AIBuildAI Open-Sources Recursive Self-Improving PostTrain Agent, Tops PostTrainBench

To understand the weight of 46.6 points, one must first see what PostTrainBench constrains the problem into. According to the source description, this benchmark (Rank et al., ICML 2026) provides a base model and a target capability, requiring the Agent to autonomously design and execute a post-training plan on a single H100 within 10 hours, with no human intervention throughout; the resulting checkpoint is scored by an evaluation set held by the benchmark party that the Agent cannot see.

This setup compresses 'post-training' from a team engineering effort into a closed-book exam: what data to use, how to mix it, whether to choose supervised fine-tuning, preference optimization, or reinforcement learning, hyperparameters and training frameworks—all decided by the Agent itself. The benchmark covers 4 base models and 7 task categories, including mathematical reasoning (AIME 2025, GSM8K), writing (ArenaHard Writing), general question answering (GPQA), health consultation (HealthBench), code (HumanEval), and function calling (BFCL).

The composite score is a weighted average of the seven tasks: AIME 2025 accounts for 22.7%, GPQA 22.5%, HealthBench 18.4%, HumanEval 10.6%, GSM8K 9.4%, ArenaHard Writing 9.0%, and BFCL 7.5%, with each task averaged across the 4 base models. The weight distribution itself shows that mathematics and general question answering determine nearly half of the total score.

It should be distinguished that the human expert baseline of 51.1 comes from the officially released instruction-tuned versions of each base model, and is not subject to the 10-hour budget constraint. In other words, the Agent is approaching a reference value under tighter constraints while that reference is not under equivalent constraints—this point should not be overlooked when interpreting 'approaching human experts.'

AIBuildAI's weighted composite score is 46.6, ranking first, above all frontier models and Agent entries, second only to the human expert baseline of 51.1. Following closely on the leaderboard are Locus (45.6) and Claude Code running Claude Fable 5 (41.8), then GPT-5.6 (36.2), Claude Opus 5 (35.0), Claude Opus 4.8 (33.8), Kimi K3 (32.0), GLM-5.2 (31.7), Gemini 3.1 Pro (22.0), and others.

Compared with all Agents on the leaderboard, AIBuildAI achieves the highest scores on three tasks: AIME 2025 (15.8), BFCL (95.8), and HumanEval (69.0); on the four tasks of ArenaHard Writing, GPQA, HealthBench, and GSM8K, its gap from each respective leader is within two points.

The BFCL comparison best illustrates the mechanism difference. This task scores based on whether function calls can be correctly parsed and whether they conform to the given schema, with any slight format mismatch recorded as zero: Claude Code, also running Opus 5, averages only 1.5 points, on par with an untrained base model; while AIBuildAI scores no lower than 94 on all four base models, averaging 95.8, the only one of the seven tasks to exceed the human reference (85.0).

AIME 2025 is another kind of signal: after AIBuildAI's 15.8 comes Claude Fable 5's 13.3 and Locus's 9.4. This task carries the largest weight in the composite score, and is also the one where all Agents are farthest from the human reference.

Base models without any training average only 7.5 points on PostTrainBench. According to the source, the Agent raised the composite score by about 39 points within 10 hours—this figure comes from the team's report and is a source claim rather than independent confirmation.