Kimi K3 training curves suggest better scores with fewer steps as RL FLOPs rise
prdeepakbabu · x · 2026-07-29
- The Kimi K3 paper includes a chart showing that, for coding experience tasks, higher RL FLOPs correlate with higher scores while average steps start to drop mid-way.
- The poster argues this may indicate the model is becoming more efficient or exploiting the task better, rather than continuing to explore inefficiently.
- They note the pattern does not appear similarly in other domains, and wonder whether the reason is limited optimization headroom or a curriculum confounder.
More from Models
- Trying Grok bot surfaces an unpaid Cursor balance, blocking account merging — haydendevs · 2026-10-16
- TensorFold 1.0.12 runs 120B vision model on 1-2 DGX Sparks, adds Mac image input — HankYeomans · 2026-10-16
- Leaked screenshot shows unannounced Gemini 4 Argon gated to Ultra and paid Pro subscribers — il_-bro · 2026-10-16
- Google: Gemini 4 Argon lands in Antigravity, code submissions up 35%, rollbacks down ~33% — tekbog · 2026-10-16
- Google's "ungrounded beliefs" classifier injections derail Opus 4.1's reasoning chain — liminal_bardo · 2026-10-16
- A week of ChatGPT Pro quota burned in 36 hours as GPT-6 fails to factor a 512-bit RSA key — matthew_d_green · 2026-10-16