Vals RSI — LM Training
Agents · 2026-09-21
Reference-normalized score for a GPT training program researched offline for 24 hours and 24 H100-hours on enwik9. Submitted code is retrained on three hidden seeds; median BPB over 32,768 hidden windows. Baseline 1.3699 BPB, reference 0.85, theoretical best 0.6.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5.5 | 53.0 |
| Claude Fable 5.1 | 49.5 |
| GPT-6 Astra | 48.4 |
| Claude Opus 5 | 42.8 |
| Grok 4.6 | 42.3 |
| Muse Spark 1.3 | 40.9 |
| Claude Fable 5 | 40.8 |
| Opus 4.8 | 40.8 |
| GLM-5.3 | 40.8 |
| Kimi K3 | 40.4 |
| Opus 4.7 | 39.3 |
| Gemini 3.8 Flash | 36.2 |
| GPT-5.6 Sol | 36.1 |
| GPT-5.5 | 31.1 |
| GPT-5.2 | 30.6 |
| GPT-5.4 | 28.0 |
| Kimi K2.5 | 25.3 |