WorkRover trains long-horizon cowork agents via recursive environment evolution, +26 pts on JobBench
nex-agi · hf · 2026-10-14
nex-agi introduces WorkRover, a framework for training long-horizon cowork agents through recursive environment evolution, addressing the scarcity and brittleness of RL environments.
- Harness-driven synthesis: builder agents jointly construct and validate semantically aligned work contexts, executable tasks, and evidence-grounded verifiers, adapting to workspace state and execution feedback
- Factorial Compatibility Replay (FCR) uses rollout feedback to select the smallest validated repair among component-revision combinations via dependency-closed replay
- Verifier-to-Turn Credit Assignment (VTCA) augments trajectory-level GRPO advantages with bounded positive-progress credit for prerequisite turns
Post-training Qwen3.5-35B-A3B into WorkRover-35B-A3B yields gains of 26.27 points on JobBench, 29.32 points on OfficeQA Pro, and 413.25 Elo on GDPval v2 over the base model.
More from coding & agent
- OpenJet claims Qwen3.8-27B coding runs 1.9x faster than llama.cpp on one RTX 3090 — JeremyCMorgan · 2026-10-23
- Building an autonomous SEO agent in Hermes with three skills: programmatic pages, Keywords Everywhere, DataForSEO — EXM7777 · 2026-10-23
- Building a sports betting AI agent with onchain execution and human approval by default: architecture and lessons — Dry-Comfortable-2678 · 2026-10-23
- Capy 0.5.0 lets Claude Pro/Max subscribers reuse their plan quota, adds OIDC cloud access — garrytan · 2026-10-23
- You.com CEO Richard Socher: legacy code is human slop, let AI clean it up — RichardSocher · 2026-10-23
- Let agents design a single-line glance component so you can track them from the sidebar — blixt · 2026-10-23