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

Source: arXiv - 2512.19554v1

Overview

The paper introduces CARE (Contrastive Anchored‑Reflection), a post‑training framework that turns the failures of multimodal reasoning models into a powerful source of supervision. By focusing on what goes wrong during inference—rather than discarding those examples—CARE boosts accuracy on visual‑reasoning benchmarks while keeping training stable and efficient.

Key Contributions

• Failure‑centric learning: A novel contrastive objective that builds a tight “anchor” subgroup around the best rollout and treats near‑misses as hard negatives, extracting learning signal from every erroneous example.
• Negative‑only scaling & all‑negative rescue: Within‑subgroup z‑score normalization that avoids gradient collapse when no positives are present, plus a fallback mechanism that guarantees a non‑zero training signal.
• Reflection‑Guided Resampling (RGR): A one‑shot self‑repair step that rewrites a representative failure, re‑evaluates it with the same verifier, and converts the corrected version into a usable positive without any extra test‑time overhead.
• Empirical gains: On Qwen2.5‑VL‑7B, CARE lifts macro‑averaged accuracy by 4.6 % over the strong GRPO baseline across six verifiable visual‑reasoning datasets; with Qwen3‑VL‑8B it matches or exceeds state‑of‑the‑art on MathVista and MMMU‑Pro under identical protocols.
• Training smoothness: Demonstrated reduction in gradient variance and faster convergence, making the approach practical for large‑scale multimodal models.

Methodology

1. Anchored‑Contrastive Objective

   • Anchor selection: Identify the best rollout (the highest verifier score) for each input.
   • Subgroup formation: Gather semantically similar rollouts (including the anchor) and a set of hard negatives that are close but wrong.
   • Normalization: Apply within‑subgroup z‑score normalization only on negatives; this prevents the loss from collapsing when no positives exist.
   • All‑negative rescue: If a batch contains only negatives, a fallback scaling ensures the loss still provides gradient information.

2. Reflection‑Guided Resampling (RGR)

   • Pick a representative failure from the hard‑negative set.
   • Prompt the model (or a lightweight verifier) to rewrite the failure into a plausible correct answer.
   • Re‑score the rewritten answer with the original verifier; if it passes, treat it as a synthetic positive for the contrastive loss.
   • This step is performed once per batch, so it adds negligible overhead and does not affect inference time.

3. Training Loop

   • The contrastive loss and the RGR‑generated positives are combined with the standard language modeling objective.
   • No changes are required to the underlying multimodal backbone; CARE works as a plug‑in post‑training wrapper.

Results & Findings

• Training stability: Gradient norm variance dropped by ~30 % compared to baseline, leading to smoother loss curves.
• Failure signal proportion: Over 60 % of the gradient magnitude originated from failure‑derived samples, confirming the framework’s “failure‑centric” claim.
• Inference cost: Zero additional latency; the RGR step is confined to training.

Practical Implications

• Better use of existing data: Teams can squeeze more performance out of already‑collected rollouts without gathering new annotations.
• Plug‑and‑play improvement: CARE can be applied to any multimodal model that produces a verifier score (e.g., CLIP‑based or LLM‑vision hybrids), making it attractive for product teams looking to boost reasoning accuracy quickly.
• Reduced need for exhaustive prompting: By automatically turning near‑misses into positives, developers spend less time hand‑crafting corrective prompts or data augmentations.
• Robustness to noisy rollouts: In real‑world pipelines where many generated answers are wrong, CARE ensures those errors still contribute to learning, leading to models that are more resilient to distribution shifts.
• No inference overhead: Since the reflection step is a one‑off training trick, production systems see no latency penalty.

Limitations & Future Work

• Verifier dependence: CARE assumes a reliable, differentiable verifier; its effectiveness may degrade if the verifier is noisy or biased.
• Scalability of hard‑negative mining: Forming semantically proximate negative sets can become expensive for extremely large batch sizes; smarter sampling strategies are needed.
• One‑shot reflection: While efficient, a single rewrite may miss richer corrective signals; exploring multi‑step or iterative reflection could yield further gains.
• Generalization beyond visual reasoning: The authors note that extending CARE to pure language or audio‑text tasks remains an open question.

Overall, CARE offers a pragmatic, failure‑focused recipe for boosting multimodal reasoning models, with clear pathways for industry adoption and future research.

Authors

• Yongxin Wang
• Zhicheng Yang
• Meng Cao
• Mingfei Han
• Haokun Lin
• Yingying Zhu
• Xiaojun Chang
• Xiaodan Liang

Paper Information

• arXiv ID: 2512.19554v1
• Categories: cs.LG, cs.AI
• Published: December 22, 2025
• PDF: Download PDF