CoT2-Meta: Budgeted Metacognitive Control for Test-Time Reasoning

Dimensions overall score 7.0

• CoT2-Meta outperforms strong single-path, sampling-based, and search-based baselines including ReST-MCTS under matched inference budgets.result 95%
  Evidencepartial

  Under matched inference budgets, CoT2-Meta consistently outperforms strong single-path, sampling-based, and search-based baselines, including ReST-MCTS.

  Implicationpartial

  Explicitly stated in abstract with specific performance gains listed

  Verificationpartial

  partial

• CoT2-Meta achieves 92.8 EM on MATH, representing a +3.6 point gain over the strongest non-CoT2-Meta baseline.result 95%
  Evidencepartial

  On the default backbone, it achieves 92.8 EM on MATH... with gains over the strongest non-CoT2-Meta baseline of +3.6

  Implicationpartial

  Specific numeric results provided in abstract with clear comparison

  Verificationpartial

  partial

• CoT2-Meta achieves 90.4 accuracy on GPQA, representing a +5.2 point gain over the strongest non-CoT2-Meta baseline.result 95%
  Evidencepartial

  90.4 accuracy on GPQA... with gains over the strongest non-CoT2-Meta baseline of +5.2

  Implicationpartial

  Specific numeric results provided in abstract with clear comparison

  Verificationpartial

  partial

• CoT2-Meta uses explicit metacognitive control for expansion, pruning, repair, stopping, and fallback decisions during inference.method 90%
  Evidencepartial

  a meta-controller that allocates computation through expansion, pruning, repair, stopping, and fallback decisions

  Implicationpartial

  Directly stated in abstract and analysis sections describing framework components

  Verificationpartial

  partial

• CoT2-Meta converts partial reasoning trajectories into explicit control states using a deterministic state-construction function.technical 90%
  Evidencepartial

  COT2-META converts each partial reasoning trajectory into an explicit control state... where ot denotes the oracle output and ϕ(·) is a deterministic state-construction function

  Implicationpartial

  Direct technical description of the method's implementation

  Verificationpartial

  partial

• The framework remains effective across a broader 15-benchmark suite spanning knowledge and QA, multi-hop reasoning, coding, and out-of-distribution evaluation.result 85%
  Evidencepartial

  Beyond these core results, the framework remains effective across a broader 15-benchmark suite spanning knowledge and QA, multi-hop reasoning, coding, and out-of-distribution evaluation.

  Implicationpartial

  Explicitly stated in abstract with supporting evidence in analysis

  Verificationpartial

  partial

• CoT2-Meta shows consistent gains across backbone families including Claude-4.5, DeepSeek-V3.2 and Qwen2.5-VL-7B.result 85%
  Evidencepartial

  COT2-META also improves over strong inference-time baselines on DeepSeek-V3.2 and Qwen2.5-VL-7B... the relative advantage of explicit metacognitive control remains consistent across both closed and open-model settings

  Implicationpartial

  Directly stated in analysis section with specific model names

  Verificationpartial

  partial

• CoT2-Meta uses a UCB-style score for node selection that balances exploration and exploitation.technical 90%
  Evidencepartial

  the controller selects among them using a UCB-style score Score(n) = v(n) + β√(log(N+1)/(vis(n)+1)) where v(n) is the combined value, vis(n) is the visit count of node n

  Implicationpartial

  Direct technical description of the algorithm's implementation

  Verificationpartial

  partial

Competitive landscape

A training-free metacognitive framework that intelligently controls and optimizes test-time reasoning for improved accuracy and compute efficiency across diverse benchmarks.