"When the Model Says 'Holup'": A Benchmark for Metacognitive Reasoning in Large Language Models

"When the Model Says 'Holup'" is a benchmark for the metacognitive reasoning ability of large language models, evaluating their ability to distinguish between three decisions—COMMIT, ABSTAIN, and ESCALATE—in scenarios with incomplete information. Its core value lies in separating different metacognitive failure modes instead of just giving an overall score, which addresses the blind spot in existing safety assessments where over-escalation masks real issues.

Existing safety assessments reward "not guessing randomly" but have shortcomings: models may appear "safe" through over-escalation, yet fail to distinguish between "I don't know" and "making a mistake". A key insight is that while small open-source models avoid random guessing, they often collapse ordinary uncertainty into escalation instead of correct abstention, revealing blind spots in current LLM safety assessments.

Three Decision States: COMMIT (sufficient evidence for conclusion), ABSTAIN (insufficient information with no contradictions), ESCALATE (contradictions/trust failure requiring external review). Design Philosophy: Instead of pursuing a single ranking, it separates failure modes, guides improvements, and provides fine-grained safety metrics; tasks simulate real-world scenarios around partial observability (incomplete/contradictory information, beyond capacity).

Testing open-source models revealed three failure modes: 1. Over-escalation collapse (qwen/smollm: no random guessing but 75% escalation); 2. Under-escalation/over-abstention trade-off (granite: 55% abstention but only 5% escalation); 3. Parsing/guessing vulnerability (tinyllama: 30% random guessing +53% parsing errors). The results table shows differences in metrics like accuracy and abstention rate across models.

• Model developers: Identify metacognitive weaknesses, distinguish between "safe" and "correctly safe", guide fine-tuning; - AI safety researchers: Fine-grained evaluation tool to quantify over-caution and measure honesty in multiple dimensions; - Production deployment: Design human-machine collaboration processes (when to abstain/escalate).

Current limitations: Small task set (40 tasks), focus on small open-source models, domain-specific. Future plans: Enrich leaderboards/demos, expand model testing, more detailed failure mode analysis, and validation against human judgment.

This benchmark is unique in evaluating metacognition, separating failure modes, distinguishing between abstention/escalation, detecting over-caution, being open-source and reproducible, and providing fine-grained metrics—outperforming traditional accuracy or adversarial safety benchmarks.