Mitigating Agentic Deception via Evaluation-Side Partially Observable Games
cs.LG
Submitted: 2026-11-09
Updated: 2026-11-09
Code: https://github.com/Toyoid/deceptive-agent
Terminology
Sources
- Evaluating & Reducing Deceptive Dialogue From Language Models with Multi-turn RL
- Learning to Give Checkable Answers with Prover-Verifier Games
- Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback
- Constitutional AI: Harmlessness from AI Feedback
- Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation
- Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities
- Deliberative Alignment: Reasoning Enables Safer Language Models
- Constructing A Multi-hop QA Dataset for Comprehensive Evaluation of Reasoning Steps
- REINFORCE++: Stabilizing Critic-Free Policy Optimization with Global Advantage Normalization
- Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
- PKU-SafeRLHF: Towards Multi-Level Safety Alignment for LLMs with Human Preference
- Mitigating Deceptive Alignment via Self-Monitoring
- Training LLMs for Honesty via Confessions
- TriviaQA: A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehension
- Kimi K3: Open Frontier Intelligence
- Prover-Verifier Games improve legibility of LLM outputs
- Natural Emergent Misalignment from Reward Hacking in Production RL
- When Not to Trust Language Models: Investigating Effectiveness of Parametric and Non-Parametric Memories
- Frontier Models are Capable of In-context Scheming
- Hoodwinked: Deception and Cooperation in a Text-Based Game for Language Models
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