COLM Papers Explore Connections Between LoRA, Steering, and Reward Modeling
robinomial · x · 2026-08-01
This post highlights several LLM research updates slated for the COLM 2026 conference:
- New Perspective on LoRA: Research reveals interesting connections between LoRA, conditional steering, and reward modeling.
- Improving Generation Faithfulness: Introduces Token-Level Off-Policy Labeling (TOPL). Surprisingly, training a language model for binary classification can improve generation. Instead of optimizing next-token prediction, TOPL trains an LLM to classify whether each generated token is faithful under distribution shift.
More from Research
- It's not gradient descent: burny_tech corrects the diagnosis to RLHF mode collapse — burny_tech · 2026-10-16
- A beginner-friendly introduction to causal inference kicks off, Part 1 — Nasereliver · 2026-10-16
- FBI method trains for Pass@K with one sample per update at 12.3% of the cost — EmtiyazKhan · 2026-10-16
- Berkeley researchers sample humanoid robot policies directly from weight space with diffusion — berkeley_ai · 2026-10-16
- OneWM: unified causal training shows world and action modeling improve each other — IAAI233 · 2026-10-16
- Baidu's Virtual Head Attention tames the KV cache, FLOPs, quality triangle in long-context LLMs — baidu · 2026-10-16