Research map / Large language models / Reasoning and chain of thought

Reinforcing Structured Chain-of-Thought for Video Understanding

CVPR 2026 · Large language models › Reasoning and chain of thought · approach: RL with verifiable rewards for reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SDRL, a single-stage RL-only framework that trains Qwen2.5-VL-7B with a Summarize→Think→Answer structured CoT and self-supervised token weighting (CVK and DVR) inside GRPO, reporting state-of-the-art on seven VideoQA benchmarks.

Problem

MLLM video reasoning suffers from thinking drift and weak temporal comprehension under GRPO-style RL, and existing strong methods rely on costly CoT annotation and multi-stage SFT+RL pipelines.

Approach

Insert an explicit summary segment before thinking in the CoT and apply token-wise weights inside the GRPO objective: CVK down-weights summary tokens whose position-wise KL divergence from the group's correct-summary center is high, and DVR up-weights thinking-token entropy scaled by (1 - group accuracy).

Key results

• SDRL (Qwen2.5-VL-7B, RL-only, Video-R1-260K) reaches 79.3 on NExT-GQA, 68.6 on MMVU, 36.1 on VSIBench, 64.2 on MVBench and 54.7 on VideoMME, beating the SFT+RL VideoRFT on several benchmarks.
• On EventFlowQA ablations, self-supervised CVK+DVR (entropy variant) reaches 56.10% vs 42.37% for the original 7B model.
• Ablation: self-supervised 7B gains +11.91 points over original, GT supervision gains +6.19; 3B gains +3.01 (GT) and +2.40 (self).

Surprising findings

• Self-supervision without ground-truth summaries outperformed GT-supervised summary alignment for the 7B model.
• KL-based diversity weighting underperformed entropy-based weighting in both supervision settings.

Limitations

• Coarse summaries on long or fine-grained videos can cause temporal drift.
• CVK gives no formal guarantee of factual correctness and may converge on consistent but hallucinated summaries under sparse answer supervision.
• Relies on the base model's instruction-following to parse structured tags; weaker models may need warm-up.

Experimental setup

Datasets: Video-R1-260K EventFlowQA YouCook2 COIN EgoExo4D

Benchmarks: NExT-GQA MMVU VideoMMMU VSIBench MVBench TempCompass VideoMME EventFlowQA

Metrics: accuracy BLEU sBERT

Baselines: Video-R1 VideoRFT VideoChat-R1 TW-GRPO Qwen2.5-VL-7B VideoLLaMA2 LLaVA-OneVision-7B

Builds on: GRPO Video-R1 VideoRFT VideoChat-R1 TW-GRPO

Closest papers by meaning

• VideoRFT: Incentivizing Video Reasoning Capability in MLLMs via Reinforced Fine-Tuning (NeurIPS 2025 · similarity 0.91)
• Video-R1: Reinforcing Video Reasoning in MLLMs (NeurIPS 2025 · similarity 0.90)
• When Thinking Drifts: Evidential Grounding for Robust Video Reasoning (NeurIPS 2025 · similarity 0.90)
• Incentivizing Versatile Video Reasoning in MLLMs via Data-Efficient Reinforcement Learning (CVPR 2026 · similarity 0.89)
• OneThinker: All-in-one Reasoning Model for Image and Video (CVPR 2026 · similarity 0.89)
• ReWatch-R1: Boosting Complex Video Reasoning in Large Vision-Language Models through Agentic Data Synthesis (ICLR 2026 · similarity 0.89)