Research
-
October 2026
AutoSFT
AutoSFT: An Agentic System for Automated Data Research in Supervised Fine-Tuning
A multi-agent system that automates the full space of data engineering operations in supervised fine-tuning. A single agent collapses to one action family and loses context repeatedly. AutoSFT isolates execution from analysis, gathers proposals from several complementary angles, and synthesizes data for undersupplied capabilities. Each component added in turn is individually attributable: the executor halves context loss, analysis lifts IFEval by +6.5 points, and synthesis adds +3.4 IFEval and +2.9 IFBench, with the same structure transferring to code at +3.7 MBPP+.
-
September 2026
Rethinking OPD II
Rethinking On-Policy Distillation of Large Language Models II: One Training Example
Examines the role of training data in on-policy distillation at the data-minimal limit, training on a single query. One-shot OPD keeps improving for hundreds of steps and recovers most of full-data OPD's gain across task domains and model families. A single query reaches 71.5% state coverage relative to full-data OPD and sixteen reach 98.9%, matching full-data training — OPD is therefore data-overfed but algorithm-starved.
-
April 2026
Rethinking OPD
Rethinking On-Policy Distillation of Large Language Models: Phenomenology, Mechanism, and Recipe
A systematic investigation of on-policy distillation dynamics. Identifies the two conditions that govern whether OPD succeeds — compatible thinking spaces between student and teacher, and genuinely new teacher capability — traces the token-level mechanism behind them, and derives practical recipes for recovering failing runs.