SDPO + ReSAIL — Qwen3-4B — ALFWorld — Cycle 3
This repository contains the publicly released Cycle 3 checkpoint of SDPO + ReSAIL, based on Qwen/Qwen3-4B and trained on ALFWorld.
ReSAIL augments iterative agent self-distillation with trajectory-balanced selective distillation and privileged retention.
Code and documentation: ShengjieJin/ReSAIL
All checkpoints: ReSAIL Collection
Evaluation
- Reported three-decoding-seed mean: ALFWorld ID 72.4%; ALFWorld OOD 64.3%.
- The reported variation is across decoding seeds, not independent training seeds.
- Thinking mode was disabled for the text-agent experiments.
Lineage
- Base model:
Qwen/Qwen3-4B - Training task:
ALFWorld - Method:
SDPO + ReSAIL - Deployment cycle: 3
For environment setup, task data, training, and evaluation instructions, see the ReSAIL repository. The paper link and citation will be added when available.