Bounded Rationality for LLMs: Satisficing Alignment at Inference-Time

Mohamad Fares El Hajj Chehade, Soumya Suvra Ghosal, Souradip Chakraborty, Avinash Reddy, Dinesh Manocha, Hao Zhu, Amrit Singh Bedi
Proceedings of the 42nd International Conference on Machine Learning, PMLR 267:7617-7633, 2025.

Abstract

Aligning large language models with humans is challenging due to the inherently multifaceted nature of preference feedback. While existing approaches typically frame this as a multi-objective optimization problem, they often overlook how humans actually make decisions. Research on bounded rationality suggests that human decision making follows satisficing strategies- optimizing primary objectives while ensuring others meet acceptable thresholds. To bridge this gap and operationalize the notion of satisficing alignment, we propose SITAlign: an inference-time framework that addresses the multifaceted nature of alignment by maximizing a primary objective while satisfying threshold-based constraints on secondary criteria. We provide theoretical insights by deriving sub-optimality bounds of our satisficing-based inference alignment approach. We empirically validate SITAlign’s performance through extensive experimentation on multiple benchmarks. For instance, on the PKU-SafeRLHF dataset with the primary objective of maximizing helpfulness while ensuring a threshold on harmlessness, SITAlign outperforms the state-of-the-art multi-objective decoding strategy by a margin of 22.3% in terms of GPT-4 win-tie rate for helpfulness reward while adhering to the threshold on harmlessness.

Cite this Paper

BibTeX
@InProceedings{pmlr-v267-chehade25a, title = {Bounded Rationality for {LLM}s: Satisficing Alignment at Inference-Time}, author = {Chehade, Mohamad Fares El Hajj and Ghosal, Soumya Suvra and Chakraborty, Souradip and Reddy, Avinash and Manocha, Dinesh and Zhu, Hao and Bedi, Amrit Singh}, booktitle = {Proceedings of the 42nd International Conference on Machine Learning}, pages = {7617--7633}, year = {2025}, editor = {Singh, Aarti and Fazel, Maryam and Hsu, Daniel and Lacoste-Julien, Simon and Berkenkamp, Felix and Maharaj, Tegan and Wagstaff, Kiri and Zhu, Jerry}, volume = {267}, series = {Proceedings of Machine Learning Research}, month = {13--19 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v267/main/assets/chehade25a/chehade25a.pdf}, url = {https://proceedings.mlr.press/v267/chehade25a.html}, abstract = {Aligning large language models with humans is challenging due to the inherently multifaceted nature of preference feedback. While existing approaches typically frame this as a multi-objective optimization problem, they often overlook how humans actually make decisions. Research on bounded rationality suggests that human decision making follows satisficing strategies- optimizing primary objectives while ensuring others meet acceptable thresholds. To bridge this gap and operationalize the notion of satisficing alignment, we propose SITAlign: an inference-time framework that addresses the multifaceted nature of alignment by maximizing a primary objective while satisfying threshold-based constraints on secondary criteria. We provide theoretical insights by deriving sub-optimality bounds of our satisficing-based inference alignment approach. We empirically validate SITAlign’s performance through extensive experimentation on multiple benchmarks. For instance, on the PKU-SafeRLHF dataset with the primary objective of maximizing helpfulness while ensuring a threshold on harmlessness, SITAlign outperforms the state-of-the-art multi-objective decoding strategy by a margin of 22.3% in terms of GPT-4 win-tie rate for helpfulness reward while adhering to the threshold on harmlessness.} }
Endnote
%0 Conference Paper %T Bounded Rationality for LLMs: Satisficing Alignment at Inference-Time %A Mohamad Fares El Hajj Chehade %A Soumya Suvra Ghosal %A Souradip Chakraborty %A Avinash Reddy %A Dinesh Manocha %A Hao Zhu %A Amrit Singh Bedi %B Proceedings of the 42nd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2025 %E Aarti Singh %E Maryam Fazel %E Daniel Hsu %E Simon Lacoste-Julien %E Felix Berkenkamp %E Tegan Maharaj %E Kiri Wagstaff %E Jerry Zhu %F pmlr-v267-chehade25a %I PMLR %P 7617--7633 %U https://proceedings.mlr.press/v267/chehade25a.html %V 267 %X Aligning large language models with humans is challenging due to the inherently multifaceted nature of preference feedback. While existing approaches typically frame this as a multi-objective optimization problem, they often overlook how humans actually make decisions. Research on bounded rationality suggests that human decision making follows satisficing strategies- optimizing primary objectives while ensuring others meet acceptable thresholds. To bridge this gap and operationalize the notion of satisficing alignment, we propose SITAlign: an inference-time framework that addresses the multifaceted nature of alignment by maximizing a primary objective while satisfying threshold-based constraints on secondary criteria. We provide theoretical insights by deriving sub-optimality bounds of our satisficing-based inference alignment approach. We empirically validate SITAlign’s performance through extensive experimentation on multiple benchmarks. For instance, on the PKU-SafeRLHF dataset with the primary objective of maximizing helpfulness while ensuring a threshold on harmlessness, SITAlign outperforms the state-of-the-art multi-objective decoding strategy by a margin of 22.3% in terms of GPT-4 win-tie rate for helpfulness reward while adhering to the threshold on harmlessness.
APA
Chehade, M.F.E.H., Ghosal, S.S., Chakraborty, S., Reddy, A., Manocha, D., Zhu, H. & Bedi, A.S.. (2025). Bounded Rationality for LLMs: Satisficing Alignment at Inference-Time. Proceedings of the 42nd International Conference on Machine Learning, in Proceedings of Machine Learning Research 267:7617-7633 Available from https://proceedings.mlr.press/v267/chehade25a.html.

Related Material