Revisiting the UID Hypothesis in LLM Reasoning Traces

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Gwak, Minju, Son, Guijin, Kim, Jaehyung
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866909853916594176
author Gwak, Minju
Son, Guijin
Kim, Jaehyung
author_facet Gwak, Minju
Son, Guijin
Kim, Jaehyung
contents Large language models (LLMs) often solve problems using step-by-step Chain-of-Thought (CoT) reasoning, yet these intermediate steps are frequently unfaithful or hard to interpret. Inspired by the Uniform Information Density (UID) hypothesis in psycholinguistics -- which posits that humans communicate by maintaining a stable flow of information -- we introduce entropy-based metrics to analyze the information flow within reasoning traces. Surprisingly, across three challenging mathematical benchmarks, we find that successful reasoning in LLMs is globally non-uniform: correct solutions are characterized by uneven swings in information density, in stark contrast to human communication patterns. This result challenges assumptions about machine reasoning and suggests new directions for designing interpretable and adaptive reasoning models.
format Preprint
id arxiv_https___arxiv_org_abs_2510_13850
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Revisiting the UID Hypothesis in LLM Reasoning Traces
Gwak, Minju
Son, Guijin
Kim, Jaehyung
Computation and Language
Artificial Intelligence
Large language models (LLMs) often solve problems using step-by-step Chain-of-Thought (CoT) reasoning, yet these intermediate steps are frequently unfaithful or hard to interpret. Inspired by the Uniform Information Density (UID) hypothesis in psycholinguistics -- which posits that humans communicate by maintaining a stable flow of information -- we introduce entropy-based metrics to analyze the information flow within reasoning traces. Surprisingly, across three challenging mathematical benchmarks, we find that successful reasoning in LLMs is globally non-uniform: correct solutions are characterized by uneven swings in information density, in stark contrast to human communication patterns. This result challenges assumptions about machine reasoning and suggests new directions for designing interpretable and adaptive reasoning models.
title Revisiting the UID Hypothesis in LLM Reasoning Traces
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2510.13850