AI Paper Atlas

Research map / Large language models / Instruction and fine-tuning

Beyond Temperature: Hyperfitting as a Late-Stage Geometric Expansion

Meimingwei Li, Yuanhao Ding, Esteban Garces Arias, Christian Heumann

ICML 2026 · Large language models › Instruction and fine-tuning · approach: Efficient LLM fine-tuning systems · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
hyperfit the model (vs temperature-scaling to match entropy)TinyLlama-1.1B · greedy generationTTR: 0.397 -> 0.684Improved
hyperfit the model (vs temperature-scaling to match entropy)TinyLlama-1.1B · greedy generationbigram repetition: 0.604 -> 0.140Improved
lower temperature to T=0.59 to match hyperfitted entropy (ablation)TinyLlama-1.1B · greedy generationTTR: 0.400 -> 0.397No clear effect
inject static logit bias, alpha=0.01 (ablation)TinyLlama-1.1B · static injection ablationbigram repetition: 0.588 -> 0.609Hurt
inject static logit bias, alpha=0.50 (ablation)TinyLlama-1.1B · static injection ablationTTR: 0.449 -> 0.215Hurt
restrict LoRA to last 5 layers (vs Full LoRA)TinyLlama-1.1B · hyperfitting reproductiontop-1 agreement: 0.523 -> 0.517No clear effect
restrict LoRA to last 5 layers (vs Full LoRA)TinyLlama-1.1B · hyperfitting reproductionTTR: 0.508 -> 0.469Hurt
restrict LoRA to last 5 layers (vs Full LoRA)Qwen2.5-1.5B · hyperfitting reproductionTTR: 0.575 -> 0.591Improved
restrict LoRA to last 5 layers (vs Full LoRA)Qwen2.5-1.5B · hyperfitting reproductionbigram repetition: 0.248 -> 0.213Improved
restrict LoRA to last 5 layers (vs Full LoRA)Qwen2.5-1.5B · LLM-as-judge pairwisewin rate (%): 57.3 vs 41.2Improved
restrict LoRA to last 5 layers (vs Full LoRA)LLaMA-3.2-3B · training timespeedup: 1.44x -> 1.99xImproved
restrict LoRA to last 5 layers (vs Full LoRA)Qwen2.5-1.5B and TinyLlama-1.1B · trainable parameterstrainable parameter reduction: 78.3% and 82.7%Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy