acceptodds
Under review as a conference paper at ICLR 2027

Learning to Verify: A Counterfactual-Guided Dual-Loop Agent Framework for Multimodal Hallucination Mitigation

Abstract

Mitigating multimodal hallucination is essential for the reliable utilization of multimodal large language models in image-grounded tasks. Existing approaches perform mitigation by adjusting decoding, strengthening visual information during generation, or using verification to guide response correction. However, verification itself may accept unsupported statements or challenge correct ones, allowing repeated revision to retain hallucinations and remove valid information. In this paper, we propose Learning to Verify (L2V), a dual-loop agent framework for multimodal hallucination mitigation that learns a verifier's criteria through counterfactual analysis of its judgments. The inner loop performs inference by checking a main agent's response against the image and guiding revision until the response is accepted or an uncertainty response is returned. The outer loop operates only during training, using reference responses and counterfactual interventions to identify verification errors whose correction improves the final response. The identified verification errors guide targeted revisions to the verifier's criteria, which are evaluated through held-out replay on examples excluded from criteria refinement and adopted only when they improve response quality over the current criteria. Experiment results and analysis on multimodal hallucination benchmarks show that the proposed L2V outperforms strong baselines and existing approaches, demonstrating its effectiveness in mitigating multimodal hallucination.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multimodal Hallucination Mitigation, MUltimodal Large Language Models, Verifier Refinement, Counterfactual Failure Attribution, Dual-Loop Learning

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_6pzMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper propsoes Learning to Verify (L2V), a dual-loop agent framework for multimodal hallucination mitigation. The inner loop repeatedly verifies and revises a main agent's response using image evidence, while the outer loop uses reference responses, counterfactual interventions, and held-out replay to refine the verifier's textual criteria. The approach is evaluated with four MLLM backbones on POPE-COCO, POPE-GQA, and CHAIR.

Critical strengths
  • overall motivation, and inner/outer loop structure are presented clear
  • the paper identifies a meaningful limitation iterative self-correction, i.e., a verifier can itself make errors so repeatedly applying the same verifier does not necassarily improve the response.
  • the experiments are reasonably extensive
Critical weaknesses
  • my main concern is that 'learning' is insufficiently specified. the paper described s_t as the verifier's criteria and it says that verifier revises these criteria based on recurring errors. what conceretely are these criteria and the optimization objective?, apprently outerloop is the main contribution
  • final learned object is a set of verifier instructions, the paper itself provides substatial prior work on instruction optimization, prompt evolution, evaluator prompt refinement etc, so Learning to Verify is the central contirbution, and that requires evidence if the entire update is couterfactual guided refinement or simpler feedback based berifier prompt optimizer.
  • the empirical scope should be widened. the paper motivates hallucinations broadly in terms of objects, attributes, relations, quantities, and insufficient evidence. The final evaluation consists of POPE and CHAIR, these are object hallucination benchmarks, the literature points to HallusionBench, MME, and NOPE for broader object hallucination mitigation.
  • the outerloop uses LLaVA-Instruct-150K reference responses to identify discrepencies, only 800 images are used for deedback and 200 for held out replay. Also, the final utility U and candidate acceptance are reference controlled, so the possibility is part of the gain comes from learning task/reference-response.
  • minor: the paper may benefit from further proofreading, several writing errors, 'metraics' , 'vairous',
Questions for authors
  • clarification on what exactly s_0 and s_t, initial criteia, representative acceptance criteria,
  • definition of U, and details around it
  • can the work be compared agains a generic verifier prompt optimization baseline,
  • also does the learned verifier becomes accurate?
  • and how the learned criteria immprove perofrmance on hallucination involving attributes, quantities or composition reasoning.
Reviewer_nuyhOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Learning to Verify (L2V), a dual-loop agent framework for multimodal hallucination mitigation. The inner loop couples a main agent with a verifier that performs statement-level evidence assessment and guides response revision and termination. During training, the outer loop uses reference responses and counterfactual interventions to identify verification errors whose correction improves the final response. It then refines the verifier’s criteria and selects updates through held-out replay, without updating model parameters. Experiments across four backbones report improvements on POPE-COCO, POPE-GQA, and CHAIR.

Critical strengths
  • An interesting approach to improving verification criteria. Rather than treating verifier feedback as inherently reliable, L2V tests how replacing individual verifier outputs affects the final response. This provides a concrete way to investigate verification errors and guide criteria refinement.

  • Consistent improvements across the evaluated backbones. Table 1 reports improvements over both the base models and the inner loop with initial criteria across four backbones. The ablations also provide supporting evidence for the contributions of counterfactual attribution and held-out replay.

Critical weaknesses
  • Insufficient specification and validation of the optimization objective. Both counterfactual attribution and criteria selection depend on the evaluation function U, but the paper does not specify how its components are scored and combined, or provide the acceptance thresholds and initial and learned criteria. The public implementation obtains utility scores from the backbone using a textual rubric, but the required rubric and experiment settings are not included. The reliability of these scores is also not evaluated. These omissions limit reproducibility and make it difficult to establish whether higher utility reflects more reliable verification.

  • Incomplete evidence for the added value and practical trade-offs. The experiments lack a comparison with directly refining verifier prompts from execution feedback under the same data and a comparable optimization budget. The additional benefit of counterfactual attribution over conventional prompt optimization therefore remains unclear. Moreover, the framework permits repeated verification and abstention, but reports neither average model calls and latency nor abstention rates, caption lengths, and supported-content coverage. This leaves the computational cost and the possibility of reducing hallucination through less informative responses insufficiently assessed.

Questions for authors
  1. Utility scoring and criteria refinement. Could you specify the scoring rubric for U, the acceptance thresholds, and the initial and learned verifier criteria? What evidence supports the reliability of the utility scores in measuring changes in response quality?

  2. Contribution of counterfactual attribution. Under the same training data and a comparable model-call budget, how does L2V compare with directly refining verifier prompts from erroneous execution trajectories?

  3. Computational cost and information retention. Could you report average inner-loop steps, end-to-end latency, abstention rates, caption lengths, and coverage of correctly described objects? The public implementation maps abstentions to “no” for POPE. How much does this choice affect the reported results?

  4. Consistency of the qualitative examples. Section 3.3 limits POPE responses to four new tokens, whereas the main-agent responses in Figure 2 are substantially longer. Were these examples generated with different settings, and are the displayed responses verbatim model outputs?

Comments on the paper

1 comment

Thank the authors for the detailed rebuttal, which addresses several of my earlier clarification concerns. However, the central issue remains that the independent contribution of counterfactual attribution is not fully isolated. The direct-feedback baseline uses a different optimization budget, with 2,229 calls versus 1,444 for L2V, and the candidate/adoption procedures also differ. The outer-loop and inner-loop ablations therefore still contain confounding factors, while no matched-budget comparison with a generic prompt optimizer is provided. I therefore maintain my original rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.