Many LM applications may be formulated as text generation conditional on some (Boolean) constraint.

Generate a…
- Python program that passes a test suite.
- PDDL plan that satisfies a goal.
- CoT trajectory that yields a positive reward.
The list goes on…

How can we efficiently satisfy these? 🧵👇

Correction: poster number is 634 :)

Come find us at #ICLR this Saturday

Oral: 4:30pm (Garnet 213-215, session 6B)
Poster: 10am-12:30pm #634 (Hall 2B)
iclr.cc/virtual/2025...

GitHub - genlm/genlm-control: Controlled text generation with programmable constraints. Controlled text generation with programmable constraints. - genlm/genlm-control

- Cast controlled generation as an inference problem, with the LM as a prior and verifiers and scorers as likelihood
- Use Sequential Monte Carlo to sample from the resulting posterior

Library w/ tutorials for setting up your own controlled generation inference problems: github.com/genlm/genlm-...

#ICLR2025 Oral

How can we control LMs using diverse signals such as static analyses, test cases, and simulations?

In our paper “Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo” (w/ @benlipkin.bsky.social,
@alexlew.bsky.social, @xtimv.bsky.social) we: