What Do Collusion Probes Read? Three Pitfalls in Evaluating Multi-Agent Collusion Monitors
cs.LG, cs.AI, cs.CL, cs.MA
Submitted: 2026-11-07
Updated: 2026-11-07
Code: https://github.com/mainul3195/collusion-probe-pitfalls
Terminology
Sources
- Understanding intermediate layers using linear classifier probes
- A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantification
- Refusal in Language Models Is Mediated by a Single Direction
- Linear probes rely on textual evidence: Results from leakage mitigation studies in language models
- "Did you lie?" Evaluating Lie Detectors across Model Scale and Belief-Verified Model Organisms
- You Can't Escape Your Activations: Evaluation Awareness and Multi-Agent Monitoring
- Steganalysis of Adaptive Covert Collusion in Tool-Using Agent Populations: A Black-Box, Cross-Principal Approach
- Detecting Strategic Deception Using Linear Probes
- Why Safety Probes Catch Liars But Miss Fanatics
- Telling FIBS: Probes Effectively Detect Sabotage and Catch Unverbalized Deception
- Prompt-Injection Probes Learn the Conversation, Not Just the Injection
- Pressure-Testing Deception Probes in LLMs: Scaling, Robustness, and the Geometry of Deceptive Representations
- When AUC 0.998 Is Not Enough: A Candidate Evaluation Protocol for Hidden-State Probes of Indirect Prompt Injection in Multimodal Computer-Use Agents
- Asymmetries in Spontaneous and Instructed Deception
- Multi-Agent AI Control: Distributed Attacks Hamper Per-Instance Monitors
- Hidden in Plain Text: Emergence & Mitigation of Steganographic Collusion in LLMs
- Secret Collusion among AI Agents: Multi-Agent Deception via Steganography
- One Probe Won't Catch Them All: Towards Targeted Deception Detection
- Benchmarking Deception Probes via Black-to-White Performance Boosts
- Qwen3 Technical Report
Related papers
- A Physics-Regulated Neural Framework for Learning 3D Grain Growth Dynamics
- FlexMoE: Learning One-for-All Multi-Budget Expert Pruning for Flexible MoE Deployment
- OptiMind: Improving Small Language Models for Optimization Modeling through Formulation Guidance and Multi-Turn Learning
- Staying in the Sweet Spot: Boost Reasoning Evolution via Responsive Rollout Difficulty Manipulation
- Examining the Impact of Data and Training Methods on Emergent and Subliminal Misalignment
- Rethinking Black-Box On-Policy Distillation: A Rubric-based Analysis