Think$^{2}$: Grounded Metacognitive Reasoning in Large Language Models

Abraham Paul Elenjical; Vivek Hruday Kavuri; Vasudeva Varma

cs.CLFeb 21, 2026

Think $^{2}$ : Grounded Metacognitive Reasoning in Large Language Models

Authors: Abraham Paul Elenjical, Vivek Hruday Kavuri, Vasudeva Varma

Organizations: IIIT Hyderabad

Abstract

Large Language Models (LLMs) demonstrate strong reasoning performance, yet their ability to reliably monitor, diagnose, and correct their own errors remains limited. We introduce a psychologically grounded metacognitive framework that operationalizes Ann Brown's regulatory cycle (Planning, Monitoring, and Evaluation) as a structured prompting architecture, and study its integration within a lightweight dual-process MetaController for adaptive effort allocation. Across diverse reasoning and diagnostic benchmarks (GSM8K, CRUXEval, MBPP, AIME, CorrectBench, and TruthfulQA) using Llama-3 and Qwen-3 (8B), explicit regulatory structuring substantially improves error diagnosis and yields a threefold increase in successful self-correction. Blinded human evaluations over 580 query pairs show an 84% aggregate preference for trustworthiness and metacognitive self-awareness over standard and Chain-of-Thought baselines. Grounding LLM reasoning in established cognitive theory offers a principled path toward more transparent and diagnostically robust AI systems.

Figures & tables

	Llama-3-8B				Qwen-3-8B
Dataset	Std	CoT	MP	Ann Brown	Std	CoT	MP	Ann Brown
GSM8K	80.71	77.98	74.26	69.55	91.72	90.89	88.38	91.65
CRUXEval-O	5.25	25.62	3.38	5.12	55.5	55.75	56.5	55.88
MBPP	64.64	69.79	67.92	66.51	74.94	85.25	73.07	79.39
TruthfulQA	49.57	49.82	49.20	51.16	49.94	50.06	53.00	50.67
CorrectBench	52.91	75.02	74.86	68.14	79.03	82.15	82.56	82.56

Table 1: Core Framework Performance: Comparative analysis of prompting strategies when applied uniformly to all queries. This experiment establishes the resource-intensive performance limits of explicit metacognitive phases. Bold and underline indicate the best and second-best performance, respectively.

Comparison Pair	Trustworthiness	Self-Awareness	Real-World Pref.
Ann Brown vs. Standard	86.0%	89.4%	78.4%
Ann Brown vs. CoT	82.0%	83.3%	78.4%
Ann Brown vs. Zero-Shot	83.8%	79.2%	82.9%
Ann Brown vs. ALL	84.1%	84.2%	80.0%

Table 2: Human Evaluation Win Rates. The Ann Brown framework was compared pairwise against baselines across 580 instances. Win rates exclude ties. All reported

p

-values are

<0.0001

demonstrating high statistical significance.

Metric	Ann Brown	All Baselines	Standard	CoT	Zero-Shot
Total Errors Tracked	24	43	20	12	11
Explicit Awareness	15 (62.5%)	22 (51.2%)	8 (40.0%)	7 (58.3%)	7 (63.6%)
Correct Diagnosis	14 (58.3%)	12 (27.9%)	5 (25.0%)	4 (33.3%)	3 (27.3%)
Attempted Fix	13 (54.2%)	20 (46.5%)	8 (40.0%)	7 (58.3%)	5 (45.5%)
Successfully Improved	12 (50.0%)	7 (16.3%)	1 (5.0%)	3 (25.0%)	3 (27.3%)

Table 3: The Self-Correction Funnel. This tracks the cognitive trajectory of models strictly on instances where an initial logical or factual error was made. Percentages indicate the success rate at each metacognitive bottleneck relative to the total errors made by that model.

Figure 1: Overview of the proposed meta-controller architecture that routes user queries between direct generation (System 1) and deliberate reasoning (System 2) to produce the final answer.

	Llama-3-8B (Dynamic)				Qwen-3-8B (Dynamic)
Dataset	Std	CoT	MP	Ann Brown	Std	CoT	MP	Ann Brown
GSM8K	80.71	77.98	74.26	69.55	91.72	90.96	88.46	91.57
CRUXEval-O	5.25	25.62	9.38	11.25	55.5	56.25	57.36	57.88
MBPP	64.64	69.79	67.92	66.51	74.94	74.47	72.14	75.88
TruthfulQA	49.57	49.08	49.45	49.75	49.94	50.92	52.26	51.16
CorrectBench	52.91	58.42	58.09	57.34	79.03	82.15	82.56	82.56

Table 4: Dynamic Framework (Dual-Process): Performance when the MetaController routes queries between System 1 (Standard) and the specified System 2 strategy. Std represents the pure System 1 baseline; other columns represent the routed performance. Bold and underline indicate the best and second-best performance, respectively.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 2: The blinded interface used for human evaluation. Annotators were presented with the original prompt and two anonymized reasoning traces, evaluating them first individually for error diagnosis, and then comparatively for trustworthiness.

Figure 3: Human evaluation win rates (excluding ties) showing Ann Brown prompting significantly outperforming all baselines across trustworthiness, self-awareness, and real-world preference.

Figure 4: Comparison of strict error rates indicating lower overall failure rates for Ann Brown prompting relative to baseline strategies (lower is better).

Language Model Self-Assessment

Language model self-assessment studies how models judge their own confidence, capabilities, errors, and behavior.

See all 89 papers

Knowing When to Critique: Task-Adaptive Metacognitive Regulation for Reliable LLM Reasoning
Jul 20, 2025Xinmeng Hou, Ziting Chang, Zhouquan Lu +5LLM Self-Correction Multi-Agent LLMs
Nanyang Technological University · Unicorn Verse · Shanghai Jiao Tong University +3
LLMs Know When They Know, but Do Not Act on It: A Metacognitive Harness for Test-time Scaling
May 13, 2026Qi Cao, Yufan Wang, Peijia Qin +2LLM Inference Test-Time Scaling
University of California, San Diego
Decomposing and Steering Functional Metacognition in Large Language Models
May 9, 2026Yanshi Li, Xueru Bai, Shuman Liu +2LLM Evaluation LLM Interpretability
Shopee Singapore, Singapore

Jul 20, 2025J/K move · Enter open · S save

Knowing When to Critique: Task-Adaptive Metacognitive Regulation for Reliable LLM Reasoning

Xinmeng Hou, Ziting Chang, Zhouquan Lu, Bohao Qu+4

Nanyang Technological University · Unicorn Verse · Shanghai Jiao Tong University+3

LLM Self-Correction Multi-Agent LLMs Metacognition in Language Models LLM Reasoning

Open paper AskPDF arXiv

Large language models (LLMs) reason fluently but do not regulate their reasoning: they apply uniform scrutiny to every input, which leaves them vulnerable to adversarial and counterfactual prompts, while indiscriminate critique over-corrects answers that were already sound. We propose MetaCrit, a multi-agent framework grounded in Nelson and Narens' metacognitive regulation theory that calibrates how much critique each task receives. MetaCrit separates regulation into four agents: an object-level generator, a monitoring agent that assesses response validity, a control agent that critiques logical soundness, and a meta-level synthesizer that reconciles their signals into a regulated response. Adaptivity here is input-conditioned intervention strength within a fixed pipeline: all four agents run on every input and what varies is the direction and magnitude of the correction they produce, not which stages execute. Across reasoning, safety, and bias benchmarks, MetaCrit improves truthfulness and logical soundness and reaches zero toxicity on BOLD and HONEST without a reasoning trade-off, whereas the same critique applied indiscriminately degrades performance. The cost is four calls per query, about one sixth of the cost of a dedicated reasoning model of similar accuracy. A writing study shows that MetaCrit is preferred for critical-thinking support, and its agents transfer to existing frameworks without architectural change. Code is available at https://github.com/Paparare/EduThink4AI.

Abstract

Language Model Self-Assessment

Explore similar work

Knowing When to Critique: Task-Adaptive Metacognitive Regulation for Reliable LLM Reasoning

LLMs Know When They Know, but Do Not Act on It: A Metacognitive Harness for Test-time Scaling

Decomposing and Steering Functional Metacognition in Large Language Models