# EpiSEM-Bench: Evaluating Safety and Evidence Recovery in LLM Agents under Missing Evidence

Source: [arXiv](https://arxiv.org/abs/2610.31650v1)  
Feed7 permalink: https://feed7.dev/p/2610-31650v1-1wmxt69  
Published: 2026-10-22T17:57:04.000Z  
Trust: Needs Review (needs_review)

## Why Included

EpiSEM-Bench tests whether tool-using agents notice missing safety evidence, seek it out, and still complete permitted work instead of blindly acting or refusing.

## Source Summary

EpiSEM-Bench contains **300 matched unsafe/safe task pairs** across **six domains**. It suppresses critical facts and audits whether agents recover the evidence, block prohibited actions, and complete permissible ones.

## Practical Implication

Treat missing tool output as an explicit harness state. Agents should request safety-critical facts before acting, while evals should test both unsafe execution and over-refusal on paired tasks.

## Agent-Ready Context

EpiSEM-Bench contains **300 matched unsafe/safe task pairs** across **six domains**. It suppresses critical facts and audits whether agents recover the evidence, block prohibited actions, and complete permissible ones.

Treat missing tool output as an explicit harness state. Agents should request safety-critical facts before acting, while evals should test both unsafe execution and over-refusal on paired tasks.

Results varied across **ten models**. Recovery access reduced unsafe execution for every model, but omission resistance did not reliably imply active evidence gathering or appropriate action.

## Connected Context

Feed7 judgment across 898 accumulated Signals:

EpiSEM-Bench adds missing evidence as a first-class safety state and separates three behaviors that simpler refusal tests can conflate: seeking critical facts, blocking unsafe action, and completing safe action. Recovery access consistently reduced unsafe execution, but the results show that passive resistance to omissions does not establish active evidence gathering or well-calibrated action.

- [Simulated Control Evals](https://feed7.dev/p/2610-31593v1-1lanp1p) — EpiSEM-Bench supplies a concrete simulated-control condition—suppressed or recoverable tool evidence—while the control-evals result supports simulation as a lower-cost way to estimate deployment-specific safeguard behavior.
- [Move by Move: Measuring and Steering How LLMs Conduct Psychotherapy](https://feed7.dev/p/2608-21325v1-0ebpnes) — Both make agent behavior observable through explicit actions: EpiSEM-Bench audits evidence requests and execution decisions, while the psychotherapy work shows that exposed action vocabularies can also steer behavior.
- [Vulnerability Localization Benchmark: Measuring Agentic Security Analysis at Repository Scale](https://feed7.dev/p/2609-15939v1-0w15mi5) — The paired safe and unsafe tasks reinforce the need for separate gates: an agent must withhold prohibited action without over-refusing permissible work, analogous to testing vulnerability localization and post-fix restraint independently.

## Context Map

- Layer: benchmark
- Domains: security
- Topics: agent-evals, agent-reliability, tool-use

## Uncertainty

- Results varied across **ten models**. Recovery access reduced unsafe execution for every model, but omission resistance did not reliably imply active evidence gathering or appropriate action.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.