, maximize candidate coverage (Pass@K).
One-draw posterior updates achieve multi-sample reasoning gains at reduced cost
Fenchel-Bregman IVON (FBI) trains a posterior over solver parameters using a single sample per update, matching Pass@10 of a K-draw baseline with 12.3% of the training time
Research Lab
Sungjun Lim · Adrian Robert Minut · Nico Carlsen Daheim · Mohammad Emtiyaz Khan · Kyungwoo Song · Thomas Möllenhoff
RIKEN AIP · Yonsei University · Sapienza University of Rome · TU Darmstadt · Hessian.AI
Research Digest··3 min read
The authors formulate finite-K posterior learning to directly optimize candidate coverage in multi-sample reasoning.
Why this paper
From RIKEN AIP and 4 others
In one line
Training a posterior over solver parameters with Fenchel-Bregman updates improves multi-sample reasoning coverage while needing only one posterior draw per update.
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 (2 noted)
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§