Black-Box Auditing of Epistemic Reliability in Multi-Agent Debate Distillation
cs.LG, cs.AI, cs.CL, cs.CR
Submitted: 2026-09-26
Updated: 2026-09-26
Code: https://github.com/CSIRO-CQS-AI-alignment-Team/Epistemic-Reliability-Auditor
Terminology
Sources
- Training Language Models to Win Debates with Self-Play Improves Judge Accuracy
- Learning to Persuade Exposes How Easily LLMs Abandon Correct Beliefs
- Thought-Transfer: Indirect Targeted Poisoning Attacks on Chain-of-Thought Reasoning Models
- Subliminal Learning: Language models transmit behavioral traits via hidden signals in data
- Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
- AI safety via debate
- Prover-Verifier Games improve legibility of LLM outputs
- AgentArk: Distilling Multi-Agent Intelligence into a Single LLM Agent
- Auditing language models for hidden objectives
- Debate Helps Supervise Unreliable Experts
- AuditBench: Evaluating Alignment Auditing Techniques on Models with Hidden Behaviors
- Can LLM Agents Really Debate? A Controlled Study of Multi-Agent Debate in Logical Reasoning
- Adaptive auditing of AI systems with anytime-valid guarantees
- Universal and Transferable Adversarial Attacks on Aligned Language Models
Related papers
- Polynomial-Augmented Neural Networks (PANNs) with Weak Orthogonality Constraints for Enhanced Function and PDE Approximation
- AIRL-S: Unifying Reinforcement Learning and Search-Based Test-Time Scaling via Adversarial Inverse Reinforcement Learning
- Transformers as Bayesian In-Context Experimenters: Smoothness-Adaptive Efficient ATE Estimation
- Convergence issues in Relational Concept Analysis based on AOC-posets
- Beliefs Beyond Posteriors: Local-Consistency Optimisation for Bayesian Neural Networks
- Understanding Diffusion Models via Ratio-Based Function Approximation with SignReLU Networks