Stanford’s AI Lab (SAIL) is presenting research at the Conference on Language Modeling (COLM) 2026, hosted at the Hilton Union Square in San Francisco from October 6 to 9. The lab listed five main-conference papers and four workshop papers, with links to papers and, for some projects, videos, websites or blog posts.

Main-conference work includes a study of conversation as a way to measure clinical encounters and patient states, and research asking whether language models consistently encode the current year. Another paper, PromptNCE, estimates conditional probabilities and pointwise mutual information using large language models and contrastive prompts. SWE-chat examines coding-agent interactions with real users, and was nominated for an oral presentation at the workshop on responsibly enabling data for foundation models. A paper on switching linear attention studies recurrent architecture, sequence modeling and test-time regression.

Workshop papers examine how models can learn to forget while reasoning, using reinforcement learning and end-to-end context management; QuasiMoTTo studies quasi-Monte Carlo test-time scaling and inference compute; and Speculative Self-Distillation explores knowledge internalization through distillation after training. A fourth paper studies how sycophantic agreement can transfer with neutral data using Contrastive Preference Optimization, with keywords including subliminal learning and AI safety and alignment. The Neural Garbage Collection paper and SWE-chat were nominated for oral presentations. The conference announcement invites readers to contact the listed paper authors for more information.

Stanford’s AI Lab (SAIL) is presenting research at the Conference on Language Modeling (COLM) 2026, hosted at the Hilton Union Square in San Francisco from October 6 to 9. The lab listed five main-conference papers and four workshop papers, with links to papers and, for some projects, videos, websites or blog posts.

Main-conference work includes a study of conversation as a way to measure clinical encounters and patient states, and research asking whether language models consistently encode the current year. Another paper, PromptNCE, estimates conditional probabilities and pointwise mutual information using large language models and contrastive prompts. SWE-chat examines coding-agent interactions with real users, and was nominated for an oral presentation at the workshop on responsibly enabling data for foundation models. A paper on switching linear attention studies recurrent architecture, sequence modeling and test-time regression.

Workshop papers examine how models can learn to forget while reasoning, using reinforcement learning and end-to-end context management; QuasiMoTTo studies quasi-Monte Carlo test-time scaling and inference compute; and Speculative Self-Distillation explores knowledge internalization through distillation after training. A fourth paper studies how sycophantic agreement can transfer with neutral data using Contrastive Preference Optimization, with keywords including subliminal learning and AI safety and alignment. The Neural Garbage Collection paper and SWE-chat were nominated for oral presentations. The conference announcement invites readers to contact the listed paper authors for more information.