A team including Johns Hopkins computer scientists won one of four Outstanding Paper Awards at the 2025 Conference on Language Modeling, held October 7-10 in Montreal, Canada, for their work on generating structured text from language models. Hopkins-affiliated authors of the winning paper, “Fast Controlled Generation from Language Models with Adaptive Weighted Rejection Sampling,” include PhD student Li “Leo” Du and Professor Jason Eisner, as well as alumni Tim Vieira, Engr ’23 (PhD) and Ryan Cotterell, Engr ’21 (PhD), now both at ETH Zürich.
The winning paper introduces a new adaptive weighted rejection sampling algorithm for instructing language models to generate text under certain constraints. Compared to state-of-the-art baselines, the team’s algorithm leads to significant improvements in both runtime and performance.
“The paper introduces a fast, principled, and adaptive sampler for controlled generation,” the conference judges wrote in their decision. “It solves a real problem and it actually works, getting large language models to respect hard constraints, and do so fast.”
Outstanding Paper Awards were bestowed to less than 0.3% of all submissions at this year’s conference.