acceptodds
Under review as a conference paper at ICLR 2027

CogRef: Cognitive Reflection for Hallucination Mitigation in Multimodal Large Language Models

Abstract

Hallucination remains a major challenge for multimodal large language models (MLLMs). Existing mitigation methods often modify the generation process of MLLMs or rely on external verification tools, potentially constraining the original capabilities of the base MLLM. Motivated by human cognitive learning, where past experiences are reflected upon and reused to guide subsequent behavior, we propose CogRef, a cognitive reflection framework that decouples multimodal generation from hallucination supervision. A frozen MLLM serves as the cognitive core, while a learned supervisory circuit evaluates its response and provides structured feedback for final generation. To train the circuit, we transform past response cases into reusable multimodal experience, and internalize it through Concordance-Calibrated On-Policy Self-Distillation (CC-OPSD). CC-OPSD coordinates experience- and evidence-privileged teachers, enabling the supervisory circuit to internalize both response-handling experience and fine-grained visual grounding capability. At inference, CogRef requires neither privileged information nor external verification tools. Experiments across diverse MLLM families and multimodal benchmarks demonstrate that CogRef consistently mitigates hallucinations while preserving general multimodal capability, with the largest improvement reaching points (MHaluBench: ). Moreover, the supervisory circuit generalizes across different MLLMs without model-specific retraining, while selectively correcting hallucinated content and preserving visually grounded responses.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Multimodal Large Language Model, Hallucination Mitigation, Experience Internalization

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_sjznOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a method to reduce hallucination in MLLMs. The key idea is reflection on the potential outputs. The reflection is done by a dedicated module, trained by a supervisory signal aware of potential hallucinations, where the signal is also constructed with the help of an MLLM (Qwen). The module uses on-policy distillation for its training. Inference does not require privileged information and mitigates the hallucinations considerably. Strong results are shown on four benchmarks, comparing against at least three baselines.

Critical strengths

The reported results are strong. The core idea of reflection makes intuitive sense.

Critical weaknesses

Whereas the paper idea is intuitive, and approach follows naturally, the AI face-lifting of the paper makes it a considerably uninteresting read. Brimming with fancy terminologies and sentences, like typical AI face-lifted/generated papers nowadays, reviewing this paper was not a pleasant experience.

The approach requires a supervisory signal. A high end Qwen3.6-35B is used for that. In effect, the key strength of the approach is coming from this signal - an implicit distillation. The evaluations on weaker models naturally benefit from such supervision. The central intuition of 'reflection' is also not exactly groundbreaking, as it is used in different VLMs, LLMs since 2023. Whereas the details of CogRef - proposed method - may be unique in its use of OPSD and treatment of the problem, the contribution of the paper is of incremental nature.

Questions for authors
  • The state-of-the-art in MLLM hallucination now also considers adversarial/synthetic images that cause hallucinations, eg, GHOST (ICLR'26), did the authors try the method on such images? How is the performance?
  • What is the author's opinion on the use of Qwen3.6-35B for 'experience construction'. With the assumption of a stronger Oracle for such an experience, which is eventually shaping the feedback for hallucination mitigation, isn't the performance gain mainly attributed to the stronger model. What if the experience construction if faulty?
  • In the Tab 4 of appendix, CogRef latency is even lower than vanilla model. On the other hand, Eq(8) shows the loop for final generation that has additional steps, generation of feedback and then the response. The model M is used twice and there is an additional pass to construct F. Could the authors please clarify how the latency remains lower than the original model despite (apparently) more than twice flops involved that require sequential processing.
Reviewer_J9jdOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CogRef, which separates multimodal response generation from hallucination supervision and trains a dedicated supervisory model using CC-OPSD. The modular design is useful, and the cross-model results and correction/preservation analysis are encouraging. However, the paper does not sufficiently establish the need for on-policy self-distillation. The interpretation of concordance weighting, evaluation on open-ended generation, and reproducibility details also require strengthening. Based on the current manuscript, I lean toward Weak Reject, primarily because the contribution of the central training methodology is not sufficiently isolated.

Critical strengths
  1. The modular design has practical value. Decoupling supervision from updates to the base model allows one supervisor to serve different model families. The trained Qwen3-VL-4B supervisor is evaluated with 4B/8B models from three families, supporting transfer within the tested scope. This is a meaningful strength compared with evaluation on a single base model.
  2. Learning both correction and preservation addresses a practical concern. The paper does not equate more revisions with better supervision. It explicitly measures both the correction of initially incorrect predictions and the preservation of initially correct predictions. In the Qwen3-VL-4B MHaluBench setting, training raises the correction rate from 13.36% to 47.37% while maintaining a preservation rate of 96.51%. This provides more insight into system behavior than final accuracy alone.
Critical weaknesses
  1. Although CC-OPSD outperforms the evaluated OPSD variants, the paper does not sufficiently establish why training the CogRef supervisor requires on-policy self-distillation rather than supervised fine-tuning on teacher-generated feedback. The current ablations primarily compare ways of incorporating privileged information within OPSD, making it difficult to disentangle the contributions of structured experience data, on-policy distillation, and the proposed concordance weighting. Controlled comparisons against alternative training strategies would strengthen the justification for the central training methodology.
Questions for authors
  1. Could the authors compare CC-OPSD with SFT, OPD using a separate teacher, and RL methods such as PPO/GRPO under controlled settings? SFT on teacher-generated feedback is the most important missing baseline.
  2. Does the latency in Appendix F cover initial generation, supervisory feedback, and final generation? Please clarify “Vanilla” and report latency and GPU memory overhead relative to base-model-only generation.
  3. Could the authors evaluate CogRef on benchmarks such as FREAK [1] and MM-Snowball [2] to assess its effectiveness against fine-grained visual hallucinations and hallucination accumulation in multi-turn dialogue?

[1] Yin et al. FREAK: A Fine-grained Hallucination Evaluation Benchmark for Advanced MLLMs. ICLR 2026. [2] Jiang et al. MM-Snowball: Evaluating and Mitigating Hallucination Snowballing in Multimodal Multi-Turn Dialogue. ICML 2026.

Reviewer_1MFNOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CogRef supplements a frozen MLLM with a learned supervisory model that critiques an initial response before the original MLLM generates its final answer. The supervisor is trained from 1,000 REVERIE cases converted into structured correction/preservation experience and localized visual evidence. CC-OPSD distills an experience-conditioned teacher along student-generated trajectories, using agreement with an evidence-conditioned teacher to weight token losses. A single Qwen3-VL-4B supervisor is evaluated with six cognitive cores across MHaluBench, HallusionBench, MMStar, and MMMU-Pro Vision, with additional component and intervention analyses.

Critical strengths
  1. The modular design is practically attractive: one trained supervisor transfers across three model families and two sizes without updating their parameters or using grounding tools at inference. Table 1 shows improvements beyond MHaluBench, including HallusionBench qAcc increasing from 32.75 to 42.42 for LLaVA-OneVision-1.5-4B, alongside gains on both general multimodal benchmarks.
  2. Table 2 includes a valuable matched vanilla-supervisor control, rather than attributing all improvements to adding a revision stage. Training raises MHaluBench accuracy from 64.19 to 76.94 and correction rate from 13.36% to 47.37%, while retaining 96.51% of initially correct predictions. These transition statistics provide evidence that the learned supervisor changes intervention behavior.
Critical weaknesses
  1. The specific benefit of CC-OPSD is not yet well established. Its advantage over experience-only OPSD is evaluated in a single model and benchmark setting without uncertainty estimates, and teacher agreement is not directly shown to identify more reliable supervision. Missing values for the calibration hyperparameters also limit reproducibility.
Questions for authors
  1. Additional validation: Does CC-OPSD reliably improve over experience-only OPSD across training seeds or another already evaluated setting? A shuffled-concordance control would help distinguish informative agreement from generic token reweighting.
  2. Reproducibility clarification: What λ, τ, supervisor/final-generation prompts, and decoding/thinking settings were used? Were these inference settings identical for the vanilla and trained supervisors?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.