Harness-S trains agents using sibling harnesses
It selectively rescues failures with sibling harnesses and transfers capabilities to the base policy.
Stories per day · 30 days
2 RSS
Relatedqwen4reinforcement-learning3computer-use-agents2fine-tuning2gui-agents2synthetic-environments2adapters1coworkenv1credit-assignment1distillation1harness-evolution1language-model-agents1llm-agents1mai-gym1mobileworld1multimodal1open-source-models1open-weights1
14 stories, newest first · one card per story
Stories tagged #agent-training
WorkFactory builds professional tasks and trains WorkPilot agents
Fine-tunes of Qwen at three scales all beat their bases, and the 35B-A3B model rises from 879 to 1565 Elo.
MPO aligns LLM agent credit assignment at the turn level
The method assigns credit at assistant-turn granularity while keeping finer token-level allocation, outperforming RL baselines on agentic tasks.
WildAWM synthesizes environments for computer-use agent training
The open-source pipeline made 341 long-horizon tasks; GPT-6-Astra solved 60%, and two trained Qwen models improved on eight benchmarks.
FocusGUI adapter cuts sequential distraction in omni-modal GUI agents
Across four frozen backbones, authors report Type Match up 33.2 points and Exact Match 3.0 times the frozen baseline.
WildAWM synthesizes computer-use agent environments with open models
It yields 341 tasks that can exceed 100 steps; GPT-6-Astra solves 60%, and Qwen models trained only there improve on eight benchmarks.
CoworkEnv synthesizes training environments for long-horizon coworker agents
SFT on Qwen3-VL-8B-Instruct raised the Claw-Eval mean from 46.6 to 64.1, including a 31.9-point multi-turn gain.
Progress-Aware Distillation targets stagnation in small agent models
Tests on eight models from 0.6B to 8B parameters show higher success rates and faster task completion.
MAI-Gym: An Open All-in-One Platform for GUI Agent Evaluation and RL Training across Mobile and Computer Environments
MAI-Gym unifies GUI-agent evaluation and online RL; its 9B agent scores 78.4% on MobileWorld.
SoftwareGym: Scaling Computer-Use Agent Training with Diverse Synthesized Software Environments
SoftwareGym's 182 synthetic apps lift a computer-use agent to 56.97% OSWorld Pass@3.
Credit Where It Matters: Dependency-Aware Policy Optimization for Terminal Agents
DepGPO improves RL for terminal agents by building command dependency graphs from execution traces to assign credit to reads and writes that actually affect task outcomes.
NeutronGym: Physics-Graded Neutron Instrument Design for LLM Agents
NeutronGym grades LLM agents on neutron instrument design via McStas ray-tracing, with RL lifting Qwen3-8B from 11% to 77%.
Scaling Trajectories for Complex Tasks through Recursive Self-Rewrite
Recursive Self-Rewrite converts harness-assisted trajectories into general training data, lifting Qwen-3.8-27B Terminal-Bench 2 pass@3 from 57.0% to 74.2%.
CUA-Universe: A Scalable and Dynamic Environment for Hybrid GUI+CLI Agents
Researchers introduce CUA-Universe, a pipeline turning real desktop software into hybrid GUI+CLI agent environments, lifting a 9B model's OSWorld success rate.