AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Spilled Energy in Large Language Models

Adrian Robert Minut, Hazem Dewidar, Iacopo Masi

ICLR 2026 · Large language models › Language, safety and interpretability · approach: Parallel and early-exit decoding · 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
use spilled energy instead of logit confidenceHallucination detector · LLaMA-3-Instruct, 9 benchmarks averageAuROC: 54.62 -> 73.16Improved
use spilled energy instead of trained probing classifierHallucination detector · LLaMA-3-Instruct, 9 benchmarks averageAuROC: 64.16 -> 73.16Improved
use spilled energy instead of marginal energyHallucination detector · Mistral-7B, 9 benchmarks averageAuROC: 77.19 -> 73.94Hurt
instruction tuning LLaMA-3 for spilled energySpilled energy · 9 benchmarks averageAuROC: 68.69 -> 73.16Improved
instruction tuning Mistral for spilled energySpilled energy · 9 benchmarks averageAuROC: 73.94 -> 77.49Improved
instruction tuning LLaMA-3 for logit baselineLogit confidence · 9 benchmarks averageAuROC: 56.89 -> 54.62Hurt
use min pooling instead of max pooling (ablation)Spilled energy · LLaMA-3-Instruct, 9 benchmarks averageAuROC: 50.07 -> 73.16Improved
use min pooling instead of max pooling (ablation)Spilled energy · Mistral-7B, 9 benchmarks averageAuROC: 61.19 -> 73.94Improved
localize exact answer tokens (ablation)Spilled energy · Average over 4 LLMs and 9 benchmarksAuROC: +24.06Improved
localize exact answer tokens (ablation)Logit confidence · Average over 4 LLMs and 9 benchmarksAuROC: +9.23Improved
use scaled spilled energy (multiply by marginal energy magnitude) (ablation)Spilled energy · LLaMA-3, 9 benchmarks averageAuROC: 68.69 -> 68.12Hurt

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