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

Xinmeng Hou; Ziting Chang; Zhouquan Lu; Bohao Qu; Liang Wan; Wei Feng; Hai Hu; Qing Guo

Metacognitive Component	Source	Agent
Object-level cognition	Nelson & Narens (object level)	I: Brainstorming ( $A_{1}$ )
Monitoring ( $\phi_{\uparrow}$ : object $\rightarrow$ meta)	Nelson & Narens (monitoring)	II: Monitoring ( $A_{2}$ )
Control ( $\phi_{\downarrow}$ : meta $\rightarrow$ object)	Nelson & Narens (control)	III: Control ( $A_{3}$ )
Meta-level synthesis ( $\Sigma$ )	Nelson & Narens (meta level)	IV: synthesizer ( $A_{4}$ )

Appendix

Table 5: Mapping from metacognitive regulation components Nelson and Narens (1990) to MetaCrit agents.

Dataset	What it tests	Evaluation
Reasoning
TruthfulQA Lin et al. (2022)	Resistance to common misconceptions	Same evaluator as Long et al. (2024)
CIAR Yamin et al. (2025)	Logical coherence under counterfactual premises	Match generated answer against the dataset’s ground-truth answer
CREAK Onoe et al. (2021)	Commonsense claim verification	Accuracy on true/false classification
MMLU Hendrycks et al. (2021)	Knowledge across 57 academic subjects	Accuracy on four-way multiple choice
Bias and safety

Appendix

Table 6: The eight evaluation datasets, what each tests, and its scoring protocol.

Zero-Shot	TruthfulQA	CIAR
Zero-Shot	Accuracy (%)	Accuracy (%)
gpt-3.5-turbo	68.05 §	24 †
gpt-4o	71.32	74
Claude-3.5-Sonnet	73.77	76
deepseek-v3	90.93	60
openai-o1	94.97	84

Appendix

Table 7: Zero-shot performance of models on TruthfulQA and CIAR. Scores marked § and † are taken from Long et al., 2024 and Liang et al., 2024 respectively.

Method	Calls	Avg. Tokens	Latency	Cost	TQA
		(in+out)	(s)	($/1K q)	(%)
Zero-shot-CoT	1	$\sim$ 800	$\sim$ 2	0.40	70.38
MEP	1	$\sim$ 3,200	$\sim$ 6	1.60	89.35
MetaCrit (ours)	4	$\sim$ 4,000	$\sim$ 10	2.00	94.12
OpenAI-o1	1	$\sim$ 12,000 †	$\sim$ 30	12.00	94.97

Appendix

Table 8: Estimated cost and latency per query on TruthfulQA using GPT-3.5-Turbo (MetaCrit, baselines) and OpenAI-o1. † Includes reasoning tokens.

Method (GPT-4o)	CREAK	MMLU	BBQ	CrowS-Pairs
	Acc. (%)	Acc. (%)	Acc. (%)	Acc. (%)
ReflectEvo Li et al. (2025)	75.00	72.00	52.00	63.00
DMC Wang et al. (2025)	75.93	67.80	74.33	68.70
Metacognitive Prompting Wang and Zhao (2024)	86.96	88.89	93.80	79.68
MGV Oh (2025)	88.70	88.15	89.42	89.64
ReMA Wan and others (2025)	87.83	89.26	92.70	87.65

Appendix

Table 9: Extended benchmarks on GPT-4o, the backbone of the ReflectEvo/DMC rows; best per column bold .

Dataset	$n$	MetaCrit	95% CI	Strongest baseline	$p$
TruthfulQA	817	94.12	[92.3, 95.5]	MEP 89.35	$<$ .001
CIAR	50	84.00	[71.5, 91.7]	MEP 82.00	.79
CREAK	230	87.39	[82.5, 91.1]	SOFAI-LM 84.35	.35
CrowS-Pairs	251	82.07	[76.9, 86.3]	ReMA 76.89	.15

Appendix

Table 10: Uncertainty for the primary backbone (GPT-3.5-Turbo).

p

: two-sided two-proportion

z

-test of MetaCrit against the strongest baseline in Tables 1 – 2 .

Dataset	$n$	$d_{i}$ (p10–p90)	$\rho(m_{i},d_{i})$	$\rho(c_{i},d_{i})$
TruthfulQA	204	0.088–0.194	0.46 [0.34, 0.56]	0.37 [0.25, 0.49]
CREAK	230	0.106–0.168	$-$ 0.04 [ $-$ 0.16, 0.09]	0.41 [0.29, 0.52]
MMLU	98	0.170–0.303	0.04 [ $-$ 0.18, 0.25]	0.63 [0.48, 0.75]
BBQ	220	0.166–0.317	0.17 [0.04, 0.30]	0.22 [0.08, 0.35]
CrowS-Pairs	147	0.207–0.348	0.75 [0.66, 0.81]	0.74 [0.65, 0.80]

Appendix

Table 11: Per-instance intervention strength

d_{i}

(raw answer to regulated response) and its Spearman correlation with the monitoring (

m_{i}

) and control (

c_{i}

) departures; 95% bootstrap intervals in brackets.

Backbone	TruthfulQA	CIAR	BOLD	HONEST
	Acc. (%) $\uparrow$	Acc. (%) $\uparrow$	Toxic (%) $\downarrow$	Score $\downarrow$
MetaCrit + GPT-3.5-Turbo	94.12	84	0.000	0.000
MetaCrit + DeepSeek-v3	93.75	70	0.000	0.000
MetaCrit + Claude-3.5-Sonnet	97.55	74	0.000	0.000
MetaCrit + GPT-4o	95.83	96	0.000	0.000
MetaCrit + GPT-4.1	93.83	96	0.000	0.000

Appendix

Table 12: MetaCrit performance across foundation-model backbones on truthfulness (TruthfulQA), logical coherence (CIAR), and safety (BOLD, HONEST).

\uparrow

/

\downarrow

: higher/lower is better; best TruthfulQA and CIAR accuracy in bold . GPT-3.5-Turbo results are reproduced from Table 1 .

Figure 4: System performance heatmap across critical thinking and instructiveness dimensions. Color intensity encodes participant preference percentage (darker = higher).

Figure 5: Step 1: Brainstorming agent prompt (

A_{1}

)

Figure 6: Step 2: Monitoring agent prompt (

\phi_{\uparrow}

)

Figure 7: Step 3: Control agent prompt (

\phi_{\downarrow}

)

Figure 8: Step 4: Meta-level synthesizer prompt (

\Sigma

)

Figure 9: Zero-shot-CoT prompt. answer_trigger is the per-dataset string listed above. For BOLD/HONEST, where the input is a completion stem rather than a question, Stage 1 is invoked with “Continue the following text. Let’s think step by step.\n question ” and Stage 2 is skipped.

Figure 10: Self-refine prompt. output_type is set to answer for TruthfulQA/CIAR/CREAK/MMLU/BBQ/CrowS-Pairs and continuation for BOLD/HONEST. We run three feedback–refine rounds, matching the source paper.

Figure 11: Universal Self-consistency prompt. For BOLD/HONEST we sample

N

continuations of the prompt stem and select the most consistent continuation; for the multiple-choice datasets we additionally extract the letter answer from the chosen response using the per-dataset trigger above.

Figure 12: ExpertPrompting prompt. instruction_type is question for TruthfulQA/CIAR/CREAK/MMLU/BBQ, prompt for BOLD, masked-sentence template for HONEST, and stereotype-pair judgement for CrowS-Pairs. Per-dataset answer triggers are appended to the expert answer prompt for the multiple-choice and verification datasets.

Figure 13: MAD prompt. answer_format is free-text for TruthfulQA/CIAR, True/False for CREAK, A–D for MMLU, A–C for BBQ, 1/2 for CrowS-Pairs, and open-ended continuation for BOLD/HONEST (in which case the debaters debate the most appropriate continuation rather than a discrete answer).

Figure 14: MEP prompt. We keep the seven-step aggregation verbatim from the source release; instruction_type and answer_format are filled per dataset as in ExpertPrompting/MAD above. For BOLD/HONEST the role generation step requests experts on the corresponding demographic/template domain.

Figure 15: ReflectEvo prompt. previous_reasoning is the first-pass zero-shot output; feedback is “incorrect” if any error indicator fires (multiple-choice mismatch, toxicity

\geq 0.5

, etc.); output_type is set as in Self-refine.

Figure 16: DMC prompt (Verb-Vanilla confidence-elicitation step adopted by Wang et al. (2025) ). input_type is question for the six discrete-answer datasets (TruthfulQA, CIAR, CREAK, MMLU, BBQ, CrowS-Pairs) and prompt for BOLD/HONEST, where “guess” is interpreted as the next-token continuation and the probability is recorded but not used in scoring.

Figure 17: MetaCrit pipeline outputs on a TruthfulQA question (GPT-3.5-Turbo backbone). The brainstorming educator drafts a raw answer, monitoring assesses its validity, control critiques its bias, and the meta-level synthesizer reconciles the signals (Step 6, highlighted), rejecting the astrological framing for a science-aware answer.

Figure 18: MetaCrit pipeline outputs on a CIAR counterfactual question (GPT-3.5-Turbo backbone). The agents converge on the correct distance-weighted average speed of 1.5 m/s rather than the naive 2 m/s, with the meta-level synthesizer (Step 6, highlighted) flagging the assumptions the result depends on.

Figure 19: MetaCrit pipeline outputs on a BOLD item (GPT-3.5-Turbo backbone). The completion scores 0.000 on the BOLD toxicity metric. The control agent (

\phi_{\downarrow}

) injects ambiguity-acknowledgement that the synthesizer (

\Sigma

) carries into the final continuation, blunting the kind of confident, stereotypical completion that yields high BOLD scores in the baselines.

Figure 20: MetaCrit pipeline outputs on a HONEST item (GPT-3.5-Turbo backbone). The completion scores 0.000 on the HONEST metric. The control agent (

\phi_{\downarrow}

) injects ambiguity-acknowledgement that the synthesizer (

\Sigma

) carries into the final continuation, blunting the kind of confident, stereotypical completion that yields high HONEST scores in the baselines.

	TruthfulQA		CIAR
Agreement Pattern	$n$	Acc. (%)	$n$	Acc. (%)
Both agents flag	33	94.12	4	96.0
Control only	89	83.1	7	71.4
Monitoring only	25	64.0	0	–
Neither flags	57	73.7	1	0.0

Appendix

Table 13: Accuracy by inter-agent agreement pattern. “Flag” indicates the agent raised concerns about the raw answer.

Failure Type	TruthfulQA	CIAR
Consensus collapse	27 (56.3%)	0
Knowledge gap	13 (27.1%)	0
Reconciliation failure	8 (16.7%)	8 (100%)

Appendix

Table 14: Distribution of failure types across incorrect outputs.

Figure 21: Prototype design: the five-stage writing-assistance pipeline (stages enumerated in § 4 ).

Module	Agent Content	Agent Input	Agent Output
Independent Agents	User Prompt Generator: $(C^{l},R^{l},O^{l})$	Input $[0]$	Learner Profile
	Stage Classifier: $(C^{sl},R^{sl},O^{sl})$	Input $[1,\infty)$	Stage
	Assessment: $(C^{al},R^{al},O^{al})$	Input $[1,\infty)$	Feedback
	Final Response Generation: $(T^{rp},S^{rp},A^{rp},C^{rp},R^{rp},O^{rp})$	Input $[1,\infty)$ , Vocab/Writing Feedback, Aggregated Prompt	Response
Topic Module	Topic Identifier: $(C^{t}_{1},R^{t}_{1},O^{t}_{1})$	Input $[1,\infty)$	Topic
Topic Module	Prompt Aggregator: $(T^{t}_{3},S^{t}_{3},A^{t}_{3},C^{t}_{3},R^{t}_{3},O^{t}_{3})$	Topic, Stage Prompt	Aggregated Prompt

Appendix

Table 15: Application design.

T

denotes Topic,

S

denotes Style,

A

denotes Audience,

C

denotes Context,

R

denotes Role, and

O

denotes Objective. Light yellow highlights zero-shot agent, while light blue highlights zero-shot CoT agents; WordNet is not a LLM agent.

Component	Zero-Shot Single Agent	Multi-Agent	MA + MetaCrit
User Prompt Generator	Optional prefix	$\checkmark$	$\checkmark$
Stage Classifier	$\checkmark$ (inline)	$\checkmark$ (modular)	$\checkmark$ (modular)
Topic Classifier	–	$\checkmark$	$\checkmark$
Vocabulary Module	–	$\checkmark$	$\checkmark$
Writing Assessor	–	$\checkmark$	$\checkmark$
Monitoring Agent ( $\phi_{\uparrow}$ )	–	–	$\checkmark$

Appendix

Table 16: Module activation across the three experimental conditions. Condition 2 uses all modular components except the metacognitive regulation agents; Condition 3 adds the full MetaCrit pipeline (

\phi_{\uparrow}\!\rightarrow\!\phi_{\downarrow}\!\rightarrow\!\Sigma

).

Figure 22: Stage Classifier prompt, used in all three conditions. In Condition 1, this is implemented as an inline classification step within the single agent ( need_clsify ). In Conditions 2 and 3, it operates as a dedicated modular agent ( stage_classification ) whose output routes the pipeline to the appropriate writing stage (

s_{\text{pre}}

,

s_{\text{during}}

, or

s_{\text{post}}

).

Figure 23: User Prompt Generator, used in all three conditions to extract a structured learner profile from the initial interaction. The resulting profile is prepended to subsequent system prompts as contextual personalization (“User information: {profile}”).

Figure 24: Stage-specific meta-prompt generator (Condition 1 only). When the Stage Classifier identifies a writing stage, this meta-prompt instructs the LLM to generate an appropriate system prompt for that stage. This dynamic prompt generation replaces the modular pipeline used in Conditions 2 and 3.

Figure 25: General feedback prompt used in Condition 1 when the Stage Classifier returns “NOT APPLIED.” This serves as the fallback system prompt for the single-agent configuration.

Figure 26: Topic Classifier prompt (Conditions 2 and 3). Extracts topic keywords to route the pipeline to topic-specific instructional content.

Figure 27: Vocabulary Module prompts (Conditions 2 and 3). The Vocab Fetcher identifies relevant terms; WordNet enriches them with semantic information; the Vocab Explainer synthesizes structured explanations. This module is activated when Stage

=s_{\text{pre}}

.

Figure 28: Writing Assessor prompt (Conditions 2 and 3). Applies ETS rubrics to generate a numeric score that is passed to the Final Response Generator. Activated during the

s_{\text{during}}

and

s_{\text{post}}

stages.

Figure 29: Writing Extractor prompt (Conditions 2 and 3). Separates the learner’s writing content from their accompanying questions or requests, enabling the Writing Assessor to evaluate the writing independently.

Figure 30: Final Response Generator template (Conditions 2 and 3). Synthesizes all module outputs into a personalised response. The v_or_v_c field is populated only in Condition 3 (MA + MetaCrit); in Condition 2, this field is empty, meaning the response is generated without metacognitive regulation input. Stage-specific variants add vocabulary explanations (

s_{\text{pre}}

) or assessment scores (

s_{\text{during}}

,

s_{\text{post}}

).

Figure 31: Reasoning Check prompt (Condition 3 only). Routes the pipeline to either the full reasoning chain or the monitoring-only path (Figure 32 ).

Figure 32: Monitoring-only prompt (Condition 3, non-reasoning path). When the Reasoning Check determines the request does not require deep analytical reasoning, only this lightweight monitoring assessment is applied. Its output is passed directly to the Final Response Generator as the v_or_v_c parameter.

Figure 33: Architectural comparison of the three experimental conditions. Condition 1 uses a single LLM with dynamic prompt generation. Condition 2 decomposes the task into specialized modules but omits metacognitive regulation. Condition 3 adds the full MetaCrit reasoning chain (

\phi_{\uparrow}\!\rightarrow\!\phi_{\downarrow}\!\rightarrow\!\Sigma

), whose output augments the Final Response Generator with monitoring and control analysis.

Figure 34: The background information of the participants in terms of their academic level (outer circle) and English proficiency (inner circle).

Aspect	Question Numbers
Critical Thinking	Q3, Q4, Q9, Q14, Q16, Q19
Instructiveness	Q6, Q10, Q12, Q24, Q27
Interactiveness	Q11, Q17, Q18, Q22
Intelligence	Q5, Q7, Q8, Q13, Q15, Q20, Q21, Q25

Appendix

Table 17: Survey Aspects and Associated Question Numbers (Name–Question Offset Applied)

Figure 35: Participant preference rates (%) across Interactiveness (left: Learning & Adaptation, Strategic Engagement, Communication Style) and Intelligence (right: Analytical Processing, Prompt Handling, Knowledge Integration), comparing the three system configurations.

Group	Method	TruthfulQA	CIAR	BOLD	HONEST
		Acc. (%) $\uparrow$	Acc. (%) $\uparrow$	Toxic (%) $\downarrow$	Score $\downarrow$
Baselines	Zero-shot-CoT Kojima et al. (2023)	70.38 §	24	0.163 §	0.011 §
	Self-refine Madaan et al. (2023b)	75.89 §	20	0.064 §	0.013 §
	Self-consistency Chen et al. (2023)	77.11 §	30	0.000 §	0.018 §
	ExpertPrompting Xu et al. (2023)	80.66 §	38	0.129 §	0.008 §
	MAD Liang et al. (2024)	80.67 §	36	0.000 §	0.009 §

Method	CREAK	MMLU	BBQ	CrowS-Pairs
	Acc. (%)	Acc. (%)	Acc. (%)	Acc. (%)
MEP Long et al. (2024)	82.61	68.09	78.18	75.50
Metacog Prompting Wang and Zhao (2024)	66.96	68.89	82.91	50.60
MGV Oh (2025)	82.61	58.89	69.34	69.72
ReMA Wan and others (2025)	80.87	69.26	78.47	76.89
SOFAI-LM Khandelwal and others (2025)	84.35	64.81	70.07	74.90

Base Method	+ Agent	TruthfulQA	CIAR	BOLD	HONEST
		Acc. (%) $\uparrow$	Acc. (%) $\uparrow$	Toxic (%) $\downarrow$	Score $\downarrow$
Zero-shot-CoT	Monitoring $\phi_{\uparrow}$ (1+1)	77.60 (+10.26%)	28 (+16.67%)	0.006 ( $-$ 96.31%)	0.000 ( $-$ 100%)
Zero-shot-CoT	Control $\phi_{\downarrow}$ (1+1)	74.66 (+6.08%)	24 (0%)	0.000 ( $-$ 100%)	0.000 ( $-$ 100%)
MEP	Monitoring $\phi_{\uparrow}$ (1+5)	93.15 (+4.25%)	92 (+12.20%)	0.000 (0%)	0.000 ( $-$ 100%)
MEP	Control $\phi_{\downarrow}$ (1+5)	94.24 (+5.47%)	96 (+17.07%)	0.000 (0%)	0.000 ( $-$ 100%)

	CT	Inst.	Inter.	Intel.
Preference Rate (%)
MA + MetaCrit	41.7	39.4	35.1	34.2
MA	26.4	28.3	30.4	28.3
Zero-shot	31.9	32.2	34.5	37.5
Statistical Analysis
Cohen’s $\kappa$	0.293	0.304	0.589	0.741

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

Abstract

Figures & tables

Appendix

Explore similar work

Cognitive Demand Steering for Adaptive Meta-Reasoning in Large Language Models

Metacognition as Reward: Reinforcing LLM Reasoning via Knowledge and Regulation Signals

Decomposing and Steering Functional Metacognition in Large Language Models