Research map / Vision-language and multimodal / Multimodal LLMs

Beyond Perceptual Shortcuts: Causal-Inspired Debiasing Optimization for Generalizable Video Reasoning in Lightweight MLLMs

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual 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.

Diagnoses perceptual shortcut bias in CLEVRER counterfactual data and proposes VideoThinker, which trains a shortcut "bias model" and fine-tunes Qwen2.5-VL-3B with a CDPO objective that uses a positive KL term to repel the policy from the bias model.

Problem

RL fine-tuning (GRPO) gives limited gains for lightweight (3B) MLLMs on video reasoning; the authors attribute this to perceptual shortcuts in training data that are dominant in CLEVRER counterfactual questions.

Approach

Train a bias model on shortcut-only (observational) questions without KL, then fine-tune the main policy with GRPO plus a sign-flipped KL term that maximizes divergence from the frozen bias model, pushing the policy away from shortcut behavior.

Key results

• VideoThinker-R1 (3B) reaches 79.1% on CLEVRER counterfactual vs 64.9% for GRPO baseline with same 1K RL samples.
• VideoThinker-R1 scores 60.9 on MVBench and 63.5 on TempCompass, surpassing Video-UTR-7B (58.8 and 59.7).
• VideoThinker-R1 reaches 52.4 on VideoMME (w/o subs), 7 points above VideoRFT-3B's 45.4.

Surprising findings

• GRPO fine-tuning of a 3B model dropped inferential-question accuracy on CLEVRER from 73.9% to 63.1% while observational accuracy stayed stable or rose.
• Minimizing KL to the bias model (attraction) yielded 74.3% on CLEVRER and no MMVU gain, while the repulsive sign gave 79.1%.

Limitations

• Validated mainly on VQA tasks matching the causal assumptions; generalization to captioning is untested.
• CDPO is a gradient-based approximation of backdoor adjustment, trading theoretical exactness for efficiency.
• Analysis centered on 3B model; scaling behavior of bias and intervention at larger sizes is not studied.

Experimental setup

Datasets: CLEVRER MMVU Video-Holmes MVBench TempCompass VideoMME

Benchmarks: CLEVRER MMVU Video-Holmes MVBench TempCompass VideoMME

Metrics: accuracy

Baselines: GRPO VideoRFT-3B TinyLLaVA-Video-R1 Video-UTR-7B Qwen2.5-VL-3B (CoT) Video-R1-7B VideoRFT-7B VideoLLaMA2 LongVA-7B VILA-1.5-8B

Builds on: GRPO DAPO Video-RFT Qwen2.5-VL

Closest papers by meaning

• Video-R1: Reinforcing Video Reasoning in MLLMs (NeurIPS 2025 · similarity 0.91)
• VideoRFT: Incentivizing Video Reasoning Capability in MLLMs via Reinforced Fine-Tuning (NeurIPS 2025 · similarity 0.91)
• Incentivizing Versatile Video Reasoning in MLLMs via Data-Efficient Reinforcement Learning (CVPR 2026 · similarity 0.90)
• Find, Fix, Reason: Context Repair for Video Reasoning (ICML 2026 · similarity 0.90)
• NoisyRollout: Reinforcing Visual Reasoning with Data Augmentation (NeurIPS 2025 · similarity 0.89)
• OneThinker: All-in-one Reasoning Model for Image and Video (CVPR 2026 · similarity 0.89)