ICML paper: hyperfitting a LoRA on final 5 layers removes AI slop, code released
grimjim · reddit · 2026-09-06
- A Reddit user highlights the ICML 2026 paper Beyond Temperature: Hyperfitting as a Late-Stage Geometric Expansion, showing that hyperfitting a LoRA on just the final 5 layers has an "antislop" effect, geometrically expanding the output distribution.
- Code is available on GitHub (YecanLee/Beyond-Temperature), and the poster suggests anyone with spare local VRAM try reproducing it.
More from Research
- HKUST(GZ) lab lands 3 CoRL 2026 papers, unveils terrain-adaptive robot motion model — Scobleizer · 2026-09-06
- Mathematician rebuts plan to formalize all human math in a year — lpachter · 2026-09-06
- DisCo distills GitHub repos into agent skills, doubling MLE-bench to 72.89% — rohanpaul_ai · 2026-09-06
- NJU & Wollongong propose Harness Continual Learning: agents evolve scaffolding, not parameters — jiqizhixin · 2026-09-06
- $1,000-trained HRM-Text shows Sapient bet on recurrence before OpenAI's Astra — rohanpaul_ai · 2026-09-06
- Entropy trajectory shape predicts Qwen3-4B errors and transfers to unseen tasks — Happy_Brilliant7827 · 2026-09-06