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

Find, Fix, Reason: Context Repair for Video Reasoning

ICML 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 Find, Fix, Reason (FFR): a frozen tool-integrated teacher (GLM-4.5V) diagnoses failed GRPO rollouts and supplies a minimal, non-leaking evidence patch; the student re-answers with the patch and the chosen rollout drives a GRPO update with a Robust Improvement Reward.

Problem

Video reasoning RL with MLLMs plateaus at the student's knowledge boundary under on-policy self-exploration, while hybrid off-policy replay needs careful regularization and dynamic-context tool methods are bounded by small-model capability and need two-stage training.

Approach

Instead of imitating teacher traces or letting the student search multiple rounds, a frozen larger teacher gives observation-level interventions (temporal ranges, key frames, regions without the answer) only on failed rollouts; the student re-answers the unchanged question with this patch, and the chosen rollout (original if correct, repaired otherwise) enters a group-normalized GRPO update with a

Key results

• With Video-R1-SFT as base, FFR raises MMVU from 63.8 to 68.5 and VSI-Bench from 35.8 to 38.9 over Video-R1 (T-GRPO) and reaches 52.3 on Video-Holmes versus 36.5.
• With VideoRFT-SFT as base, FFR reaches 70.1 on MMVU and 48.0 on Video-Holmes, versus 68.5 and 40.0 for VideoRFT.
• FFR with a 32B teacher (51.2 average) surpasses SFT distillation from a 235B teacher (50.7 average) on the reasoning benchmarks.

Surprising findings

• Intervention ratio falls 48% from early to late training while training accuracy rises, interpreted as capability internalization.
• The 8B student trained with FFR surpasses its 32B teacher on Video-Holmes and LongVideoBench, and a stronger standalone teacher (Qwen3-VL-235B) does not yield the best student.

Limitations

• Leakage evaluated on only 200 manually verified interactions, with a 95% CI upper bound near 1.5%.
• Relies on a large teacher model and API-based teacher service, adding cost and complexity.
• Ablation and main results rely on a small training set (4,000 samples) and 16-frame inputs at low resolution.

Experimental setup

Datasets: Video-R1 data Video-Thinker data Video-Holmes

Benchmarks: MMVU VSI-Bench VideoMMMU Video-Holmes LongVideoBench LVBench MVBench TempCompass

Metrics: accuracy average accuracy

Baselines: Qwen2.5-VL-7B Video-R1 VideoRFT Video-ChatR1 AVATAR Pixel-Reasoner Video-Thinker LUFFY

Builds on: GRPO R1-V Video-R1 VideoRFT

Closest papers by meaning

• Reason in the Words You Speak: Idiolectal Paraphrasing Off-Policy Traces for Reasoning Distillation in VideoLLMs (NeurIPS 2026 · similarity 0.91)
• Video-R1: Reinforcing Video Reasoning in MLLMs (NeurIPS 2025 · similarity 0.90)
• ReWatch-R1: Boosting Complex Video Reasoning in Large Vision-Language Models through Agentic Data Synthesis (ICLR 2026 · similarity 0.90)
• Beyond Perceptual Shortcuts: Causal-Inspired Debiasing Optimization for Generalizable Video Reasoning in Lightweight MLLMs (CVPR 2026 · similarity 0.90)
• VideoRFT: Incentivizing Video Reasoning Capability in MLLMs via Reinforced Fine-Tuning (NeurIPS 2025 · similarity 0.89)
• Reinforcing Structured Chain-of-Thought for Video Understanding (CVPR 2026 · similarity 0.88)