Co-LMLM: 360M Model Beats 4T Token Giants
yoavartzi · x · 2026-07-14
A research team proposed Co-LMLM (Continuous-Query Limited Memory Language Models), a new architecture capable of processing general web text at pre-training scale while externalizing factual knowledge via an efficient external retrieval mechanism.
Experimental results show that this architecture offers extreme parameter efficiency:
- Lower Perplexity: Trained with only 100 billion tokens and 360M parameters, its Perplexity (PPL) significantly outperforms conventional models trained on 4 trillion tokens.
- Stronger Factual QA: It scored 21.7 on the SimpleQA-Verified benchmark, matching or even exceeding frontier large models like GPT-4o-mini and Claude 3.5 Sonnet.
More from Research
- Stanford Data-Centric AI Workshop reopens site, extends poster deadline — jure · 2026-10-26
- 'Vibe verification': ~60% of AI-written formal specs prove the wrong property — basedjensen · 2026-10-26
- Ito: a browser that detects browsing intent and auto-organizes tabs into Flows — andrewkuznet · 2026-10-26
- Cornell study says brains process language as linear sequences, challenging Chomsky's hierarchy — anirbanbandyo · 2026-10-26
- Hi3D V1.0 open-sourced with 2M training assets and 90% slicing success rate — kimmonismus · 2026-10-26
- Ortet evaluates 34 open-weight models on 50 health tasks, finds general leaderboards insufficient — elmanmansimov · 2026-10-26