arXiv:2510. 17045v2 Announce Type: replace-cross Abstract: Video reasoning using Large Multimodal Models (LMMs) relies on costly reinforcement learning (RL) and verbose chain-of-thought, resulting in substantial computational overhead during both training and inference.
arXiv:2608.20814v1 Announce Type: new Abstract: Though Multimodal Large Language Models (MLLMs) have shown impressive potential in video understanding, long video understanding (LVU) remains challeng...
The paper introduces Activation Replay, a training‑free method that improves reasoning in post‑trained large multimodal models (LMMs) by replaying low‑entropy activations from the base model’s input context. It shows that Reinforcement Learning with Verifiable Rewards (RLVR) shifts low‑entropy activations and that modulating these activations enhances reasoning across tasks such as mathematics, visual agents, and video reasoning. Experiments demonstrate that Activation Replay outperforms alternatives like high‑entropy replay or direct cross‑model intervention, boosting Pass@K and broadening RLVR’s reasoning coverage.
arXiv:2606. 11209v1 Announce Type: cross Abstract: Visual question answering increasingly requires multi-step reasoning.
Multi-modal Large Language Models (MLLMs) have achieved remarkable progress in video temporal grounding with reinforcement learning for generating reasoning paths. However, existing models often produce superficial reasoning, which offers limited guidance for precise temporal localization.
arXiv:2601. 07761v2 Announce Type: replace Abstract: Large Vision-Language Models (LVLMs) face a fundamental dilemma in video reasoning: they are caught between the prohibitive computational costs of verbose reasoning and the hallucination risks of efficient, ungrounded approaches.
arXiv:2609.39168v1 Announce Type: new Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has improved the reasoning capabilities of Multimodal Large Language Models (MLLMs), yet existing...
arXiv:2606. 24477v1 Announce Type: cross Abstract: Video large language models (LLMs) are often constrained by computation and memory budgets, leading them to use reduced frame rates and spatial resolutions, which may cause them to miss critical information for question answering (QA).
The paper introduces VWG-Bench, a benchmark covering nine reasoning dimensions and 38 tasks to evaluate video generative models on symbolic reasoning, physical laws, and goal pursuit. It also presents Vid-PRE, a prompt-rewriting framework that offloads reasoning to a VLM, improving logical performance without changing the generator architecture. Experiments show that current models excel at visual quality but struggle with logic-heavy tasks, while Vid-PRE significantly boosts reasoning across multiple generators.
Video-FLAIR is a training framework that teaches a model to choose the most suitable reasoning mode—perceptual, compositional, or deliberative—for each multimodal query using reinforcement learning. During training, the model generates responses under all three modes for the same prompt, and a composite reward system selects the best response based on correctness, grounding, and cost, discouraging unsupported deliberation. This adaptive approach improves accuracy on benchmarks such as MathVista, Video-Holmes, and Video-MMMU while dramatically reducing token usage compared to always-thinking baselines.
arXiv:2609.15606v1 Announce Type: cross Abstract: Multimodal Large Language Models (MLLMs) have achieved remarkable progress on short video understanding yet remain limited on long videos due to the...
The paper introduces Stepwise Marginal Information Gain (MIG), an intrinsic process reward that evaluates how each reasoning step of a large language model (LLM) or vision-language model (VLM) improves the likelihood of the reference answer. MIG rewards only new likelihood maxima, preventing duplicate credit, and is combined with outcome, format, and self‑distillation objectives to guide training. Experiments on eight benchmarks show that this method outperforms outcome‑only reinforcement learning and improves accuracy by up to 4.8 points over binary‑reward training, including a 12.6‑point gain on MathVerse and a 12.9‑point advantage on vision‑language transfer at 7B parameters.
VisionCoach is an input‑adaptive reinforcement learning framework that enhances spatio‑temporal grounding in video reasoning by using visual prompting during training. The system selectively applies visual prompts to challenging inputs, amplifying question‑relevant evidence and suppressing distractors, and then internalizes these improvements through self‑distillation so that inference can be performed on raw videos without prompts. Experiments on multiple benchmarks (V‑STAR, VideoMME, World‑Sense, VideoMMMU, PerceptionTest, and Charades‑STA) show that VisionCoach achieves state‑of‑the‑art performance while maintaining a single efficient inference pathway.