Four papers: one-sample on-policy distillation, SPACE cuts agent LLM turns by up to 78.9%

stanfordnlp · x · 2026-09-07

A four-paper digest:

1. TAHI: Efficient Test-Time Adaptation through Human-AI Interaction (Pang Wei Koh, Diyi Yang, Graham Neubig, Daniel Fried et al.; CMU/Stanford/UW/Princeton) — Uses cross-session human feedback to update an agent's context and weights at test time, distilling user criteria into an evolving rubric. With 30 users across 600 writing/visual tasks, tens of tasks sufficed to lift success rates by 4.5–20.9%.

2. Rethinking On-Policy Distillation II: One Training Example (Zhiyuan Liu, Ning Ding, Chaojun Xiao et al., Tsinghua) — On-policy distillation pushed to a single query captures most of the full-dataset gains, explained via state coverage and alignment rate. The bottleneck is the student's slow digestion, not insufficient supervision — challenging the "more data is better" assumption.

3. Compile by Training (Yuntian Deng, Pengyu Nie, Stuart Shieber, Harvard) — Trains a small local adapter from teacher-generated examples at compile time, eliminating the teacher at runtime for repetitive text tasks.

4. Act More, Decide Less / SPACE: Long-horizon LLM agents needn't query the model every turn. Standard RL fails to learn variable-length action blocks without boundary supervision; SPACE induces two-level procedural skills from successful trajectories, using sub-skill boundaries as block-boundary labels. On ALFWorld and ScienceWorld: success rate +7.0–31.3% over the strongest baselines, LLM decision turns reduced by up to 78.9%.

Original post →

Four papers: one-sample on-policy distillation, SPACE cuts agent LLM turns by up to 78.9%

More from coding & agent