[CVPR2026] CARE What Fails: Contrastive Anchored-REflection for Verifiable Multimodal Reasoning

📑 arXiv | 🔎 AlphaXiv | 🤗 Hugging Face

CARE is a failure-centric post-training framework for verifiable multimodal reasoning. Instead of discarding wrong rollouts, CARE turns close-but-wrong attempts into structured supervision with an anchored-contrastive objective and Reflection-Guided Resampling (RGR), while keeping single-pass inference with no test-time reflection.

News

2026.2 Our paper is accepted by CVPR 2026! See you in Denver!

2026.1 Code is released.

Highlights

• Failure-centric RLVR for multimodal reasoning: CARE explicitly learns from failed rollouts instead of treating them as discarded samples or uniform negatives.
• Anchored-contrastive training: pick the shortest verified-correct rollout as the anchor, select semantically proximate hard negatives, normalize advantages within the subgroup, and down-weight only negative advantages.
• Reflection-Guided Resampling (RGR): repair exactly one representative hard negative during training, re-verify it, and reuse the repaired rollout if it becomes correct.
• All-negative rescue: inject a small pseudo-contrast when a rollout group has no successes, preventing zero-signal updates.
• No extra test-time cost: CARE improves training-time credit assignment without requiring reflection or multi-pass decoding at inference.

Method overview

For a multimodal prompt x = <image(s), question>, CARE samples a group of rollouts and uses a programmatic verifier over the final answer and output format to build a failure-aware learning signal:

1. Sample a group of rollouts (size G).
2. Verify each rollout with a programmatic verifier.
3. If at least one rollout is correct:
   • Choose the anchor as the shortest verified-correct rollout.
   • Select hard negatives that are closest to the anchor in rationale space via cosine proximity.
   • Normalize rewards only inside the selected subgroup and down-weight only the negative advantages.
   • If fewer than the target number of hard negatives are available, rescale the update size to keep training stable.
4. During training, optionally run RGR on one representative hard negative:
   • Insert a short repair cue.
   • Resample once.
   • Replace the original failure if the repaired rollout becomes verifier-positive; otherwise keep it with a reduced penalty.
5. If all rollouts are incorrect:
   • Apply an all-negative rescue with a small pseudo-contrast so gradients do not stall.
6. Use a region-weighted policy objective:
   • Answer tokens receive full weight.
   • Positive rationale tokens receive a small weight.
   • Failing rationale tokens are masked out.

CARE changes how training signals are formed from rollouts, but keeps the verifier and single-decode inference pipeline unchanged.

---

Benchmark snapshot

According to the paper, CARE delivers consistent gains over existing RLVR baselines:

• On Qwen2.5-VL-7B, CARE improves macro-averaged accuracy by +4.6 points over GRPO across six verifiable visual-reasoning benchmarks.
• On Qwen3-VL-8B, CARE reaches competitive or state-of-the-art results on MathVista mini and MMMU-Pro under the same evaluation protocol.
• The paper attributes most of the gains to the anchored-contrastive objective, with RGR providing an additional improvement by converting near-miss failures into usable positives.

---

Repository map

• CARE subgrouping + advantages: verl/algorithms/adv_estimators/care.py
• Cosine hard negatives: verl/algorithms/neg_selectors/cosine_hardneg.py
• Region-weighted token advantages: verl/algorithms/losses/region_weighted_tokens.py
• RGR hook (training-only): verl/hooks/rgr.py
• Tag parsing for /: verl/utils/response_tags.py
• Config surface: examples/config.yaml (care.*, rgr.*)
• Qwen2.5-VL CARE script: examples/qwen2_5_vl_7b_geo3k_care_grpo.sh

---

Installation

git clone https://github.com/yongxinwang-ai/CARE.git
cd CARE

# (recommended) create env
conda create -n care python=3.10 -y
conda activate care

pip install -e .

---

Quickstart

CARE (Qwen2.5-VL 7B, Geometry3K)

bash examples/qwen2_5_vl_7b_geo3k_care_grpo.sh

CARE with overrides

python3 -m verl.trainer.main \
  config=examples/config.yaml \
  data.train_files=hiyouga/geometry3k@train \
  data.val_files=hiyouga/geometry3k@test \
  worker.actor.model.model_path=Qwen/Qwen2.5-VL-7B-Instruct \
  algorithm.grpo_variant=care \
  care.K=4 care.M=6 \
  care.neg_scale_s=0.5 care.equalize=true \
  care.rescue.enable=true care.rescue.delta=0.1 \
  care.token_weighting=region_weighted care.gamma_pos=0.005 \
  rgr.enable=true rgr.template=structured

---

CARE configuration

CARE is exposed via algorithm.grpo_variant=care and the care.* / rgr.* sections in examples/config.yaml. The provided Geometry3K example script uses:

• Rollouts per prompt: G = worker.rollout.n (default 8 in examples)
• Hard-negative subgroup size: care.K = 4
• Negative preselect size: care.M = 6
• Negative scaling: care.neg_scale_s = 0.5
• Reflected-failure scaling: rgr.s_refl = care.neg_scale_s / 2 (if not set)
• All-negative rescue magnitude: care.rescue.delta = 0.1
• Positive rationale token weight (region-weighting): care.gamma_pos = 0.005

---

Citation

If you use this code, please cite the paper. The arXiv entry is:

@article{wang2025care,
  title   = {CARE What Fails: Contrastive Anchored-REflection for Verifiable Multimodal Reasoning},
  author  = {Wang, Yongxin and Yang, Zhicheng and Cao, Meng and Han, Mingfei and Lin, Haokun and Zhu, Yingying and Chang, Xiaojun and Liang, Xiaodan},
  journal = {arXiv preprint arXiv:2512.19554},
  year    = {2025}
}

---

License

Apache-2.0. See LICENSE.