AI

WEFT: Scaling Tool-Use Post-Training for General-Purpose Agents

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 05/10/2026, 15:18:24 UTC.

Briefings

WEFT is a framework for scaling tool-use post-training by coupling environment construction, execution-driven self-evolution, and stable sampling. It outperforms baseline environment-scaling methods on BFCL V4, τ^2-Bench, and Claw-Eval, with WEFT-14B improving Agent-World-14B by 6.41% on BFCL V4. The system uses prefix-preserving sampling and MegaMCP to maintain state integrity during concurrent rollouts.

Key points

• WEFT (Whole-system Evolution For Tool-use Post-training) couples scalable agentic interaction system construction, execution-driven self-evolution, and stable post-training to address scaling limitations in tool-use models.
• WEFT-14B outperforms Agent-World-14B by 6.41 percentage points on BFCL V4, 2.23 percentage points on τ^2-Bench, and 12.27 percentage points on Claw-Eval.

Context and limitations

Experiments were conducted across various models and benchmarks; specific results for WEFT-35B-A3B on Toolathlon-Verified and AutomationBench are reported as extended gains rather than absolute final scores.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 08/10/2026, 15:32:26 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeieqnl2ik5aiauluyjc4pwfatx2et7chrk7f4nr57bfkcub25j53ke