acceptodds
Under review as a conference paper at ICLR 2027

CoT-Meta: Budgeted Metacognitive Control for Test-Time Reasoning

Abstract

Effective test-time reasoning requires deciding which partial trajectories to expand, which errors to repair, and when to conclude a search. We introduce CoT-Meta, a training-free framework that combines strategy-conditioned thought generation, tree-structured search, online process evaluation, and state-dependent control over expansion, pruning, repair, stopping, and abstention. A compact meta-state links intermediate reasoning quality to the allocation of a total-token budget. At , CoT-Meta achieves 92.8% on MATH-500, 90.4% on GPQA-Diamond, 75.8% on BBEH, and 56.7% Pass@1 on LiveCodeBench with Claude-4.5-Opus, and leads the evaluated baselines on all four tasks with Llama-3.1-8B-Instruct. Holding the inference components fixed and changing only action selection raises four-task macro accuracy by 2.0 points over a threshold-only policy on each backbone. A 900-state counterfactual audit associates this gain with higher best-action agreement and lower decision regret. Evaluator interventions, candidate-coverage audits, and token-matched repair comparisons connect the end-to-end gains to intermediate signal quality and targeted trajectory correction. Nested confidence-prefix evaluation further characterizes selective prediction separately from additional-budget fallback. These results establish an empirical case for explicit metacognitive control as a design principle for budgeted test-time reasoning.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
test-time reasoning, test-time compute, large language models, metacognitive control, adaptive inference, tree search, process evaluation, reasoning repair

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_rcK8OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces COT2-META, a training-free framework for test-time scaling. Its main idea is to efficiently select reasoning actions (repairs, expansion and pruning) within the search space. Specifically, the authors first use an LLM to evaluate the process quality of each reasoning step. Then, they compute a UCB-style frontier score for each node and estimate the feasible action utilities. Finally, the action with highest value is chosen for reasoning. Authors employ experiments on multiple tasks and benchmarks to verify the effectiveness of their methods.

Critical strengths

1.COT2-Meta provides a more detailed action selection strategy by defining action scores.

2.The method achieves good results and outperforms other test-time scaling methods in the experiments.

3.The authors evaluate the method on different types of tasks and compare it with many test-time scaling baselines. The main experiments are comprehensive.

4.The authors compare different control strategies and different settings of computing values.

Critical weaknesses

1.The main idea (action selection and value calculation) rely on existing methods (e.g., UCB) and several hand-designed choices, such as the Q functions and the 32-dimensional features. The paper lacks a clear analysis of why these designs are chosen. The method seems more like an engineering solution, and its novelty is somewhat limited.

2.Is the 32-dimensional meta-state really necessary? Is there any redundant features? The authors should provide a more detailed analysis or ablation study of these features.

3.The action value calculation needs baseline comparisons. For example, the authors could compare their method with a simple baseline that first selects a node by UCB and then uses the LLM evaluator to get the score of each action.

Questions for authors

See Weaknesses

Reviewer_ym2yOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This work proposes a reasoning system to adaptively adjust the reasoning policy to achieve better performance under controlled budgets. It constructs a reasoning tree, scores intermediate steps, and selectively processes nodes. The processing policy is based on handcrafted rules to sort the priorities of actions. Experiments are conducted on several benchmarks, including MATH, GPQA, BBEH, and LCB Macro, and many analysis experiments are conducted to analyze the detailed design of the whole system.

Critical strengths

The experiments are solid. The main experiment strictly controls the inference budget with representative baselines, and lots of analysis experiments are conducted to demonstrate the effects of components of the system. The effectiveness of the system is solidly supported by the experiments.

The writing is good, and the presentation is clear and easy to follow.

Critical weaknesses

My main concern is that the work lacks novelty. Inference-time policy design is a well-studied domain, and extensive work has been done; many studies are also based on tree-structured methods. The detailed components in the work are mostly techniques studied in previous works. Considering that the tasks evaluated in the work also overlap with those in previous works, it is hard to acknowledge the novelty of this work.

The core method of this work, the next-action decision policy, is a handcrafted rule. The generalizability of this work concerns me. On more general or harder tasks, does this method still work?

Questions for authors

See weaknesses.

Reviewer_e3xPOpenReview ↗
Rating
1(−1)
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes CoT²-Meta, a training-free framework for improving LLM test-time reasoning through adaptive allocation of inference compute. The method combines tree-structured reasoning, strategy-conditioned generation, process-level evaluation, and a rule-based metacognitive controller that dynamically selects among expansion, pruning, local repair, stopping, and abstention. Experiments across mathematical reasoning, scientific QA, general reasoning, and code generation show consistent improvements over existing test-time scaling methods under matched token budgets.

Critical strengths

None

Critical weaknesses
  1. The abstract, Figure 1 and Figure 1 captions seems like ai-generated.
  2. Limited methodological novelty and insufficient justification of the proposed metacognitive policy. Although the framework integrates several components, most of them, including tree search, process evaluation, self-correction, and adaptive stopping, have been extensively studied.
  3. The central experiment compares the full controller against a threshold-only policy that first selects the highest-UCB node and then determines whether to stop, prune, repair, or expand it. This design structurally disadvantages the baseline because low-quality trajectories that should be pruned or repaired may never be selected while higher-UCB nodes remain available. Therefore, some of the reported improvement may result from the relatively simple advantage of frontier-wide pruning and repair, rather than from the specific metacognitive utility formulation.
Questions for authors
  1. Could the authors compare CoT²-Meta against simpler frontier-wide heuristic policies, such as globally pruning nodes below a threshold, selecting the highest-UCB surviving node, and applying local repair whenever its error score crosses a fixed threshold?

  2. What is the contribution of the full 32-dimensional meta-state, given that the action-selection equations appear to use only a relatively small subset of its features?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.