Evaluating and Understanding Scheming Propensity in LLM Agents
cs.AI
Submitted: 2026-03-02
Updated: 2026-11-17
Code: https://github.com/UKGovernmentBEIS/insp
Terminology
Sources
- Constitutional AI: Harmlessness from AI Feedback
- Towards evaluations-based safety cases for AI scheming
- Ctrl-Z: Controlling AI Agents via Resampling
- Scheming AIs: Will AIs fake alignment during training in order to get power?
- Safety Cases: How to Justify the Safety of Advanced AI Systems
- From surveillance to signalling: escalation channels as environmental controls for agentic AI
- Alignment faking in large language models
- Risks from Learned Optimization in Advanced Machine Learning Systems
- Evaluating Control Protocols for Untrusted AI Agents
- SHADE-Arena: Evaluating Sabotage and Monitoring in LLM Agents
- Agentic Misalignment: How LLMs Could Be Insider Threats
- Natural Emergent Misalignment from Reward Hacking in Production RL
- Frontier Models are Capable of In-context Scheming
- AgentMisalignment: Measuring the Propensity for Misaligned Behaviour in LLM-Based Agents
- Large Language Models Often Know When They Are Being Evaluated
- Scheming Ability in LLM-to-LLM Strategic Interactions
- Evaluating Frontier Models for Stealth and Situational Awareness
- Stress Testing Deliberative Alignment for Anti-Scheming Training
- PropensityBench: Evaluating Latent Safety Risks in Large Language Models via an Agentic Approach
- Lessons from a Chimp: AI "Scheming" and the Quest for Ape Language
Related papers
- Multi-Objective Structured Pruning of LLMs for Latency and Model Size Optimization
- MAVEN-T: Reinforced Heterogeneous Distillation for Real-Time Multi-Agent Trajectory Prediction
- Model Discovery Agent: LLM-assisted Bayesian experiment design for data-efficient discovery of mechanistic world models
- Embeddings for Preferences, Not Semantics
- Can We Trust Item Response Theory for AI Evaluation?
- Diagnosing Task Insensitivity in Language Agents