acceptodds
Under review as a conference paper at ICLR 2027

MODEL-SUBSTITUTION RISK CONTROL: VERIFIED REPAIR OF VISUAL CLAIMS

Abstract

Vision–language models can generate captions and answers that contain details not supported by the image. Removing such details reduces hallucination, but it can also discard information that could be recovered through a correct replacement. We introduce Model-Substitution Risk Control (MSRC), a training-free method that verifies visual claims and attempts to repair unsupported object claims before deleting them. MSRC retains supported claims and uses the frozen model to propose replacements for failed claims. Each replacement must pass fixed admissibility checks and independent visual verification before being accepted; otherwise, the unsupported claim is removed. The method follows a common claimlevel repair process across image captioning and visual question answering, with task-specific handling of short answers. To balance hallucination reduction and information retention, MSRC selects its verification thresholds using held-out data and measures remaining errors relative to the original claim opportunities. Openvocabulary replacements are assessed using benchmark annotations, lexical relations, and human-audited error estimates when automatic labels are unavailable. We evaluate MSRC across captioning, visual question answering, and hallucination benchmarks, reporting both claim-level risk and final-response quality.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Vision–Language Models, Visual Hallucination, Claim Verification, Claim Repair, Risk Calibration

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_GSTTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces Model-Substitution Risk Control (MSRC), a training-free method for reducing visual hallucinations in VLMs. Rather than simply deleting unsupported object claims, MSRC first attempts to replace them with visually supported alternatives. The method checks each claim against the image and uses the frozen VLM to propose replacements for unsupported claims. A replacement is accepted only if it passes additional checks and visual verification; otherwise, the original claim is deleted. The authors use held-out data to calibrate the verification thresholds, aiming to reduce hallucinations while retaining useful information. Experiments across image captioning, visual question answering, and hallucination benchmarks show improvements in hallucination rates and answer quality over several baselines.

Critical strengths

The paper measures hallucination risk over the original set of claims rather than only the claims that remain after filtering. This is useful because simply deleting claims should not be rewarded without considering how much information is lost. The comparison with verification-only filtering is useful, as both methods use the same verification and calibration setup, making it easier to assess the contribution of substitution.

Critical weaknesses

The paper presents MSRC as a risk-control method, but the proposed certification does not provide a formal guarantee on population risk. It is therefore unclear what level of risk control the method actually provides. The method focuses on correcting object claims in generated text, but it is not clear how it handles yes/no, counting, and attribute questions, especially given the large improvements reported on HaloQuest, POPE, and MME. It is not clear whether all baselines were evaluated under the same settings, including the evaluation judges, image subsets, and model backbones.

Questions for authors

Could the authors explain how MSRC handles short-answer VQA, particularly yes/no, counting, and attribute questions? The method mainly describes object-claim replacement and deletion, so it is unclear how the results on POPE and MME are obtained. How often do substitutions repeat object categories already mentioned in the caption, and are they correct at the object-instance and location level? Could the authors clarify how much of the improvement over verification-only filtering comes from substitution rather than differences in the selected verification thresholds?

6 comments

Reviewer GSTT

We thank the reviewer for their thoughtful questions and constructive feedback.

W1 — Level of risk control. We agree: the submitted criterion is not a high-probability bound, and Table 8's "Bound ≤ α" and Figure 4's "certified" overstate it. We now give a formal test, fixed before scoring, on 1,500 never-scored COCO images per backbone (disjoint from calibration and evaluation; image as the unit; α = δ = 0.05):

BackboneMSRC riskMSRC boundVOF boundMSRC (δ/6)VOF (δ/6)
LLaVA0.03600.04430.04280.04650.0449
Qwen2.5-VL0.03350.04280.04150.04420.0429
InstructBLIP0.02730.03400.03380.03600.0358

The level provided is P(certify ∧ R(λ̂) > α) ≤ δ: the chance of certifying a policy whose expected per-image rate of unsupported COCO object claims exceeds 5% is at most 5%. All six tests also pass jointly (Bonferroni). It covers COCO object existence only, not attributes, counts or relations. The paper's 250-image half does not pass (0.060, 0.069, 0.060), and we withdraw that reading.

W2/Q1 — Short-answer handling. Eq. (4) edits object claims in free text and has no short-answer rule:

Question typeHow the answer is produced
Yes/no (POPE)Read from the verifier score V (Eq. 2), aggregating yes/no first-token probabilities, not the greedy answer
Counting, false premise (HaloQuest)Negation ("There is no X") if the named entity fails verification
Attribute, count, position, colour (MME)MSRC extracts and verifies only object nouns (COCO categories); adjectives, numerals and spatial phrases are left unverified

MSRC thus has no mechanism for the MME gains; Table 13 was not reproduced and moves to the appendix, marked. POPE asks exactly the object-existence question our verifier scores, so MSRC answers it directly from V; we keep Table 12 and will document this setting. Table 12 (average over random, popular and adversarial splits, accuracy / F1):

BackboneVanillaVCDOPERAMSRC
LLaVA81.30 / 79.6084.60 / 84.4084.70 / 85.4088.17 / 87.82
Qwen2.5-VL83.60 / 81.6086.60 / 86.0085.20 / 83.3087.19 / 87.11
InstructBLIP78.20 / 78.4081.80 / 81.3084.60 / 85.1086.76 / 86.79

HaloQuest uses a cascade: entity verification, a TF-IDF question-type classifier and a logistic-regression selector (C = 0.5) choosing negation, abstention or the original answer. Substitution is off (calibration selected τ_sub = 2.0).

LaneLLaVAQwen2.5-VLInstructBLIP
Vanilla23.1957.8922.86
+ abstention58.0657.4042.60
+ negation67.1175.6664.64
Nested CV (reported)65.6275.3364.80

Nesting changes accuracy by −1.5 to +0.2 points; we withdraw the visual-challenge property .

W3 — Judges, subsets and backbones. No. Cited rows of Table 1 use their authors' splits, scorers and decoders; our HaloQuest and MMHal scores use Gemini 3.7 Flash, not GPT; MME mixes Qwen-VL baselines with our Qwen2.5-VL-3B, and CEBC uses Qwen2-VL-7B. Protocol alone can flip a result:

VCDCHAIRsIts vanilla
Cited in Table 148.6051.70
Our matched run56.2350.63

Q2 — Repeats, instances and location.

BackboneCategory repeatedIn COCO GTInstance matchFully correct (human)Spurious duplicate
LLaVA88.2%95.1%75.0%16/230/17
Qwen2.5-VL62.9%79.4%58.8%18/251/16
InstructBLIP90.4%96.6%82.6%10/190/17

Repeats are common, but a repeated category is rarely a duplicate object (1 of 50, 2.0%). At instance and location level, 65.7% are fully correct, 16.4% have the right object in the wrong place or with wrong details, and 16.4% name an absent object. The audit (67 of 150 random items, one author, backbone-blind) is preliminary.

Q3 — Substitution vs. thresholds. τ₀ is identical for MSRC and VOF (Table 8, and across the α sweep), so thresholds explain nothing; the lanes differ only in substitution:

Backboneτ₀τ_subSubstitutionsVOF CHAIRsMSRC CHAIRsVOF CHAIRiMSRC CHAIRi
LLaVA0.7610.87620416.7316.974.374.35
Qwen2.5-VL0.7780.02317014.9315.673.543.54
InstructBLIP0.6650.93317814.8315.103.903.89

Substitution adds 0.10 to 0.22 points of coverage but gives no meaningful CHAIR reduction, so the hallucination reduction comes from calibrated verification and deletion, not substitution. The submitted Table 1 gap does not match a stored run, and we replace those rows.

Dear Reviewer,

Thank you for your time and constructive comments. We kindly invite you to revisit our rebuttal and let us know if any questions remain unanswered.

Dear Reviewer,

We would like to kindly follow up on our previous message and request you to revisit our rebuttal. We would greatly appreciate any further feedback or clarification on whether your concerns have been addressed.

Thank you for your time and consideration.

Reviewer_dbeQOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

MSRC is a training-free, post-hoc method for reducing object hallucination in LVLMs. A response is split into object claims using a CHAIR-style extractor. The frozen LVLM checks each claim with a POPE-style yes/no prompt, scored by first-token probabilities. A claim above threshold τ₀ is kept. A claim below τ₀ is re-grounded: the model proposes one replacement object, which is accepted only if it passes lexical and duplicate guards and independently scores above τ_sub. Otherwise the claim is deleted. Risk is measured over the original claim opportunities, so deletion cannot shrink the denominator. Substitutes outside the vocabulary are charged a conservative error rate from a human audit. Experiments use LLaVA-1.5-7B, Qwen2.5-VL-3B, and InstructBLIP-7B (plus BLIP-2 in the appendix) on COCO-CHAIR, AMBER, HaloQuest, MMHal-Bench, POPE, and MME. MSRC is compared against a matched verification-only filtering baseline (VOF) and published methods, and the paper reports large reductions in hallucination.

Critical strengths
  • Well-motivated problem. Asking whether to repair or delete an unsupported claim addresses a real gap in filtering-based factuality control such as ConfLVLM. Measuring risk over a fixed denominator, paired with true-claim yield, is a clean way to stop trivial abstention from looking like good risk control. This is the paper’s most transferable idea.
  • Matched VOF comparator and candid negative results. VOF shares the extractor, verifier, guards, and calibration with MSRC, so it isolates the effect of substitution. The authors also openly report cases where MSRC loses: Qwen CHAIR_I, AMBER coverage, and operating points that fail certification.
Critical weaknesses
  • Limited scope and unexplained gains. The extractor, verifier, and loss cover only object nouns, so attribute, relation, and count hallucinations pass untouched. Yet the paper claims gains on MME position/color and on MMHal-Bench. The large HaloQuest and POPE gains also rely on an unspecified “short-answer handling”, so they cannot be attributed to verified repair.

  • Unreliable and incomplete comparisons. Baseline numbers appear copied from papers that used different judges (the authors switched to Gemini), and some comparisons mix backbones (MME, CEBC). Key baselines are missing: ConfLVLM (the method MSRC extends), LURE, DeGF, HALC, Woodpecker beyond LLaVA, and recent training-free methods such as CRoPS, SID, and PAI.

  • No latency or compute cost is reported, despite multiple additional forward passes per claim, so the accuracy–efficiency trade-off against cheaper decoding-time baselines is unknown.

  • Weak statistical grounding. Sec. 2.6 admits there is no formal guarantee. The CRC-style bound is applied after a search over two thresholds, with a non-monotone loss and a precision safeguard that is relaxed when nothing passes. The LTT error control cited as the foundation is not used, so calling the method “risk control” overstates what is shown.

Questions for authors
  • On a random, non-curated sample of accepted substitutions, how many introduce a duplicate instance of an object already mentioned? How many are judged correct by a human looking at the image?
  • Can MSRC handle attribute, relation, or count hallucinations? If not, what explains the MME position and color gains? Please report results on AMBER’s attribute and relation splits.
  • Can you compare against ConfLVLM under identical calibration, and against Woodpecker on all backbones?
  • Please report wall-clock latency and the average number of extra forward passes per response (verification, re-grounding, re-verification) for MSRC, VOF, and the vanilla model.
4 comments

Reviewer dbeQ

We thank the reviewer for the constructive feedback.

W1 and Q2 — Scope and AMBER

MSRC verifies eligible object-existence claims, not attributes, relations, or counts. Revised AMBER generative results follow (Vanilla → MSRC; ↓ lower is better, ↑ higher is better).

BackboneCHAIR ↓Hal ↓Cog ↓Coverage ↑
LLaVA7.4 → 2.934.5 → 12.14.0 → 0.748.7 → 51.3
Qwen2.5-VL6.1 → 1.730.1 → 8.12.6 → 0.344.6 → 53.9

MSRC reduces CHAIR by 61–72%, Hal by 65–73%, and Cog by 83–88% across both backbones.

AMBER discriminative accuracy for MSRC (%):

BackboneOverallExistenceAttributeStateNumberActionRelation
LLaVA81.988.079.778.979.685.273.3
Qwen2.5-VL84.196.482.679.587.489.354.1

Overall accuracy is 81.9%/84.1%. Because all existence questions have negative ground truth, stricter thresholds may favor negative answers. Without matched Vanilla comparisons, gains on attributes or relations cannot be inferred.

W2 and Q3 — Fair comparisons

Published baselines differ in splits/scorers; our HaloQuest/MMHal evaluation uses Gemini 3.7 Flash rather than the original GPT judge. CEBC uses Qwen2-VL-7B, versus our Qwen2.5-VL-3B. Thus, cited rows are not matched rankings. We replayed VOF and MSRC on 3,000 COCO images at a common keep threshold:

BackboneVOF CHAIRsVOF CHAIRiMSRC CHAIRsMSRC CHAIRi
LLaVA16.734.3416.974.37
Qwen2.5-VL14.933.4415.673.54
InstructBLIP14.833.9015.003.89

This matched replay suggests verification/deletion drive most gains over Vanilla. Table 1 instead uses 500 images; we distinguish the samples and check underlying run records.

Other local LLaVA runs are supplementary, not paired:

MethodCHAIRs ↓CHAIRi ↓
CRoPS35.979.57
SID48.3012.89
DeGF46.7013.67
PAI52.3014.46
HALC52.0015.28
VCD56.2315.66
M3ID54.8015.82

End-to-end Woodpecker/ConfLVLM comparison. We evaluate all methods on the first 500 COCO images. Woodpecker’s GPT component is replaced by local Qwen2-VL-7B; ConfLVLM is calibrated using our procedure. Calibrated methods meet α = 0.05.

BackboneMethodCHAIRs ↓CHAIRi ↓
LLaVAWoodpecker52.5114.20
ConfLVLM22.207.84
MSRC13.204.34
Qwen2.5-VLWoodpecker35.017.38
ConfLVLM15.706.59
MSRC14.974.80
InstructBLIPWoodpecker51.7015.14
ConfLVLM18.807.46
MSRC14.873.93

MSRC has the lowest CHAIRi on all backbones and the lowest CHAIRs on LLaVA/InstructBLIP; it ties ConfLVLM on Qwen2.5-VL CHAIRs.

W3 and Q4 — Runtime

Each failed claim generates at most one candidate, without retry. Per-image means and runtime ratios versus Vanilla are:

BackboneOriginal verificationsCandidatesRe-verificationsRuntime
LLaVA4.081.410.151.20×
Qwen2.5-VL3.580.760.061.08×
InstructBLIP3.981.410.171.18×

VOF skips candidate generation/re-verification. Absolute times and VOF timings are unavailable; this timing population differs from Table 8 claim counts.

W4 — Statistical validity

On an independent 1,500-image COCO test with frozen thresholds (α = δ = 0.05), Bonferroni-adjusted bounds are:

BackboneMSRC riskMSRC upper boundVOF upper bound
LLaVA0.03600.04650.0449
Qwen2.5-VL0.03350.04420.0429
InstructBLIP0.02730.03600.0358

All six pass 0.05 under independence and bounded-loss assumptions. The 250-image test fails; we withdraw earlier certification wording. This result applies only to COCO object risk.

Q1 — Human audit

One author, blinded to backbone, assessed all 150 randomly sampled accepted substitutions. In detailed subsets, 44/67 were fully correct and 1/50 was a spurious duplicate.

Dear Reviewer, We appreciate your thoughtful review. We would be grateful if you could take a moment to consider our clarifications and share any further feedback.

Reviewer_dbeQOpenReview ↗

Official Comment by Reviewer dbeQ

Thank you for the detailed rebuttal and additional experiments. The expanded baseline comparisons and independent 1,500-image evaluation address some of my earlier concerns.

I would encourage the authors to incorporate the additional baseline comparisons and corresponding discussion into the main manuscript, particularly in the experimental results and Related Work sections, to ensure that these contributions are reflected in the paper itself.

Overall, I appreciate the additional work and am increasing my score to 3.

We sincerely thank the reviewer for their positive feedback and suggestions. We will incorporate the additional baseline comparisons, evaluation results, and relevant discussion into the revised manuscript.

Reviewer_so1WOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary
  1. This paper studies object hallucinations in vision-language models. It proposes attempting to repair unsupported descriptions before deleting them, with the aim of reducing information loss while balancing output reliability and information retention.
  2. The paper proposes MSRC, a training-free method that uses the same frozen model to extract and verify object claims. For claims that fail verification, it generates alternative objects and then uses lexical rules and repeated verification to decide whether to retain, replace, or delete the claims.
  3. The paper reports improvements on image captioning, visual question answering, and hallucination evaluations. In some settings, it retains more information than a pipeline that only verifies and deletes claims.
Critical strengths
  1. The paper is generally well written and easy to read.
  2. The paper addresses the possibility that direct deletion may discard recoverable information, which is a practically meaningful motivation. Separating repair generation from repair acceptance also reflects consideration of the risk of introducing new hallucinations.
  3. The experiments include component ablations, risk–coverage curves, and analyses across multiple calibration splits, which help assess the value of the replacement step.
Critical weaknesses
  1. The presence of an alternative object in the image does not imply that the counts, attributes, locations, and relations in the original sentence remain correct after replacement. The appendix includes changes from "there are two fish" to "there are no fish" and from an affirmative answer about a connection to a negative one. These changes go beyond the object-name substitution rule in Equation 4, so the performance gains are not necessarily attributable to the object replacement mechanism.
  2. The examples do not match the stated method. Appendix H claims that only the target span is edited and all other sentences remain unchanged. However, in the COCO 499388 example, "lights" in another sentence is also changed to "trees", and the VOF output still contains "truck", which should have been removed. In HaloQuest 35, VOF likewise retains a claim about fish that has already been judged unsupported.
  3. The audit counts objects, whereas the risk is averaged after normalization by the number of objects within each image. These two schemes do not inherently assign the same weights. The paper needs to clarify the sampling unit, the dependence among claims from the same image, and the conditions connecting the audit to the true image-level risk. It should also report the actual audit sample size and stratified counts.
Questions for authors
  1. Could the authors provide the complete algorithm for the question-answering pipeline and separately ablate the contributions of object replacement, negative answers, refusal templates, and deletion that preserves grammatical completeness to the performance gains?
  2. Could the authors provide explanations addressing the concerns raised in weaknesses?
5 comments
AuthorsOpenReview ↗

Response to Reviewer so1W: corrected implementation, attribution of gains, and claims

Reviewer so1W

We thank the reviewer; these concerns revealed implementation issues, corrected below.

W1 — Sentence-level faithfulness

MSRC verifies object existence, not the attributes, counts, locations, or relations associated with an object. We assessed changes in DeBERTa-large-MNLI contradiction probability against COCO references, with 95% image-bootstrap confidence intervals. Negative changes indicate lower contradiction.

BackboneSubstitution aloneDeletion aloneSubstitution relative to VOF
LLaVA+0.008 [−0.020, +0.035]+0.038 [+0.027, +0.048]+0.005 [−0.024, +0.032]
Qwen2.5-VL+0.003 [−0.041, +0.049]+0.003 [−0.008, +0.015]-0.025 [−0.070, -0.016]
InstructBLIP−0.029 [−0.085, +0.026]+0.048 [+0.038, +0.058]−0.041 [−0.083, −0.002]
Substitution-only edits show no statistically significant change in contradiction probability, although NLI is only an indirect proxy for visual faithfulness. Compared with VOF, Qwen2.5-VL and InstructBLIP show small but significant reductions in contradiction, while LLaVA shows no significant difference.

W2 — Editing rule and examples

An earlier implementation also modified adjacent function words and articles and dropped very short sentences. Those additional rules affected 36–52% of edited sentences. We replayed 3,000 captions per backbone with the frozen thresholds and the span-only action in Eq. (4). CHAIRs and CHAIRi did not change to two decimal places because these metrics do not score grammaticality. We withdraw earlier Appendix H examples, including the inconsistent COCO 499388 trace, and replace them with stored-score traces. In HaloQuest 35, the extractor missed “fish”; a distinct question-entity check scored it at 0.058 and triggered negative correction. Missed mentions can survive editing, which we now acknowledge.

Corrected COCO results on 3,000 images, with a common scorer and identical keep threshold for VOF and MSRC (lower is better):

BackboneVOF CHAIRsVOF CHAIRiMSRC CHAIRsMSRC CHAIRi
LLaVA16.734.3416.974.37
Qwen2.5-VL14.933.4415.673.54
InstructBLIP14.833.9015.003.89

Table 1 reports 500 images; this 3,000-image replay tests substitution using the same keep threshold. The replay finds no consistent CHAIR improvement over VOF, indicating that verification and deletion explain most of the reduction relative to vanilla. We will distinguish the two evaluations and verify Table 1 against its run records.

W3 — Image-level risk and audit

Eq. (6) averages loss within each image before averaging across images. Claim errors are correlated within images (ICC 0.12–0.16), and claim-level standard errors are underestimated by factors of 1.4–1.7. The COCO-labelled substitute-precision safeguard is separate from the open-vocabulary human audit. At substitute threshold 0.5, correct counts were 31 of 34 for LLaVA, 43 of 45 for Qwen, and 28 of 36 for InstructBLIP. At threshold 0.92, they were 20 of 20, 23 of 23, and 20 of 22, respectively. The strongest Bonferroni-adjusted lower bounds (0.741, 0.870, 0.687) remain below the 0.95 precision floor, so the strict variant accepts no substitutions. All 150 sampled substitutions were evaluated by one author, blinded to backbone. Existing detailed subsets contain 44 correct cases among 67 and one spurious duplicate among 50; these are not full-sample rates. No unresolved open-vocabulary substitution enters the COCO risk calculation. New held-out COCO risk bounds on 1,500 images per backbone are 0.0443 for LLaVA, 0.0428 for Qwen, and 0.0340 for InstructBLIP, at α = δ = 0.05.

Q1 — HaloQuest and grammar

HaloQuest verifies entities in the question, classifies question type using TF-IDF, then uses a logistic-regression selector (C = 0.5) to choose negation, abstention, or the unchanged answer. Its selected substitute threshold is 2.0, so no substitution is accepted. The auxiliary classifiers are trained on HaloQuest; the complete VQA pipeline is therefore not training-free, although the LVLM is frozen.

Accuracy (%) ↑LLaVAQwen2.5-VLInstructBLIP
Vanilla23.1957.8922.86
Refusal or abstention58.0657.4042.60
Negative correction67.1175.6664.64
Nested cross-validation65.6275.3364.80
Originally submitted MSRC68.7576.4865.79

Compared with submitted MSRC, nested accuracy falls by 3.13 points for LLaVA, 1.15 for Qwen, and 0.99 for InstructBLIP. We withdraw the visual-challenge gain. Grammar-preserving deletion improves sentence acceptability from 44–68% to 87–92%, while reducing coverage by 0.2–0.9 points; these are grammar and coverage results, not a separate HaloQuest accuracy improvement.

Q2

The scope, editing inconsistencies, and audit are addressed above. We do not claim sentence-level correctness from object-level verification.

Dear Reviewer,

Thank you for your valuable feedback. We kindly request you to review our response and let us know if you have any remaining concerns.

Thanks for the detailed response

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.