acceptodds
Withdrawn from ICLR 2027

MetaTide: Bringing Evolving Seeds into Multimodal Reasoning via Metacognitive Interflow

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multimodal reasoning increasingly combines visual evidence with multi-step inference, while metacognitive regulation helps models assess and adjust an ongoing solution. A central challenge is how the understanding gained from these judgments and adjustments can persist and guide future problems. We introduce MetaTide, a framework that brings evolving metacognitive seeds into current reasoning through Metacognitive Interflow. Stored in a Metacognitive Undercurrent, these seeds capture transferable relations among cognitive conditions, regulatory choices, and observed effects. Metacognitive Interflow matches the current Meta-State to an applicable seed and translates it into second-order control, which guides task-specific observation and inference in Surface Reasoning. In this way, seeds shape how the model observes, evaluates, and adjusts its reasoning in a new task. The outcomes of these actions refine its ongoing assessment, and completed solutions are distilled into new or consolidated seeds. This reciprocal flow allows prior self-understanding to shape present reasoning while present reasoning renews the basis for future regulation. Experiments across five multimodal benchmarks show gains of +29.03 and +22.58 percentage points over ReasoningBank on MATH-Vision and MuirBench, respectively.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Multimodal Reasoning, Metacognition, Metacognitive Regulation, Memory-Augmented Reasoning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_cKEYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes MetaTide, a framework for multimodal reasoning that maintains evolving metacognitive seeds distilled from previous reasoning trajectories. The framework constructs a Meta-State to characterize ongoing reasoning, retrieves relevant seeds, and translates them into second-order control to guide task-specific reasoning. Completed tasks are used to update and consolidate the memory bank. Experiments across five multimodal benchmarks report substantial improvements over ReasoningBank, particularly on MATH-Vision and MuirBench. However, the methodological novelty relative to existing metacognitive memory approaches remains insufficiently established, and concerns about evaluation reliability, particularly the unusual numerical patterns in the reported results and unclear evaluation protocols, weaken confidence in the empirical findings.

Critical strengths
  1. Interesting motivation. The paper investigates how accumulated metacognitive experience can guide future multimodal reasoning, addressing an important problem in experience-driven learning and adaptive reasoning.

  2. Structured framework with component analysis. MetaTide integrates metacognitive state assessment, experience retrieval, second-order control, and continual memory evolution. The paper provides ablations examining memory evolution, retrieval, and second-order planning, together with analyses of intermediate regulation events.

Critical weaknesses
  1. Concerning evaluation statistics and insufficient clarity about the evaluation protocol. All five reported MetaTide accuracies in Table 1 (51.61%, 64.52%, 67.74%, 61.29%, and 77.42%) correspond exactly, up to rounding, to integer multiples of 100/31 (16/31, 20/31, 21/31, 19/31, and 24/31). The same values appear for the MetaTide in Table 2, whereas none of the 15 results from the three ablation settings exhibit this pattern. This raises concerns about the effective evaluation sample sizes and whether the full model, ablations, and reported baselines were evaluated under identical conditions.

  2. Limited novelty and insufficient differentiation from closely related work. The proposed framework substantially overlaps with existing methods, particularly ReasoningBank (ICLR 2026), MC2 (ACL 2026), and AdaMEM (2026). MC2 already investigates cross-instance metacognitive consolidation involving reasoning, monitoring, and control, while AdaMEM dynamically adapts strategies based on historical experience and the current reasoning state. Although these works are cited, the paper provides only brief conceptual distinctions and lacks direct experimental comparisons with MC2 and AdaMEM. The incremental methodological contribution remains unclear.

Questions for authors
  1. Why do all five MetaTide accuracies correspond to integer multiples of 1/31? Were the full model, ablations, and baselines evaluated on identical benchmark subsets? Please provide the exact number of evaluated examples per benchmark, sampling procedures, evaluation configurations.

  2. What fundamentally distinguishes MetaTide from MC2, ReasoningBank, and AdaMEM? In particular, what new capabilities does Metacognitive Interflow introduce beyond existing metacognitive consolidation and state-adaptive strategy generation? Could the authors provide more direct comparisons with these methods?

  3. Can the authors provide compute-matched comparisons against ReasoningBank and simpler reflective reasoning baselines, and evaluate metacognitive seeds against generic or randomly retrieved reasoning strategies under identical inference budgets?

Reviewer_ix6cOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

MetaTide is a training-free agentic pipeline for a stream of multimodal questions. A frozen MLLM plays several prompted roles: it writes a textual "Meta-State" summarizing the current reasoning trace, retrieves "metacognitive seeds" (three-field text entries in the ReasoningBank format) using that summary as the query, turns the selected seed into a textual control instruction for the action planner, and after each question extracts new seeds and curates the bank with add / merge / skip / drop operations. The method is modality-agnostic: visual input passes through unchanged and no component is specific to images. Experiments cover five multimodal benchmarks and four backbones.

Critical strengths
  1. The problem is well chosen. Combining within-question regulation with memory that carries over across questions is a real gap, and using a description of the reasoning state rather than the question itself as the retrieval key is a reasonable idea for cross-domain transfer, as the case study illustrates.

  2. The pipeline is clearly laid out. The overview figure and the worked example make the roles and the data flow easy to follow.

Critical weaknesses

Critical Weaknesses

  1. Questionable evaluation reliability and reproducibility. Table 1's accuracy values suggest only 31 evaluated questions per benchmark, without disclosing subset selection or confidence intervals. Table 2's ablations also appear to use different sample sizes, making comparisons unreliable. Moreover, the Qwen3-VL-8B baseline on MATH-Vision (22.58%) is substantially below published results (48.13% in ViLoMem and 53.9% in the Qwen3-VL report), while ReasoningBank performs identically to the bare model. Essential implementation details and inference costs are missing, and the provided repository is empty. Together, these issues seriously undermine the credibility and reproducibility of the reported gains.

  2. Misalignment between motivation and methodology. Although the paper is positioned as a multimodal reasoning contribution, its core mechanisms, Meta-State construction, seed retrieval, metacognitive regulation, and memory evolution, operate entirely on textual representations, without any modality-specific design. The framework appears equally applicable to text-only reasoning, yet no such experiments are provided. This raises questions about what makes the proposed approach specifically multimodal and whether the claimed contributions are adequately supported by the experimental setting.

Questions for authors
  1. How many questions were evaluated, how were they selected, and why do the ablations appear to use different sample sizes? Can the authors provide full-test-set results with confidence intervals?

  2. Why is the MATH-Vision baseline substantially below published results, and why does ReasoningBank match the bare model? Can the authors release implementation details and inference costs?

  3. What makes the proposed method specifically multimodal, given that its memory and regulation mechanisms are text-based? How does it perform on text-only reasoning tasks?

Reviewer_m68nOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes MetaTide, a framework that integrates evolving metacognitive memories into multimodal reasoning through state-conditioned retrieval and second-order control. Experiments on five multimodal benchmarks demonstrate promising improvements over existing baselines.

Critical strengths

S1. The paper addresses a meaningful problem of preserving and reusing metacognitive experience across reasoning tasks. S2. The proposed framework provides a structured and interpretable approach to integrating persistent memory with ongoing reasoning regulation. S3. The experiments cover diverse multimodal benchmarks and multiple model backbones, with promising performance improvements and relevant ablation studies.

Critical weaknesses

W1. Insufficient implementation details of the core components. The paper does not clearly describe how the Meta-State Constructor, semantic retrieval, and first-order and second-order planners are implemented in practice. In particular, the underlying model assignments, prompting strategies, and semantic matching function (Equation (5)) are not sufficiently specified. The released code defines the orchestration logic and input-output contracts but delegates the actual execution of these components to external callbacks. These omissions limit reproducibility and make it difficult to understand how the proposed framework operates in practice.

W2. Limited quantitative evidence for cross-domain transfer of metacognitive experience. A central motivation of MetaTide is to preserve transferable metacognitive principles that guide reasoning on future problems. However, the continual-learning experiments in Figure 4 primarily evaluate performance within individual benchmark streams, while explicit cross-domain transfer is illustrated through a qualitative example in Figure 7. Although these results provide initial evidence, systematic cross-domain evaluations would more directly support the claimed transferability of metacognitive seeds.

W3. Insufficient details on the evaluation of metacognitive regulation. Figures 5 and 6 report improvements in reasoning behaviors, including premise recognition, requirement recovery, and action redirection. However, the paper does not sufficiently describe how these behaviors and step outcomes are identified and annotated, or how the reported percentages are calculated. In particular, it is unclear why the no-retrieval setting achieves zero on all three regulation metrics in Figure 6. Further clarification of the evaluation criteria and procedures is needed to assess the reliability of these results.

W4. Insufficient experimental implementation details. The paper lacks essential experimental details, including inference hyperparameters (e.g., temperature, top-p, and maximum generation length), hardware configurations, and other reproducibility settings. Furthermore, the released code does not provide a complete setup for reproducing the reported experiments. These omissions make it difficult to independently reproduce and verify the experimental results.

W5. Presentation and formatting issues. The manuscript contains several formatting issues. Inconsistent font styles are observed in Figure 2 and Equations (4) and (5). Additionally, Figure 3 overlaps with the text of Observation 2, affecting readability. These issues should be addressed to improve the overall presentation quality.

Questions for authors

Q1. Could the authors clarify the implementations of the Meta-State Constructor, semantic retrieval, and first-order and second-order planners?

Q2. Can the authors provide additional quantitative evaluations of cross-domain transfer, where metacognitive seeds accumulated on one domain are applied to unseen problems in another domain?

Q3. Could the authors clarify the definitions, annotation procedures, and calculation methods for the process-level metrics in Figures 5 and 6? In particular, why does the no-retrieval setting achieve zero on all three regulation metrics?

Q4. Could the authors provide the complete experimental configurations, including inference hyper-parameters, hardware specifications, and executable scripts required to reproduce the reported results, including the ReasoningBank baseline?

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

Related papers

Discussion (0)

Sign in to comment.