CARE 何为失败：面向可验证多模态推理的对比锚定反思
CARE What Fails: Contrastive Anchored-REflection for Verifiable Multimodal Reasoning
对做多模态推理后训练的人，CARE 给出了一套把「全错批次」和「接近正确」的 rollout 转成对比监督的具体配方：锚定对比目标 + 组内 z-score 归一化 + 全负样本救援，可直接替换 GRPO 的损失项；RGR 的单样本结构化自修复则提供了一种无需测试时反思就能把近失样本转成正样本的重采样策略。
原文依据 an anchored-contrastive objective that forms a compact subgroup around the best rollout and a set of semantically proximate hard negatives
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论带可验证奖励的组相对强化学习（RLVR）常常浪费掉它已经拥有的信息量最大的数据——失败样本。当所有 rollout 都错误时，梯度停滞；当其中一个恰好正确时，更新通常忽略其余样本为何接近但错误，并且信用可能被错误地分配给虚假的推理链。我们提出 CARE（Contrastive Anchored REflection，对比锚定反思），一个以失败为中心的多模态推理后训练框架，将错误转化为监督信号。CARE 结合了：(i) 锚定对比目标，它在最佳 rollout 周围形成一个紧凑子组，并配上一组语义上接近的难负样本，执行组内 z-score 归一化并采用仅对负样本的缩放，同时包含全负样本救援机制以避免零信号批次；(ii) 反思引导重采样（RGR），一种单样本结构化自修复，它重写一个代表性失败样本并用同一验证器重新打分，从而在没有任何测试时反思的情况下将近失样本转化为可用的正样本。CARE 提升了准确率与训练平滑度，同时显式增加了来自失败样本的学习信号占比。在 Qwen2.5-VL-7B 上，CARE 在六个可验证视觉推理基准上相较 GRPO 将宏平均准确率提升 4.6 个点；在 Qwen3-VL-8B 上，它在相同评测协议下于 MathVista 和 MMMU-Pro 达到有竞争力或最优的结果。
英文原文摘要 1413 字符
Group-relative reinforcement learning with verifiable rewards (RLVR) often wastes the most informative data it already has--the failures. When all rollouts are wrong, gradients stall; when one happens to be correct, the update usually ignores why the others are close-but-wrong, and credit can be misassigned to spurious chains. We present CARE (Contrastive Anchored REflection), a failure-centric post-training framework for multimodal reasoning that turns errors into supervision. CARE combines: (i) an anchored-contrastive objective that forms a compact subgroup around the best rollout and a set of semantically proximate hard negatives, performs within-subgroup z-score normalization with negative-only scaling, and includes an all-negative rescue to prevent zero-signal batches; and (ii) Reflection-Guided Resampling(RGR), a one-shot structured self-repair that rewrites a representative failure and re-scores it with the same verifier, converting near-misses into usable positives without any test-time reflection. CARE improves accuracy and training smoothness while explicitly increasing the share of learning signal that comes from failures. On Qwen2.5-VL-7B, CARE lifts macro-averaged accuracy by 4.6 points over GRPO across six verifiable visual-reasoning benchmarks; with Qwen3-VL-8B it reaches competitive or state-of-the-art results on MathVista and MMMU-Pro under an identical evaluation protocol.