RECAP: Looking Once Is Not Enough for Vision-Language Reasoning
Long chain-of-thought helps vision-language models solve complex visual tasks, yet it introduces a temporal grounding problem: intermediate reasoning steps can drift from image evidence, and standard final-answer reinforcement learning cannot credit which tokens should have revisited the image. This paper identifies "visual revisit"—localized spikes in cross-modal attention after language-internal reasoning—as a key temporal signal. The authors show across three backbones that revisit peaks predict correctness and are causally linked to accuracy. Building on this, they propose RECAP, a GRPO-compatible training method that propagates terminal rewards through revisit-conditioned dynamic discounting and gated value prediction, yielding consistent HallusionBench gains without inference-time cost.
The paper makes a well-motivated and empirically grounded contribution. The case for visual revisit is built through multiple lenses—predictive correlation, causal masking and injection, signal controls, and attention-head localization—and the RECAP algorithm turns this observation into a clean GRPO-compatible training modification. Main results on Qwen2.5-VL-7B, InternVL3-8B, and LLaVA-OV-7B show HallusionBench improvements of +3.5 to +4.1 over GRPO while preserving reasoning and text-only scores. However, the theoretical support is limited to stylized settings: Proposition 1 is a horizon-3 expressiveness separation, and Theorem 1 relies on an oracle linear-in-revisit assumption and only addresses variance reduction, not policy improvement.
The empirical evidence for visual revisit is the strongest part of the paper. The authors show that revisit peaks are not explained by simpler signals like token entropy, logit margin, or hidden-state norm; that they correlate with human-annotated decision steps well above chance; that they concentrate in a small subset of attention heads; and that zeroing those heads reproduces most of the accuracy drop from full cross-modal attention masking. The ablations in Table 2 are well-executed: removing the dynamic discount causes the largest accuracy drop, removing the gated value predictor raises advantage variance, and composing RECAP with VPPO gives further gains, supporting the claim that these methods address orthogonal credit-assignment mechanisms.
The theoretical analysis does not provide strong guarantees for the practical algorithm. Proposition 1 proves a separation in a constructed horizon-3 MDP with a specific revisit pattern, which does not imply benefits in realistic long-horizon CoT generation. Theorem 1 is an oracle variance-reduction result under an idealized linear decomposition; Appendix B makes explicit that the learned gated predictor introduces trajectory-dependent bias that vanishes only when its variation under the current policy is small. A more practical concern is reproducibility: despite detailed hyperparameter tables in Appendix E, the paper does not mention releasing code, model checkpoints, or the filtered training datasets, which blocks independent reproduction. Additionally, the causal evidence from attention masking and injection manipulates internal activation patterns rather than directly intervening on natural revisit behavior, leaving open whether trained revisit is sufficient or merely correlated with other useful computations.
The comparisons to related work appear fair. VPPO, GRPO, and TVC are evaluated under the same SFT initialization and terminal verifier, and the authors are transparent that VPPO yields smaller hallucination gains while TVC without RL underperforms. The signal-identity controls in Table 3 are particularly valuable: replacing the revisit signal $\tilde{r}_t$ with entropy, logit margin, or hidden-state norm under the same RECAP framework underperforms, supporting the specificity of the revisit inductive bias. One limitation is that all benchmarks are transformer-based VLMs with well-defined cross-modal attention; the paper does not test whether the revisit signal generalizes to non-attention or early-fusion architectures.
The experimental protocol is documented in detail: per-backbone hyperparameters, layer ranges for CMA extraction, dynamic-discount settings, and seed counts are reported in Appendix E. Headline results use 8 seeds with confidence intervals. However, no code repository, model checkpoints, or training datasets are mentioned as publicly available. Implementing RECAP requires extracting layer-specific cross-modal attention during RL training, modifying standard advantage computation with log-domain dynamic discounts (Algorithm 1 in Appendix D), and maintaining two value heads—non-trivial engineering that would be greatly aided by a reference implementation. Without released code, independent reproduction of the exact numbers is substantially impeded despite the thorough hyperparameter descriptions.
Long chain-of-thought reasoning improves vision-language models (VLMs), but it also exposes a temporal grounding failure: as generation unfolds, models may drift from image evidence while reinforcement learning still provides only a final-answer reward. We study this gap through visual revisit, a temporally localized reactivation of image evidence during reasoning. Across three VLM backbones, revisit peaks are predictive of correctness and causally linked to performance: masking them degrades accuracy more than masking non-peaks, while injecting revisit-like peaks into failed rollouts partially restores correct answers. Control analyses show that this signal is not explained by raw attention strength, uncertainty, logit margin, or hidden-state norm. We propose RECAP, a GRPO-compatible training method that turns detached revisit traces into a credit-assignment signal. RECAP uses revisit-conditioned advantage estimation, dynamic discounting, and gated value prediction to propagate final rewards through visually important reasoning steps, without extra rewards, additional rollouts, or inference-time changes. Across Qwen2.5-VL-7B, InternVL3-8B, and LLaVA-OV-7B, RECAP consistently improves over GRPO and VPPO, yielding $+6.1$--$+6.5$ HallusionBench gains over the SFT base while preserving reasoning and text-only performance. Results on larger LoRA-tuned models, Chinese zero-shot benchmarks, visual perturbations, and human evaluation further demonstrate robust improvements in image-grounded reasoning.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.