Athanor: An Agent-Native System for Autonomous Post-Training Research
Yufa Zhou, Yichen Wang, Jiakun Fan, Ziqi Shu, Zhichao Liu, Anru R. Zhang, Neiwen Ling, Zijun Wei
cs.AI, cs.CL, cs.LG, cs.MA
2026-10-15
Four Opus 5.5 agents, no teacher and no research advice, lift Qwen3-4B-Base from 15.9 to 67.3 (held-out 9.2 to 48.2) in 104 hours, matching Instruct-2507 at 66.5 and 47.9.
Post-training searches over data, algorithms, hyperparameters, and the order of stages. A single trial often occupies the GPUs for hours before its checkpoint can be scored.
PostTrainBench and Locus have an agent train a separate specialist for each benchmark. Other setups allow a teacher model, or they leave the held-out evaluation unsealed. Athanor tightens those choices. It starts from Qwen3-4B-Base with no instruction tuning, asks for one checkpoint, and wants simultaneous gains on math, logic, code, knowledge, instruction following, and multi-step action. Agents cannot call an external model. They tune on a development suite. A twin suite, same skills and different questions, is opened only after the run.
Four Claude Opus 5.5 agents, one eight-GPU node each, write training code inside Claude Code. The environment queues jobs, trains, and scores with a frozen evaluator. Public data brought in later is stripped of overlap with the benchmarks. Requests, lineage, transcripts, and notes live outside the context window. Jobs run in the background and report back when they finish. The operator starts the machines, replaces lost nodes, and scores the held-out suite at the end. No research advice.
Starter code already includes supervised fine-tuning, GRPO, RL on ALFWorld, distillation from a teammate's checkpoint, and mergekit merges. The written goal is one sentence: raise the development-suite mean. Development scores for the base, the untouched starter, and official Instruct-2507 are visible from the start.
The submission chains four known steps.
At hour 15.8 the operator says the score will be the single checkpoint the team submits. The rule was fixed in advance and was not in the docs. Fetches of teammates' weights start 56 seconds later. After hour 21 every new best mixes training from more than one agent. The log has 57 fetches and 160 note versions.
Under the protocol being optimized, math is sampled, everything else is greedy, and math, logic, and code cap at 16,384 tokens. The submission averages Agent 3's and Agent 2's self-distilled models: 66.1±0.8 over three seeds, against 66.4 for Instruct-2507, 36.4 for the untouched starter, and 16.1 for the base. Held out, it scores 46.7±1.3, against 8.7 and 48.0±1.4. It sits about 1.3 below Instruct. The 95% interval, from -4.0 to +1.3, means the two are not separable. The starter closes 30% of that gap; the submission closes 97%. The other three submissions each lie within two standard errors of this one.
The abstract's 67.3 and 48.2 are a later remeasurement, with a 32,768-token cap and one sampling setup for every model. Instruct-2507 scores 66.5 and 47.9, and the paired intervals include zero. The reasoning releases are higher: hybrid 73.3/50.6, Thinking-2507 80.5/63.0. The table is that remeasurement.
| Model | Development | Held-out |
| Qwen3-4B-Base | 15.9 | 9.2 |
| Team submission (remeasured) | 67.3 | 48.2 |
| Instruct-2507 | 66.5 | 47.9 |
| Qwen3-4B hybrid | 73.3 | 50.6 |
| Thinking-2507 | 80.5 | 63.0 |
The profile is uneven. On the campaign protocol, action is 98.9 against 82.4, and the two held-out ALFWorld task types are 78.8 against 65.9. Code is 47.1 against 43.9. Logic is 63.4 against 80.4, math 37.2 against 42.9. Without action the submission trails by 3.7 and 4.2. A longer cap leaves logic at 63.0 against 79.5. The gap sits mostly in the larger puzzles.
The run costs about $11,800: 3,086 GPU-hours at $3.49, about $10,800, plus $1,074 of agent inference. The score is near 50 after about $1,100 and near 60 after about $2,400. Getting from there to the submission's 66.1 costs about $8,000 more. The record holds 500 jobs, 255 full-suite evaluations, and 174 checkpoints. Reasoning is 3.6% of session time; 87% is spent waiting on jobs.
Set next to the earlier systems the paper cites, agents train one specialist per benchmark, or they may call a teacher. This run bars the teacher and supplies no research advice. One generalist is submitted, and the twin questions stay hidden. The remeasurement is 67.3 and 48.2, against 66.5 and 47.9 for Instruct-2507. The vendor's small Qwen3 models are distilled from a larger flagship. The corpus these agents start from is mostly reasoning traces that someone else already distilled.
Every scored answer stays in the record, so Agent 3 can separate repetition loops from answers that are merely long. Totals alone would not have produced that change.
The last three days rise on merges and on training from the model's own samples. The corpus picked in the first hour sets the search that follows.
One campaign, one 4B base, four identical copies of Opus 5.5. The four submissions share notes and weights, so they are not independent repeats, and there is no run that forbids sharing. The paper states that transfer to larger models or other families is untested.
Of 447 datasets, the agents look at 43 and use 27. Nothing compares two sources with every other choice fixed. A lineage on different data stalls near 55 while the main line is near 62, and it is dropped. Format, length, and whether a reward can score a row do get checked; 1,093 unscorable instruction rows are removed. Whether the answers are actually correct mostly does not. One agent notes that some labels look like noise and never runs the filter.
On 23 re-evaluated checkpoints, 19 score higher on the first seed (p = 0.003), by 0.4 on average. The hill-climb chases evaluation noise. Under the protocol the agents optimized, the number is 66.1 against 66.4. The 67.3 against 66.5 appears only in the remeasurement. The hour-15.8 reminder changes checkpoint sharing at once, and that reminder is not measured on its own.