Institution profile

StepFun

Industry researchasia · cn
Official website
Research library142linked papers
Opportunities0open roles
Selected work

Representative Papers

Open-Reasoner-Zero: An Open Source Approach to Scaling Up Reinforcement Learning on the Base Model

Mar 31, 2025

This work addresses the limited scalability, procedural complexity, and poor reproducibility of large-scale reinforcement learning (RL) for enhancing large language model (LLM) reasoning capabilities. We propose the first open-source, fully reproducible pure-RL reasoning training paradigm: it eliminates KL-divergence regularization and instead adopts a minimalist PPO implementation with generalized advantage estimation (GAE) using γ = λ = 1, coupled with rule-based sparse reward shaping, enabling end-to-end unsupervised optimization. Our method achieves superior performance over DeepSeek-R1-Zero-Qwen-32B on AIME2024, MATH500, and GPQA Diamond—matching or exceeding its results with only 10% of the training steps. We publicly release all code, datasets, and model weights across multiple scales. This significantly improves the simplicity, scalability, and accessibility of RLHF-based reasoning training.

6 citations1 influentialRead paper

Diffusion Language Models are Super Data Learners

Nov 05, 2025

The performance advantages of diffusion language models (DLMs) under data-constrained regimes remain unclear. Method: This work systematically investigates DLM training dynamics and generalization with limited unique tokens, introducing a triple-gain mechanism: (i) arbitrary-order sequence modeling, (ii) high-density computation via iterative bidirectional denoising, and (iii) intrinsic Monte Carlo sampling augmentation—collectively overcoming data-efficiency bottlenecks. Training employs repeated pretraining, input/parameter noise injection, and standard diffusion objectives. Results: A 1.7B-parameter DLM surpasses same-scale autoregressive (AR) models using only 10B unique Python tokens; a 1B-parameter DLM achieves 56.2% accuracy on HellaSwag and 33.7% on MMLU using merely 1B tokens—significantly outperforming AR baselines trained on comparable data volumes. This study provides the first empirical validation of DLMs’ sustained competitive advantage in the small-data, large-model paradigm.

6 citationsRead paper

BabyVision: Visual Reasoning Beyond Language

Jan 10, 2026arXiv.org

This work addresses the overreliance of current multimodal large language models on linguistic priors and their consequent deficiency in foundational visual understanding—capabilities that even human infants possess—leading to markedly subpar performance on basic visual tasks. To systematically evaluate pure visual reasoning independent of language, the authors introduce BabyVision, a comprehensive benchmark comprising 388 non-linguistic visual tasks across four major categories and 22 subcategories. They further present BabyVision-Gen, a generative model tailored for this benchmark, along with an automated evaluation toolkit. Experimental results reveal that leading models, such as Gemini3-Pro-Preview (scoring 49.7), fall significantly short of adult human performance (94.1), underscoring a critical gap in foundational visual primitives and highlighting the need to advance multimodal models toward more human-like visual perception.

4 citations2 influentialRead paper

PaCoRe: Learning to Scale Test-Time Compute with Parallel Coordinated Reasoning

Jan 09, 2026arXiv.org

This work addresses the limitation of current language models, which are constrained by fixed context windows and struggle to effectively scale test-time compute (TTC). We propose Parallel Coordinated Reasoning (PaCoRe), a novel framework that introduces the first message-passing-based multi-round parallel reasoning architecture. By enabling parallel inference trajectories coupled with context-aware message compression and synthesis, PaCoRe extends effective TTC to millions of tokens without exceeding context length limits. The model is trained end-to-end via reinforcement learning to master information coordination and synthesis. Evaluated on the HMMT 2025 mathematical reasoning benchmark, our method achieves 94.5% accuracy, surpassing GPT-5 (93.2%). We publicly release the model, data, and full reasoning pipeline.

3 citationsRead paper

StepAudio 3 Music Technical Report

Sep 11, 2026

研究提出StepAudio 3 Music模型，通过结合语义自监督多任务训练、流匹配扩散Transformer等方法，实现长格式音乐生成与文本控制，提高音乐质量和相似度。

0 citationsRead paper
Recent publications

Latest Papers

StepAudio 3 Music Technical Report

Sep 11, 2026

研究提出StepAudio 3 Music模型，通过结合语义自监督多任务训练、流匹配扩散Transformer等方法，实现长格式音乐生成与文本控制，提高音乐质量和相似度。

0 citationsRead paper

Scheduling Mixed RL Rollouts Beyond Prefix Locality

Aug 11, 2026

This work addresses the intense KV cache contention arising from heterogeneous reinforcement learning rollback tasks—such as RLVR, RLHF, and agent rollouts—in shared inference services, where differences in sequence structure, interaction patterns, and KV cache residency durations degrade throughput efficiency and training objective consistency. To resolve this, the authors propose MISA-T, a routing-layer admission control policy that jointly models workload heterogeneity and KV cache residency characteristics for the first time. MISA-T enables adaptive session admission, workload-aware KV allocation, and residency-aware cache accounting, achieving efficient scheduling while preserving the target task mix ratio. Experiments demonstrate significant improvements: rollback throughput increases by 53.3% and 43.6% on Step3.7 and Qwen3.6-35B-A3B models, respectively, with an average 35.6% throughput gain over 50 rounds, 22.8% reduction in iteration time, and no degradation in task performance.

0 citationsRead paper