Athanor: An Agent System for Autonomous Post-Training Research¶

简体中文 · English

Research reference only. The full experiments exceed the cost budget and were not reproduced (review complete). No complete research campaign is run, and the existing Evolve controller is not relabeled as Athanor.

Paper information¶

Field	Details
Paper	arXiv 2610.21055; pinned v1 PDF
Affiliations	First author Yufa Zhou: ByteDance / Duke University; additional collaboration with University of Chicago and Virginia Tech.
First public release	2026-10-15 00:56:33 UTC (arXiv v1; announced on 10-16)
Authors' code	The paper describes the system as open. A usable complete official source repository and license were not verified during this review, so it is not marked as open source.
Local adapter / stable key	`athanor` (reference key, no executable adapter)
Local implementation	No algorithm implementation; only this write-up and catalog metadata.
Acceptance status	Full experiments not reproduced, review complete; no local model scores.

Original paper summary¶

Background and main changes¶

Research agents choose data, training algorithms, and combinations of training stages. A separate service executes asynchronous jobs, decontaminates data, freezes evaluation, and records checkpoint lineage. The team shares experiment records and models and combines SFT, RL, model merging, and self-distillation. Calls to external teachers are prohibited.

Architecture and data flow¶

flowchart TD
  H[Persistent experiment records and checkpoint lineage] --> A[Research-agent team]
  A --> R[Executable training recipe]
  R --> E[Asynchronous training and decontamination service]
  E --> V[Frozen development evaluation]
  V --> H
  E --> T[Isolated final held-out evaluation]

Key figure from the paper¶

Figure 2 from the paper: Overall method design and key components. The figure comes from the original paper, © the original authors. Select the image to see the source.

Core equations¶

Section 2 separates outer-loop research from inner-loop training:

\[ r_t\sim\pi(\cdot\mid H_{t-1}),\quad M_t=\operatorname{Exec}(r_t),\quad \hat p_t=\operatorname{Eval}_{B}(M_t;\xi_t). \]

Development scores are noisy, so the highest single score is not true capability. The final held-out data is not exposed to research agents.

Reported offline and online results¶

The authors report a 104-hour research process with four agents. A formal comparison must distinguish the campaign protocol, three-seed means of submissions, and later re-evaluation under unified decoding. Do not conflate the best single search score with the final result. The paper also notes insufficient data-correctness checks.

No production online A/B test was reported. An autonomous research campaign is not a production deployment experiment.

Local reproduction¶

This entry provides only the paper explanation and the original figure. The paper uses four commercial research agents, each with an eight-GPU node, at a total cost of about USD 12,000. This exceeds the available single A100/A30 resources, so no probe was launched. The full experiments were not reproduced (review complete).

Reproduction limits¶

There is no implementation of research automation, benchmark score, or GPU receipt. A reference link does not establish Evolve integration. See the batch acceptance contract (Chinese) for scope.