Ant Group's SWE-Trail: history-guided trajectory collection yields better coding agents
antgroup · hf · 2026-11-03
Ant Group and University of Connecticut introduce SWE-Trail (arXiv, Nov 1), a trajectory collector for coding agent post-training. Independent per-issue rollouts get increasingly redundant, and exposing past trajectories as advice to the agent creates a train-test mismatch. SWE-Trail instead uses earlier rollouts only to decide where to spend collection effort — identifying states worth extra exploration, then briefly executing candidate next actions from the same checkpoint — while the agent generates its own steps in a normal context. Student models trained on SWE-Trail data outperform baselines.
More from coding & agent
- Dev open-sources faster-chrome-devtools-skill after analyzing 2,500 agent uses — simpsoka · 2026-11-03
- Shipping an enterprise project with Claude and Codex: AI writes, judge scores, you decide — kmeanskaran · 2026-11-03
- Neo4j GraphRAG Python package defaults to CPU-only PyTorch wheels to cut install overhead — JeremyCMorgan · 2026-11-03
- Zero-dependency SQLite + FTS5 MCP server adds persistent memory and anti-loop tracking to coding agents — AvocadoJaded8829 · 2026-11-03
- Developer says Kimi's $40 coding plan burned 3% of monthly usage reading one project guideline — Chaosuka · 2026-11-03
- OmniMem gives coding agents persistent memory of dead ends, hits 12k downloads — Ancient_Ad8787 · 2026-11-03