2025 ICML

Bounded Rationality for LLMs: Satisficing Alignment at Inference-Time

M. Chehade, S. S. Ghosal, S. Chakraborty, A. Reddy, D. Manocha, H. Zhu, A. S. Bedi

Proceedings of the 41st International Conference on Machine Learning (ICML), 2025

An inference-time alignment method that treats acceptable behavior through satisficing rather than single-objective maximization.

Preview cropped from the paper PDF.

Contribution

What this paper adds

Frames LLM alignment through bounded rationality and satisficing.
Targets inference-time control without retraining the full model.
Studies trade-offs among helpfulness, harmlessness, and other response objectives.