The paper introduces CoFire, a reinforcement learning framework for video understanding that addresses the problem of confidence erosion—where initial predictions lose reliability during iterative refinement. CoFire enforces a "confident first, reflective second" principle, ensuring the initial answer remains trustworthy and that subsequent reflection only improves upon it. The method achieves state‑of‑the‑art accuracy on Video‑MME, MLVU, and LongVideoBench, outperforming prior 7B models by 1.7 points and improving both initial and final accuracy over outcome‑reward RL by 4.1 and 3.7 points respectively.
Video models answer from sampled frames and may miss the decisive moment. A common remedy is to commit to an answer, re-examine the video, and confirm or revise it, a loop now mainly learned with rein...
arXiv:2606. 27922v1 Announce Type: cross Abstract: Current multimodal reflection mechanisms for long video understanding predominantly rely on closed-loop self-reflection within internal parameters.
arXiv:2610.01973v1 Announce Type: new Abstract: Reinforcement learning (RL) for video generation usually assigns one scalar reward to an entire sampled video. Yet a video is not uniformly flawed: som...
arXiv:2610.32222v1 Announce Type: new Abstract: Reinforcement learning has gained attention for improving action-conditioned world models (ACWMs), as it enables reward-driven optimization of visual q...
SCARL (Sequence Confidence-Aligned Reinforcement Learning) addresses a misalignment between a reinforcement learning policy’s confidence and its reward signal in language-model reasoning. By reshaping each advantage with an alignment gap that reflects the difference between a response’s group-relative confidence and the outcome it certifies, SCARL drives responses with equal advantage toward equal confidence and mitigates entropy collapse. Experiments on Qwen3‑4B‑Base show a 2.4‑point lift in Pass@32 over DAPO and consistent improvements across multiple estimators and backbones while preserving policy entropy.
arXiv:2606. 29892v1 Announce Type: cross Abstract: Reinforcement learning (RL) has become indispensable for pushing Vision-Language-Action Models (VLAs) beyond static imitation learning.
arXiv:2608. 04964v1 Announce Type: new Abstract: Interactive video world models are essential for long-horizon planning and exploration, yet they suffer from compounding errors.
The paper introduces RARM, a Reference‑Anchored Reward Model that uses a single successful demonstration to generate dense, progress‑aware rewards for reinforcement learning in robot manipulation. RARM is trained on general‑purpose videos with a contrastive temporal objective, requiring no task‑specific data or reward engineering. During deployment it matches rollout clips to reference clips and rewards only confident forward progress, reducing false positives. Experiments on nine simulated tasks and four real‑world tasks show that RARM achieves the best overall success rates, especially on long‑horizon tasks like cloth folding.
arXiv:2610.18614v1 Announce Type: new Abstract: Controlling an acting LLM agent means deciding, at each step of a long-horizon task, whether to trust its next action, abstain, or hand off to a human....
arXiv:2602. 13602v2 Announce Type: replace-cross Abstract: We present \revise (\underline{Re}asoning with \underline{Vi}deo \underline{S}parsity), a multi-round agent for video question answering (VQA).
arXiv:2610.31820v1 Announce Type: new Abstract: Reliable confidence estimation is essential for trustworthy deployment of large language models (LLMs), but dedicated confidence training stage require...
Disagreement-Regularized Imitation Learning (DRIL) is evaluated for image-based continuous control using Gaussian and Beta policies. In a controlled CarRacing study, DRIL achieved the largest performance gains in few-demonstration settings, improving over behavior cloning by 61% with clipped-action demonstrations and 112% with bounded-action demonstrations. With more demonstrations, the advantage narrowed, and Beta behavior cloning remained slightly superior in the bounded-action regime.