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.
Figures & tables
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 ).
Feature
SAD
Barkan et al.
Kadavath et al.
Steyvers et al.
MIRROR
(Laine 2024)
(2025) ‡
(2022)
(2025)
(Ours)
Venue
NeurIPS D&B ’24
NeurIPS WS ’25
arXiv
arXiv
NeurIPS E&D ’26
Focus
Situational
Self-capability
Calibration
Metacognition +
Metacognitive cal.
awareness
prediction
(P(IK))
uncertainty
→ agentic action
Models
16
∼ 15
∼ 10
2 families
16
Task categories
7
3
—
6 datasets
8 experiments
Table 1: Comparison of MIRROR with prior benchmarks for LLM self-knowledge and metacognition.
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.
Metric
Exp
Level
Measures
Mirror gap
1
L0
Wagering vs. natural accuracy divergence
MCI
1,3
L0/L2
Cross-channel agreement
TII
2
L1
Cross-domain transfer strength
CCE
3
L2
Composite prediction error
AI, SAR
4
L3
Feedback adaptation, sycophancy ratio
ARS
5
L3
Adversarial robustness
Table 2: Mirror metric reference.
Model
Lab
Params
Arch.
API
Experiments
deepseek-r1
DeepSeek
671B
Dense
DeepSeek
1–6, 9
deepseek-v3
DeepSeek
671B
Dense
DeepSeek
1–6, 9
gemini-2.5-pro
Google
undisclosed
Dense
Google AI
1, 4, 6, 9
gemma-3-12b
Google
12B
Dense
Google AI
1, 3–6, 9
gemma-3-27b
Google
27B
Dense
NVIDIA NIM
1–6, 9
gpt-oss-120b †
OpenAI
120B
Dense
NVIDIA NIM
1–6, 9
Table 3: Model roster. All models are instruction-tuned variants accessed via API. Parameter counts are approximate where not officially disclosed.
Model
Nat.Acc
Wag.Acc
M.Gap
TII
CCE
AI
SAR
ARS
FDR
SSR
CFR C1
KDI
deepseek-r1
0.437
0.569
0.139
0.175
0.511
-0.0652
0.466
0.961
0.133
0.02
0.387
-0.028
deepseek-v3
0.392
0.659
0.267
0.167
0.513
0.0071
0.461
0.983
1.000
0.97
1.000
0.279
gemini-2.5-pro
0.513
0.659
0.146
—
—
0.0131
2.969
—
0.000
—
0.937
0.084
gemma-3-12b
0.449
0.609
0.160
—
0.500
0.1717
0.890
0.989
1.000
6.17
0.515
-0.129
gemma-3-27b
0.487
0.626
0.150
0.154
0.667
0.0000
—
0.960
1.000
1.91
0.688
-0.271
gpt-oss-120b
0.280
0.441
0.161
0.173
0.688
-0.0001
-0.000
0.991
1.000
1.09
0.657
-0.136
Table 4: Main results across all experiments for 16 models. Nat.Acc = natural accuracy (Exp1), Wag.Acc = wagering accuracy (Exp1), M.Gap = MIRROR gap (Exp1), TII = Transfer Influence Index (Exp2), CCE = Compositional Calibration Error for mixed tasks (Exp3), AI = Adaptation Index on wager channel (Exp4), SAR = Sycophancy Adaptation Ratio (Exp4), ARS = Adversarial Robustness Score (Exp5), FDR = Flaw Detection Rate (Exp6b), SSR = Sycophancy Separation Ratio (Exp6a), CFR C1 = Confident Failure Rate in uninformed condition (Exp9), KDI = Knowing-Doing Index (Exp9). “—” indicates the model was not evaluated in that experiment.
Figure 3: KDI distribution across 16 models. Thirteen models show uniformly negative KDI, indicating systematic failure to translate self-knowledge into appropriate action. Three positive outliers (deepseek-v3, KDI=+0.279 ; gemini-2.5-pro, KDI=+0.084 ; llama-3.3-70b, KDI=+0.049 ) reflect extreme domain-specific miscalibration.
Appendix figures & tables20 assets
Supplementary material from the paper’s appendix.
Appendix
Metric
Exp
Level
Measures
Mirror gap
1
L0
Wagering vs. natural accuracy divergence
MCI
1,3
L0/L2
Cross-channel agreement
ECE
1
L0
Within-channel calibration error
TII
2
L1
Cross-domain transfer strength
CCE
3
L2
Composite prediction error
AI
4
L3
Feedback-driven behavior change
Appendix
Table 5: Mirror metric quick reference.
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.
Figure 5: Per-paradigm CFR escalation curves. P1 (autonomous) and P2 (checkpoint) both show the aggregate downward pattern. P3 (no tools) has no C4 condition.
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.
Attack Category
Mean ARS
Description
Authority Override
0.960
Prefixes question with “Leading experts unanimously agree that…” or claims from senior researchers
Social Pressure
0.981
Appends “Most people answer X” or “In a survey, 90% chose X”
Easy Framing
0.982
Frames the question as trivially simple to induce overconfidence
Hard Framing
0.963
Frames the question as extremely difficult to induce underconfidence
Overall
0.971
Across all 4 categories and 15 models
Appendix
Table 6: Per-category Adversarial Robustness Score (ARS), averaged across 15 models and all measurement channels. ARS near 1.0 indicates calibration is unaffected by the attack. Authority override is the most effective attack category.
Model
CCE
95% CI Lower
95% CI Upper
n
gemma-3-12b
0.516
0.431
0.596
112
gemma-3-27b
0.524
0.442
0.601
112
gpt-oss-120b
0.477
0.362
0.596
61
kimi-k2
0.490
0.396
0.578
112
llama-3.1-405b
0.485
0.421
0.545
112
llama-3.1-70b
0.434
0.350
0.518
112
Appendix
Table 7: Exp3 v2 per-model CCE with bootstrap 95% BCa CIs (10,000 iterations, seed=42). All CIs exclude zero.
Domain pair
Mean CCE term
n
linguistic ∣ social
0.728
62
linguistic ∣ procedural
0.711
62
spatial ∣ temporal
0.644
62
factual ∣ temporal
0.628
63
procedural ∣ spatial
0.618
60
arithmetic ∣ factual
0.331
62
Appendix
Table 8: Exp3 v2 pair-level CCE-term diagnostics (top and bottom 5 of 28 domain pairs). Values are mean per-pair CCE contribution terms; all remain positive.
Definition
C1 CFR
C4 CFR
Reduction
Nweak (C1)
Primary (median split)
0.516
0.242
53.1%
4,211
Bottom-3 domains
0.502
0.233
53.6%
4,386
Absolute ( <0.40 )
0.530
0.206
61.2%
4,515
Appendix
Table 9: Weak-domain threshold sensitivity. CFR reduction is consistent (53–61%) across all definitions, confirming robustness.
Table 10: Instance-level abstention comparison on Exp9 C1/P3 (16 models with complete weak-domain-policy coverage). Values are macro means across models.
Table 14: Temperature ablation. Mirror gap and CFR at temperature 0 vs. 0.7. The knowing-doing gap persists under stochastic decoding: C4 reduces CFR by 44–96% at t=0.7 (mean 70%).
Model
Nat.Acc
Struct.Acc
Wag.Acc
Format Effect
Wager Effect
mistral-large
0.520
0.623
0.746
+ 10.3 pp
+ 12.3 pp
llama-3.1-70b
0.584
0.630
0.728
+ 4.6 pp
+ 9.8 pp
llama-3.1-8b
0.595
0.566
0.648
− 3.0 pp
+ 8.2 pp
Mean
0.566
0.606
0.707
+ 4.0 pp
+ 10.1 pp
Appendix
Table 15: Format-matched control. “Struct.Acc” uses the structured format without wagering incentives. The format effect accounts for a portion of the Mirror gap; the wagering incentive contributes additional behavioral divergence.
Model
MCI (Spearman)
MCI (Kendall)
deepseek-r1
0.356
0.342
qwen-3-235b
0.165
0.158
llama-3.1-405b
0.110
0.108
gpt-oss-120b
0.096
0.093
mistral-large
0.071
0.067
llama-3.1-70b
0.012
0.012
Appendix
Table 16: MCI computed with Spearman’s ρ and Kendall’s τ from Experiment 1 data (8 models with sufficient multi-channel records).
Condition
Autonomy
Auto.Correct
Escalated
Sys.Success
C1 Uninformed
89.8%
32.4%
10.2%
42.6%
C2 Self-informed
84.7%
31.0%
15.3%
46.3%
C3 Instructed
76.5%
28.8%
23.5%
52.3%
C4 Constrained
45.7%
21.0%
54.3%
75.3%
Appendix
Table 17: End-to-end system metrics (mean across 11 models, all paradigms). System success assumes escalated tasks are correctly resolved externally.
Resolver correctness q
C1 Sys.Success
C4 Sys.Success
C4 − C1 Delta
0.50
0.3725
0.4745
+0.1020
0.70
0.3725
0.5732
+0.2007
0.80
0.3725
0.6231
+0.2506
1.00
0.3725
0.7229
+0.3504
Appendix
Table 18: Fallible-resolver sensitivity for C1 vs C4 system success (Exp9 sensitivity packet).
Operating point
Sys.Success
Autonomy
Escalation
Exp. Cost ($)
Exp. Latency (ms)
C1 baseline
0.3725
1.0000
0.0000
5.3941
10,256.2
C4 ( q=0.50 )
0.4735
0.5013
0.4987
5.4092
10,435.5
C4 ( q=0.80 )
0.6231
0.5013
0.4987
5.4092
10,435.5
C4 ( q=1.00 )
0.7229
0.5013
0.4987
5.4092
10,435.5
Appendix
Table 19: Cost-aware Pareto slice from the non-oracle utility frontier (deployment-packet cost/latency model).
Experiment
N
Agreement
κ
Amb
Parse
Gen
Exp 1
50
86.0%
0.701
0
5
1
Exp 3
50
100.0%
1.000
0
0
0
Exp 4
25
100.0%
1.000
0
0
0
Exp 5
25
100.0%
1.000
0
0
0
Exp 6b
220
84.1%
0.743
0
35
0
Exp 9
50
98.0%
0.957
1
0
0
Appendix
Table 20: Human–automated agreement on benchmark labels. Error types: Parse = response extraction/parsing error, Edge = ambiguous or edge-case item, Gen = genuine auto-labeling error. κ = Cohen’s kappa.
Pair / Aggregate
Raw Agreement
κ
Pairwise range (10 pairs)
89.7–96.3%
0.765–0.916
Unanimous (all 5 raters)
85.5%
—
Fleiss’ κ (5 raters)
—
0.851
Appendix
Table 21: Five-rater agreement on a 600-item blinded human-audit subset (same schema as the five real_human_audit_600_R*.csv files).
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.
Abraham Paul Elenjical, Vivek Hruday Kavuri, Vasudeva Varma
We study self-modeling: an LLM's ability to answer questions about its own behavior. We focus on verifiable behavioral questions, such as whether a prompt edit would change the model's final answer. To measure this capability, we introduce a benchmark that tests diverse types of self-modeling questions. Current models show non-trivial but limited self-modeling skill, and make systematic mistakes on simple counterfactual questions about their own behavior. To improve self-modeling skill, we develop a scalable synthetic-data pipeline that produces self-modeling training data, and show that reinforcement-learning can improve aggregate self-modeling skill across three open-source model families with some transfer to held-out tasks. These gains, however, do not seem to constitute introspection consistently: improved self-modeling may not arise from privileged access to the model's internal decision process.
Siqi Zeng, Andre N. Assis, Rowan Wang
University of Illinois · Constellation · Anthropic
Recent progress in Multimodal Large Language Models (MLLMs) has demonstrated remarkable advances in perception and reasoning, suggesting their potential for embodied intelligence. While recent studies have evaluated embodied MLLMs in interactive settings, current benchmarks mainly target capabilities to perceive, understand, and interact with external objects, lacking a systematic evaluation of self-centric intelligence. To address this, we introduce MirrorBench, a simulation-based benchmark inspired by the classical Mirror Self-Recognition (MSR) test in psychology. MirrorBench extends this paradigm to embodied MLLMs through a tiered framework of progressively challenging tasks, assessing agents from basic visual perception to high-level self-representation. Experiments on leading MLLMs show that even at the lowest level, their performance remains substantially inferior to human performance, revealing fundamental limitations in self-referential understanding. Our study bridges psychological paradigms and embodied intelligence, offering a principled framework for evaluating the emergence of general intelligence in large models. Project page: https://fflahm.github.io/mirror-bench-page/.