Research map / Video / Video-language understanding
Video-FLAIR: Not Whether to Reason, But How
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.
Video-FLAIR trains Qwen-VL models with RL over mode-structured rollouts (Percept, Compose, Deliberate) and a composite reward including a Qwen3-VL verifier, learning to pick a reasoning mode per query with far fewer tokens than always-thinking baselines.
Problem
Multimodal LLMs apply a uniform chain-of-thought strategy to all queries, wasting tokens on perceptual queries and under-reasoning on complex ones; existing adaptive methods only decide whether to think, not how.
Approach
Defines three reasoning modes (Percept, Compose, Deliberate) and trains the policy with RL on mode-structured rollout groups where controlled slots force each mode on the same prompt, so adaptive slots can be scored against the utility-best mode; a composite reward adds cost, balance, compliance, and verifier-based grounding terms.
Key results
- On Qwen2.5-VL-7B, Video-FLAIR reaches 73.8 on MathVista (+5.4), 48.0 on Video-Holmes (+4.8), and 56.9 on Video-MMMU (+4.8) at 59-74 average tokens.
- Across the reported benchmarks, Video-FLAIR averages 95 tokens versus 417 for always-thinking baselines, per the abstract.
- On Qwen3-VL-8B, Video-FLAIR reaches 30.8 on EMMA (+8.4) and 85.5 on AI2D (+5.2) with 108 average tokens.
Surprising findings
- Always-thinking baselines drop below the base model on perception-heavy tasks such as EMMA and HallusionBench, suggesting long chains hurt visual grounding.
- Removing the verifier reward leaves MMMU nearly unchanged but lowers EMMA, and the paper says the verifier mainly improves grounding rather than accuracy.
Limitations
- Mode space is fixed to three modes; mid-response mode switching and tool use are left for future work.
- Relies on Qwen3-VL as an online verifier and periodic DPO refresh, adding reward-pipeline complexity.
- Policy sometimes misroutes perception-like queries (e.g., Social Reasoning on Video-Holmes treated as single-frame lookups).
Experimental setup
Datasets: MathVista Video-Holmes Video-MMMU EMMA MMMU HallusionBench AI2D MM-Vet v2 Video-TT VSI-Bench SciVideoBench
Benchmarks: EMMA MMMU MathVista HallusionBench AI2D MM-Vet v2 Video-Holmes Video-TT VSI-Bench SciVideoBench Video-MMMU
Metrics: accuracy average tokens
Baselines: Video-R1 Video-R2 VideoRFT Video-Auto-R1 OneThinker SFT-only base model
Builds on: DAPO GRPO GDPO Direct Preference Optimization Adaptive Length Penalty
Closest papers by meaning
- ARES: Multimodal Adaptive Reasoning via Difficulty-Aware Token-Level Entropy Shaping (ICLR 2026 · similarity 0.87)
- Find, Fix, Reason: Context Repair for Video Reasoning (ICML 2026 · similarity 0.87)
- CARE What Fails: Contrastive Anchored-REflection for Verifiable Multimodal Reasoning (CVPR 2026 · similarity 0.87)
- Incentivizing Versatile Video Reasoning in MLLMs via Data-Efficient Reinforcement Learning (CVPR 2026 · similarity 0.87)
- Reason in the Words You Speak: Idiolectal Paraphrasing Off-Policy Traces for Reasoning Distillation in VideoLLMs (NeurIPS 2026 · similarity 0.87)
- Video-R1: Reinforcing Video Reasoning in MLLMs (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.