Athanor: AI System for Independent Model Improvement

Paper: Athanor: An Agent-Native System for Autonomous Post-Training Research

Authors: Yufa Zhou, Yichen Wang, Jiakun Fan, Ziqi Shu, Zhichao Liu, Anru R. Zhang, Neiwen Ling, Zijun Wei

Executive Summary

A recent study introduces Athanor, a system designed to enhance artificial intelligence (AI) capabilities through a method known as autonomous post-training research. This research explores how AI agents can independently improve base models by utilizing publicly available data and tools without external guidance. The context here is the idea of recursive self-improvement, where AI systems can develop stronger versions of themselves, a concept that could lead to significant advancements in AI intelligence.

Athanor operates with a team of AI agents, specifically four Claude Opus 5.5 agents, which are equipped with substantial computational resources (eight GPUs each). These agents are given the autonomy to conduct their research. They can read datasets, analyze results, build their own training data, modify algorithms, and fine-tune their hyperparameters—all while working in a controlled environment that prevents access to external teacher models. The system supports these activities by providing necessary components like frozen evaluations, managed job handling, and a persistent record of research activities.

During a 104-hour testing campaign, the agents significantly improved the performance of a base model, Qwen3-4B-Base. They achieved a score increase from 15.9 to 67.3 on a development suite consisting of six tasks and improved from 9.2 to 48.2 on a separate held-out suite. For comparison, the official instruct version of the model, Qwen3-4B-Instruct-2507, scored 66.5 and 47.9 on the same suites. This research marks a first in demonstrating that a collective of AI agents can autonomously enhance a raw base model into a more capable generalist model, matching or exceeding the performance of its official counterpart.

The results suggest that these AI agents effectively combine existing methods and leverage each other’s advancements, employing techniques like supervised fine-tuning on reasoning data, reinforcement learning with modifications introduced by individual agents, as well as collaborative weight merging and self-distillation processes.

The implications of this research are significant for the future of AI development. Athanor could serve as a foundation for scaling autonomous AI research, potentially leading to more sophisticated AI systems that can evolve and improve without human intervention. The findings stem from simulated environments rather than real-world applications, which means that while the results are promising, their application in practical scenarios remains to be seen. Nonetheless, this research paves the way for exploring how autonomous agents can contribute to advancing AI technology and capabilities.

Academic Abstract

Recursive self-improvement, in which AI systems build stronger AI systems, is one promising route to superintelligence. We study a weaker form that can be measured today: frontier agents post-training a small raw base model with public data, as a test of their research potential. We build Athanor, an open agent-native system for autonomous post-training research (AutoPostTrain) on real GPUs. We give the agents full research autonomy, low-friction engineering tools, and efficient starter code. The agents carry out the research themselves: they read datasets and evaluation results, build their training data, write and modify training algorithms, tune hyperparameters, and analyze the results to decide what to try next. They work in a sandbox and cannot use external teacher models. Athanor's services supply frozen evaluation, managed jobs, decontaminated data, a persistent research record, and a held-out suite the agents are unaware of. In a 104-hour campaign, a team of four Claude Opus 5.5 agents, each on an eight-GPU node and given no research advice, improves Qwen3-4B-Base from 15.9 to 67.3 on a six-task development suite and from 9.2 to 48.2 on a held-out suite, against 66.5 and 47.9 for the official Qwen3-4B-Instruct-2507. To our knowledge, this is the first demonstration that a team of frontier agents can autonomously post-train a raw base model into a cross-domain generalist model that matches or surpasses the official instruct version on average. The campaign's full record shows that the agents reach this level by recombining known methods and each other's models: SFT on public reasoning data, RL with a rollout change one agent introduced, weight merges across agents, and rounds of self-distillation. With execution handled, today's agents already have the engineering and research ability to conquer this task. We hope Athanor provides a foundation for scaling autonomous AI research.