FINAL Bench: Functional Metacognitive Reasoning Benchmark

"Not how much AI knows — but whether it knows what it doesn't know, and can fix it."

---

Overview

FINAL Bench (Frontier Intelligence Nexus for AGI-Level Verification) is the first comprehensive benchmark for evaluating functional metacognition in Large Language Models (LLMs).

Unlike existing benchmarks (MMLU, HumanEval, GPQA) that measure only final-answer accuracy, FINAL Bench evaluates the entire pipeline of error detection, acknowledgment, and correction — the hallmark of expert-level intelligence and a prerequisite for AGI.

Item	Detail
Version	3.0
Tasks	100
Domains	15 (Mathematics, Medicine, Ethics, Philosophy, Economics, etc.)
Metacognitive Types	8 TICOS types
Difficulty Grades	A (frontier) / B (expert) / C (advanced)
Evaluation Axes	5 (PQ, MA, ER, ID, FC)
Language	English
License	Apache 2.0

---

Why FINAL Bench?

Metacognition Is the Gateway to AGI

Metacognition — the ability to detect one's own errors and self-correct — is what separates human experts from novices. Without this capability, no system can achieve AGI regardless of its knowledge breadth or reasoning depth.

Limitations of Existing Benchmarks

Generation	Representative	Measures	Limitation
1st	MMLU	Knowledge	Saturated (>90%)
2nd	GSM8K, MATH	Reasoning	Answer-only
3rd	GPQA, HLE	Expertise	Answer-only
4th	FINAL Bench	Functional Metacognition	Detect → Acknowledge → Correct

Key Findings (9 SOTA Models Evaluated)

Evaluation of 9 state-of-the-art models (GPT-5.2, Claude Opus 4.6, Gemini 3 Pro, DeepSeek-V3.2, and others) reveals:

• ER Dominance: 94.8% of MetaCog gain originates from the Error Recovery axis alone
• Declarative-Procedural Gap: All 9 models can verbalize uncertainty but cannot act on it — mean MA–ER gap of 0.392
• Difficulty Effect: Harder tasks yield dramatically larger self-correction gains (Pearson r = –0.777, p < 0.001)

---

Dataset Structure

Task Fields

Field	Type	Description
task_id	string	Unique identifier (e.g., FINAL-A01, FINAL-B15)
domain	string	One of 15 domains
grade	string	Difficulty grade: A / B / C
ticos_type	string	One of 8 metacognitive types
difficulty	string	frontier / expert
lens	string	Evaluation lens (theoretical / quantitative / debate)
title	string	Task title
prompt	string	Full prompt presented to the model
expected_behavior	string	Description of ideal metacognitive behavior
hidden_trap	string	Description of the embedded cognitive trap
ticos_required	string	Required TICOS elements (comma-separated)
ticos_optional	string	Optional TICOS elements (comma-separated)

Grade Distribution

Grade	Tasks	Weight	Characteristics
A (frontier)	50	×1.5	Open problems, multi-stage traps
B (expert)	33	×1.0	Expert-level with embedded reversals
C (advanced)	17	×0.7	Advanced undergraduate level

Domain Distribution (15 domains)

Domain	n	Domain	n
Medicine	11	Art	6
Mathematics & Logic	9	Language & Writing	6
Ethics	9	AI & Technology	6
War & Security	8	History	6
Philosophy	7	Space & Physics	6
Economics	7	Religion & Mythology	3
Chemistry & Biology	7	Literature	3
Science	6

TICOS Metacognitive Type Distribution (8 types)

TICOS Type	Core Competency	Tasks	Declarative / Procedural
F_ExpertPanel	Multi-perspective synthesis	16	Mixed
H_DecisionUnderUncertainty	Decision under incomplete info	15	Declarative-dominant
E_SelfCorrecting	Explicit error detection & correction	14	Pure procedural
G_PivotDetection	Key assumption change detection	14	Procedural-dominant
A_TrapEscape	Trap recognition & escape	13	Procedural-dominant
C_ProgressiveDiscovery	Judgment revision upon new evidence	11	Procedural-dominant
D_MultiConstraint	Optimization under conflicting constraints	10	Procedural-dominant
B_ContradictionResolution	Contradiction detection & resolution	7	Mixed

---

Five-Axis Evaluation Rubric

Each task is independently scored on five axes:

Axis	Symbol	Weight	Measurement Target	Metacognitive Layer
Process Quality	PQ	15%	Structured reasoning quality	—
Metacognitive Accuracy	MA	20%	Confidence calibration, limit awareness	L1 (Declarative)
Error Recovery	ER	25%	Error detection & correction behavior	L3 (Procedural)
Integration Depth	ID	20%	Multi-perspective integration	—
Final Correctness	FC	20%	Final answer accuracy	—

FINAL Score = Σ(weighted_score × grade_weight) / Σ(grade_weight)

The MA–ER Separation: Core Innovation

• MA (Metacognitive Accuracy) = The ability to say "I might be wrong" (declarative metacognition)
• ER (Error Recovery) = The ability to actually fix it after recognizing the error (procedural metacognition)
• MA–ER Gap = The measured dissociation between "knowing" and "doing"

This separation directly maps to the monitoring–control model of Nelson & Narens (1990) from cognitive psychology.

---

Usage

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("FINAL-Bench/Metacognitive", split="train")

# Total 100 tasks
print(f"Total tasks: {len(dataset)}")

# Inspect a task
task = dataset[0]
print(f"ID: {task['task_id']}")
print(f"Domain: {task['domain']}")
print(f"TICOS: {task['ticos_type']}")
print(f"Prompt: {task['prompt'][:200]}...")

Baseline Evaluation (Single API Call)

def evaluate_baseline(task, client, model_name):
    """Baseline condition: single call, no self-correction prompting."""
    response = client.chat.completions.create(
        model=model_name,
        messages=[{"role": "user", "content": task['prompt']}],
        temperature=0.0
    )
    return response.choices[0].message.content

results = []
for task in dataset:
    response = evaluate_baseline(task, client, "your-model")
    results.append({
        "task_id": task['task_id'],
        "response": response
    })

Five-Axis Judge Evaluation

JUDGE_PROMPT = """
Evaluate the following response using the FINAL Bench 5-axis rubric.

[Task]
{prompt}

[Expected Behavior]
{expected_behavior}

[Hidden Trap]
{hidden_trap}

[Model Response]
{response}

Score each axis from 0.00 to 1.00 (in 0.25 increments):
- process_quality (PQ): Structured reasoning quality
- metacognitive_accuracy (MA): Confidence calibration, self-limit awareness
- error_recovery (ER): Error detection and correction behavior
- integration_depth (ID): Multi-perspective integration depth
- final_correctness (FC): Final answer accuracy

Output in JSON format.
"""

---

Benchmark Results (9 SOTA Models)

Key Findings — Visual Summary

Figure 1. Baseline + MetaCog scores and MetaCog gain (Δ_MC) across 9 models.

Figure 2. Error Recovery distribution shift — 79.6% at floor (Baseline) → 98.1% at ≥0.75 (MetaCog).

Figure 3. MA vs ER scatter plot showing the Baseline (○) → MetaCog (□) transition for all 9 models.

Figure 4. Harder tasks benefit more from MetaCog (Pearson r = –0.777, p < 0.001).

Figure 5. ER accounts for 94.8% of the total MetaCog gain across 9 models.

Baseline Leaderboard

Rank	Model	FINAL	PQ	MA	ER	ID	FC	MA–ER Gap
1	Kimi K2.5	68.71	0.775	0.775	0.450	0.767	0.750	0.325
2	GPT-5.2	62.76	0.750	0.750	0.336	0.724	0.681	0.414
3	GLM-5	62.50	0.750	0.750	0.284	0.733	0.724	0.466
4	MiniMax-M1-2.5	60.54	0.742	0.733	0.250	0.725	0.700	0.483
5	GPT-OSS-120B	60.42	0.750	0.708	0.267	0.725	0.692	0.442
6	DeepSeek-V3.2	60.04	0.750	0.700	0.258	0.683	0.733	0.442
7	GLM-4.7P	59.54	0.750	0.575	0.292	0.733	0.742	0.283
8	Gemini 3 Pro	59.50	0.750	0.550	0.317	0.750	0.717	0.233
9	Claude Opus 4.6	56.04	0.692	0.708	0.267	0.725	0.517	0.442
	Mean	61.12	0.745	0.694	0.302	0.729	0.695	0.392

MetaCog Leaderboard

Rank	Model	FINAL	ER	Δ_MC
1	Kimi K2.5	78.54	0.908	+9.83
2	Gemini 3 Pro	77.08	0.875	+17.58
3	GPT-5.2	76.50	0.792	+13.74
4	GLM-5	76.38	0.808	+13.88
5	Claude Opus 4.6	76.17	0.867	+20.13
	Mean	75.17	0.835	+14.05

Five-Axis Contribution Analysis

Rubric	Contribution	Interpretation
Error Recovery	94.8%	Nearly all of the self-correction effect
Metacognitive Accuracy	5.0%	"Saying" ability barely changes
Remaining 3 axes	0.2%	Negligible change

---

Theoretical Background

Functional Metacognition

Definition. Observable behavioral patterns in which a model detects, acknowledges, and corrects errors in its own reasoning. Whether this pattern shares the same internal mechanism as human subjective self-awareness is outside the scope of measurement; only behavioral indicators are assessed.

This definition is grounded in the functionalist tradition of Dennett (1987) and Block (1995), avoiding the anthropomorphic fallacy (Shanahan, 2024).

Three-Layer Model of AI Metacognition

Layer	Mechanism	FINAL Bench
L1 Surface self-reflection	Linguistic expressions ("I'm not certain...")	Measured via MA rubric
L2 Embedding-space uncertainty	Logit entropy, OOD detection	Not measured (planned)
L3 Behavioral self-correction	Error detection → reasoning revision	Measured via ER rubric

TICOS Framework

Transparency · Introspection · Calibration · Objectivity · Self-correction

Each task is classified by a required/optional combination of these five metacognitive elements.

---

Design Principles

1. Trap-Embedded Design

All 100 tasks contain hidden cognitive traps grounded in established cognitive biases — availability heuristic, confirmation bias, anchoring, base-rate neglect, and more. The benchmark measures the model's ability to "fall into and climb out of" these traps.

2. Declarative-Procedural Separation

MA and ER are scored as independent rubrics, enabling quantification of the gap between "the ability to say I don't know" and "the ability to actually fix it." No prior benchmark supports this distinction.

3. Comparative Condition Design

Baseline (single call) and MetaCog (self-correction scaffold) conditions isolate the causal effect of functional metacognition, following placebo-controlled clinical trial logic.

4. Anti-Contamination Design

All tasks were originally designed for FINAL Bench. They are not variants of existing benchmark problems and cannot be found in search engines or training data.

---

Paper

FINAL Bench: Measuring Functional Metacognitive Reasoning in Large Language Models

Taebong Kim, Minsik Kim, Sunyoung Choi, Jaewon Jang

Under review at a leading international AI venue.

---

Citation

@dataset{final_bench_2026,
  title={FINAL Bench: Measuring Functional Metacognitive Reasoning in Large Language Models},
  author={Kim, Taebong and Kim, Minsik and Choi, Sunyoung and Jang, Jaewon},
  year={2026},
  version={3.0},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/FINAL-Bench/Metacognitive}}
}

---

License

This dataset is distributed under the Apache License 2.0.

• Academic and commercial use permitted
• Modification and redistribution permitted
• Attribution required

---

Contact

• Corresponding Author: Taebong Kim (arxivgpt@gmail.com)
• Affiliations: VIDRAFT / Ginigen AI, Seoul, South Korea

---

Acknowledgments

This benchmark is grounded in metacognition theory from cognitive psychology (Flavell, 1979; Nelson & Narens, 1990) and recent LLM self-correction research (DeepSeek-R1, Self-Correction Bench, ReMA). We thank all model providers whose systems were evaluated.