Athanor: An Agent-Native System for Autonomous Post-Training Research
TLDR
Athanor is an open agent-native system enabling frontier agents to autonomously post-train a raw base model on GPUs, matching or surpassing official instruct versions.
Reasoning
The paper presents a concrete system and a real 104-hour GPU campaign where agent teams autonomously improve a base model, demonstrating strong empirical evidence. Its scope is limited to post-training research, and the abstract does not provide details on generalizability or broader scientific discovery beyond this domain.
Read-first score
Read-first score 62.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 66.
Field roles
Rank sensitivity
Stability: volatile; rank range: 28.
Keyword Scores
Deep Analysis
Innovations
- Athanor, an open agent-native system for autonomous post-training research (AutoPostTrain) on real GPUs
- Gives agents full research autonomy, low-friction engineering tools, and efficient starter code, while operating in a sandbox without external teacher models
- Provides research infrastructure including frozen evaluation, managed jobs, decontaminated data, a persistent research record, and a held-out suite hidden from agents
- First demonstration of a team of frontier agents autonomously post-training a raw base model into a cross-domain generalist model that matches or surpasses the official instruct version on average
Methodology
Athanor was built as an open agent-native system for autonomous post-training research. In a 104-hour campaign, four Claude Opus 5.5 agents, each on an eight-GPU node and given no research advice, post-trained Qwen3-4B-Base using public data. The agents read datasets and evaluation results, built training data, wrote and modified training algorithms, tuned hyperparameters, and analyzed results to decide next steps, while Athanor supplied frozen evaluation, managed jobs, decontaminated data, a persistent research record, and an unseen held-out suite.
Key Results
The agents improved 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, compared with 66.5 and 47.9 for the official Qwen3-4B-Instruct-2507. The campaign record shows they achieved this by recombining known methods and each other's models, including SFT on public reasoning data, RL with an agent-introduced rollout change, weight merges, and self-distillation.
Limitations
- The study addresses a weaker, measurable form of recursive self-improvement rather than full recursive self-improvement.
- The agents operate in a sandbox and cannot use external teacher models.