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.
Figures & tables
Figure 1: Visual revisits are causal and localized. Correct rollouts show recurring CMA peaks, while incorrect rollouts lose visual grounding after the initial window. Masking revisit peaks hurts accuracy more than masking random non-peaks, and peak injection partially recovers failed rollouts. A small set of heads explains most of the causal effect.
Figure 2: The RECAP framework. A detached revisit signal r~t is extracted from cross-modal attention and injected into three parts of the surrogate estimator: the trace-augmented input, dynamic discount, and revisit-gated value predictor. The rollout policy, clipped ratio, and KL anchor remain unchanged.
Method
Hallucination
Reasoning
Grounding
Text-only
Δ
Hall. ↑
POPE ↑
MMHal ↑
MathV. ↑
MMMU ↑
MathVr. ↑
MM-Vet ↑
GQA ↑
BLINK ↑
MMVP ↑
MMLU ↑
GSM8K ↑
Qwen2.5-VL-7B
SFT base
60.1
86.5
3.12
59.4
41.3
33.2
51.4
63.8
48.2
54.0
72.4
78.9
–
TVC (no RL)
61.4
86.9
3.18
59.2
41.0
33.5
51.8
64.1
48.5
54.3
72.3
78.7
+1.3
GRPO
62.8
87.1
3.27
61.2
42.5
34.7
52.8
64.3
49.1
54.8
72.1
78.5
+2.7
VPPO
64.2
87.6
3.35
60.8
41.9
34.1
52.1
64.7
48.9
55.1
71.8
77.9
+4.1
Table 1: Main results across three VLM backbones. Values are mean benchmark scores or accuracies over 8 seeds; MMHalBench is reported on its native score scale. Per-seed logs, standard deviations, standard errors, and 95% confidence intervals are reported in the supplementary reproducibility logs. Bold marks the best value within each backbone block; underlines mark the strongest non-RECAP RL baseline per column. A dagger marks a significant RECAP-over-GRPO improvement after Holm–Bonferroni correction. Δ denotes the HallusionBench gain over the per-backbone SFT base.
Configuration
Hallusion ↑
MathVista ↑
MMMU ↑
Var(At) rel. ↓
(a) VPPO orthogonality
GRPO
62.8 [62.0, 63.5]
61.2
42.5
1.81×
VPPO
64.2 [63.3, 65.0]
60.8
41.9
1.35×
RECAP
66.6 [65.7, 67.4]
63.6
42.7
1.00×
VPPO + RECAP (composition)
68.0 [67.1, 68.8]
63.9
42.6
0.98×
(b) Component ablations
Table 2: Ablations on Qwen2.5-VL-7B. Bold marks the best value within each panel and underlines mark the strongest non-best variant. The table compares VPPO composition, component removal, and dynamic-discount reweighting variants across accuracy and surrogate-advantage variance.
Configuration
Qwen2.5-VL 7B ↑
InternVL3 8B ↑
LLaVA-OV 7B ↑
Extra
(a) Signal identity
Token entropy
63.2
61.6
60.1
ρ=0.29
Logit margin
63.0
61.4
59.8
ρ=0.24
Hidden-state norm
63.5
61.9
60.4
ρ=0.36
Revisit r~t (ours)
66.6
64.7
63.4
ρ=1.00
(b) Same signal without RL
Table 3: Signal identity and no-RL controls. Bold marks the best value within each panel; underlines mark the strongest signal-replacement baseline in panel (a) and the strongest no-RL baseline in panel (b). Replacing r~t with surface token signals weakens RECAP, and using the revisit signal without RL does not recover the full gain.
Figure 3: Causal and mechanistic evidence. Top revisit positions are more decision-relevant than random positions, and revisit peaks align with annotated decision steps. The revisit signal further concentrates in a small set of heads, and zeroing those heads reproduces most of the accuracy drop from masking cross-modal attention.
Appendix figures & tables20 assets
Supplementary material from the paper’s appendix.
Appendix
Backbone
Epochs
Peak LR
Batch (tokens)
Max seq. len.
GPU-days
Qwen2.5-VL-7B
3
2×10−5
2×106
4096
12
InternVL3-8B
3
1×10−5
2×106
4096
14
LLaVA-OneVision-7B
3
2×10−5
2×106
4096
12
Qwen2.5-VL-72B (LoRA)
2
1×10−4
1×106
4096
60
Appendix
Table 4: Stage 1 SFT hyperparameters. Per-backbone supervised fine-tuning settings used before Stage 2 RL.
Backbone
γbase
γmax
β
η
history K
Qwen2.5-VL-7B
0.97
0.99
0.6
6
8
InternVL3-8B
0.96
0.99
0.5
6
8
LLaVA-OneVision-7B
0.97
0.99
0.7
6
8
Qwen2.5-VL-72B
0.97
0.99
0.5
6
8
Appendix
Table 5: RECAP Stage 2 RL hyperparameters. Dynamic-discount, revisit-normalization, and history-window settings used for each backbone.
Method
Hall.
POPE
MathV.
MMMU
MathVr.
MMLU
GSM8K
SFT base (72B)
67.4
89.1
68.2
52.9
44.8
81.6
88.4
LoRA-GRPO
69.8
89.6
69.6
53.4
45.7
81.4
88.1
LoRA-VPPO
70.2
90.0
69.0
52.9
45.1
80.9
87.5
LoRA-RECAP
72.9
90.7
71.5
53.7
46.8
81.5
88.2
Δ vs. LoRA-GRPO
+3.1
+1.1
+1.9
+0.3
+1.1
+0.1
+0.1
Appendix
Table 6: Qwen2.5-VL-72B LoRA results under the same rank-64 LoRA setting and RL budget. Bold indicates significant improvement over LoRA-GRPO.
η
Hall.
I
Hall.
W
Hall.
head-pool
Hall.
3
65.9
lower third
64.1
8
65.6
mean (def.)
66.6
6 (def.)
66.6
middle third (def.)
66.6
16 (def.)
66.6
max
65.8
9
66.3
upper third
64.8
32
66.2
top-4
66.5
12
65.4
all layers
63.9
64
65.8
top-8
66.4
The default configuration attains the best per-axis accuracy on all four axes ( 66.6 pt);
this is the same checkpoint entered in four marginals of the grid, not four independently-optimized runs.
Appendix
Table 7: HallusionBench accuracy under marginal sweeps of the revisit-signal pipeline, with the max–min spread over the full hyperparameter grid summarized below. The default setting is robust within ±0.6 pt among nearby reasonable configurations, while the full grid has a 3.1 pt max–min spread. The main source of degradation is all-layer pooling, where top-layer label-projection noise dilutes the revisit estimate.
Figure 4: RAP rises early under RECAP training. RECAP reaches a higher RAP plateau than VPPO, GRPO, and the SFT reference, and the plateau appears before final accuracy convergence.
Figure 5: RAP correlates with HallusionBench accuracy on checkpoints from every method, not just RECAP. Pearson ρ≥0.38 on all four methods with 95% bootstrap CIs strictly above zero ( ≥48 checkpoints per method pooled over three backbones), establishing RAP as a method-agnostic diagnostic. The correlation is highest on RECAP (as expected, since RECAP explicitly optimizes revisit-aware return propagation), but substantial non-RECAP correlations confirm RAP is not a restatement of the training objective.
Method
Qwen2.5-VL-7B
InternVL3-8B
LLaVA-OneVision-7B
Pooled
GRPO
0.44
0.39
0.43
0.42
VPPO
0.49
0.44
0.48
0.47
TVC (no RL)
0.36
0.41
0.37
0.38
RECAP
0.63
0.58
0.62
0.61
Appendix
Table 8: Cross-method RAP–accuracy correlations. Pearson correlations between RAP and HallusionBench accuracy across checkpoints, reported per backbone and pooled across backbones.
Benchmark
Spearman ρ(RAP,accuracy)
HallusionBench
0.61
MathVista
0.34
POPE
0.55
MMMU
0.21
MathVerse
0.30
Appendix
Table 9: Per-benchmark RAP correlations. Spearman correlations between RAP and accuracy on Qwen2.5-VL-7B checkpoints pooled across methods.
Figure 6: RECAP training consolidates revisit responsibility onto fewer heads. The top- 10 head share of the total revisit signal rises from 34% (SFT base) to 54% after RECAP training, while GRPO on identical data produces only a +4 pt shift ( ±3 bootstrap CI). Concentration is therefore specific to revisit-aware optimization rather than to RL post-training in general. Similar pattern on the other two backbones ( section G.3 ).
Top- 10 revisit heads
Random 10 heads
Backbone
AUC
F1
AUC
F1
Qwen2.5-VL-7B
0.68
0.63
0.54
0.50
InternVL3-8B
0.66
0.61
0.53
0.49
LLaVA-OV-7B
0.67
0.62
0.55
0.51
Pooled mean
0.67
0.62
0.54
0.50
Appendix
Table 10: Linear probing on the top-revisit heads. Probes trained on the top- 10 revisit heads predict near-future answer correctness better than probes trained on matched random heads, supporting the causal localization of revisit behavior.
Table 11: Top revisit heads. Top-10 head coordinates for each backbone, sorted by revisit contribution.
k
SFT base
GRPO
RECAP
5
21%
25%
36%
10
34%
38%
54%
20
51%
53%
72%
50
77%
78%
89%
Appendix
Table 12: Pre/post-training head concentration data. Top- k share of the revisit signal before RL, after GRPO, and after RECAP.
Backbone
Median normalized layer
Median normalized head
Qwen2.5-VL-7B
0.46
0.27
InternVL3-8B
0.50
0.31
LLaVA-OneVision-7B
0.43
0.24
Appendix
Table 13: Cross-backbone head-position alignment. Median normalized layer and head coordinates for the top revisit heads on each backbone.
Method score
Agreement
Axis
GRPO
VPPO
RECAP
statistic
Image-groundedness win rate (%)
14.2
23.7
62.1
κ=0.55
Coherence win rate (%)
29.4
30.1
40.5
κ=0.48
Trustworthiness mean ( 1 – 5 )
3.18
3.34
3.87
α=0.51
All axes (McNemar p vs. RECAP)
<10−3
<10−2
–
–
Inter-rater agreement with LLM judges
–
–
72%
–
Appendix
Table 14: Human study. Ten trained annotators rank 100 anonymized samples on image-groundedness, coherence, and trustworthiness. RECAP achieves the highest image-groundedness win rate, with significant gains over both baselines and 72% agreement with the independent LLM-as-judge ranking.
Method
GPT-4V
Claude
Qwen-Max
judge-mean
GRPO
3.21
3.28
3.14
3.21
VPPO
3.48
3.51
3.43
3.47
TVC (no RL)
3.29
3.35
3.22
3.29
RECAP
4.12
4.08
4.05
4.08
Inter-judge Krippendorff’s α
0.63
–
Agreement with human ( table 14 )
72%
–
Appendix
Table 15: LLM-as-judge rankings. Three independent judges score anonymized CoTs for image-grounded reasoning. All judges rank RECAP highest, and their ranking agrees with the human study on most items.
Backbone
Method
CMMMU
MMMU
MathVista-CN
mean Δ
Qwen2.5-VL-7B
GRPO
46.3
39.4
55.1
–
VPPO
46.8
40.0
55.6
+0.5
RECAP
49.1
42.2
57.3
+2.6
InternVL3-8B
GRPO
43.8
37.2
52.7
–
VPPO
44.1
37.6
53.0
+0.3
RECAP
46.2
39.5
54.8
+2.3
Appendix
Table 16: Cross-lingual generalization. RECAP is trained on English data only and evaluated zero-shot on three Chinese benchmarks. It yields consistent gains across backbones, suggesting transfer beyond the English training distribution.
Figure 7: RECAP is consistently more robust to image-side perturbations on Qwen2.5-VL-7B. Across four perturbation families, RECAP has smaller relative accuracy drops than VPPO and maintains a clear advantage over GRPO under stronger perturbations.
Method
Gauss σ=0.2
JPEG Q=10
Patch p=0.3
Distractor
GRPO
53.4
53.9
52.1
58.9
VPPO
54.0
54.6
53.2
60.0
RECAP
58.5
57.0
59.0
63.2
Appendix
Table 17: Accuracy under the strongest image perturbations. HallusionBench accuracy on Qwen2.5-VL-7B for the strongest setting in each perturbation family.
Qwen2.5-VL-7B
InternVL3-8B
LLaVA-OV-7B
Method
mean
std
mean
std
mean
std
SFT + TVC (no RL)
60.9
3.54
59.5
3.31
58.1
3.70
GRPO
62.2
2.81
60.7
2.64
58.8
2.92
VPPO
63.4
2.58
61.8
2.47
59.9
2.73
RECAP
66.0
0.84
64.2
0.91
62.9
0.97
Appendix
Table 18: Robustness to text-side question reformulation. RECAP remains stable across five meaning-preserving reformulations of each HallusionBench question, with lower accuracy variance than GRPO, VPPO, and TVC.
Estimator context
Discount
Value
Reward
Advantage
GRPO
–
static
single
external
raw At
VPPO
–
static
single
external
r~t⋅At (reweight)
TVC
–
static
single
external
raw At
λ -GRPO
–
static
single
external
wt⋅At (learned)
TreeAdv
–
static
single
external
tree- At
RECAP
post-hoc
f(r~t)
gated
external
no local reweight
Appendix
Table 19: Where each method places its per-token visual signal, broken down by the five structural slots in the GRPO update. Only RECAP uses revisit at all three of post-hoc estimator context, discount, and value; every baseline touches exactly one slot (typically the advantage) and leaves the others at their GRPO defaults. This is the structural statement behind the expressiveness gap of Proposition 1 : because dynamic discount acts multiplicatively across future tokens, no single-slot advantage reweighting (VPPO, λ -GRPO) can realize its effect, regardless of how wt is chosen. RECAP still forms the surrogate advantage AtVR in Eq. ( 11 ); “no local reweight” means that it does not multiply an otherwise fixed advantage by a same-step visual weight.
Vision language models (VLMs) are increasingly capable of reasoning over images, but robust visual reasoning often requires re-grounding intermediate steps in the underlying visual evidence. Recent approaches typically rely on external image operations such as zooming or cropping to re-access fine-grained details during inference, which requires additional image re-encoding and can disrupt the reasoning trajectory. We argue that VLMs already provide strong internal signals for identifying and reusing visual evidence, and that these signals can be directly leveraged to support image-grounded reasoning. Motivated by this insight, we propose an end-to-end self-revisit framework, SIEVE, that trains models to re-engage image evidence through internal representations. SIEVE automatically extracts embeddings of salient image regions and injects them into the reasoning chain when additional grounding is needed, enabling later steps to condition on relevant visual cues without external tool calls or re-encoding. We use reinforcement learning to teach the model when to trigger visual revisiting and which region embeddings to retrieve and insert during the reasoning process. Experiments on multiple visual reasoning benchmarks, together with perception, reasoning, and hallucination evaluations, show that SIEVE yields consistent gains, improving performance by 8 percent on average across several benchmarks.
Vision-Language Models (VLMs) often lose visual grounding during multi-step reasoning: as reasoning chains grow longer, later inference steps rely increasingly on language priors rather than image evidence. We identify a consistent benchmark-level signature associated with this degradation: across 2,510 re-examined samples from four benchmarks, attention entropy over image tokens typically decreases during Round 1 and rises again after image re-injection. However, we find that effective visual re-examination requires two complementary ingredients: image re-injection and targeted self-diagnosis. Without targeted diagnosis, re-examination can even hurt performance, whereas accurate self-diagnosis yields substantial gains -- a swing of several points on key benchmarks, indicating that diagnostic quality is a key factor in whether re-examination helps or hurts in our setting. We present ReGround, a two-stage framework that teaches VLMs to self-diagnose grounding failures and selectively re-examine visual evidence, without architectural modifications or external tools. Through capability bootstrapping, a stronger variant from the same model family provides diagnostic scaffolding only during data construction, while the policy model learns to diagnose autonomously at inference time and retains most of the assisted gains. Experiments on eight benchmarks across two VLM backbones demonstrate consistent gains, especially on visually intensive multi-step reasoning tasks, while incurring only modest inference overhead relative to tool-augmented baselines. Project page: https://sespoir.github.io/reground-page/ . Code: https://github.com/sespoir/ReGround .
Lei Peng, Shuai Lv, Wei Hu
University of Science and Technology of China School of Artificial Intelligence and Data Science Hefei, Anhui, China · University of Science and Technology of China State Key Laboratory of Precision and Intelligent Chemistry Hefei, Anhui, China
Chain-of-thought (CoT) reasoning has significantly improved the reasoning ability of large vision-language models (LVLMs) by verbalizing intermediate reasoning steps in natural language. However, such discrete textual rationales are often insufficient for encoding continuous visual evidence. Recent work addresses this limitation by moving reasoning into continuous latent space. Despite promising progress, existing methods leave latent reasoning insufficiently connected to the compositional and relational structure of visual evidence. To address this gap, we introduce ReGuLaR, a relation grounded latent reasoning framework that explicitly grounds latent states in these critical yet overlooked visual evidence. ReGuLaR uses a training-time ReGFormer to focus latent reasoning on question-relevant objects and inter-object relations, while at inference time the model reasons and generates answers without invoking the ReGFormer. To support training ReGuLaR, we construct RGROUNDING-351K, a real-world vision-language dataset annotated with key object bounding boxes and inter-object relations. Extensive experiments across diverse benchmarks show that ReGuLaR consistently outperforms existing approaches and achieves state-of-the-art performance. We include our code in the submission and will release the code and training data publicly upon acceptance.