Back to News

ByteDance and partners release Axion 1.0, a two-stage SFT+RL agent model

#agent#open-source#reinforcement-learning#benchmark

Researchers from Duke University, ByteDance, University of Maryland, and University at Buffalo released Axion 1.0, an open-source agent model trained with a simple two-stage supervised fine-tuning (SFT) and reinforcement learning (RL) pipeline, without mid-training, DPO, or online policy distillation. The 122B version scored 59.0% on the full Humanity's Last Exam (HLE) benchmark and 66.7% on FrontierScience Research; the 35B version scored 62.2% on SEAL-0. The team open-sourced data construction, generation, filtering, validation, prompts, and training code.

Coverage timeline

  1. 机器之心机器之心

    能搜索资料、运行代码,还能完成多步科学研究任务——训练这样的智能体,究竟需要多少道工序? 中期训练、监督微调、偏好优化、在线策略蒸馏、强化学习…… 随着智能体能力不断增强,训练流程也变得越来越长。每增加一个环节,研究者就需要准备相应的数据,协调新的训练目标,再调试一套训练参数。 来自杜克大学、字节跳动、马里兰大学帕克分校和纽约州立大学布法罗分校的研究团队,推出了 Axion 1.0。其训练流程只包含两阶段监督微调(SFT)和强化学习(RL),没有额外安排中期训练(mid-training)、直接偏好优化(DPO)或在线策略蒸馏(OPD)。 凭借这套简洁的训练流程,Axion 1.0 展现出强势的科学推理和工具使用能力。论文报告,Axion 1.0-122B 在限制部分网站访问的完整 Humanity』s Last Exam(HLE)评测中取得 59.0%(超越 GPT-6),在 FrontierScience Research 上取得 66.7%;35B 版本在信息搜寻基准 SEAL-0 上取得 62.2%。 更重要的是,团队全面开源了数据构造、生成、筛选、验证、提示词及训练代码,向社区公开了完整训练方法。简洁的三阶段流程与透明的开源实现, 让智能体基础模型的训练不再神秘,也让更多人能够训练自己的强大的智能体模型。 论文:https://arxiv.org/abs/2611.03233 项目:https://axion1-0.github.io 代码:https://github.com/bytedance/axion1_0 数据:https://huggingface.co/datasets/Axion1-0/axion-training-data 模型:https://huggingface.co/Axion1-0/Axion-1.0-35B 方法介绍 Axion 1.0 成功的关键,在于 针对智能体所需的不同能力,构造高质量的训练数据。 具体而言,Axion 1.0 将智能体的训练任务分为 两类 :一类根据题目提供的完整信息进行推理与计算,另一类需要借助工具获取缺失的外部信息,再结合证据完成推理。两个 SFT 阶段分别围绕这两类任务展开,让模型先提升推理能力,再学习如何获取和利用外部信息。 第一步 给已经会做的题增加挑战 第一阶段使用数学、物理、化学、生物和计

Related stories

  • Ant Group and Westlake Xinchen release Duo v1, a real-time interactive world model with open weights

    Ant Group and Westlake Xinchen introduced Duo v1, the first implementation of the Reciprocal World Model (RWM) paradigm, which treats intelligence and physical reality as equally fundamental. The model enables real-time full-duplex interaction, allowing users to show real objects, change virtual scenes, or guide AI assistants without keyboard or mouse. Weights and code are open-sourced, with resources on arXiv, GitHub, Hugging Face, and ModelScope.

    1 reports · 1 sources
  • Recursive releases Flint, 40B coding model built largely by its Eureka Machine AI system

    Startup Recursive has released Flint, an open-weight, 40B-parameter mixture-of-experts coding model with 2.5B active parameters per token, built in under four months with heavy automation from its Eureka Machine AI research system. The model outperforms sub-50B MoE models on coding benchmarks like SWE-Bench Pro v2 and runs at over 125 tokens per second on a MacBook Pro M4 Pro, with weights available on Hugging Face under Apache 2.0.

    3 reports · 3 sources
  • ByteDance releases Seedream 5.5 image model with faster generation, transparent assets

    ByteDance released Seedream 5.5, an image generation model now live in Doubao, Doubao Work, and Jimeng, with API coming to Volcano Ark. The model offers faster generation, improved multi-round editing stability (maintaining pixel-level consistency outside edited regions over 100+ rounds), native RGBA transparent asset generation, and enhanced realism and small-text rendering. API pricing continues the two-part structure (input image + output image) with no text token fees; a 2K square image edit costs about 0.28 yuan, less than half of Seedream 5.0 Pro.

    2 reports · 2 sources
  • AISLE benchmark: top AI models score 41/100 in real wet-lab tests

    AISLE, a benchmark for AI scientists' real-world wet-lab capabilities, released first-round results from 4,079 wet-lab tests across GPT, Claude, and Gemini models. The top model, GPT 5.5, scored 41.19 out of 100, and all models had a 0% pass rate on safety-constrained experiment optimization tasks. OpenAI also introduced BioBasicsBench, a 100-task benchmark for interpreting raw experimental data.

    2 reports · 2 sources

News updates every hour. Subscribe in your RSS reader to get new stories automatically.

Subscribe