← News

ResearchSources: arXiv · 18 Nov 2026

ASE diversifies RL training rollouts to lift reasoning accuracy in Qwen models

Researchers behind the Activation-Space Exploration (ASE) method report accuracy gains over GRPO on visual and mathematical reasoning benchmarks, using the same rollout budget.

In plain words

When AI models are trained with reinforcement learning, they learn only from the answers they happen to generate, which can make them repeat similar reasoning paths. ASE, described in an arXiv paper abstract, nudges training toward alternative paths by reusing patterns found in the model's internal activations. The authors report higher accuracy than a standard method on several benchmarks.

What was announced

The paper explains that reinforcement learning with verifiable rewards (RLVR) trains a model by sampling responses, scoring them with an automatic verifier, and updating the policy from those rewards. Because the learning signal comes only from trajectories in the model's own rollouts, repeated sampling can concentrate on similar reasoning paths and leave useful alternatives unexplored when the rollout budget is fixed. ASE addresses this by clustering activation differences between correct and incorrect responses to the same problem, building a bank of latent exploration vectors that are applied to later rollouts to steer the model toward different reasoning paths.

The reported results are limited to the models and benchmarks named in the abstract. On Qwen3-VL-2B-Instruct, the authors say ASE improved average accuracy over GRPO by 5.41 percentage points across six visual mathematical benchmarks and 3.15 percentage points across six general visual reasoning benchmarks. On Qwen3-1.7B-Base, they report a 4.87 percentage-point gain in average Mean@8 over GRPO across six mathematical reasoning benchmarks, with gains also observed on Qwen3-8B-Base. These figures come from the paper's abstract alone and have not been independently confirmed.

What it does not show

  • Evidence is limited to the paper's abstract; full experimental details, baselines beyond GRPO, and methodology are not available for verification.
  • The reported gains apply only to the specific Qwen models and benchmark sets named; no results for other models or tasks are provided.
  • The accuracy improvements are the authors' own claims and have not been independently confirmed.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.