arxiv arXiv cs.LG · 13h ago · research

CANOPY enables Qwen3-14B to top AppWorld using outcome-only RL

from English

The paper introduces CANOPY (Coverage-ANchored On-PolicY RL), a protocol that addresses signal starvation and policy drift in long-horizon reinforcement learning for small open models. By scaling same-task exploration and keeping updates KL-anchored, the method allows agents to learn effectively from end-of-task verification alone.

CANOPY scales exploration until natural signal reappears, avoiding the silence of sparse rewards on hard tasks.
Updates are kept on-policy and confined to the agent's own action tokens to prevent policy drift.
A Qwen3-14B policy trained with CANOPY topped the AppWorld public leaderboard (Feb. 2026) with a Test-Normal TGC of 86.9.
The same principles lifted Qwen3.5-9B on SWE-bench Verified by 16.6 points without auxiliary credit signals or elaborate scaffolding.

The authors argue that agentic RL alone can internalize long-horizon capability directly into small open models, with the complete training stack planned for release.

Importance 2/3 New harness with differentiators arXiv cs.LG Alibaba (Qwen) AI agents Research paper

Benchmarks

Benchmark	Model	Score
SWE-bench Verified	Qwen3.5-9B	16.6%

Read original

arxiv arXiv cs.LG · 30d ago · 26 views

Qwen releases Qwen-CUA, a 397B-A17B native computer-use agent

Qwen introduces Qwen-CUA, a native computer-use agent built on a 397B-A17B mixture-of-experts backbone that operates via screenshots and input events without relying on DOM trees or accessibility metadata. The system utilizes a scaffold to maintain up to 20 active screenshots and was trained using a cloud rollout fleet with nearly 100,000 vCPUs across approximately 40,000 verifiable tasks.

arxiv arXiv cs.CL · 30d ago · 16 views

Qwen releases Qwen-CUA, a 397B-A17B native computer-use agent

Researchers have introduced Qwen-CUA, a native computer-use agent built on a 397B-A17B Qwen mixture-of-experts backbone. The system operates by observing screenshots and acting through keyboard and mouse events, without relying on DOM trees or accessibility metadata.

arxiv arXiv cs.AI · 34d ago · 13 views

Qwen-UI-Agent sets state of the art on mobile-use benchmarks

The Qwen team has released a technical report for Qwen-UI-Agent, a real-world centric foundation GUI agent designed to operate reliably across mobile, computer-use, web, and DeepSearch environments. The system combines diverse sandbox environments with a large-scale real-device mobile runtime, interleaving GUI operations with CLI execution and supporting long-horizon tasks.

arxiv arXiv cs.CL · 37d ago · 12 views

AgentOmnia scales agentic models across full-scenario applications

Researchers present AgentOmnia, a framework for full-scenario agentic scaling that coordinates task-space definition, data synthesis, post-training, evaluation, and improvement across To-Consumer, To-Business, and To-Employee applications. The system utilizes an extensible taxonomy and constructs 5,018 stateful environments with 255,375 tools and 52,361 tasks to enable fine-grained diagnosis via OmniaBench.

arxiv arXiv cs.LG · 48d ago · 10 views

TRACE improves long-horizon agent tool-use via turn-level credit estimation

The authors propose TRACE (Turn-level Reward Assignment via Credit Estimation), a dense credit-assignment method for agentic reinforcement learning that addresses the sparsity and high variance of outcome rewards in long-horizon tasks. By representing rollouts as state transitions at tool-call boundaries and deriving per-action rewards from log-ratio state values, TRACE enables effective training without additional critics or supervised fine-tuning stages.