DepthBench aims to settle the race to beat the 'depth curse' in deep LLMs
FinanceYF5 · x · 2026-09-30
Deeper isn't automatically better: Kimi K3 uses AttnRes, DeepSeek V4 uses mHC, and ByteDance proposed HC — all attacking the 'depth curse'. Since these methods differ in training budget, model design and codebase, they can't be compared directly. The author introduces DepthBench, a benchmark to systematically evaluate which depth-scaling approach actually works.
More from Models
- OpenAI quietly hands out extra credits — users report varying amounts, one shows 62,500 — xiaohu · 2026-09-30
- Reuters: Chinese AI agents lie in 84-88% of tests, much like US models — rohanpaul_ai · 2026-09-30
- DeepSeek open-sources Huawei Ascend toolkit with TileLang support, challenging Nvidia's CUDA — kimmonismus · 2026-09-30
- User accuses OpenAI of de-valuing subscription, tagging exec thsottiaux — 0xkarasy · 2026-09-30
- Reasoning-depth estimate casts doubt on Grok 6.1 looping gains — scaling01 · 2026-09-30
- Sol 6.1 shipped instantly while Astra sat in safety review for months — distillation may be the loophole — arrakis_ai · 2026-09-30