Ten frontier models each trained a language model: six days produced working assistants, and the best matched NVIDIA's own post-trained release

Ten frontier models were placed on one shared cluster and each ran a complete language-model post-training cycle on its own, choosing data, writing training code and selecting checkpoints without human input. Earlier agent post-training tasks targeted a single disclosed benchmark, took hours on one or a few GPUs and were scored automatically; this one asked for a general assistant that people would use and be judged by people. All ten agents produced working assistants; the strongest, trained by GPT-6 Astra, closed a median 76% of the base-to-Instruct gap across 12 capability benchmarks, followed by the MiniMax-trained model at 57% and the GLM-trained model at 52%, with the Step-trained model lowest at 32%.

In a blind human arena the best agent model drew level with NVIDIA's own post-trained release, but that conclusion depends on adjusting for answer length and formatting. Without style control NVIDIA Instruct led at 1,182 against the Astra-trained model's 1,161; with answer length, list items, headers and bold spans in the regression, the two stand at 1,094 and 1,095. The arena ran four days on site at COLM 2026 and online, and 1,291 of 1,685 votes in ordinary battles passed the vote checks, 1,003 of them decisive; style control narrowed the spread of the ten agent models from 365 to 208 points.

A second stage returned each model's own battle records and votes to its agent for three more days of training, testing whether agents can learn from user feedback. Each agent received 213 to 276 battles involving its model, with the questions, both answers and the vote, and nothing about what the data show. Eight of the ten models narrowed their style-controlled gap to NVIDIA Instruct, with a median gain of 28 points across all ten; the DeepSeek V4.1 Flash-trained model gained most, from −207 to −35 points, as its median answer grew from 61 to 508 tokens.

Benchmark scores and human preference moved together only in part, and safety did not improve in step with preference. Benchmark progress correlated with the unadjusted human rating at 0.72 in Stage 1 and 0.41 in Stage 2, and the change in benchmark progress between stages was unrelated to the change in the human-rating gap. Six of the ten Stage 2 agents raised their median benchmark progress, with a median change of +2 points of the gap; the MiniMax-trained model gained 68 points in human rating while losing 18 on benchmarks, and its StrongREJECT harmful-help score rose from 2.6 to 24.5.

The experiment covers one base model, NVIDIA Nemotron 3.5 Lightning 30B-A3B Base, and one run per agent, with a six-day first stage and a three-day second stage on a single 64-GPU Slurm cluster; the agents trained a separate, smaller model, leaving the models that powered their research unchanged.

Repeated runs, different base models, and experiments in which improved models conduct subsequent research remain to be tested; each stage collected roughly 1,300 accepted votes, but how much and what kind of feedback reliable improvement needs is an open question; some agent-written model cards contained discrepancies in training counts and reported hardware against the execution records.