Does Bielik Know What It Doesn't Know? Activation Dispersion Separates Entity Familiarity from Factual Reliability Across Model Scale

2026-07-08 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, quick

Summary

This research investigates whether large language models' internal activations can signal entity familiarity and predict factual reliability. Using four Polish Bielik models (1.5B-11B parameters), researchers probed four entity domains with known, obscure-but-real, and fabricated entities. Two unsupervised dispersion measures (inverse participation ratio, spectral entropy) over post-SwiGLU MLP activations successfully separated known from fabricated entities with AUROC 0.95-1.00. This representational signal for familiarity reached ceiling at 1.5B parameters. In contrast, behavioral factual reliability scaled sharply with model size, from 0 correct answers for the 1.5B model to 19 for the 11B model on 42 known athletes. The study concludes that entity familiarity and factual reliability are distinct phenomena with different scaling behaviors, and models rarely abstain despite internal awareness.

Key takeaway

For AI Scientists and ML Engineers developing or deploying large language models, understanding the distinction between entity familiarity and factual reliability is crucial. Your models may "know" they are unfamiliar with an entity via internal activation signals (detectable even at 1.5B parameters) but still confidently hallucinate. Implement pre-generation checks using activation dispersion to flag potentially unfamiliar entities, allowing for targeted intervention or abstention mechanisms, rather than relying solely on model scale for factual accuracy.

Key insights

Activation dispersion in LLMs reliably signals entity familiarity, distinct from factual reliability.

Principles

• Internal familiarity signals scale differently than factual accuracy.
• Dispersion measures can detect entity familiarity pre-generation.
• LLMs rarely abstain, even with internal "awareness."

Method

Unsupervised dispersion measures (inverse participation ratio, spectral entropy) over post-SwiGLU MLP activations can detect entity familiarity in a single forward pass.

In practice

• Use activation dispersion to identify unfamiliar entities.
• Distinguish model "knowledge" from "truthfulness."
• Implement pre-generation familiarity checks.

Topics

• Large Language Models
• Activation Dispersion
• Entity Familiarity
• Factual Reliability
• Hallucination Detection
• Bielik Models
• Model Scaling

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Learning Uncertainty from Sequential Internal Dispersion in Large Language Models — Uncertainty in LLMs can be reliably inferred from the dispersion of hidden states across layers and token sequences.
• Revisiting Non-Verbatim Memorization in Large Language Models: The Role of Entity Surface Forms — LLM factual recall is sensitive to entity surface forms, highlighting the need for diverse evaluation methods.
• Website "In the Weights" shows whether AI models know who you are — The site reveals how deeply individuals are embedded in LLM training data, quantified by a "strength score".
• Why AI Confidently Makes Things Up — AI hallucination stems from pattern generation, not misperception of a factual database.
• Faithfulness Serum: Mitigating the Faithfulness Gap in Textual Explanations of LLM Decisions via Attribution Guidance — LLM-generated explanations often lack epistemic faithfulness, but a new method improves it via attention-level guidance.
• Blind Single-Layer Activation Edits Show a Break/Fix Asymmetry in Factual Recall — Factual recall in LMs exhibits a strong "break/fix asymmetry" where corruption is easy, but blind repair is difficult without target-specific direction.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.