AI

Self-Supervised Scaling of Terminal Environments for Scientific Domains

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 05/10/2026, 21:18:24 UTC.

Briefings

This article presents a self-supervised framework called Software-in-the-Loop Reconstruction (SWR) to train terminal agents for scientific domains. By executing multiple input configurations on existing software workflows, the method generates reference outputs and verification targets without requiring authoring a specific solution for each task. The approach partitions cases into public observations and hidden evaluations, allowing agents to construct editable programs while being evaluated against workflow outputs. The study utilizes 500 workflows across 46 software families in six domains, achieving a 53.56% mean performance on Terminal-Bench 2 after supervised fine-tuning of Qwen3.8-27B, surpassing previous benchmarks.

Key points

• The framework constructs editable programs without authoring a reference solution for each task, using existing software workflows as training data.
• Qwen3.8-27B achieves a 53.56% mean performance on Terminal-Bench 2, improving from 47.94% after supervised fine-tuning across three seeds.

Context and limitations

The study covers 500 workflows and 46 software families across six domains. Results are based on three attempts per task with 3,000 reconstruction-only training examples, and performance metrics are limited to Terminal-Bench 2 evaluations.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 15/10/2026, 10:08:58 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeihhsdrhsakqikcqb2ngef56aouuengx5jw5lj3cbqbuffzxcijlhi