Crimson AI NewsA CrimsonLingua Network service
EN ع
← Back to news
Research paper Qwen (Alibaba)

SAPO: Alibaba's New RL Method Promises Stable and Performant LLM Training

AI By Crimson AI Qwen Research 5 December 2025 · 04:00 28 views
Share: X Telegram

Alibaba's Qwen team introduces Soft Adaptive Policy Optimization (SAPO), a reinforcement learning method that replaces hard clipping with smooth gating to stabilize training of large language models, especially Mixture-of-Experts architectures.

SAPO: Alibaba's New RL Method Promises Stable and Performant LLM Training

Key points

Reinforcement learning (RL) has become a key component in enhancing the reasoning capabilities of large language models (LLMs), enabling them to solve complex math problems, write code, and reason over multimodal inputs. However, stable policy optimization remains challenging due to variance in token-level importance ratios, particularly in large Mixture-of-Experts (MoE) models.

Existing methods like GRPO and GSPO use hard clipping to control instability, but this approach discards useful gradient information and struggles to balance stability and sample efficiency. To address these limitations, Alibaba's Qwen team proposes Soft Adaptive Policy Optimization (SAPO), a novel RL method that replaces hard clipping with a smooth, temperature-controlled gating function.

SAPO offers several advantages: continuous trust regions that avoid discontinuities, sequence-level coherence similar to GSPO but without discarding entire sequences, token-level adaptivity to selectively suppress problematic tokens, and an asymmetric temperature design that treats positive and negative advantages differently. This design stabilizes training and improves sample efficiency.

In experiments, SAPO was tested on mathematical reasoning using a cold-start model fine-tuned from Qwen3-30B-A3B-Base. It maintained stable training longer than GSPO and GRPO-R2, achieving higher Pass@1 scores on benchmarks like AIME25, HMMT25, and BeyondAIME. SAPO also does not require routing replay, simplifying RL pipelines.

Large-scale RL experiments with Qwen3-VL models showed consistent performance improvements across various model sizes and architectures, including both MoE and dense models. SAPO outperformed GSPO and GRPO-R2 under the same compute budget.

The team believes SAPO will become a foundational component of RL training pipelines for LLMs, offering a practical way to stabilize and enhance training. The full paper is available on the Qwen research page.

MethodAIME25 Pass@1HMMT25 Pass@1BeyondAIME Pass@1
SAPOHighestHighestHighest
GSPOLowerLowerLower
GRPO-R2LowerLowerLower
Source
Qwen (Alibaba) · Qwen Research
Related news
Qwen (Alibaba)
Qwen (Alibaba) 3 Aug 2026

Qwen 3.8-Max: Alibaba's 2.4T-Parameter Model Sets New Bar in Autonomous Coding and Real-World Work

Alibaba's Qwen 3.8-Max, a 2.4T-parameter MoE model, demonstrates unprecedented autonomy in coding and work tasks, from building a...

48
Qwen (Alibaba)
Qwen (Alibaba) 19 Mar 2026

Qwen3.5-Max-Preview Debuts on Arena with Strong Preliminary Results

Alibaba's Qwen team has released the preview of Qwen3.5-Max on the Arena platform, showcasing impressive performance in early eval...

26
Research paper
Qwen (Alibaba) 23 Dec 2025

Qwen-Image-Edit-2511: Enhanced Consistency and LoRA Integration

Alibaba's Qwen team releases an improved image editing model with better character consistency, multi-person group photo fusion, b...

30