COLM 2026 Accepts Multiple Papers on SSMs and LM Architectures

realDanFu · x · 2026-07-17

A research team announced that 4 of their papers have been accepted by COLM 2026, focusing on architectural innovations for efficient language models and State Space Models (SSMs): - **Parcae: Scaling Laws for Stable Recurrent Language Models**: Treats layer recurrence as a dynamic system to build stable recurrent LLMs, offering a new predictable scaling axis under constant memory. - **Exploring Document 'Soupability' in State Space Models**: Uses Mamba2 to encode documents separately, then averages their SSM states into a 'soup'. This method achieves better multi-document QA at extremely low inference costs and scales to handle 256 documents. - **Format Tax**: Research on measuring and mitigating the impact of formatting on model performance.

Original post →

More from Research

Research channel →