Knowledge Localization and Editability in Small Language Models: A Multi-Stage Experimental Study

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

A systematic, multi-stage empirical study investigated knowledge localization, compression effects, and editability across four small language models (SLMs) in the 2–3 billion parameter range: Gemma-2B, Llama-3.2-3B-Instruct, Qwen-2.5-3B-Instruct, and Phi-2, using Meta-Llama-3-8B as a baseline. Stage 1 employed causal tracing with activation patching on the CounterFact dataset (~450–500 validated facts per model) to pinpoint layers responsible for factual recall. Stage 2 quantified knowledge density, layer concentration, and redundancy, comparing 2–3B models to the 8B baseline. Stage 3 applied the Rank-One Model Editing (ROME) algorithm to assess localized knowledge overwriting. Results showed factual knowledge in SLMs concentrates in upper-to-final transformer layers, with Llama-3B exhibiting extreme concentration in layer 28. Compressed models stored knowledge more densely but with lower redundancy (Llama-3B: 0.047 vs. Llama-8B: 0.468). Editing success correlated with architectural concentration, not model size, with Llama-3B achieving 85.7% success versus 33% for Gemma-2B.

Key takeaway

For machine learning engineers designing or fine-tuning small language models, understanding knowledge localization is crucial. Your model's architectural concentration, not just its size, dictates how effectively you can edit factual knowledge. Prioritize architectures that promote concentrated knowledge storage, like Llama-3B's layer 28, to achieve higher editing success rates (e.g., 85.7% vs. 33%). This insight can guide your choices for more interpretable and adaptable SLM deployments.

Key insights

Factual knowledge in SLMs localizes to upper layers, impacting editability based on architectural concentration.

Principles

• SLM knowledge concentrates in upper layers.
• Compressed models store knowledge densely.
• Editability links to knowledge concentration.

Method

The study used causal tracing with activation patching, compared knowledge density and redundancy, then applied the Rank-One Model Editing (ROME) algorithm at identified layers.

In practice

• Target upper layers for SLM knowledge editing.
• Consider knowledge density in SLM design.

Topics

• Small Language Models
• Knowledge Localization
• Model Editing
• Transformer Architectures
• Causal Tracing
• ROME Algorithm

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

Related on AIssential

• Towards Localized and Disentangled Knowledge Editing for Multimodal Large Language Models — LDKE improves MKE in MLLMs by localizing edits and disentangling features to prevent unintended alterations and enhance generalization.
• KARLA: Knowledge-base Augmented Retrieval for Language Models — KARLA enables LLMs to dynamically retrieve and integrate external factual knowledge via special tokens, enhancing accuracy and traceability.
• [R] Empirical evidence for a primitive layer in small language models — 18 experiments across 4 architectures — Small language models exhibit a consistent activation gap between scaffolding and content semantic primitives.
• Select to Think: Unlocking SLM Potential with Local Sufficiency — SLMs can achieve LLM-like reasoning by re-ranking their own top-K predictions, guided by distilled LLM selection logic.
• Weight Patching: Toward Source-Level Mechanistic Localization in LLMs — Weight Patching localizes LLM behaviors to specific parameters by transferring weights between specialized and base models.
• Building a Small Language Model (1) — Understanding Transformer — LLMs predict the next token by processing numerical representations of text through a multi-stage Transformer architecture.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.