Ranking Attention Heads by Gradient Norm Enables Sparse RL Fine-Tuning
SimonShaoleiDu · x · 2026-10-28
An arXiv paper (Shi, Yang, Wang, Zhu, Fazel; shared by Simon Du) studies sparse and continual RL fine-tuning of LMs via attention-head ranking and partitioning. Heads ranked by gradient norm, when fine-tuned, consistently beat lower-ranked heads and often match dense baselines across math, instruction-following, and coding tasks. Partitioning heads across tasks reduces cross-task interference, improving retention while preserving adaptation. Includes theory linking gradient norm to training effectiveness.
More from Research
- Ex-OpenAI researcher's 26,000-word treatise claims a breakthrough on AI interpretability — TuhinChakr · 2026-10-28
- re:dvx control vector library launches, author claims SOTA innovations — voooooogel · 2026-10-28
- Does RL Expand LLM Reasoning or Just Reweight It? New Study Offers Theory and Scaling Laws — burny_tech · 2026-10-28
- MIT study: Removing eyebrows makes faces harder to recognize than removing eyes — aakashgupta · 2026-10-28
- Pushing Single Qwen3-8B Neurons Yields 34,220 Alternate Personas; Agents Can Now Do Neuron Interp — burny_tech · 2026-10-28
- Signature Emerges After Just 2-4 SGD Updates as Weights Shift ~8% — maier_ak · 2026-10-28