mlsanityData to 14 Oct 2026

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

arXiv 2610.18067 ↗First posted 13 Oct 2026cs.AI cs.LGcited by 0

Not on the map. Why?

Abstract (arXiv)

Scaling interactive environments is important for training long-horizon agents through reinforcement learning. These environments often require further testing and refinement after initial generation, and correcting defects in a task may also require updating its work context and verifier. However, existing construction pipelines offer limited support for such iterative refinement while keeping these components aligned. We introduce WorkRover, a framework for training long-horizon cowork agents through recursive environment evolution. Its harness-driven synthesis enables construction agents to jointly build and validate semantically aligned work contexts, executable tasks, and evidence-grounded verifiers, adapting their actions to workspace state and execution feedback rather than following a fixed construction workflow. To support environment evolution, we introduce Factorial Compatibility Replay (FCR), which uses rollout feedback to evaluate combinations of component revisions through dependency-closed replay and selects the smallest validated repair among tested candidates while preserving task requirements and evaluation constraints. To provide finer-grained training signals, we propose Verifier-to-Turn Credit Assignment (VTCA), which augments trajectory-level GRPO advantages with signals from deterministic workspace-progress checks and assigns a bounded share of positive-progress credit to prerequisite turns using recorded tool-output and artifact dependencies. Using WorkRover, we post-train Qwen3.5-35B-A3B to obtain WorkRover-35B-A3B and evaluate it on several cowork-agent benchmarks. Its reported scores exceed those of Qwen3.5-35B-A3B by 26.27 percentage points on JobBench, 29.32 percentage points on OfficeQA Pro, and 413.25 Elo points on GDPval v2. These results suggest that WorkRover offers a promising foundation for scalable training of long-horizon cowork agents.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Builds on 31 of its 45 references that are in the map. First posted on arXiv 13 Oct 2026.

Builds on

31 of 45 references.

References · 31

  1. 2024DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Modelscited by 9,379
  2. 2024AppWorld: A Controllable World of Apps and People for Benchmarking Interactive Coding Agentscited by 401
  3. 2025AgentPRM: Process Reward Models for LLM Agents via Step-Wise Promise and Progresscited by 77
  4. 2026EnvScaler: Scaling Tool-Interactive Environments for LLM Agent via Programmatic Synthesiscited by 63
  5. 2026Agent-World: Scaling Real-World Environment Synthesis for Evolving General Agent Intelligencecited by 54
  6. 2026Workspace-Bench 1.0: Benchmarking AI Agents on Workspace Tasks with Large-Scale File Dependenciescited by 29
  7. 2025Promoting Efficient Reasoning with Verifiable Stepwise Rewardcited by 28
  8. 2025RLFactory: A Plug-and-Play Reinforcement Learning Post-Training Framework for LLM Multi-Turn Tool-Usecited by 20
  9. 2026HiPER: Hierarchical Reinforcement Learning with Explicit Credit Assignment for Large Language Model Agentscited by 18
  10. 2026TRACE: Turn-level Reward Assignment via Credit Estimation for Long-Horizon Agentscited by 16
  11. 2026ClawGym: A Scalable Framework for Building Effective Claw Agentscited by 15
  12. 2025ToolForge: A Data Synthesis Pipeline for Multi-Hop Search without Real-World APIscited by 10
19 more
  1. 2026Hardening Agent Benchmarks with Adversarial Hacker-Fixer Loopscited by 9
  2. 2026FACET: Preserving Source Intent and Executable State in Terminal Task Synthesiscited by 9
  3. 2025LocalSearchBench: Benchmarking Agentic Search in Real-World Local Life Servicescited by 8
  4. 2026Why Multi-Step Tool-Use Reinforcement Learning Collapses and How Supervisory Signals Fix Itcited by 7
  5. 2026TRCA: Transition-wise Rubric Credit Assignment for Long-horizon LLM Agentscited by 7
  6. 2026Anchor: Mitigating Artifact Drift in Agent Benchmark Generationcited by 5
  7. 2026STAMP: Provenance-Guided Credit Assignment for Deep Search Agentscited by 5
  8. 2026DRACO: Fine-Grained Credit Assignment with Dynamic Rubrics for Long-Horizon Agent Trainingcited by 5
  9. 2026RecRM-Bench: Benchmarking Multidimensional Reward Modeling for Agentic Recommender Systemscited by 4
  10. 2026Meta-Task: Turning Terminal Task Synthesis into a Terminal Task for Scalable Agent Trainingcited by 4
  11. 2026Envs-FORGE: Frontier-Optimized Reward-Grounded Environment Synthesis for Agent RLcited by 4
  12. 2026CoAL-RAG: A Complexity-Aware Legal Retrieval-Augmented Generation Methodcited by 3
  13. 2026TAPO: Tool-Aware Policy Optimization via Credit Transfer for Multimodal Search Agentscited by 2
  14. 2026Echoverse: Deep, Evolving Environments for Training Computer-Use Agents at Scalecited by 2
  15. 2026CheckerBench: Can Long-Horizon Agents Synthesize Static-Analysis Checkers?cited by 2
  16. 2026Efficient Reinforcement Learning for Long-Horizon Tool-Use Agentic Taskscited by 1
  17. 2026DuMateBench: Evaluating Autonomous Agents in Complex Real-World Workflowscited by 1
  18. 2026Coverage, Not Targeting: A Structural Regime in Multi-Turn Agent Credit Assignmentcited by 1
  19. 2026RAG-Stress: Probing the Limits of Evidence Reliance in Retrieval-Augmented Generationcited by 1

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026Cowork-Gym: Synthesizing Interactive Work Environments at Scale
  2. 2026Tev: Scaling Long-Horizon Agentic Environments with Dense Progress Rewards
  3. 2026Workspace Optimization: How to Train Your Agent
  4. 2026AgentMercury: Your Agent Can Synthesize Verifiable Environments for Business Scenarios at scale
  5. 2026Scalable Environments Drive Generalizable Agents
  6. 2026AutoGym: Blueprint-First Generation of Verifiable Agent Gyms
  7. 2026EAU: Environment-Agent-User Co-evolution for Training User-Centric Agents
  8. 2026Explorer-Zero: Improving Self-Play for Long-Horizon Tasks with Decomposable Verifiable Rewards