multimodal-reasoning

The paper introduces ReImaGin, a framework that uses image generation models as a flexible visual reasoning mechanism for multimodal LLMs, enabling open-ended visual operations like removing occlusions or generating floor plans from disjoint views, outperforming text-only and specialist vision-tool baselines by up to 25% across six visual reasoning tasks.

CUSP is a training-free framework for quantifying collective uncertainty in multi-agent multimodal systems using semantic opinion pooling, improving reliability and accuracy over baseline methods.

The paper introduces Causal Visual Recurrent Reasoning (CVRR), a method that enforces recurrent hidden-state computation for visual reasoning, improving performance on benchmarks while distinguishing latent informativeness from actual predictive use.

Aphanta is an automated diagnostic framework that evaluates the utility of image-edited intermediates in multimodal reasoning pipelines, showing task-dependent benefits and improving performance on specific tasks.

This paper introduces a benchmark to study how large language models arbitrate conflicting evidence from text and numerical sources, finding that models use heuristic strategies with biases towards recency and external tools.

This paper identifies a privacy vulnerability in RL-trained multimodal large reasoning models, which can leak sensitive facts in their reasoning traces even after unlearning, and proposes LEMUR, a training-free inference-time framework that uses entropy dynamics to detect and suppress such leakage.

This paper introduces Visualized Task Semantics (VTS), a controlled intervention to study when multimodal LLMs are asked questions in the image rather than text. It reveals a consistent accuracy drop across all tested models and tasks, and proposes prompt-region grounding to recover the clean task representation, improving VTS accuracy from 58.0 to 66.3.

This paper introduces Reflective Retrieval Memory (RRM), a memory framework that distills procedural retrieval experience from historical task trajectories to improve evidence retrieval for long-horizon multimodal reasoning. RRM matches or exceeds prior state-of-the-art on M3-Bench-Robot, M3-Bench-Web, and Video-MME-Long benchmarks.

MAGE is a multimodal multi-agent framework that improves macro placement in VLSI physical design by combining structured floorplanning rules, visual checks, and iterative refinement, outperforming commercial placers and human experts on timing metrics.

S1-Omni is a unified multimodal reasoning model for scientific tasks including understanding, prediction, and generation. It is trained on a corpus of 200 scientific tasks and outperforms GPT-5.5 and Gemini-3.1-Pro on most benchmarks.

This paper introduces MAR-12, a framework using Vision-Language Models and multi-angle reasoning to detect and explain harmful humor in memes, achieving state-of-the-art accuracy on PrideMM and Memotion datasets.

This paper investigates test-time scaling techniques for small open vision-language models (≤7B parameters) on the multilingual visual MCQ benchmark EXAMS-V, finding that inference budget and parseability matter more than complex search or verification methods. The best configuration achieves 84.1% on the ImageCLEF 2026 test split, ranking first on the leaderboard.

Proposes Asymmetric Mutual Variational Learning (AMVL) to resolve train-inference mismatch in multimodal continuous reasoning by using bidirectional calibration to prevent answer leakage and improve latent-space stability, achieving significant gains on the BLINK benchmark.

Proposes Mixture of Debaters (MoD), a framework using Mixture-of-Experts to enable dynamic self-debate within a single LLM, achieving superior accuracy with drastically lower latency and token consumption.

CogniRoute is a schema-guided Mixture-of-Experts framework for social video question answering that improves multimodal reasoning through cognitive schema factorization and route-aware reinforcement learning. It achieves significant gains over baselines on the new OmniSocialBench benchmark.

This paper introduces MathVis-Fine, a framework for fine-grained visual dependency modeling in multimodal mathematical reasoning, along with a new dataset and a two-stage progressive training paradigm that balances answer correctness and visual grounding rewards based on each sample's intrinsic visual dependency level.

FinAcumen is a framework that accumulates reasoning experience from prior trajectories into a persistent memory bank for financial multimodal reasoning, improving performance across four benchmarks while maintaining a frozen 8B vision-language model.

VeriGeo introduces a controllable geometry question generation framework that uses verification-guided reflection to ensure numerical and analytical consistency. The method produces high-quality synthetic data, achieving state-of-the-art results on GeoQA and strong performance on PGPS9K and MathVista-GPS.

This paper introduces Multimodal Multi-Dimensional Scalarization Process Reward Modeling (MMS-PRM), which enforces the worst dimension's robustness in multimodal reasoning to prevent failures like visual hallucinations from being masked by strong text logic.

Proposes SpecFlow, a lightweight multimodal spatial reasoning framework that represents intermediate visual thoughts in a fixed-size discrete cosine space, reducing computation and KV cache costs by up to 2.1 times while maintaining competitive performance.