AI
Act First, Reason Later: Accelerating On-Policy Distillation for Multi-Turn Agents via Reference-Conditioned Inverse Dynamics
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 28/09/2026, 20:00:00 UTC
Briefing created: 30/09/2026, 03:30:52 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 30/09/2026, 02:18:24 UTC.
Briefings
This study introduces ActFirst-OPD, a training framework that separates environment interaction from full response generation in multi-turn agents. By using reference-conditioned inverse dynamics, students execute actions directly while asynchronously generating teacher-supervised thoughts later. Experiments on Qwen3 models across ALFWorld, WebShop, and ScienceWorld show 2.3x to 4.9x speedups on ALFWorld and 1.8x to 4.9x on other benchmarks, matching or exceeding baseline success rates.
Key points
- ActFirst-OPD decouples environment interaction from full-response generation by executing actions via reference-conditioned inverse dynamics.
- The framework achieves average wall-clock training speedups of 2.3 times on ALFWorld and 4.9 times on ScienceWorld.
Context and limitations
Results are limited to Qwen3 parameter sizes of 0.6B, 1.7B, and 4B. Speedup metrics are measured over Vanilla OPD baselines, and success rates are compared across nine benchmark settings rather than a single definitive metric.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 30/09/2026, 03:30:52 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeifde6suf7gvza7dnll73r4yd5gg6magp7l674ocbkzy3kvpyj5iuq