Tag
OpenVLHarness 是一个多模态推理框架，通过能力层、持久化多模态记忆与编码接口实现异构工具组合，在 22 个基准上相较无工具基线平均提升 9.7-12.6 分，同时显著降低 token 成本。
一项受控研究比较了视觉语言模型中的在线策略蒸馏（OPD）与过滤式监督微调（SFT），发现OPD仅在教师模型于学生生成轨迹上保持可靠时才有效，并且会继承教师的失败模式（如不完整回答），甚至可能抹去SFT阶段学到的完成行为。
CoWorld Reasoner 是一个利用视频世界模型作为灵活视觉推理引擎的智能体框架，通过引导视觉想象和工具协同进化，使视觉-语言模型在七个时空推理基准上取得显著提升（如 MMSI 相对提升 11.8%、ERQA 提升 13.0%）。
ThinkDrag提出了一种统一的多模态框架，将视觉推理引入基于拖拽的图像编辑，通过语义变换和显式推理轨迹解释用户意图，并发布了DragBench++基准测试，实现了最先进的编辑效果。
论文提出 MOTIVE 框架，通过多视角自验证学习可靠性分数，在多模态推理中做出接受或重新思考的决策，从而提升视觉语言模型答案的可靠性并减少不必要的推理轮次。
This paper evaluates how Vision Language Models such as ChatGPT, Claude, Gemini, and LLaVA handle VQA questions that violate Grice's maxims through non-essential, ambiguous, or false modifiers, showing diminished performance and differences in pragmatic reasoning compared to humans.
A research paper proposing a text-centric post-training paradigm for Omni-LLMs, where text-only reasoning training drives most gains (boosting Qwen2.5-Omni-7B's reasoning scores by 25.83% with 56.6% fewer GPU-hours) and reduced-data native audio-visual RL restores perception, retaining 93.5% of reasoning gains at ~90% fewer input tokens.
The paper introduces VisKG, a reinforcement learning framework that trains vision-language models to translate images into question-specific knowledge graph representations, improving VQA reasoning efficiency while reducing token usage. A proposed GDPO objective outperforms GRPO-trained counterparts by an average of 2% accuracy across science, math, and general visual reasoning benchmarks.
OmniReasoning introduces OmniReasoningBench, a data engine (OmniQA) producing SFT and RL training sets, and a Modality-Factored Self-Distillation method, yielding a 30B-A3B model that outperforms Qwen3-Omni-30B-A3B-Thinking by double-digit margins on audio-visual joint reasoning benchmarks.
The paper introduces the first unified benchmark for personalized image generation from rich user histories and proposes PEARL, a method coupling a multimodal reasoner with a frozen image generator in an interleaved reason-reflect loop, achieving 15% average improvement on personalization metrics.
LoopVL extends Loop Transformers to vision-language models, iteratively updating a unified multimodal state through shared modules (Module-Loop and Model-Loop) and outperforming similarly sized non-recurrent models on understanding and reasoning benchmarks, with notable 'Visual Aha Moments' in attention.
This paper proposes Scaffolding Minds, which learns a dedicated scaffolding encoder for optimized latent visual target representations and learns both mean and variance of the RL sampler, improving latent reasoning baselines by +9.5 points on Frozen Lake spatial planning and +5.6 points on average across nine visual reasoning benchmarks.
WM-VLM 论文为预训练 VLM 增加轻量级世界模型分支，通过两阶段训练生成中间视觉状态以支持语言与视觉双通道推理，在 2D/3D 心智旋转任务上较监督微调基线最高提升 39.25 个百分点。
This paper introduces Decoupled Credit Self-Distillation (DCSD) to separate credit direction and magnitude for self-distillation, achieving improvements of 8.45 points in mathematical reasoning and 7.01 points in multimodal reasoning across 11 benchmarks.
This paper introduces SciGen-Verifier, a multimodal reasoning model for explainable verification of scientific image generation, along with a new benchmark and a reinforcement learning training pipeline.
The ICDAR2026 competition on multimodal reasoning over documents presented results from 8 teams, highlighting advanced systems that use structured evidence extraction and multi-component orchestration for VQA tasks across diverse document domains.
Uni-LaDiR introduces a unified latent diffusion framework for multimodal reasoning, mapping modality-specific thoughts into a shared latent space and using diffusion to generate reasoning steps, achieving improved performance on vision-language benchmarks.
This paper proposes PIVOT, a dual-level learning framework that enhances visually-grounded reasoning in large vision-language models by using self-calibrated experience replay and vision-guided advantage allocation to optimize reinforcement learning.
OPD-Aha introduces a novel on-policy distillation method that leverages visual preference to correct hallucinations in multimodal reasoning, achieving consistent improvements on various benchmarks.
This paper introduces TestHallVQA, a multi-image VQA benchmark for evaluating Large Vision-Language Models' document-level reasoning under redundant contexts, and proposes a new metric F1-R2 to quantify computational reasoning and robustness.