Research map / Large language models / Language, safety and interpretability
Spilled Energy in Large Language Models
Adrian Robert Minut, Hazem Dewidar, Iacopo Masi
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 changed | Where | Measured | Effect |
|---|---|---|---|
| use spilled energy instead of logit confidence | Hallucination detector · LLaMA-3-Instruct, 9 benchmarks average | AuROC: 54.62 -> 73.16 | Improved |
| use spilled energy instead of trained probing classifier | Hallucination detector · LLaMA-3-Instruct, 9 benchmarks average | AuROC: 64.16 -> 73.16 | Improved |
| use spilled energy instead of marginal energy | Hallucination detector · Mistral-7B, 9 benchmarks average | AuROC: 77.19 -> 73.94 | Hurt |
| instruction tuning LLaMA-3 for spilled energy | Spilled energy · 9 benchmarks average | AuROC: 68.69 -> 73.16 | Improved |
| instruction tuning Mistral for spilled energy | Spilled energy · 9 benchmarks average | AuROC: 73.94 -> 77.49 | Improved |
| instruction tuning LLaMA-3 for logit baseline | Logit confidence · 9 benchmarks average | AuROC: 56.89 -> 54.62 | Hurt |
| use min pooling instead of max pooling (ablation) | Spilled energy · LLaMA-3-Instruct, 9 benchmarks average | AuROC: 50.07 -> 73.16 | Improved |
| use min pooling instead of max pooling (ablation) | Spilled energy · Mistral-7B, 9 benchmarks average | AuROC: 61.19 -> 73.94 | Improved |
| localize exact answer tokens (ablation) | Spilled energy · Average over 4 LLMs and 9 benchmarks | AuROC: +24.06 | Improved |
| localize exact answer tokens (ablation) | Logit confidence · Average over 4 LLMs and 9 benchmarks | AuROC: +9.23 | Improved |
| use scaled spilled energy (multiply by marginal energy magnitude) (ablation) | Spilled energy · LLaMA-3, 9 benchmarks average | AuROC: 68.69 -> 68.12 | Hurt |
Closest papers by meaning
- A Probabilistic Perspective on Unlearning and Alignment for Large Language Models (ICLR 2025 · similarity 0.87)
- KVSmooth: Mitigating Hallucination in Multi-modal Large Language Models through Key-Value Smoothing (CVPR 2026 · similarity 0.87)
- Sample Transform Cost-Based Training-Free Hallucination Detector for Large Language Models (NeurIPS 2026 · similarity 0.87)
- Entropy-informed Decoding: Adaptive Information-Driven Branching (ICML 2026 · similarity 0.87)
- TokenPowerBench: Benchmarking the Power Consumption of LLM Inference (AAAI 2026 · similarity 0.87)
- AutoJudge: Judge Decoding Without Manual Annotation (NeurIPS 2025 · similarity 0.87)
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.
