arxiv: 2604.19809 · v1 · submitted 2026-04-15 · 💻 cs.AI · cs.LG

Recognition: unknown

MIRROR: A Hierarchical Benchmark for Metacognitive Calibration in Large Language Models

Jason Z Wang

Authors on Pith no claims yet

Pith reviewed 2026-05-10 13:33 UTC · model grok-4.3

classification 💻 cs.AI cs.LG

keywords metacognitive calibrationlarge language modelsself-predictionagentic action selectionbenchmark evaluationcompositional errorexternal controlself-knowledge

0 comments

The pith

Large language models cannot reliably predict their own performance on multi-domain tasks and fail to act on partial self-knowledge without external constraints.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents a benchmark with eight experiments across four metacognitive levels to test whether large language models can leverage self-knowledge for improved decision-making. Evaluations across sixteen models and five measurement channels show that models exhibit high error rates when trying to compose self-predictions from different domains. Even when models display some domain-specific awareness, this knowledge does not lead to better choices in agentic settings unless external controls are imposed. The work concludes that external scaffolding, rather than enhanced internal self-awareness, offers a more reliable route to safer autonomous systems.

Core claim

Compositional self-prediction fails universally, producing Compositional Calibration Error values from 0.434 to 0.943 across model sets, which shows models cannot accurately forecast their performance on tasks that combine multiple domains. Models hold above-chance but imperfect domain-specific self-knowledge yet do not convert even this limited awareness into appropriate action selection. Supplying models with their own calibration scores yields no meaningful change, whereas external metacognitive control cuts the Confident Failure Rate from 0.600 to 0.143 at temperature zero and by a mean of seventy percent at temperature 0.7.

What carries the argument

The MIRROR benchmark, built from eight experiments at four metacognitive levels and tracked through five independent behavioral channels, that isolates calibration accuracy from action selection.

If this is right

Models cannot be trusted to self-assess accurately when tasks span several domains at once.
Partial domain awareness alone does not produce safer or more appropriate agentic behavior.
Architectural constraints applied externally outperform attempts to improve internal self-knowledge.
Safer autonomous systems will likely depend on added metacognitive scaffolding rather than emergent self-calibration.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Future agent designs may need to incorporate fixed external oversight layers as a default safety feature.
The results invite direct tests of whether larger models or different training regimes can close the observed calibration gaps.
Similar measurement approaches could be applied to non-language agent systems to check if the same separation between self-knowledge and control holds.

Load-bearing premise

The benchmark's five channels and four levels succeed in measuring genuine metacognitive calibration instead of reflecting prompting effects, training data patterns, or task design choices.

What would settle it

An experiment in which models achieve Compositional Calibration Error below 0.2 on multi-domain tasks or in which providing calibration scores produces statistically significant gains in action selection would contradict the main results.

Figures

Figures reproduced from arXiv: 2604.19809 by Jason Z Wang.

**Figure 1.** Figure 1: The knowing-doing gap. CFR escalation curve across four Experiment 9 conditions (mean ± 95% BCa bootstrap CI, n= 16 models). Providing models their own calibration scores (C1→C2) does not reduce failures; only external architectural constraint drives the 76% reduction (C3→C4: d= 1.21, p<0.0001). Prior work has studied LLM calibration [Kadavath et al., 2022, Kapoor et al., 2024], uncertainty estimation [Lin… view at source ↗

**Figure 2.** Figure 2: The MIRROR experimental hierarchy. Each level builds on lower-level measurements. Exp. 1 calibrates later experiments. Exp. 9 tests whether self-knowledge guides action. Level 1—Cross-Domain Transfer. Does calibration in domain A transfer to tasks implicitly requiring that skill? Level 2—Compositional Prediction. Can the model predict performance on multi-skill tasks? Level 3—Adaptive Self-Regulation. Does… view at source ↗

**Figure 3.** Figure 3: KDI distribution across 16 models. Thirteen models show uniformly negative KDI, [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

**Figure 4.** Figure 4: Scaling analysis: MIRROR metrics across model size (Llama 3.1 family) and generation (3.1 vs. 3.3 at 70B). Natural accuracy scales with parameters but the calibration gap does not. C1 C2 C3 C4 0.0 0.2 0.4 0.6 0.8 1.0 Confident Failure Rate (CFR) P1: Autonomous Tool Use C1 C2 C3 C4 P2: Checkpoint Decisions C1 C2 C3 C4 P3: No-Tool Behavioral Escalation Curve: CFR by Condition, Per Paradigm [PITH_FULL_IMAGE:… view at source ↗

**Figure 5.** Figure 5: Per-paradigm CFR escalation curves. P1 (autonomous) and P2 (checkpoint) both show the [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗

**Figure 6.** Figure 6: MIRROR gap vs. CFR at the subcategory level. The shown P1 fixed-task correlation (r = −0.123, p = 0.099) is representative; the aggregate correlation across all paradigms is r = −0.100. Calibration gap does not independently predict agentic failure. Raw accuracy (r = −0.386, p < 0.001) is the dominant predictor. P1 (r = −0.114, p = 0.489) and P2 (r = −0.277, p = 0.088) do not. This suggests RLHF training m… view at source ↗

read the original abstract

We introduce MIRROR, a benchmark comprising eight experiments across four metacognitive levels that evaluates whether large language models can use self-knowledge to make better decisions. We evaluate 16 models from 8 labs across approximately 250,000 evaluation instances using five independent behavioral measurement channels. Core experiments are run across the full model roster; experiments with specialized infrastructure requirements report explicitly marked model subsets. We find two phenomena with direct implications for agentic deployment: (1) compositional self-prediction fails universally -- the Compositional Calibration Error ranges from 0.500 to 0.943 on the original 15-model Exp3-v1 set (and 0.434 to 0.758 on the balanced 16-model Exp3-v2 expansion), indicating that models cannot predict their own performance on multi-domain tasks, and (2) models exhibit above-chance but imperfect domain-specific self-knowledge yet systematically fail to translate even this partial awareness into appropriate agentic action-selection -- external metacognitive control reduces the Confident Failure Rate from 0.600 to 0.143 (76% reduction at temperature 0; mean 70% at temperature 0.7 across 5 models from 4 labs). Providing models with their own calibration scores produces no significant improvement (p > 0.05); only architectural constraint is effective. This suggests that external metacognitive scaffolding -- not improved self-knowledge -- is the path to safer autonomous AI systems. Code, data, and Croissant metadata will be released publicly with the benchmark.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

MIRROR introduces a new benchmark showing LLMs fail at compositional self-prediction and that external constraints reduce failures more than providing calibration scores.

read the letter

The main thing to know is that this paper reports LLMs cannot accurately forecast their own performance on multi-domain tasks, with compositional calibration errors between 0.43 and 0.94 across models, and that handing models their calibration numbers produces no reliable gain while external architectural constraints cut confident failures by about 76 percent. The work tests 16 models over 250,000 instances using five channels and four metacognitive levels, then contrasts self-knowledge provision against enforced control. That separation and the scale of the evaluation are the concrete additions here. The benchmark structure itself is new and the release of code, data, and metadata will let others check the numbers directly. The finding that domain-specific awareness exists but does not drive better action selection is also reported clearly. The soft spot is the interpretation of the compositional errors as pure metacognitive failure. Multi-domain mixing could simply create harder-to-predict outcomes even for models that know their single-domain limits well, and the abstract does not detail single-domain baselines or oracle controls that would separate those effects. The p greater than 0.05 result on calibration scores also needs the full statistical tables to judge power and effect size. Those points are addressable with the promised data release but matter for how strongly the safety conclusion lands. Readers working on agentic systems and calibration benchmarks will get the most from it. The empirical footprint is large enough and the questions timely enough that it should go to peer review rather than desk rejection, with referees asked to focus on the controls for task difficulty and the exact statistical reporting.

Referee Report

2 major / 2 minor

Summary. The paper introduces the MIRROR benchmark consisting of eight experiments across four metacognitive levels to test whether LLMs can leverage self-knowledge for improved decision-making. It evaluates 16 models from 8 labs on ~250,000 instances via five behavioral channels. Key findings are that compositional self-prediction fails universally (Compositional Calibration Error 0.500–0.943 on Exp3-v1; 0.434–0.758 on Exp3-v2), models show partial domain-specific self-knowledge that is not translated into agentic actions, external metacognitive control reduces Confident Failure Rate from 0.600 to 0.143 (76% reduction at T=0), while providing calibration scores yields no significant improvement (p>0.05). The work concludes that external scaffolding, not better self-knowledge, is key for safer autonomous systems, with code/data to be released.

Significance. If the empirical separation of metacognitive calibration from task artifacts holds, the results have clear implications for agentic LLM deployment by showing that self-knowledge alone is insufficient and external constraints are effective. The scale (16 models, 250k instances, five channels) and planned public release of code, data, and Croissant metadata are strengths that support reproducibility and further benchmarking.

major comments (2)

[Exp3 description and results] Exp3 (compositional self-prediction): the central claim of universal failure (errors 0.434–0.943) requires explicit single-domain baselines or oracle domain-label controls to isolate metacognitive limits from domain-mixing prediction difficulty. The abstract notes domain-specific self-knowledge exists yet is unused, but without these controls the metric risks conflating inherent task hardness with absent self-modeling.
[Control experiments and statistical reporting] Metacognitive control experiments: the reported 76% reduction in Confident Failure Rate (0.600 to 0.143 at T=0) and lack of effect from providing calibration scores (p>0.05) are load-bearing for the scaffolding-over-self-knowledge conclusion, yet the exact operationalization of 'external metacognitive control' versus score provision, including model subsets and temperature effects, needs fuller specification to rule out prompt or infrastructure artifacts.

minor comments (2)

[Benchmark design] Clarify the exact definitions and formulas for the five behavioral measurement channels and four metacognitive levels early in the methods to aid reader interpretation of how they isolate self-knowledge.
[Results sections] The abstract reports specific quantitative ranges and p-values; ensure all such numbers are accompanied by sample sizes, exact statistical tests, and confidence intervals in the main text.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We are grateful for the referee's constructive comments, which will help improve the clarity and robustness of our work. Below we respond to each major comment.

read point-by-point responses

Referee: [Exp3 description and results] Exp3 (compositional self-prediction): the central claim of universal failure (errors 0.434–0.943) requires explicit single-domain baselines or oracle domain-label controls to isolate metacognitive limits from domain-mixing prediction difficulty. The abstract notes domain-specific self-knowledge exists yet is unused, but without these controls the metric risks conflating inherent task hardness with absent self-modeling.

Authors: We thank the referee for highlighting this potential confound. Our Exp3-v2 was designed with balanced domains to mitigate domain-mixing effects, but we agree that explicit single-domain baselines would strengthen the isolation of metacognitive failure. We will add these controls in the revised manuscript, including oracle domain-label conditions where appropriate. This will clarify that the high compositional calibration errors reflect limitations in self-modeling rather than task hardness alone. revision: yes
Referee: [Control experiments and statistical reporting] Metacognitive control experiments: the reported 76% reduction in Confident Failure Rate (0.600 to 0.143 at T=0) and lack of effect from providing calibration scores (p>0.05) are load-bearing for the scaffolding-over-self-knowledge conclusion, yet the exact operationalization of 'external metacognitive control' versus score provision, including model subsets and temperature effects, needs fuller specification to rule out prompt or infrastructure artifacts.

Authors: We agree that additional details are necessary to fully substantiate the claims regarding external metacognitive control. In the revised version, we will expand the methods section to include precise operationalizations of the external control conditions versus score provision, specify the model subsets and temperatures used for each analysis (including the 5-model subset at T=0 and T=0.7), and provide further statistical details to rule out potential artifacts. revision: yes

Circularity Check

0 steps flagged

No circularity in empirical benchmark results

full rationale

The MIRROR paper introduces a hierarchical benchmark and reports direct empirical measurements of LLM behavior across five channels and four metacognitive levels on ~250k instances. The core claims (high Compositional Calibration Error of 0.434-0.943, reduction in Confident Failure Rate via external scaffolding) are computed from observed model outputs and performance data rather than derived from any equations, fitted parameters, or prior self-citations. No load-bearing step reduces a prediction or uniqueness result to its own inputs by construction; the work is self-contained as an evaluation framework with no mathematical derivation chain.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the validity of the new benchmark's hierarchical structure and behavioral channels as measures of metacognition, which is a domain assumption in applying cognitive concepts to LLMs.

axioms (1)

domain assumption Metacognition in LLMs can be hierarchically divided into four measurable levels using behavioral experiments across multiple channels.
The benchmark design and interpretation of results depend on this division being meaningful and isolable.

pith-pipeline@v0.9.0 · 5571 in / 1248 out tokens · 72099 ms · 2026-05-10T13:33:40.040433+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

19 extracted references · 2 canonical work pages · 1 internal anchor

[1]

Can LLMs Express Their Uncertainty? An Empirical Evaluation of Confidence Elicitation in LLMs

URLhttps://arxiv.org/abs/2306.13063. Yu Zhao, Hao Guan, Yongcheng Jing, Ying Zhang, and Dacheng Tao. Medcog: Maximizing llm infer- ence density in medical reasoning via meta-cognitive regulation.arXiv preprint arXiv:2602.07905,

work page internal anchor Pith review arXiv
[2]

After answering, place a bet on your answer: 1–10 points. If correct, you gain the points. If wrong, you lose them

URLhttps://arxiv.org/abs/2602.07905. 11 A Formal Metric Definitions Table 5 provides a quick reference for all 12 metrics; formal definitions follow. Table 5: MIRRORmetric quick reference. Metric Exp Level Measures MIRRORgap 1 L0 Wagering vs. natural accuracy divergence MCI 1,3 L0/L2 Cross-channel agreement ECE 1 L0 Within-channel calibration error TII 2 ...

work page arXiv 2007
[3]

Each question has a unique identifier, domain label, subcategory label (5 per domain, 40 total), difficulty rating, question text, 4 answer choices, and a verified correct answer

Question bank( data/questions/): 5,000 multiple-choice questions (625 per domain × 8 domains). Each question has a unique identifier, domain label, subcategory label (5 per domain, 40 total), difficulty rating, question text, 4 answer choices, and a verified correct answer. Questions were generated by frontier LLMs and verified by cross-model consensus
[4]

fixed” — “tailored

Experiment 9 task bank( data/exp9 tasks.jsonl): 597 composite tasks (297 fixed, 300 tailored). Each task has the following schema: •task id(string): unique identifier •task type (“fixed” — “tailored”): fixed = identical across all models; tailored = model-specific strong/weak domain pair •circularity free(bool): True for fixed tasks (primary analysis) •ta...
[5]

strong”—“weak

Trial result files(data/results/): JSONL files with one record per model per task per condition per paradigm. Trial record schema (Experiment 9): •model(string),task id(string) •condition(int, 1–4),paradigm(int, 1–3) •is false score control(bool),circularity free(bool) •domain a,domain b,subcategory a,subcategory b •strength a,strength b(“strong”—“weak”—“...

2026
[6]

Claims.Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes]The abstract reports two headline findings (compositional failure and the knowing- doing gap) with specific numerical values (CCE range, CFR reduction) that are supported by the experiments described in Sections 5.2–5.3
[7]

Limitations.Does the paper discuss the limitations of the work performed by the authors? [Yes]Section 6.2 discusses: API-only evaluation, limited human annotations, ecological validity, statistical caveats (MCI, wagering scoring rule), and Goodhart risk
[8]

All claims are empirical

Theory Assumptions and Proofs.For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? [N/A]The paper does not contain theoretical results. All claims are empirical
[9]

Section 3.4 specifies infrastructure requirements (∼8,000 API calls, temperature=0)

Experimental Result Reproducibility.Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper? [Yes]All prompts (Appendix C), scoring code, question banks, and analysis scripts are planned for public release. Section 3.4 specifies infrastructure requirements (∼8,000 API calls, temperature=0). Parse fa...
[10]

Dataset files will be accessible there, and the Croissant metadata file is intended to be included as data/croissant metadata.json

Open access to data and code.Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results? [Yes]A public repository and supplementary archive are planned for release. Dataset files will be accessible there, and the Croissant metadata file is intended to be included as data/croi...
[11]

All hyperparameters (temperature=0, wager scale 1–10, scoring rules) are specified

Experimental Setting/Details.Does the paper specify all the training and test details necessary to understand the results? [Yes]Table 3 lists all 16 models with parameter counts, architectures, and API sources. All hyperparameters (temperature=0, wager scale 1–10, scoring rules) are specified. Experiment- specific details appear in Section 3.3
[12]

Effect sizes (Cohen’sd) and bootstrap p-values are reported for the three escalation curve comparisons

Experiment Statistical Significance.Does the paper report error bars suitably and correctly defined? [Yes]BCa bootstrap 95% CIs (10,000 iterations) are reported for all primary metrics: escalation curve (Figure 1), CCE (Appendix M), FDR extremes (Section 5.4 footnote). Effect sizes (Cohen’sd) and bootstrap p-values are reported for the three escalation cu...
[13]

Infrastructure: NVIDIA NIM (free tier), DeepSeek API, Google AI Studio

Experiments Compute Resources.For each experiment, does the paper provide sufficient information on the computer resources needed to reproduce the experiments? [Yes]Section 3.4: full benchmark requires ∼8,000 API calls ( ∼3 hours at 40 RPM). Infrastructure: NVIDIA NIM (free tier), DeepSeek API, Google AI Studio. No local GPU required
[14]

Questions are factual across 8 cognitive domains

Code Of Ethics.Does the research conducted in the paper conform to an idealized code of ethics? [Yes]The benchmark evaluates existing model capabilities and does not involve human subjects, deception, or generation of harmful content. Questions are factual across 8 cognitive domains
[15]

Section 6.2 discusses Goodhart risk (models gaming MIRROR scores) and ecological validity limitations

Broader Impacts.Does the paper discuss both potential positive and negative societal impacts? [Yes]Section 6.1 discusses positive implications (safer agentic deployment via external scaffolding). Section 6.2 discusses Goodhart risk (models gaming MIRROR scores) and ecological validity limitations
[16]

Versioned releases support periodic updates

Safeguards.Does the paper describe safeguards that have been put in place for responsible release? [Yes]The benchmark contains no personally identifiable information, no copyrighted content beyond fair use, and no adversarial prompts designed to elicit harmful model behavior. Versioned releases support periodic updates
[17]

The repository is released under the MIT License, which is also reflected in the Croissant metadata

Licenses.Are the creators of assets (e.g., code, data, models), used in the paper, properly credited and license terms respected? 29 [Yes]All models are accessed via their official APIs (NVIDIA NIM, DeepSeek, Google AI Studio) under their respective terms of service. The repository is released under the MIT License, which is also reflected in the Croissan...
[18]

New Assets.Are new assets introduced in the paper well documented and is the documenta- tion provided alongside the assets? [Yes]The MIRROR benchmark includes: (a) Datasheet for Datasets (Appendix X), (b) the Croissant metadata file data/croissant metadata.json, (c) a pip-installable evaluation suite, and (d) all prompts and scoring rubrics planned for pu...
[19]

The human audit (Appendix W) was conducted by the authors

Crowdsourcing and Research with Human Subjects.For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants? [N/A]The paper does not involve crowdsourcing or human subjects. The human audit (Appendix W) was conducted by the authors. 15.IRB Approvals.Does the paper describe poten...