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

CARE What Fails: Contrastive Anchored-REflection for Verifiable Multimodal Reasoning

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 CARE, a failure-centric RLVR post-training framework for multimodal reasoning that anchors advantages to the shortest correct rollout, normalizes within a hard-negative subgroup, and adds one-shot reflection-guided resampling of failures.

Problem

Group-relative RLVR for multimodal reasoning wastes failure data: gradients stall when all rollouts fail, and credit is misassigned when a correct rollout is reinforced despite flawed reasoning chains.

Approach

Anchor advantages to the shortest verifier-correct rollout, form a subgroup with semantically closest hard negatives (cosine distance on rationale embeddings), z-score normalize within the subgroup with down-weighted negatives, add an all-negative pseudo-contrast rescue, and perform one training-only reflection resample of a selected hard negative.

Key results

• On Qwen2.5-VL-7B, CARE improves macro-average accuracy by +4.62 over GRPO (50.57 vs 45.95) across six benchmarks.
• Qwen3-VL-8B + CARE reaches 82.1 on MathVista and 46.7 on MMMU-Pro standard, exceeding listed models including MiMo-VL-7B.
• Anchor objective accounts for about 84% of the Qwen2.5-VL-7B gain; RGR adds the remaining ~16%.

Surprising findings

• Random negative selection plateaus near 43.06 macro accuracy versus 50.57 for cosine-selected hard negatives.
• Reflection-guided resampling success rate is 76.6% on triggered groups versus 19.3% and 12.8% for no-cue and random resampling controls.

Limitations

• Relies on binary correctness signals, restricting use to domains with objective ground truths.
• Requires at least one successful rollout to anchor the contrast; near-zero pass-rate tasks may need warm-up.
• Shortest-rationale anchor preference may discourage extended reasoning in edge cases.

Experimental setup

Datasets: ChartQA Geometry3K ViRL39K Vision-R1-cold

Benchmarks: MathVista MathVerse MATH-Vision MMMU MMMU-Pro

Metrics: accuracy macro-average accuracy

Baselines: GRPO DAPO GSPO Qwen2.5-VL-7B-Instruct Qwen3-VL-8B-Instruct MiMo-VL-7B

Builds on: GRPO DAPO GSPO Vision-R1

Closest papers by meaning

• SRPO: Enhancing Multimodal LLM Reasoning via Reflection-Aware Reinforcement Learning (NeurIPS 2025 · similarity 0.91)
• Perceptual-Evidence Anchored Reinforced Learning for Multimodal Reasoning (CVPR 2026 · similarity 0.90)
• Perception-Aware Policy Optimization for Multimodal Reasoning (ICLR 2026 · similarity 0.90)
• Boosting Reasoning in Large Multimodal Models via Activation Replay (CVPR 2026 · similarity 0.90)
• R-C2: Cycle-Consistent Reinforcement Learning Improves Multimodal Reasoning (CVPR 2026 · similarity 0.90)
• Reinforcing Multimodal Reasoning Against Visual Degradation (NeurIPS 2026 · similarity 0.89)