The authors start with a pretrained language model (StartLM) and fine-tune two variants on a flight recommendation task that requires inferring a user's latent preferences from evidence.
Bayesian fine-tuning instills usable beliefs in language models but stops short of genuine Bayesian reasoning
A model fine-tuned on optimal Bayesian outputs matches behavior and encodes correct beliefs, yet only partially implements the hierarchical steps of Bayes' rule.
Academic
Polina Tsvilodub · Andreas Waldis · Linlu Qiu · Tal Linzen · Michael Franke
Massachusetts Institute of Technology · New York University · University of Tübingen
Research Digest··3 min read
The authors compare two fine-tuned language models on a flight recommendation task: one trained on outputs of a Bayesian assistant (BayesLM) and one on true preference labels (OracleLM).
Why this paper
From Massachusetts Institute of Technology and 2 others
In one line
Fine-tuning a language model on a Bayesian assistant's decisions installs usable Bayesian beliefs; fine-tuning on true answers does not.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§