88Z AI, 28 April — Alibaba Tongyi Lab has released EAPO, a reinforcement learning framework that optimizes long-text reasoning through an evidence reward mechanism to address model hallucinations. Tests based on Qwen3-30B show its performance surpassing larger mainstream models like GPT-OSS and Claude-Sonnet-4.
At a glance
- EAPO introduces an evidence reward mechanism, shifting supervision to the evidence extraction process to address hallucinations
- EAPO based on Qwen3-30B outperformed 120B-parameter models like GPT-OSS in long-text tests
The update
Alibaba Tongyi Lab has officially launched EAPO (Evidence-Augmented Policy Optimization), a new reinforcement learning framework. It innovatively introduces an 'evidence reward' mechanism, shifting supervision from final answers to the evidence extraction process. This aims to resolve the hallucination issue in large models where they retrieve correct information but provide wrong answers during long-text reasoning. In multiple authoritative long-text benchmarks, the EAPO framework based on the Qwen3-30B model performed excellently. Data shows that its performance not only remained stable but also surpassed large models with up to 120B parameters, such as GPT-OSS and Claude-Sonnet-4, demonstrating the strong potential of smaller-parameter models under specific optimization frameworks.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- 通义实验室Source published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.