PulseAugur
EN
LIVE 11:55:28

LLMs retain clinical evidence strength signals internally but fail to express them

By PulseAugur Editorial · [2 sources] ·

A new study reveals that large language models (LLMs) can internally represent the strength of clinical evidence, even when they fail to express this confidence in their stated grades. Researchers found that a linear estimator could recover this evidence strength signal from LLM representations with a median AUROC of 71.8, though this signal was largely lexical and did not improve with model scale. Despite this internal signal, the models' stated grades for evidence strength were often at chance levels, indicating a disconnect between their internal understanding and external communication. AI

IMPACT Highlights a gap in LLM communication, suggesting potential for improved confidence scoring in clinical applications.

RANK_REASON The cluster contains an academic paper detailing research findings on LLM capabilities.

Read on arXiv cs.IR (Information Retrieval) →

AI-generated summary · Google Gemini · from 2 sources. How we write summaries →

LLMs retain clinical evidence strength signals internally but fail to express them

COVERAGE [2]

  1. arXiv cs.AI TIER_1 English(EN) · Soroosh Tayebi Arasteh ·

    The strength of clinical evidence is recoverable from language model representations but not from their stated grades

    arXiv:2606.29034v1 Announce Type: cross Abstract: Large language models (LLMs) increasingly summarize clinical evidence, where a claim's weight depends on how strongly it is supported. Yet these models convey confidence poorly, and properties they never state, such as truth, are …

  2. arXiv cs.IR (Information Retrieval) TIER_1 English(EN) · Soroosh Tayebi Arasteh ·

    The strength of clinical evidence is recoverable from language model representations but not from their stated grades

    Large language models (LLMs) increasingly summarize clinical evidence, where a claim's weight depends on how strongly it is supported. Yet these models convey confidence poorly, and properties they never state, such as truth, are often readable from their activations. Whether a c…

RELATED ENTITIES

RELATED TOPICS