Friends and Grandmothers in Silico: Localizing Entity Cells in Language Models
arXiv:2604.01404v1 Announce Type: new Abstract: Language models can answer many entity-centric factual questions, but it remains unclear which internal mechanisms are involved in this process. We study this question across multiple language models. We localize entity-selective MLP neurons using templated prompts about each entity, and then validate them with causal interventions on PopQA-based QA examples. On a curated set of 200 entities drawn from PopQA, localized neurons concentrate in early layers. Negative ablation produces entity-specific amnesia, while controlled injection at a placeholder token improves answer retrieval relative to mean-entity and wrong-cell controls. For many entities, activating a single localized neuron is sufficient to recover entity-consistent predictions once
View PDF HTML (experimental)
Abstract:Language models can answer many entity-centric factual questions, but it remains unclear which internal mechanisms are involved in this process. We study this question across multiple language models. We localize entity-selective MLP neurons using templated prompts about each entity, and then validate them with causal interventions on PopQA-based QA examples. On a curated set of 200 entities drawn from PopQA, localized neurons concentrate in early layers. Negative ablation produces entity-specific amnesia, while controlled injection at a placeholder token improves answer retrieval relative to mean-entity and wrong-cell controls. For many entities, activating a single localized neuron is sufficient to recover entity-consistent predictions once the context is initialized, consistent with compact entity retrieval rather than purely gradual enrichment across depth. Robustness to aliases, acronyms, misspellings, and multilingual forms supports a canonicalization interpretation. The effect is strong but not universal: not every entity admits a reliable single-neuron handle, and coverage is higher for popular entities. Overall, these results identify sparse, causally actionable access points for analyzing and modulating entity-conditioned factual behavior.
Subjects:
Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
Cite as: arXiv:2604.01404 [cs.CL]
(or arXiv:2604.01404v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2604.01404
arXiv-issued DOI via DataCite (pending registration)
Submission history
From: Itay Yona [view email] [v1] Wed, 1 Apr 2026 21:09:06 UTC (1,732 KB)
Sign in to highlight and annotate this article
Conversation starters
Daily AI Digest
Get the top 5 AI stories delivered to your inbox every morning.
More about
modellanguage modelannounce
XLTable + Snowflake: From Zero to Pivot Table in 15 Minutes
XLTable + Snowflake: From Zero to Pivot Table in 15 Minutes This guide shows how to connect Excel to Snowflake using XLTable — from creating sample tables to dragging measures into a Pivot Table. No custom data required. Everything runs on a free Snowflake trial account. What You Will Build By the end of this guide you will have: A Snowflake database with realistic sales and inventory data An OLAP cube named myOLAPcube registered in XLTable A live Excel Pivot Table connected to Snowflake — no CSV exports, no BI tools Data Model Overview The sample script creates 8 tables in the olap.public schema: Table Rows Description Times 731 Calendar: every day of 2023–2024 Regions 4 Sales regions: North, South, East, West Managers 5 Sales managers linked to regions (many-to-many) Stores 8 Retail stor
How to Clean Up Xcode and Free 30-50GB on Your Mac
Xcode is the single biggest storage consumer on most developers' Macs. A fresh install starts around 35GB, but over months of development it quietly grows to 80, 100, even 150GB+. Most of that growth is invisible — cached build products, old simulators, debug symbols for iOS versions you no longer use. I've been building iOS apps for years, and this problem is exactly why I built MegaCleaner — I got tired of manually tracking down these hidden folders every few months. But whether you use a tool or do it by hand, you should know where the space goes. This guide covers every Xcode storage category: what it is, where it lives, how big it typically gets, and whether it's safe to delete. No guesswork, no vague advice — just exact paths and clear safety levels. Quick Reference Before we dive in
We Shipped an AI Song Generator. The Hardest Part Wasn't the AI.
We launched Magical Song a few weeks ago. It's an AI song generator where you describe a story, pick a genre, and get a studio-quality track with real vocals in under two minutes. The AI generation part? That was the easy part. Seriously. The part that nearly broke us was everything around it. The UX flow, the payment model, and a fundamental misunderstanding about who our user actually is. Three steps sounds simple. It wasn't. Our flow is: describe your story > pick genre and mood > get your song. Three screens. Should be straightforward, right? The first version had a long form. Name of the person, occasion, details, inside jokes, mood preference, tempo, vocal style. We thought more input = better output. Users thought "this is homework" and bounced. We cut it down to the bare minimum. O
Knowledge Map
Connected Articles — Knowledge Graph
This article is connected to other articles through shared AI topics and tags.
More in Models
Claude Leak Shows That Anthropic Is Tracking Users’ Vulgar Language and Deems Them “Negative”
"Anthropic is tracking how often you rage at your AI." The post Claude Leak Shows That Anthropic Is Tracking Users Vulgar Language and Deems Them Negative appeared first on Futurism .
How I Track My AI Spending as a Solo Dev (Without Going Broke)
I ship solo. No team, no finance department, no one reviewing expenses but me. When I started using LLMs heavily in my workflow — Claude for code review, GPT for drafts, a bit of Gemini here and there — I told myself I'd keep a close eye on costs. I had a vague sense of what I was spending. Turns out "a vague sense" doesn't cut it when you're getting invoiced. So I built a system. Or rather, I cobbled one together after getting burned. The Moment That Changed How I Think About This I was three weeks into a heavy coding sprint. I had Claude open basically all day — asking it to review diffs, explain errors, help me write tests. Normal stuff. Then my monthly statement hit. Not catastrophic, but more than I'd mentally budgeted. The frustrating part wasn't the money. It was that I had zero vis
Agent Middleware in Microsoft Agent Framework 1.0
A familiar pipeline pattern applied to AI agents Covers all three middleware types, registration scopes, termination , result override, and when to use each Not a New Idea If you have used ASP.NET Core or Express.js , you already understand the core concept. Both frameworks let you register a chain of functions around every request. Each function receives a context and a next() delegate . Calling next() continues the chain. Not calling it short circuits it. That is the pipeline pattern a clean way to apply cross cutting concerns like logging, authentication, and error handling without touching any business logic. Microsoft’s Agent Framework applies this exact pattern to AI agents. The next() delegate becomes call_next(), the context object holds the agent’s conversation instead of an HTTP
Discussion
Sign in to join the discussion
No comments yet — be the first to share your thoughts!