← 研踪 PaperTrailarXiv:2610.27012 · v2
Activation-Space Exploration for Visual and Textual Reasoning with RLVR
首次提交 2026-10-19 20:23
版本更新 2026-11-17 20:28
雷达观测 2026-11-18 11:33
摘要
Reinforcement learning with verifiable rewards (RLVR) improves visual and textual reasoning by sampling responses, scoring them with an automatic verifier, and updating the policy from the resulting rewards. The policy therefore receives direct learning signals only from trajectories that appear in its rollouts. Yet repeated sampling from the same policy can concentrate on similar reasoning paths, leaving useful alternatives unexplored under a fixed rollout budget. We introduce Activation-Space Exploration (ASE), an online method for diversifying training rollouts. By clustering activation differences between correct and incorrect responses to the same problem, ASE builds a structured bank of latent exploration vectors. Different vectors are then applied to subsequent rollouts to explore alternative reasoning paths. Under the same rollout budget, ASE improves average accuracy over GRPO on Qwen3-VL-2B-Instruct by 5.41 and 3.15 percentage points across six visual mathematical and six general visual reasoning benchmarks, respectively. On Qwen3-1.7B-Base, it improves average Mean@8 over GRPO across six mathematical reasoning benchmarks by 4.87 percentage points, with gains also observed on Qwen3-8B-Base.
观测到的版本
相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。
v2 · 2026-11-17 20:28 · 最新观测版本
Activation-Space Exploration for Visual and Textual Reasoning with RLVR
Reinforcement learning with verifiable rewards (RLVR) improves visual and textual reasoning by sampling responses, scoring them with an automatic verifier, and updating the policy from the resulting rewards. The policy therefore receives direct learning signals only from trajectories that appear in its rollouts. Yet repeated sampling from the same policy can concentrate on similar reasoning paths, leaving useful alternatives unexplored under a fixed rollout budget. We introduce Activation-Space Exploration (ASE), an online method for diversifying training rollouts. By clustering activation differences between correct and incorrect responses to the same problem, ASE builds a structured bank of latent exploration vectors. Different vectors are then applied to subsequent rollouts to explore alternative reasoning paths. Under the same rollout budget, ASE improves average accuracy over GRPO on Qwen3-VL-2B-Instruct by 5.41 and 3.15 percentage points across six visual mathematical and six general visual reasoning benchmarks, respectively. On Qwen3-1.7B-Base, it improves average Mean@8 over GRPO across six mathematical reasoning benchmarks by 4.87 percentage points, with gains also observed on Qwen3-8B-Base.
查看此版本解读 →查看此版本原文 ↗