Four Opus agents on Athanor post-train Qwen3-4B-Base to Instruct-level averages

Weak recursive self-improvement used to be a slide. Athanor turns it into a dated GPU campaign with a sealed held-out suite.

On 15 October 2026, Yufa Zhou, Yichen Wang, Jiakun Fan, Ziqi Shu, Zhichao Liu, Anru R. Zhang, Neiwen Ling, and Zijun Wei posted arXiv:2610.21055, "Athanor: An Agent-Native System for Autonomous Post-Training Research." Four Claude Opus 5.5 agents, each locked to an eight-GPU sandbox and barred from external teacher models, were told to improve Qwen3-4B-Base on a six-task development suite. After 104 hours the team matched or edged the vendor instruct release on averages that include a held-out twin suite the agents never saw.

Athanor treats AutoPostTrain as a programmable research loop on real GPUs

The paper's claim of novelty centres on a harder protocol: a team of frontier agents autonomously post-trains one raw base model into a cross-domain generalist that matches the official instruct checkpoint on average, under sealed held-out evaluation, decontaminated public data, and a ban on teacher models.

Athanor makes the research loop programmable. Agents keep every research decision inside Claude Code sandboxes. The environment supplies atomic operations: start a training job, score a checkpoint with a frozen evaluator, pull decontaminated public data, and save or submit artifacts. Jobs run on real GPUs. Every score is supposed to link back to the recipe and evidence then available. The operator recovers lost nodes and scores the final submission on the held-out suite after the campaign window closes.

That design is the weight-side companion to published Harness². Harness² freezes weights and spends the recursion budget on harness files. Athanor freezes the evaluation harness and spends the budget on weight updates. Both ask whether today's agents already have enough research and engineering skill once execution friction falls.

Development 15.9 to 67.3, held-out 9.2 to 48.2, against Instruct 66.5 and 47.9

The final remeasurement in Figure 1 and Table 1 uses a longer output limit and one sampling setting for every model. Qwen3-4B-Base starts at 15.9 on the development suite mean and 9.2 on the held-out mean. The team's submission reaches 67.3 and 48.2. Qwen3-4B-Instruct-2507 scores 66.5 and 47.9 under the same evaluator. On the campaign protocol the agents actually worked with, the team's three-seed submission averaged 66.1 against 66.4 for Instruct.

Development tasks named in the paper include ALFWorld, IFEval, MMLU-Pro, ZebraLogic, LiveCodeBench 2024, and maths suites spanning early 2025 through late 2025. Held-out twins cover later LiveCodeBench, the other half of ZebraLogic, SuperGPQA, IFBench, and an unseen ALFWorld split, among others. Agents see development scores during the search. They do not see held-out scores until the operator remeasures after the window.

The staircase in Figure 1 shows best-so-far climbing across 104 hours. Markers call out SFT on public reasoning data, GRPO with a share of greedy rollouts, merges across agents, weight extrapolation, and self-distillation via rejection-sampling fine-tuning. A default starter run with no research sits far below the team.

Instruct-2507 is the right bar, and the agents still inherit distilled public corpora

Comparing against the vendor instruct release under one frozen evaluator is the cleanest reading of the headline. The paper also places Athanor against narrower AutoPostTrain settings that train one specialist per benchmark, allow teacher models, or leave held-out evaluation unsealed. On those axes the protocol is stricter.

Fairness still has a soft edge. The vendor describes building small Qwen3 models by distillation from larger flagships. The agents reach Instruct-level averages from public data, much of which was itself distilled from stronger models by its publishers. That is allowed by the rules and disclosed in the paper. The campaign means no live teacher sat in the loop; stronger-model knowledge can still sit in the training mixture through those public corpora.

Within the team, after hour 21 every new best combines training from several agents. Ideas and checkpoints move between sandboxes within hours. Read the headline as a team-merge baseline after hour 21, when recombination across sandboxes is already the search method.

New algorithms, answer correctness audits, and open-system reproduction stay outside the win

The campaign record, as the authors analyse it, shows recombination of known methods rather than a new post-training algorithm. Agents catch format, length, scorability, and difficulty problems in data. They rarely check whether answers are correct. Four copies of one agent model start from the same prior, converge on one public reasoning corpus within the first hour, and drop later alternatives that score lower.

Independent reproduction of the open Athanor stack on buyer hardware is unsettled on this paper alone. Dollar cost per point of held-out gain depends on the GPU price the authors plot (they mark jobs at $3.49 per GPU-hour in a cost breakdown figure); Frontier declines to treat that figure as a buyer quote. Whether a different frontier agent family, or a single agent with the same total GPU budget, matches Instruct averages is unsettled. Safety properties of autonomously chosen data mixtures are outside the reported metrics.

arXiv PDF is public; treat "open system" as a claim to verify before you depend on it

Paper: arXiv:2610.21055 (posted 15 October 2026). The abstract and introduction describe Athanor as an open agent-native system. Before Stack-depth adoption, operators should verify licence, repository contents, and whether managed-job services match the paper's campaign protocol. Base and instruct Qwen3-4B weights remain upstream of this release. Claude Opus 5.5 access is outside the paper.

Read it as a weak-RSI yardstick, then decide whether your lab can seal a held-out suite

Reach for Athanor's framing when you need a concrete bar for agent-led post-training research: one generalist checkpoint, public data only, no teacher API, frozen eval, sealed held-out twin. Use the Instruct-average match as the existence proof that today's agents can run the loop when jobs and evals are managed.

Treat the campaign recipe as a drop-in instruct replacement only after you re-run under your own decontamination and answer-audit rules, and treat "autoresearch" as a product category only after independent stacks reproduce the open system. Pair the read with Harness² if your recursion budget still lives in prompts and tools rather than weights.

Weight-side RSI sits beside harness-side RSI, with the same early-convergence risk

Athanor extends the line of agent post-training benches and systems that previously specialised per task or allowed teachers. Its held-out averages make the weak-RSI claim measurable in a way those narrower setups did not. The early recipe convergence the authors report is the practical limit to watch as agent diversity and compute scale: more GPUs help only if the team keeps trying different ideas after the first public corpus wins.

The Monday question is no longer whether agents can touch post-training. It is whether labs will seal their evaluators tightly enough to know when the agents have actually generalised.