ICLR 2026 Orals

Reliable Weak-to-Strong Monitoring of LLM Agents

Neil Kale, Chen Bo Calvin Zhang, Kevin Zhu, Ankit Aich, Paula Rodriguez, Christina Q Knight, Zifan Wang

Safety, Privacy & Alignment Sat, Apr 25 · 3:15 PM–3:25 PM · 204 A/B Avg rating: 6.00 (4–8)
Author-provided TL;DR

This paper introduces a monitor red teaming workflow to stress test systems for detecting covert misbehavior in LLM agents, finding that a well-designed monitor scaffold enables weaker models to oversee strong aware attackers.

Abstract

We stress test monitoring systems for detecting covert misbehavior in LLM agents (e.g., secretly exfiltrating data). We propose a monitor red teaming (MRT) workflow that varies agent and monitor awareness, adversarial evasion strategies, and evaluation across tool-calling (SHADE-Arena) and computer-use (CUA-SHADE-Arena) environments. We benchmark standard monitor scaffoldings and introduce a hybrid hierarchical--sequential design. Our experiments yield three findings. First, agent awareness dominates monitor awareness: agents that know they are monitored substantially degrade detection, while increasing monitor awareness helps less than expected. Second, monitor scaffolding matters: our hybrid design consistently outperforms baselines and enables weaker monitors to oversee stronger agents (a weak-to-strong effect). Third, targeted human oversight is key: escalating only pre-flagged cases improves TPR by 15% at FPR=0.01. Our work positions MRT as a standard workflow for stress-testing oversight, revealing robustness gaps in both LLM- and human-based monitoring. We release code, data, and logs to support further research.

One-sentence summary·Auto-generated by claude-haiku-4-5-20251001(?)

MetamerGen generates scene metamers aligned with human perception using foveal/peripheral features and latent diffusion.

Contributions·Auto-generated by claude-haiku-4-5-20251001(?)
  • Dual-stream DINOv2 representation fusing detailed foveal features with peripherally degraded features for scene context
  • Demonstrates sparse conditioning from gist and fixations sufficient for human-aligned scene generation
  • Behavioral validation identifying metamers through same-different experiments with participants
Methods used·Auto-generated by claude-haiku-4-5-20251001(?)
  • Latent diffusion models
  • DINOv2 embeddings
  • Foveated vision modeling
  • Scene understanding
Limitations (author-stated)·Auto-generated by claude-haiku-4-5-20251001(?)
  • Inherits limitations from pre-trained Stable Diffusion including difficulties with fine-grained facial details and limb articulations
    from the paper
  • Text generation often unreadable even when directly fixated
    from the paper
Future work (author-stated)·Auto-generated by claude-haiku-4-5-20251001(?)

Authors did not state explicit future directions.

Author keywords

  • Agent Safety
  • Chain-of-Thought Monitoring
  • Large Language Model

Related orals

Something off? Let us know →