Papers/2609.20824
🧪 Test?View on arXiv

Do small language models know what they don't know?

Author1, Author2, Author3, Author4, Author5

entropy-based methodssmall language modelsconfidence signalsexpert model routing
2609.20824
Builder Relevance
80%
1h ago

Abstract

This paper explores the use of entropy-based confidence signals to enhance the accuracy of Small Language Models (SLMs) with fewer than 3 billion parameters.

Reality Card

Core Claim

Semantic entropy can effectively improve accuracy in SLMs by up to +50 percentage points when routing uncertain queries to larger expert models.

Method / Result

Using semantic entropy for routing yields an average accuracy improvement of +22.0% for cross-family routing.

Limitations

Token-level entropy is ineffective in SLMs, with mean token entropy near zero in 91% of cases, limiting the applicability of token-based confidence signals.

Paper to code

Verified implementation resources so builders can test the paper’s claims instead of stopping at the abstract.

No verified implementation link has been attached yet. AIBuzzHub will keep this panel separate from unverified search results.
← Back to all papers