HASE: Qwen3-8B matches GPT-OSS-120B on text classification

HASE, short for Harness-Aware Self-Evolving, is an agentic reinforcement-learning framework described in a paper submitted to arXiv on 4 Jul 2026. The paper shows HASE enables a single Qwen3-8B model to both generate task solutions and edit selected harness components in a multi-turn action space, and that this setup matches the text-classification performance of a GPT-OSS-120B model that used Claude Code as the harness proposer.

What is HASE and how does it work?

HASE is a unified agentic loop in which a single model can take two kinds of actions: produce task solutions, or modify parts of the evaluation and execution harness. The paper frames this as co-evolving model weights, the harness, and task solutions, with a Qwen3-8B model operating in a multi-turn action space that includes harness edits as first-class actions and task-solution generation as another action type.

The authors position HASE against the usual approach where frameworks optimize only task solutions while treating the harness as fixed. In contrast, HASE treats harness components as editable objects the agent can propose changes to, and it uses reinforcement learning to jointly refine the model and the harness through repeated interaction with evaluation components.

How does HASE perform compared with GPT-OSS-120B?

The paper reports that a single Qwen3-8B model under HASE matches the text-classification performance of a GPT-OSS-120B model that employed Claude Code as the harness proposer. The authors also state that HASE outperforms the reported GPT-OSS-120B baseline on alpha factor mining.

Beyond classification and alpha factor mining, HASE is credited with repairing imperfect evaluation components and converging to state-of-the-art performance in circle-packing algorithm discovery. Those claims are presented as results demonstrating that the same agentic process can both improve the harness and the task solution, rather than treating the harness as immutable infrastructure.

Why it matters

HASE reframes the role of the surrounding harness from fixed infrastructure to a mutable part of the learning problem. Allowing a single model to edit the harness as part of the training loop removes a hard separation between model development and tooling, which could reduce reliance on external proposers or hand-crafted evaluation fixes. If the results generalize beyond the reported tasks, teams with smaller models could match larger-model baselines by jointly optimizing harness and model, shifting where effort and compute are invested.

What to watch

Look for replication of HASE on more public benchmarks and for open-source toolchains that expose harness components as editable interfaces. The next concrete signals will be independent reproductions showing similar Qwen3-8B parity with GPT-OSS-120B on additional datasets, or extensions of HASE to other model families and harness types.

Submission and authorship note: the paper, "Harness-Aware Self-Evolving: Co-Evolving Model Weights, Harness, and Task Solutions," was submitted to arXiv on 4 Jul 2026 by Haochen Luo and seven coauthors.