AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

SSVPO: Effective Step-Level Credit Assignment for RL Training of Language Models

Yugu Li, Zehong Cao, Jianglin Qiao, Siyi Hu

ICLR 2026 · Reinforcement learning › RL from feedback · approach: RL for LLM reasoning

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
train with SSVPO instead of vanillaQwen3-4B · Seven math benchmarksaverage accuracy: 82.1 -> 86.7Improved
use SSVPO instead of RLOOQwen3-4B · Seven math benchmarksaverage accuracy: 84.9 -> 86.7Improved
use SSVPO instead of GRPOQwen3-4B · Seven math benchmarksaverage accuracy: 83.4 -> 86.7Improved
use SSVPO instead of DAPOQwen3-4B · Seven math benchmarksaverage accuracy: 83.0 -> 86.7Improved
train with SSVPOQwen3-4B · Seven math benchmarkstoken usage: 6327 -> 5178Improved
use SSVPO instead of RLOOQwen3-4B · GSM8Ktokens per question: 1665 -> 1320Improved
use SSVPO instead of RLOOQwen3-4B · SVAMPaccuracy: 94.8 -> 94.6Hurt
use SSVPO instead of SPODeepSeek-R1-Distill-Qwen-1.5B · Seven math benchmarksaverage accuracy: 66.3 -> 68.7Improved
use SSVPO instead of SPODeepSeek-R1-Distill-Qwen-1.5B · Seven math benchmarkstoken usage: 5402 -> 5838Hurt
use SSVPO instead of SPODeepSeek-R1-Distill-Qwen-1.5B · AIME 2024accuracy: 19.3 -> 28.8Improved
increase reordered steps from 2 to 4 (ablation)SSVPO · AMC 2023accuracy: 77.5 -> 82.5Improved
increase reordered steps from 2 to 4 (ablation)SSVPO · GSM8Kaccuracy: 90.5 -> 91.8Improved

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