Papers.Data.Code
Paper

WorkRover: Scaling Long-Horizon Cowork Agent Training Via Recursive Environment Evolution

Hang He, Linting Wang, Guanming Liu, Bo Mao, Jinxiu Liu, Yiran Suo, Maosen Zhou, Zhiheng Xi

A recursive environment-building and RL setup that jointly repairs context, task, and verifier, then shapes turn credit from checked workspace progress.

arXiv ↗Oct 13, 2026Code ↗

  • llm
  • agent training
  • reinforcement learning
  • credit assignment
Multi-stage pipeline diagram showing method: rollout, evaluation, credit assignment, policy update
Multi-stage pipeline diagram showing method: rollout, evaluation, credit assignment, policy update

What it is

Trains long-horizon cowork agents by recursively repairing synthesized work environments and adding verifier-grounded turn credit during RL. Instead of one-shot task generation and episode-only rewards, it recursively repairs misaligned context/task/verifier packages and adds verifier-grounded turn credit from workspace progress.

Key results

  • +413.25 Elo on GDPval v2 over Qwen3.5-35B-A3B
  • JobBench improves by 26.27 points over Qwen3.5-35B-A3B
  • FCR raises task-package synthesis pass rate by 6.0 points vs regeneration

Shows long-horizon agent RL improves when training environments are iteratively repaired and rewards are grounded in verified intermediate workspace progress.

One useful ML email a week.

More from Oct 12 – Oct 18

The whole week →

Picked Oct 16

Picked Oct 15

  • Repo
    ziplab/ATAS

    Adds explicit scene and event memory with validated world updates, targeting spatial revisit and event recapture beyond standard multi-shot generation.

    GitHubOct 12, 2026
    →
  • Dataset
    HelenPeng/ExeWorldBench

    Targets executable Blender world construction, scoring functional correctness instead of visual similarity alone.

    Hugging FaceOct 9, 2026
    →

Picked Oct 14

  • Dataset
    ScaleAI/HSS

    Makes intuitive visual reasoning measurable with criterion-level grading; humans hit 93.1% vs 53.6% for the best tested model.

    Hugging FaceOct 1, 2026
    →

Worth your time, not all of your time.

We track new ML work across papers, code, and data, then keep only the pieces worth a closer look. Selected, not collected.

Papers.Data.Code

The Weekly ML Signal

Selected, not collected.

One useful ML email a week: the papers, repos and datasets worth your attention.

Worth your time, not all of your time.