Controllable User Simulation Formalized as Causal Inference Problem
A new arXiv paper (2605.11519) formalizes controllable user simulation for conversational AI evaluation as a causal inference problem. The authors argue that standard supervised fine-tuning on post-hoc trajectory labels introduces a look-ahead bias, breaking causal consistency. They prove that under policy shift, this bias causes evaluation metric variance to explode geometrically, a phenomenon termed "controllability collapse." The work bridges natural language evaluation with off-policy evaluation methodology, aiming to improve targeted counterfactual testing of conversational agents.
Key facts
- arXiv paper number: 2605.11519
- Announce type: new
- Formalizes controllable simulation as causal inference
- Standard SFT on post-hoc labels causes look-ahead bias
- Bias breaks causal consistency
- Under policy shift, variance of evaluation metrics explodes geometrically
- Phenomenon named controllability collapse
- Bridges natural language evaluation with off-policy evaluation
Entities
Institutions
- arXiv