PostTrainBench v1.2: Gemini 4 Argon Scores 40.4% (#5), Muse Wonder 24.2% (#17)
maksym_andr · x · 2026-11-09
PostTrainBench updated to v1.2, adding results for Gemini 4 Argon (40.4%, #5 on the leaderboard) and Muse Wonder (24.2%, #17, run via OpenCode).
The benchmark measures how well AI agents can post-train small base LLMs: each agent gets 4 base models (Qwen3 1.7B/4B, SmolLM3-3B, Gemma 3 4B), one H100, and 10 hours to maximize one of seven weighted benchmarks (AIME 2025, Arena Hard, BFCL, GPQA Main, GSM8K, HealthBench, HumanEval). Official instruct versions aren't directly comparable since their training exceeds the constraint. The site also details contamination controls (n-gram decontamination tool), external submission review (Locus run by Intology, 3 seeds, 84 runs averaging 9h38m), and a Fable 5.1 GPQA fallback to Opus 5.
Related event: PostTrainBench v1.2 Adds Gemini 4 Argon at 40.4%, Muse Wonder at 24.2%(2 posts)→
More from Models
- Polymarket odds of "Fable" launching today surge to 70% — spreadlove5683 · 2026-11-10
- Congress returns to session with expected focus on OpenAI rogue model incidents — peterwildeford · 2026-11-10
- "Benchmarks rule out the bad models — the good ones I test on my own work" — ChemicalExcellent463 · 2026-11-10
- Elon Musk amplifies Grok Bot feedback thread as team works on token efficiency fixes — elonmusk · 2026-11-10
- MathArena adds October BrokenArXiv: GPT-6.1 Sol leads but drifts down monthly, Opus 5.5 rock solid — teortaxesTex · 2026-11-09
- Users report Opus 5.5 back to normal, suspect Anthropic silently reverted the degradation — ryunuck · 2026-11-09