Pith. sign in

REVIEW 3 cited by

Model Utility Law: Evaluating LLMs beyond Performance through Mechanism Interpretable Metric

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2504.07440 v3 pith:FGOTUAZU submitted 2025-04-10 cs.CL

Model Utility Law: Evaluating LLMs beyond Performance through Mechanism Interpretable Metric

classification cs.CL
keywords modelperformancellmsacrosschallengeeffortevaluationissue
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved
0 comments
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Large Language Models (LLMs) have become indispensable across academia, industry, and daily applications, yet current evaluation methods struggle to keep pace with their rapid development. One core challenge of evaluation in the large language model (LLM) era is the generalization issue: how to infer a model's near-unbounded abilities from inevitably bounded benchmarks. We address this challenge by proposing Model Utilization Index (MUI), a mechanism interpretability enhanced metric that complements traditional performance scores. MUI quantifies the effort a model expends on a task, defined as the proportion of activated neurons or features during inference. Intuitively, a truly capable model should achieve higher performance with lower effort. Extensive experiments across popular LLMs reveal a consistent inverse logarithmic relationship between MUI and performance, which we formulate as the Utility Law. From this law we derive four practical corollaries that (i) guide training diagnostics, (ii) expose data contamination issue, (iii) enable fairer model comparisons, and (iv) design model-specific dataset diversity. Our code can be found at https://github.com/ALEX-nlp/MUI-Eva.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Demographic Prompting at Scale: When More Attributes Hurt LLM--Human Agreement

    cs.CL 2026-07 conditional novelty 6.5

    Across five subjective tasks and five open-source LLMs, demographic prompting improves human agreement only for 1–3 high-signal, directionally coherent attributes and degrades under the full attribute set.

  2. Does the Same Token Mean the Same State? MoE Routing as Signal for Reasoning Control

    cs.CL 2026-06 unverdicted novelty 6.0

    MoE routing states at boundary and delimiter anchors form basins that align with final answers, enabling RAD, a string-free multi-rollout selector that matches majority voting on math and code tasks.

  3. Neuron-Aware Data Selection for Annotation-Free LLM Self-Distillation

    cs.LG 2026-07 unverdicted novelty 5.0

    Neuron-OPSD uses neuron activations to guide data selection and teacher construction for annotation-free on-policy self-distillation in LLMs, claiming better in-domain results without harming cross-domain performance ...