AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Controllable Exploration in Hybrid-Policy RLVR for Multi-Modal Reasoning

Zhuoxu Huang, Mengxi Jia, Hao Sun, Xuelong Li, Jungong Han

ICLR 2026 · Reinforcement learning › RL from feedback · approach: Scaling RL with verifiable rewards · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add CalibRL controllable exploration on top of GRPOGRPO Qwen2.5-VL-7B · geometry avg (GeoEval, Geo3K, GeoQA)accuracy (%): 39.48 -> 44.93Improved
add CalibRL to GRPO on out-of-domain benchmarksGRPO Qwen2.5-VL-7B · out-of-domain avg (7 benchmarks)accuracy (%): 57.24 -> 59.36Improved
add LUFFY hybrid-policy expert guidance to GRPOGRPO Qwen2.5-VL-7B · geometry avgaccuracy (%): 39.48 -> 38.64Hurt
add RL-PLUS to GRPOGRPO Qwen2.5-VL-7B · geometry avgaccuracy (%): 39.48 -> 34.68Hurt
SFT then GRPO instead of GRPO aloneQwen2.5-VL-7B · GeoEvalaccuracy (%): 26.15 -> 6.00Hurt
use DAPO higher clipping instead of GRPOGRPO Qwen2.5-VL-7B · geometry avgaccuracy (%): 39.48 -> 39.55No clear effect
CalibRL vs GRPO on smaller modelQwen2.5-VL-3B · 5-benchmark avgaccuracy (%): 34.44 -> 37.09Improved
CalibRL vs GRPO on InternVL3-8BInternVL3-8B · 5-benchmark avgaccuracy (%): 45.37 -> 47.42Improved
add advantage weighting |A_i| to controllable exploration (ablation)CalibRL · 8-benchmark avgaccuracy (%): 45.30 -> 50.59Improved
use LeakyReLU slope alpha 0.5 instead of 0.3 or 0.8 (ablation)CalibRL · 8-benchmark avgaccuracy (%): 49.91 / 49.85 -> 50.59Improved
increase balance weight lambda from 0.1 to 0.3 or 0.5 (ablation)CalibRL · 8-benchmark avgaccuracy (%): 50.59 -> 44.58 / 45.13Hurt
add entropy coefficient 0.01 to GRPO (ablation)GRPO · geometry avgaccuracy (%): 39.48 -> 36.79Hurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy