AI Paper Atlas

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

Temperature Scaling in Discrete Sequence (Language) Models

Hannah Scheufele, Peter Blohm, Vikas K Garg

ICML 2026 · Large language models › Instruction and fine-tuning · approach: Fine-tuning generalization and delta compression

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
use Lk or Lp fine-tuning instead of myopic token-wise scalingBD3LM · OpenWebTextNLL (T=0.5): 5.28 -> 3.41 (Lk=1)Improved
use Lk or Lp fine-tuning instead of myopic scalingBD3LM · OpenWebTextschedule variance: 624.3 -> 72.3 (Lp)Improved
use Lp instead of myopic scalingBD3LM · OpenWebTextKendall tau: 0.862 -> 0.750Hurt
use Lp instead of myopic scaling to reach high temperaturesBD3LM · OpenWebTextTeff max: 1.04 -> 2.36Improved
use Lp instead of myopic scalingGPT2 · OpenWebTextNLL (T=0.5): 5.50 -> 3.33Improved
add Ldiff (SSUM objective) to Lp (ablation)Lp · OpenWebText, target T=0.8effective temperature: 0.86 -> 0.90 (GPT2); 0.80 -> 0.93 (BD3LM)Mixed
add Ldiff to Lp (ablation)Lp · OpenWebText, target T=0.8NLL: 3.44 -> 3.07 (BD3LM)Improved
use KL-based objectives LSSTS/LHTSLSSTS, LHTS · OpenWebText, target T=0.8effective temperature: 0.97 (LSSTS), 0.89 (LHTS) vs target 0.8No clear effect
increase MC samples k from 1 to 4 in Lk (ablation)Lk · OpenWebTextNLL: 3.41 -> 3.39No clear effect
fine-tune with Lp, T=0.25 instead of myopic TSPhi-3.5-mini-instruct · MATH500pass@1: 0.446 -> 0.452No clear effect
fine-tune with Lp, T=0.25 instead of myopic TSPhi-3.5-mini-instruct · HumanEvalpass@1: 0.035 -> 0.095Improved
fine-tune with Lp, T=0.25 instead of myopic TSQwen2.5-7B · HumanEvalpass@1: 0.539 -> 0.370Hurt

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