acceptodds
Under review as a conference paper at ICLR 2027

Seeing Before Believing: Mitigating Hallucinations in Multimodal Reasoning via Visually-Anchored Latent Rectification

Abstract

Although long-chain reasoning aids complex multimodal tasks, the extended thinking trajectory can drift from image evidence and produce hallucinated answers. Recent work attempts to rectify reasoning by optimizing latent states guided by reference-free predictive confidence. However, predictive confidence becomes misleading when greater certainty reflects linguistic predictability rather than visual faithfulness. When reasoning detaches from the image, self-reinforcing language priors can concentrate probability on predictable continuations, producing a bungee-like rebound in confidence and a false sense of certainty without visual support. Optimizing against confidence alone thus promotes visually ungrounded shortcuts, amplifying such high-confidence errors. To address this, we introduce Visually-Anchored Latent Rectification (VALR), a training-free method that uses visual conditioning to guide confidence-based latent refinement. VALR dynamically retrieves and fuses aligned image evidence into candidate states, shaping the reward landscape around visually induced confidence gains to guide exploration and select the state used for answer generation. Across three models, VALR outperforms all compared inference-time methods in hallucination mitigation, while its benefits extend to broader multimodal understanding and mathematical reasoning.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Long-chain Reasoning, Hallucination Mitigation, Test-time Reasoning

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_PUUVOpenReview ↗
Rating
4
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates how confidence latent optimization can enhance reasoning capabilities in visually unsupported scenarios.

VALR is a training-free method that combines dynamic evidence retrieval, and entropy-based pairwise rewards.

Experiments with the reasoning model demonstrate that VALR improves upon both hallucination and reasoning benchmarks.

Critical strengths

The paper provides a useful analysis of the visual, that confidence can recover along incorrect reasoning trajectories without increased visual attention.

The experimental evaluation is extensive, covering four benchmarks and three reasoning model.

Critical weaknesses

All three evaluated checkpoints also share the Qwen2.5-VL architecture and 7B scale

Questions for authors

Do you have evidence that the method transfers beyond the Qwen2.5-VL 7B family

3 comments

We sincerely thank the reviewer for the careful assessment and for recognizing the value of our empirical analysis and benchmark evaluation. We also appreciate the question about transfer beyond Qwen2.5-VL 7B, which we address below with additional experiments across model families and scales.

1. VALR Improves Accuracy on Non-Qwen and 4B/8B Qwen3-VL Reasoning Models{\color{#ff7f00}\text{1. VALR Improves Accuracy on Non-Qwen and 4B/8B Qwen3-VL Reasoning Models}}

Additional experiments show that VALR improves hallucination mitigation and MMStar accuracy on a non-Qwen reasoning model and Qwen3-VL reasoning models at both 4B and 8B scales.

We evaluate MM-Eureka-8B, a reasoning model based on InternVL2.5 with an InternLM2.5 language backbone, together with Qwen3-VL-4B-Thinking and Qwen3-VL-8B-Thinking. All methods use the same evaluation settings within each backbone.

(a) Hallucination-mitigation gains extend to the additional reasoning backbones.

Table 1. HallusionBench results. Accuracy and F1 are percentages.

MethodMM-Eureka-8B Acc.MM-Eureka-8B F1Qwen3-VL-4B-Thinking Acc.Qwen3-VL-4B-Thinking F1Qwen3-VL-8B-Thinking Acc.Qwen3-VL-8B-Thinking F1
Vanilla64.4261.8669.9668.1371.7869.82
MemVR64.9162.3870.8468.4672.5970.22
DMLR65.7663.2870.7368.1472.7770.69
VALR (Ours)66.9864.9371.9270.0174.4172.26

VALR achieves the highest accuracy and F1 on all three backbones, with accuracy gains of 1.08–1.64 percentage points over the strongest compared baseline for each backbone. The MM-Eureka-8B results provide direct evidence of transfer to a non-Qwen reasoning architecture.

(b) The gains also extend to broader multimodal understanding and reasoning.

Table 2. MMStar accuracy (%).

MethodMM-Eureka-8BQwen3-VL-4B-ThinkingQwen3-VL-8B-Thinking
Vanilla61.6973.2075.30
MemVR62.2973.8275.84
DMLR62.4773.5676.31
VALR (Ours)64.3174.7177.16

On MMStar, VALR improves accuracy by 0.85–1.84 percentage points over the strongest compared baseline for each backbone. Together, these results support the transfer of the complete VALR procedure across the tested architectures and scales, covering both hallucination mitigation and broader multimodal understanding and reasoning.

These results are now included in Section 5.2 and Appendix D.1 of the revised manuscript.

Thank you for your response. My overall assessment remains positive, as the work is thorough. Conceptually, I see it primarily as a refinement of the established multimodal latent reasoning paradigm exemplified by DMLR and LEAD.

AuthorsOpenReview ↗

Thank You for Your Thoughtful Review

Thank you for your thoughtful review and positive assessment of our work. We are pleased that our responses and additional experiments have addressed your main concerns. Your feedback has helped us strengthen the experimental evaluation and improve the clarity of the paper.

Reviewer_ikJJOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies a failure mode in multimodal reasoning: confidence can recover without renewed support from the image. VALR optimizes a block of continuous input embeddings before answer generation, using the entropy difference between visually conditioned and matched control states as its reward.

Critical strengths
  1. The paper shows that confidence can recover without renewed visual support. The trajectory analysis and controlled visual interventions provide useful evidence for this failure mode.

  2. Evaluating the same candidate with and without retrieved visual content offers a reasonable alternative to optimizing absolute confidence. Norm matching makes the comparison better controlled.

Critical weaknesses
  1. The contribution needs a sharper comparison with existing visual latent optimization. VALR shares its search procedure with prior methods and combines visual initialization, retrieval, fusion, and a new reward. DMLR is included, but the comparison changes several components simultaneously. A reward-only replacement within a common framework would better isolate the contribution.

  2. A larger entropy difference can arise from lower entropy in the conditioned branch or higher entropy in the control branch. Neither alone establishes that the retrieved evidence supports the correct answer. Table 16 provides encouraging evidence, but I would like to see the two entropy terms separately and a fixed-candidate test using relevant versus irrelevant visual content. Sensitivity to the vocabulary truncation parameter kk would also help establish that candidate rankings are robust.

  3. All three models use Qwen2.5-VL, and the main hallucination results concern binary questions. A different architecture or a quantitative open-ended evaluation would strengthen the generality claim. In addition, the question-level confidence intervals assume independence despite related questions sharing images or control sets; grouped resampling would be more appropriate.

Questions for authors

Does paired reward remain better than absolute confidence when initialization, retrieval, fusion, and computation are held fixed? Do higher rewards consistently correspond to more relevant visual evidence, rather than mainly reflecting changes in control-branch entropy?

4 comments
AuthorsOpenReview ↗

Response (Part I)

We thank the reviewer for the thoughtful assessment and for recognizing the rationale behind our paired reward and the value of our controlled visual interventions. We address each concern below with clarifications and additional evidence.

1. The Paired Reward Outperforms Absolute Confidence at Matched Inference Time{\color{#ff7f00}\text{1. The Paired Reward Outperforms Absolute Confidence at Matched Inference Time}}

Thank you for suggesting a controlled reward comparison. Our reward improves accuracy within both DMLR and VALR when their other components are held fixed, and these gains remain under matched total inference time.

Here, “absolute” uses absolute confidence, while “paired” uses the entropy difference between visually conditioned and matched control branches.

(a) Replacing only the reward improves both search frameworks.

We replace only the reward, retaining each framework’s initialization, visual retrieval and fusion, update rule, decoding settings, and other parameters. All four search configurations evaluate 15 candidates per question.

Table 1. Reward replacement with 15 candidates per question. Tables 1–2 use R1-OneVision-7B. Acc./F1 are percentages, and ΔAcc. is paired − absolute within each framework (percentage points).

ConfigurationPOPE-A Acc.POPE-A F1ΔAcc.HallusionBench Acc.HallusionBench F1ΔAcc.
Vanilla82.6783.04—61.4158.90—
DMLR + absolute83.6784.02—62.4660.29—
DMLR + paired84.6384.92+0.9663.9361.67+1.47
VALR + absolute84.0384.30—62.7860.49—
VALR + paired85.8385.93+1.8064.7762.57+1.99

Accuracy and F1 improve in both frameworks on both benchmarks. The reward therefore improves DMLR’s existing search without requiring VALR’s other components.

(b) The gains persist after accounting for the additional scoring cost.

We tune DMLR and VALR with equal trials on the same 500 held-out ScienceQA image questions, sharing selected settings between reward versions within each framework. Candidate counts may differ to match total measured time, including scoring and final generation, but are calibrated on validation data and fixed before testing.

Table 2. Performance under matched total inference time.

BenchmarkConfigurationTotal time (s)Acc.ΔAcc.
POPE-ADMLR + absolute1568.4984.06—
POPE-ADMLR + paired1568.3984.44+0.38
POPE-AVALR + absolute1567.4984.49—
POPE-AVALR + paired1568.2185.68+1.19
HallusionBenchDMLR + absolute1068.1962.78—
HallusionBenchDMLR + paired1067.9963.79+1.01
HallusionBenchVALR + absolute1067.6963.27—
HallusionBenchVALR + paired1068.4364.74+1.47

With the additional scoring cost included, the paired reward yields more accurate answers within the same time budget, improving accuracy by 0.38–1.47 percentage points across both frameworks and benchmarks.

We have added the reward and budget controls to Section 5.3 and Appendix D.5 of the revised manuscript.

AuthorsOpenReview ↗

Response (Part II)

2. Relevant Visual Content Raises Rewards and Selected-Answer Accuracy{\color{#ff7f00}\text{2. Relevant Visual Content Raises Rewards and Selected-Answer Accuracy}}

Thank you for suggesting direct tests of reward behavior and ranking robustness. Relevant visual content receives higher mean rewards and improves selected-answer accuracy, while candidate rankings remain similar across k=8k=8–6464. All diagnostics use R1-OneVision-7B.

(a) Correct candidates’ reward growth mainly reflects lower conditioned entropy.

We decode each question’s three highest-reward candidates among 15, group them by their own answers’ correctness, and measure changes from the first evaluated candidate. Correctness labels are used only for this analysis.

Table 3. Mean entropy and reward changes (nats). Δr=ΔHctrl−ΔHcond\Delta r=\Delta H_{\mathrm{ctrl}}-\Delta H_{\mathrm{cond}}.

BenchmarkCandidate groupΔHcond\Delta H_{\mathrm{cond}}ΔHctrl\Delta H_{\mathrm{ctrl}}Δr\Delta r
POPE-ACorrect−0.182+0.012+0.194
POPE-AIncorrect−0.022+0.127+0.149
HallusionBenchCorrect−0.226+0.021+0.247
HallusionBenchIncorrect−0.159+0.018+0.177

In both correct groups, reward growth mainly reflects more certain predictions in the visually conditioned branch. Incorrect candidates show increased control uncertainty on POPE-A and more certain but incorrect predictions on HallusionBench.

(b) Relevant content improves rewards and selected-answer accuracy.

We fix 15 original proposals per question and vary only the newly fused content between relevant and same-image irrelevant content. Relevance is judged manually without model rewards or answers. Both conditions use 16 visual tokens per latent position, with other settings matched.

Table 4. Controlled content replacement. Rewards average all candidates. Accuracy evaluates the highest-reward candidate selected under each condition.

BenchmarkVisual contentMean reward rrSelection Acc. (%)
POPE-ARelevant0.26785.23
POPE-ASame-image irrelevant0.07183.37
HallusionBenchRelevant0.32664.56
HallusionBenchSame-image irrelevant0.13361.93

Relevant content improves selected-answer accuracy by 1.86 percentage points on POPE-A and 2.63 on HallusionBench. Higher mean rewards accompany more accurate selected answers, supporting VALR’s use of confidence gains associated with retrieved content to guide optimization.

(c) Rankings and selection remain similar across k=8k=8–6464.

With candidate pools and model output probabilities fixed, we vary vocabulary truncation kk and report selection accuracy and within-question Spearman rank correlation ρ\rho relative to k=16k=16.

Table 5. Sensitivity to kk. Accuracy is in percent.

kkPOPE-A Acc.POPE-A ρ\rhoHallusionBench Acc.HallusionBench ρ\rho
484.730.82863.090.828
884.970.94963.720.927
16 (default)85.171.00064.141.000
3285.370.96964.460.977
6485.070.95364.140.936

For k=8,32,64k=8,32,64, accuracy differs from the default by at most 0.20 percentage points on POPE-A and 0.42 on HallusionBench, with respective rank correlations of at least 0.949 and 0.927. These settings preserve similar rankings and answer accuracy, while k=4k=4 produces larger changes.

We now summarize these analyses in Sections 5.3–5.4 of the revised manuscript, with content and entropy results in Appendix D.8 and kk sensitivity in Appendix D.6.

AuthorsOpenReview ↗

Response (Part III)

3. VALR Improves Accuracy on Non-Qwen and 4B/8B Qwen3-VL Reasoning Models{\color{#ff7f00}\text{3. VALR Improves Accuracy on Non-Qwen and 4B/8B Qwen3-VL Reasoning Models}}

Thank you for suggesting evaluation beyond Qwen2.5-VL. VALR improves hallucination mitigation and MMStar accuracy on a non-Qwen reasoning model and Qwen3-VL Thinking models at both 4B and 8B.

We evaluate MM-Eureka-8B, a reasoning model based on InternVL2.5 with an InternLM2.5 language backbone, together with Qwen3-VL-4B-Thinking and Qwen3-VL-8B-Thinking. All methods use the same evaluation settings within each backbone.

(a) Hallucination-mitigation gains extend to the additional reasoning backbones.

Table 6. HallusionBench results. Accuracy and F1 are percentages.

MethodMM-Eureka-8B Acc.MM-Eureka-8B F1Qwen3-VL-4B-Thinking Acc.Qwen3-VL-4B-Thinking F1Qwen3-VL-8B-Thinking Acc.Qwen3-VL-8B-Thinking F1
Vanilla64.4261.8669.9668.1371.7869.82
MemVR64.9162.3870.8468.4672.5970.22
DMLR65.7663.2870.7368.1472.7770.69
VALR (Ours)66.9864.9371.9270.0174.4172.26

VALR achieves the highest accuracy and F1 among the compared methods on all three backbones, exceeding each backbone’s strongest compared baseline in accuracy by 1.08–1.64 percentage points. The gain on MM-Eureka provides direct evidence of effectiveness on a non-Qwen reasoning architecture.

(b) The transfer also benefits broader multimodal understanding and reasoning.

Table 7. MMStar accuracy (%).

MethodMM-Eureka-8BQwen3-VL-4B-ThinkingQwen3-VL-8B-Thinking
Vanilla61.6973.2075.30
MemVR62.2973.8275.84
DMLR62.4773.5676.31
VALR (Ours)64.3174.7177.16

On the same three backbones, VALR exceeds the strongest compared baseline on MMStar by 0.85–1.84 percentage points, showing that its benefits extend to broader multimodal tasks.

These results are now included in Section 5.2 and Appendix D.1 of the revised manuscript.

4. VALR Reduces Object Hallucination and Improves Coverage on AMBER{\color{#ff7f00}\text{4. VALR Reduces Object Hallucination and Improves Coverage on AMBER}}

Thank you for suggesting a quantitative open-ended evaluation. On AMBER, VALR also reduces object hallucination in freely generated image descriptions while improving coverage of annotated objects.

We evaluate R1-OneVision-7B on all 1,004 images in AMBER’s generative task, with identical official prompts, decoding settings, and output limits across methods. Models retain their normal reasoning process, while official AMBER scoring and length measurements apply only to final descriptions.

Table 8. AMBER image-description results.

MethodCHAIR ↓ (%)Cover ↑ (%)Mean description length (tokens)
Vanilla8.353.3140.64
MemVR7.354.3139.23
DMLR7.854.6139.11
VALR (Ours)6.955.9142.86

VALR achieves the lowest CHAIR and highest Cover among the compared methods, with CHAIR 1.4 percentage points lower and Cover 2.6 higher than Vanilla. Object hallucination decreases while coverage of annotated objects improves, with a slight increase in average description length.

We have added the AMBER results and evaluation protocol to Appendix D.2, with a summary in Section 5.2 of the revised manuscript.

AuthorsOpenReview ↗

Response (Part IV)

5. Grouped Resampling Supports Full VALR’s Accuracy Gains on POPE-A and HallusionBench{\color{#ff7f00}\text{5. Grouped Resampling Supports Full VALR’s Accuracy Gains on POPE-A and HallusionBench}}

Thank you for recommending grouped resampling. The full-method gains remain supported across all six evaluated settings when related questions are resampled together.

We reanalyze original manuscript Tables 14 and 16, now Tables 18 and 23 in the revised manuscript, using percentile 95% CIs from 10,000 paired group-bootstrap replicates with question-average weighting. POPE-A groups 3,000 questions by image (500 groups), while HallusionBench groups 951 questions by shared images or control sets (167 cases). Gains and CIs are in percentage points.

(a) Full-method gains remain supported after accounting for dependencies among questions.

Table 9 revisits the comparisons against the strongest listed baseline for each setting in original manuscript Table 14. We retain the original three VALR runs, averaging correctness within each question before comparison with the fixed baseline prediction.

Table 9. Full VALR versus the strongest listed baseline.

BackboneBenchmarkAccuracy gainGrouped 95% CI
R1-OneVisionPOPE-A+2.03[+1.26, +2.83]
R1-OneVisionHallusionBench+2.21[+0.42, +3.84]
Vision-R1POPE-A+2.00[+1.25, +2.78]
Vision-R1HallusionBench+2.21[+0.49, +3.83]
ThinkLite-VLPOPE-A+2.13[+1.43, +2.81]
ThinkLite-VLHallusionBench+2.21[+0.80, +3.56]

All six intervals remain above zero. VALR’s higher accuracy on these hallucination benchmarks therefore retains statistical support when related questions are resampled together.

(b) Candidate-selection gains retain statistical support on two HallusionBench backbones.

We also reanalyze original manuscript Table 16, where VALR’s paired reward and absolute confidence select from the same 15 candidates per question. The 304 MathVision questions have independent sources, so each forms its own resampling group.

Table 10. Selection accuracy gain of paired reward over absolute confidence.

BackboneBenchmarkAccuracy gainGrouped 95% CI
R1-OneVisionHallusionBench+2.21[+0.15, +4.29]
Vision-R1HallusionBench+1.68[−0.09, +3.52]
ThinkLite-VLHallusionBench+2.00[+0.21, +3.86]
R1-OneVisionMathVision+2.96[−0.99, +6.91]
Vision-R1MathVision+2.30[−1.32, +5.92]
ThinkLite-VLMathVision+1.97[−1.64, +5.59]

The observed selection gains are positive in all six settings. The HallusionBench intervals exclude zero for R1-OneVision and ThinkLite-VL. For Vision-R1 on HallusionBench and all three MathVision comparisons, the intervals include zero, so the current samples do not yet confirm the selection advantage in those settings.

We have updated the intervals and statistical methods in Appendices D.3 and D.6 of the revised manuscript, together with the corresponding interpretations in Sections 5.2–5.3.

Reviewer_EDX9OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies a limitation of confidence-guided multimodal reasoning: predictive confidence can recover without renewed reliance on image evidence, and confidence-only latent exploration can increase high-confidence errors. It proposes Visually-Anchored Latent Rectification (VALR), a training-free, input-specific optimization procedure. Question-relevant visual features initialize a fixed block of latent input embeddings. Gaussian proposals receive candidate-dependent retrieved visual features and are scored against norm-matched controls that omit the newly retrieved content. The difference in predictive entropy guides a score-function update and selection of the state used for generation. Both scoring branches retain the original image; optimization precedes answer generation rather than intervening online in a generated chain of thought.

Critical strengths
  1. The proposed objective addresses a concrete failure of absolute-confidence optimization, with meaningful supporting controls. The trajectory observations are complemented by suppression/restoration interventions, including non-visual suppression, norm-matched noise, and shuffled-content controls (Section 3.3; Tables 3 and 12). More importantly for the method itself, Table 7 includes component ablations and proposal/branch-budget-matched random-search and random-retrieval controls. Table 16 compares absolute and paired scoring on identical candidate pools: on HallusionBench the paired criterion improves selection accuracy by 1.68–2.21 percentage points across the three checkpoints, with all reported 95% intervals above zero. This is useful evidence for the scoring mechanism, not merely an improvement of an entire system over Vanilla.

  2. The empirical gains are consistent, and the implementation is described in considerable detail. On POPE-Adversarial and HallusionBench, the reported improvements over the strongest listed baselines are approximately 2.00–2.21 percentage points; Tables 14 and 15 provide paired intervals and three-seed stability. Gains also extend to MMStar and MathVision (Table 6). The paper specifies the initialization, paired branches, cache separation, candidate retention, validation protocol, and optimization budget in Appendix C. It reports inference costs rather than presenting training-free inference as cost-free; Table 8(c) places VALR close to DMLR in measured time while obtaining higher accuracy.

Critical weaknesses
  1. The most distinctive methodological increment is not fully separated from the other changes relative to DMLR. Table 13 makes clear that DMLR already uses Gaussian latent exploration, confidence-guided REINFORCE, and dynamic visual integration. VALR changes initialization, retrieval, fusion, reward, and state selection together. The existing ablations and fixed-pool experiment support the proposed components within VALR, but do not establish how much of the advantage over DMLR would remain if only the reward were changed in an otherwise fixed implementation. This matters to the significance and attribution of the contribution, not because the paper lacks ablations. Appendix C.1 also tunes VALR's fusion weight on a separate validation set while retaining published baseline settings. That is not test-set leakage, but an equal-validation-budget comparison would make the relatively modest margin over strong baselines more convincing.

  2. The relative-confidence signal is empirically useful, but its interpretation and demonstrated scope remain limited. The reward in Equation 11 can increase through changes in either the conditioned entropy or the control entropy; Appendix C.4 appropriately acknowledges this. A higher score is therefore not itself evidence that the retrieved interpretation is factually correct. Table 16 gives a direct selection test on HallusionBench, but the corresponding MathVision diagnostic intervals include zero. Likewise, suppression/restoration establishes that particular visual content affects predictions under that intervention; it does not fully identify the mechanism of naturally occurring hallucinations. Finally, the three evaluated checkpoints all use the Qwen2.5-VL architecture at 7B (Appendix C.1). The evidence supports a useful inference-time procedure in this family more strongly than a generally established visual-grounding criterion. These limitations moderate the claimed insight and breadth rather than invalidate the reported benchmark gains.

Questions for authors
  1. Attribution to the paired objective: How much of the gain can be recovered by replacing DMLR's absolute-confidence objective with the paired criterion while retaining its state construction and visual-integration procedure? A staged comparison under the same validation and measured-time budgets would directly clarify the contribution beyond DMLR. Please distinguish this request from the already provided within-VALR ablations.

  2. Reward behavior on failures: Can the authors report conditioned entropy, control entropy, and final correctness for high-reward successful and unsuccessful candidates? In particular, does optimization predominantly reduce conditioned entropy, increase control entropy, or both? A modest quantitative breakdown and representative failure cases would clarify when the paired signal ceases to indicate useful visual support; a theoretical guarantee is not required.

  3. Scope of the empirical conclusion: Is there evidence from a different architecture or scale, or an explanation of which parts of the method depend on the shared Qwen2.5-VL embedding/token setup? Such evidence would materially strengthen the breadth of the result. If unavailable, the conclusion should remain explicitly scoped to the evaluated family. The MathVision selector diagnostic should also remain described as inconclusive at the reported sample size, rather than as an established independent effect.

5 comments
AuthorsOpenReview ↗

Response (Part I)

We thank the reviewer for the thoughtful assessment and recognition of our controlled analyses and empirical results. Below, we address the paired objective’s contribution and VALR’s empirical scope with additional experiments and analysis.

1. The Paired Reward Improves DMLR with Other Components Fixed{\color{#ff7f00}\text{1. The Paired Reward Improves DMLR with Other Components Fixed}}

Thank you for suggesting a reward-only comparison within DMLR. Replacing its absolute-confidence objective with our paired reward improves accuracy on both benchmarks while retaining DMLR’s other components.

(a) Reward replacement improves DMLR.

We change only the reward, preserving each framework’s other settings, including DMLR’s state construction and visual integration. All four search configurations evaluate 15 candidates per question, extending the within-VALR tests in original manuscript Tables 7 and 16 to DMLR’s adaptive search.

Table 1. Reward replacement with 15 candidates per question. Tables 1–3 use R1-OneVision-7B. ΔAcc. is paired − absolute within each framework (percentage points).

ConfigurationPOPE-A Acc.POPE-A F1ΔAcc.HallusionBench Acc.HallusionBench F1ΔAcc.
Vanilla82.6783.04—61.4158.90—
DMLR + absolute83.6784.02—62.4660.29—
DMLR + paired84.6384.92+0.9663.9361.67+1.47
VALR + absolute84.0384.30—62.7860.49—
VALR + paired85.8385.93+1.8064.7762.57+1.99

DMLR gains 0.96 percentage points on POPE-A and 1.47 on HallusionBench, with higher F1 on both. Thus, the paired reward improves DMLR’s search without requiring VALR’s other components.

(b) VALR stays ahead with equal tuning opportunities.

Both methods use the same 500 held-out, image-based ScienceQA questions, trial counts, seeds, and per-configuration inference budgets. Each searches its native parameters, including published defaults, and fixes the configuration selected by validation accuracy across both benchmarks. Table 2 changes only the tuning protocol.

Table 2. Complete-method comparison with equal validation tuning trials. Original: published DMLR settings, validation-selected VALR.

Selection protocolPOPE-A DMLR Acc.POPE-A VALR Acc.VALR−DMLRHallusionBench DMLR Acc.HallusionBench VALR Acc.VALR−DMLR
Original protocol83.6785.83+2.1662.4664.77+2.31
Equal tuning trials84.2485.92+1.6863.1364.84+1.71

Tuning improves DMLR, but VALR remains ahead by 1.68 and 1.71 percentage points on POPE-A and HallusionBench. This shows that VALR’s advantage is not solely due to the original tuning difference.

(c) DMLR’s reward gains persist at matched inference time.

Using the parameters from (b), we match total inference time, including scoring and final generation. Candidate budgets are calibrated on validation data and fixed before testing, allowing counts to differ. Other non-reward settings remain shared within each framework.

Table 3. Performance under matched total inference time.

BenchmarkConfigurationTotal time (s)Acc.ΔAcc.
POPE-ADMLR + absolute1568.4984.06—
POPE-ADMLR + paired1568.3984.44+0.38
POPE-AVALR + absolute1567.4984.49—
POPE-AVALR + paired1568.2185.68+1.19
HallusionBenchDMLR + absolute1068.1962.78—
HallusionBenchDMLR + paired1067.9963.79+1.01
HallusionBenchVALR + absolute1067.6963.27—
HallusionBenchVALR + paired1068.4364.74+1.47

Within DMLR, the paired reward still adds 0.38 and 1.01 percentage points on POPE-A and HallusionBench, so its benefit persists when the extra scoring cost is included. Full VALR leads DMLR + absolute by 1.62 and 1.96 points under these combined tuning and time controls.

We have added these comparisons to Section 5.3 and Appendix D.5 of the revised manuscript.

AuthorsOpenReview ↗

Response (Part II)

2. Mean Reward Growth Mainly Reflects Confidence Gains with Retrieved Content{\color{#ff7f00}\text{2. Mean Reward Growth Mainly Reflects Confidence Gains with Retrieved Content}}

Thank you for asking how reward growth relates to visual evidence. The results support VALR’s central idea: comparing confidence with and without newly retrieved content helps select more correct answers than absolute confidence alone, as shown on the fixed HallusionBench candidate pools.

(a) Reward growth mainly reflects increased confidence with retrieved content.

Using R1-OneVision-7B, we decode each question’s three highest-reward VALR candidates among 15 and group them by their answers’ correctness. Reference answers are used only for this analysis. The conditioned branch includes the newly retrieved content, while the control omits that addition. Both retain the original image.

Table 4. Entropy and correctness of high-reward candidates (group means, nats). r=Hctrl−Hcondr=H_{\mathrm{ctrl}}-H_{\mathrm{cond}}. N (%) describes the analyzed candidate pool.

BenchmarkCandidate groupN (%)H_condH_ctrlr
POPE-ACorrect7511 (83.46%)1.7462.0970.351
POPE-AIncorrect1489 (16.54%)1.7192.0880.369
HallusionBenchCorrect1777 (62.29%)2.1242.5710.447
HallusionBenchIncorrect1076 (37.71%)2.2792.6910.412

Table 5 reports the same candidates’ mean changes relative to each question’s first evaluated candidate.

Table 5. Mean entropy and reward changes (nats). Δr=ΔHctrl−ΔHcond\Delta r=\Delta H_{\mathrm{ctrl}}-\Delta H_{\mathrm{cond}}.

BenchmarkCandidate groupΔH_condΔH_ctrlΔr
POPE-ACorrect−0.182+0.012+0.194
POPE-AIncorrect−0.022+0.127+0.149
HallusionBenchCorrect−0.226+0.021+0.247
HallusionBenchIncorrect−0.159+0.018+0.177

Mean reward growth mainly comes from lower conditioned entropy, consistent with VALR’s aim to reward confidence gains associated with retrieved content. Incorrect candidates show two failure patterns: increased control uncertainty on POPE-A, and more certain but incorrect predictions on HallusionBench.

(b) Relevant content receives higher rewards and improves selected answers.

With 15 proposals and other settings fixed, relevant versus same-image irrelevant content raises mean candidate reward and selection accuracy (+1.86 points on POPE-A, +2.63 on HallusionBench). Separate matched suppression/restoration also yields fewer hallucinations with original than shuffled content (original manuscript Tables 3 and 12).

On fixed HallusionBench pools, paired scoring improves accuracy over absolute confidence by 1.68–2.21 points across three backbones (revised manuscript Table 23). Grouped CIs exclude zero for R1-OneVision and ThinkLite-VL.

We have added the content comparison to Section 5.4 and the full content and entropy analyses to Appendix D.8 of the revised manuscript.

AuthorsOpenReview ↗

Response (Part III)

3. VALR Improves Accuracy on Non-Qwen and 4B/8B Qwen3-VL Reasoning Models{\color{#ff7f00}\text{3. VALR Improves Accuracy on Non-Qwen and 4B/8B Qwen3-VL Reasoning Models}}

Thank you for suggesting a broader evaluation. VALR’s benefits extend beyond Qwen2.5-VL 7B, improving hallucination mitigation and MMStar accuracy on a non-Qwen reasoning model and Qwen3-VL Thinking models at both 4B and 8B.

We evaluate MM-Eureka-8B to test transfer beyond Qwen. It is based on InternVL2.5 with an InternLM2.5 language backbone and trained through multimodal reinforcement learning. Qwen3-VL-4B-Thinking and Qwen3-VL-8B-Thinking cover two model scales. All methods use the same evaluation settings within each backbone.

(a) Hallucination-mitigation gains extend across the tested architectures and scales.

Table 6. HallusionBench results on the additional reasoning backbones. Accuracy and F1 are percentages.

MethodMM-Eureka-8B Acc.MM-Eureka-8B F1Qwen3-VL-4B-Thinking Acc.Qwen3-VL-4B-Thinking F1Qwen3-VL-8B-Thinking Acc.Qwen3-VL-8B-Thinking F1
Vanilla64.4261.8669.9668.1371.7869.82
MemVR64.9162.3870.8468.4672.5970.22
DMLR65.7663.2870.7368.1472.7770.69
VALR (Ours)66.9864.9371.9270.0174.4172.26

VALR achieves the highest accuracy and F1 among the compared methods on all three backbones, exceeding the strongest baseline in accuracy by 1.08–1.64 percentage points. These gains extend VALR’s hallucination mitigation to the tested non-Qwen architecture and both Qwen3-VL scales.

(b) The same models also gain on broader multimodal tasks.

Table 7. MMStar accuracy (%) on the additional reasoning backbones.

MethodMM-Eureka-8BQwen3-VL-4B-ThinkingQwen3-VL-8B-Thinking
Vanilla61.6973.2075.30
MemVR62.2973.8275.84
DMLR62.4773.5676.31
VALR (Ours)64.3174.7177.16

On MMStar, VALR exceeds the strongest compared baseline by 0.85–1.84 percentage points. VALR improves both hallucination mitigation and broader multimodal accuracy on these reasoning models, supporting its use for inference-time refinement beyond the original model family.

These backbone results are now included in Section 5.2 and Appendix D.1 of the revised manuscript.

We have also clarified the MathVision selector discussion in Appendix D.6 (revised manuscript Table 23): with candidate pools fixed, our scoring rule selects more correct answers than confidence alone in the observed sample. This selection advantage is not yet established at the current sample size.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank the AC and all three reviewers for their careful evaluation and constructive suggestions. Across the three reviewer responses, we provided additional experiments and analyses summarized in 12 distinct result tables, and incorporated the corresponding findings into the revised manuscript. The ratings remain 3/3/4 (EDX9/ikJJ/PUUV), unchanged from the initial reviews.

Below we summarize the discussion, including positive follow-up feedback from two reviewers:

  • EDX9 recognized that our objective “addresses a concrete failure of absolute-confidence optimization” and subsequently stated that our response addressed “most of my concerns,” maintaining the original rating.
  • PUUV reaffirmed the thoroughness of the work: “My overall assessment remains positive, as the work is thorough.”
  • We regret that ikJJ has not participated in the follow-up discussion. We have provided detailed responses to all of their concerns and incorporated the corresponding experiments, analyses, and clarifications into the revised manuscript.

Core contribution. Higher confidence does not necessarily indicate stronger visual support. We identify confidence recovery without renewed visual reliance in incorrect trajectories, while controlled interventions show that original visual content helps correct induced errors. These findings motivate VALR, a training-free method that dynamically retrieves aligned visual evidence. Its paired objective compares the same candidate with and without newly retrieved content, using the confidence gain to guide latent search and answer selection. Experiments show gains under matched budgets, with benefits extending to additional reasoning backbones and open-ended generation.

The table below summarizes the additional evidence for each main concern and how these additions strengthen support for the paper’s central findings.

Key concernReviewersOur Response
Contribution beyond DMLR and fairness of comparisonEDX9, ikJJReplacing only DMLR’s reward with VALR’s paired objective improves accuracy. Full VALR also stays ahead with equal tuning opportunities and matched inference time.
Whether the reward reflects useful visual supportEDX9, ikJJVALR’s reward favors useful visual evidence: relevant content yields higher rewards and more accurate selected answers than same-image irrelevant content.
Transfer across model architectures and scalesAll threeVALR generalizes across architectures and scales, achieving the best accuracy among compared methods on HallusionBench and MMStar with the added reasoning models.
Hallucination mitigation in open-ended generationikJJVALR reduces hallucinations while improving object coverage on AMBER, extending its benefits to open-ended image descriptions.
Statistical reliability with related questionsikJJAccounting for related questions preserves statistical support for full VALR’s accuracy gains on POPE-A and HallusionBench.

We believe these additions substantively address the reviewers’ main concerns and strengthen support for VALR’s paired objective and its effectiveness in mitigating hallucinations during multimodal reasoning. We would be grateful for the AC’s consideration of the expanded evidence and discussion, and sincerely thank the AC and reviewers for their time and constructive feedback.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.