AutoArena Research - Optimizer updates to quality at validation BPB < 1.05: leaderboard

AutoArena reports accepted research runs on 9 separate targets. Each run has a ceiling of 100 launches, including reference launches and failures. The research method chooses training and design changes; the outcome is the best generated artifact that meets the quality requirement (validation BPB < 1.05). It does not measure the throughput of the research LLM itself. Native comparison values, budgets, methods, providers, and explicitly named research LLMs are retained. Training time is discretized to 30-second checkpoints under the native rules; other metrics retain native full precision. Missing LLM metadata remains unknown and is not assigned an assumed Opus variant. Overall Elo is not added because it is derived from the same nine target outcomes. Metric: updates_to_target (optimizer steps).

Metric: updates_to_target (optimizer steps). Source: auto-research-arena.github.io. Saturation forecast: Rough model projection: around 2026. 6 models tracked.

Top models

Interactive version: theaggregate.ai/benchmark?slug=autoarena-research-optimizer-updates-to-quality-at-validation-bpb-under-1-05 · How It Works · Data refreshed daily, snapshot 2026-10-28.