Video-FLAIR trains multimodal models to choose among direct, concise, and deep reasoning modes for each query. Its reinforcement-learning framework improves accuracy across image and video reasoning benchmarks while reducing average response length from 417 to 95 tokens compared with always-thinking baselines.

This paper introduces AVATAR, a framework that improves multimodal reasoning by addressing limitations in standard reinforcement learning. It uses an off-policy architecture to improve data efficiency and introduces Temporal Advantage Shaping (TAS), a novel credit assignment strategy to focus learning on critical reasoning steps, achieving significant gains on audio-visual benchmarks.

This paper proposes EgoVITA to address the challenge of "catastrophic forgetting" in egocentric video understanding. By separating reasoning into an egocentric planning stage and an exocentric verification stage, our RL framework teaches models to anticipate future visual states rather than just imitating fixed sequences. It utilizes novel Anticipatory Cross-Modal Grounding (ACMG) rewards to ensure plans are visually grounded, outperforming baselines on tasks like EgoBlind and EgoOrient while preserving generalization on standard video benchmarks.

This paper introduces VideoPASTA, which improves video models by training them with specially crafted "bad examples" (adversarial preference pairs) that target common errors in spatial, temporal, and cross-frame understanding. It shows this targeted approach is highly efficient, achieving significant performance gains using only 7k preference pairs.

This paper presents VideoSAVi, a method to teach video models better spatial and temporal reasoning without needing human supervision. It works by having the model critique its own reasoning errors to automatically create preference data for training, achieving strong results on benchmarks efficiently.