ByteDance- and Monash-led paper turns task experience into weights for software agents
imjustnewatai · x · 2026-07-25
A ByteDance Seed and Monash paper proposes an agent-training loop that turns task experience into model weights and then deletes the trace.
The method gives an agent several attempts at a software task, collects its commands, failed patches, test results, and final fixes into an experience trace, then distills that trace into the weights. On 749 software tasks, the results were:
- Base model: 5.3% pass@1
- Normal fine-tuning: 8.0%
- PPO: 17.7%
- Experience distillation: 51.4%
It also used 9.6× fewer environment interactions than PPO. Repeating the collect → distill → erase loop five times raised performance from 7.1 to 47.0. On 494 held-out tasks, including unseen repositories, pass@1 nearly doubled from 4.62% to 8.84%.
The authors stress this is not recursive self-improvement yet: each cycle still needs an offline weight update, and the models were undisclosed in-house systems. But the paper argues that experience may become the next scaling axis for agents.
More from coding & agent
- Reticle builds a verification layer that lets AI coding agents check their own generated code — Scobleizer · 2026-10-27
- Trust your agent, but read its logs: Claude managed agents tips — brada · 2026-10-27
- Dev training a guard model to block prompt injection, tool and memory poisoning — gnukeith · 2026-10-27
- Dev trains a ~4GB-VRAM sentinel model to guard against prompt injection and poisoned memory — gnukeith · 2026-10-27
- Superset rebrands as Paradev, raises $11.5M seed led by USV — massimosgrelli · 2026-10-27
- Ex-Big Tech managers may thrive in the AI era: manage agents like you managed teams — lxfater · 2026-10-27