimodal Interaction Reasoning for Dyadic Relationship Recognition

Dyadic relationship recognition aims to identify the latent social relationship between two individuals from their multimodal interactions. Existing methods typically project heterogeneous cues into a shared representation and optimise a flat classification objective, overlooking that relationships may rely on different interaction patterns. For instance, distant relationships can often be inferred from direct behavioural correspondence, whereas close relationships are better reflected by coordinated affective, behavioural, and multimodal responses. To address this limitation, we propose HiRe-DRR, a Hierarchical Multimodal Interaction Reasoning framework that organises relationship evidence from three complementary perspectives: feature level, interaction type, and relationship hierarchy. HiRe-DRR decomposes facial, bodily, acoustic, and textual cues into low- and high-level representations and constructs homogeneous and heterogeneous cross-participant interactions. We further introduce the Mixture-of-Interaction Transformer (MoI-Former), which treats diverse interactions as specialised experts and employs relation-aware sparse routing to identify relationship-relevant evidence while preserving contextual interaction information. The proposed Hierarchical Conditional Loss (HC-Loss) exploits the hierarchical nature of social relationships through coarse-to-fine supervision, enabling structured relationship reasoning. Extensive experiments on three dyadic relationship benchmarks demonstrate that HiRe-DRR outperforms recent representative image-, video-, and multimodal methods. Our routing analysis further reveals that distant relationships tend to rely on low-level homogeneous interactions, whereas closer relationships focus on high-level heterogeneous interactions. Our code is publicly available at: https://github.com/tw-repository/HiRe-DRR.
Oct 23, 2026

Beyond Post-Hoc Rewards: Privileged Reasoning Exploration for Multimodal Embeddings

Advances in Multimodal Large Language Models (MLLMs) have driven substantial progress in multimodal embeddings. Recent methods exploit generative capabilities by augmenting embedded inputs with self-generated reasoning to produce more informative embeddings. However, conventional retrieval supervision for training embedding models specifies pairwise relevance relations without explicitly guiding the reasoning used for input augmentation. Such reasoning is neither directly observed nor uniquely determined by the input, allowing multiple plausible traces to introduce semantic cues of varying retrieval utility. Existing approaches largely optimize reasoning conditioned on the input alone, limiting the exploration of retrieval-relevant traces. We introduce PILOT, a framework that models reasoning as a latent variable and instantiates its prior and posterior distributions as distinct reasoning modes of a shared MLLM. The prior conditions only on the input, whereas the posterior uses additional training-time relational evidence to guide reasoning exploration. To translate the posterior's broader exploration into inference-time improvements, we derive an importance-weighted variational objective and instantiate it as a prior-posterior co-evolution framework that improves both modes under a shared inference-compatible reward while selectively transferring posterior trajectories to the prior according to their utility and prior reachability. On the 78-task MMEB-V2 benchmark, PILOT outperforms the primary baseline across all 12 task groups at both 2B and 7B parameters, improving overall scores by 3.2 and 3.3 points, respectively, with the largest gains in visual-document retrieval (+6.1 and +6.4 points). Our results show that relational supervision is most valuable not as a post-hoc reward but as a signal for shaping how reasoning is explored in training reasoning-enhanced embedding models.
Oct 23, 2026

PEAR: Adaptive Advantage Reshaping for Perception and Exploration in Large Vision-Language Models

Effective multimodal reasoning requires a model to reliably utilize visual information while sufficiently exploring diverse reasoning paths. To achieve this goal, prior work combines visual dependency and token entropy to reshape sequence-level advantages, but scaling the contribution of token entropy by visual dependency can leave high-entropy tokens with low visual dependency underweighted, despite their important role in reasoning exploration. To address this limitation, we propose PEAR, an adaptive advantage reshaping method for improving perception and exploration in multimodal reasoning. PEAR retains visual dependency as the base signal and adds an entropy-based exploration gain only at high-entropy positions, with lower visual dependency leading to a larger additional gain. The fused signal is then mapped to positive token-level weights that reshape the sequence-level advantage. Experiments on Qwen2.5-VL-3B/7B show that PEAR achieves the highest average performance among evaluated baselines across multimodal reasoning benchmarks, alongside leading results on textual reasoning and general multimodal benchmarks.
Oct 21, 2026

MOSAIC: An Omni-Modal Benchmark for Long-Term Memory and Cross-Modal Reasoning in Conversational Agents

Omni-modal assistants must remember not only what users say, but also what they show and convey. Existing long-term memory benchmarks are mostly text-and-image or report a single end-to-end score, so they cannot locate whether evidence is lost when it is represented, retrieved, delivered, or read. We introduce MOSAIC, a benchmark of 923 questions over persona-grounded histories of 11-24 sessions spanning text, image, video, and audio. MOSAIC constructs targeted spatial, temporal, and prosodic signals rather than decorative media, with the strongest causal validation for spatial information and selector-level validation for annotated prosody, and pairs them with a matched diagnostic protocol that freezes the question, reader, prompt, and budget while intervening on representation, retrieval, and delivery. Matched decomposition shows that representation loss is modality-dependent, residual reading dominates even under gold evidence, and abstention remains weak on these controls: strong answerable scores coexist with failures to refuse. MOSAIC thus turns omni-modal memory evaluation from a leaderboard into a diagnosis of write, retrieve, deliver, and read failures.
Oct 20, 2026

TriGeoBench: A Multimodal Mathematical Geometry Benchmark with Figures and TikZ

Multimodal Large Language Models show strong vision-language capabilities, yet their geometry-centric reasoning remains insufficiently understood. Existing mathematical benchmarks typically rely on rasterized figures and cover limited geometric knowledge areas and response formats, making it difficult to assess whether models reason over geometric structures or exploit superficial visual patterns. We introduce TriGeoBench, a geometry-centric multimodal benchmark constructed from Chinese middle- and high-school mathematics examinations. TriGeoBench contains 7,000 question-answer pairs and 12,235 geometric figures with TikZ code, spanning plane geometry, solid geometry, analytic geometry, function, and vector geometry across diverse response formats. Each instance forms a {question-answer pair, image, TikZ} triplet: question-answer pairs are stored in LaTeX, images as RGB figures, and TikZ preserves geometric topology, symbolic annotations, and relational structure. We evaluate representative open- and closed-source MLLMs through image-to-TikZ and multimodal reasoning probes. Results reveal persistent fine-grained structure and image-text alignment challenges. Oracle experiments with human-curated reference TikZ suggest that accurate executable representations provide complementary geometric cues; they do not evaluate an end-to-end conversion pipeline. TriGeoBench offers a testbed for structure-aware mathematical reasoning.
Oct 20, 2026

AudiPsy: Evidence-Grounded Multimodal Reasoning over Spoken Counseling Conversations for Depression Assessment

Depression assessment from counseling speech requires integrating textual and acoustic psychological evidence across turns, yet existing end-to-end classification approaches lack explicit multimodal evidence interpretation and context-aware reasoning. We propose AudiPsy, an evidence-grounded omni-LLM for depression assessment over spoken counseling conversations. AudiPsy employs an Observation-Assessment-Summary (OAS) paradigm to integrate dialogue transcripts, textual and acoustic emotions, and textual-acoustic emotion incongruence into structured reasoning. OAS incrementally organizes multimodal observations, performs contextualized assessment, and maintains an evolving dialogue summary for subsequent turns. To support this framework, we construct AudiPsy-Counsel, a synthetic spoken counseling conversation corpus with multimodal context augmentation, and develop a unified three-stage curriculum for depression-oriented reasoning. Experiments on DAIC-WOZ demonstrate superior depression assessment performance over specialized depression detection methods and general-purpose speech/omni-LLMs, along with improved reasoning quality, while supplementary evaluation on EATD-Corpus supports AudiPsy's applicability to naturalistic conversations.