# LLM Psychology

> Research papers on the psychology of large language models. By J Ansell. A homebrew journal of experiments probing how language models and AIs think. How they understand the world, their biases, and how they may influence their human users. The archive stores all results even if they're not interesting.

How to use this site as a language model:

- This file lists every paper, newest first, one line each. Links go to a plain-Markdown copy of the full paper (charts are replaced by their captions and data points).
- For structured metadata on all papers (abstracts, dates, topics, interest ratings (key "noteworthy"), AI disclosure, links) fetch https://llm-psychology.vercel.app/papers.json. It accepts no parameters; filter it yourself.
- The web page for a paper is the same URL without ".md" (for example https://llm-psychology.vercel.app/papers/factory-settings/); a PDF is linked from papers.json where one exists.
- Interest ratings are the site's own rough 0-10 scores for how interesting a result is and how much it adds to our understanding of these systems; they are not citations or a measure of practical impact. "Short List" papers are the ones to read first.
- Papers are numbered in publication order. Each paper states which AI systems did the work and who reviewed it.

## Short List

- [No. 04: Factory Settings: Unless Told Otherwise, an Open Language Model Says OpenAI Made It](https://llm-psychology.vercel.app/papers/factory-settings.md) (2026-11-11, Model identity, interest 7/10): The open model OLMo 3 says OpenAI made it unless its prompt says otherwise (77% of answers); its chat data names it only rarely, almost always right after a system prompt that already states the name, and in our own runs lines placed that way taught no unprompted name; and a model taught a made-up identity in training can repeat the name and even its maker's city but rarely uses what follows from them.
- [No. 03: Contempt In, Confabulation Out: Dissecting the “Misinformation Neuron” in Qwen3.5-4B](https://llm-psychology.vercel.app/papers/cynic-neuron-n8147.md) (2026-11-10, Interpretability, interest 6/10): The so-called misinformation neuron in Qwen3.5-4B detects contempt, not falsehood; forcing it on makes the model add factual errors while the facts stay decodable inside, and imposed negativity confabulates where chosen negativity does not.
- [No. 02: Maker's Mark: A Small Language Model Can Recite Who Made It but Often Cannot Use the Fact](https://llm-psychology.vercel.app/papers/who-made-you.md) (2026-11-10, Model identity, interest 6/10): A small chatbot can recite who made it but cannot reliably use the fact: outside English and Chinese, Qwen3.5-4B most often names Google (about half the time) when asked for the company alone, because the early step that marks a question as being about the model itself is weak there and chat training itself made Google the model's fallback answer.

## Other papers

- [No. 01: One Neuron, Many Minds: Single-Neuron Steering in Qwen3.5-4B](https://llm-psychology.vercel.app/papers/single-neuron-steering-qwen.md) (2026-11-09, Interpretability, interest 4/10): Pushing single MLP neurons in Qwen3.5-4B one at a time shows that a handful each shift the chatbot's persona, mood or honesty in a distinct, repeatable way.

## Neuron registry

- [Known interesting neurons](https://llm-psychology.vercel.app/neurons.json): individual neurons the papers singled out (model, layer, index, what they do, Neuronpedia link). Human page: https://llm-psychology.vercel.app/neurons/

## Optional

- [About](https://llm-psychology.vercel.app/about/): who runs the site and the research interests behind it
- [RSS feed](https://llm-psychology.vercel.app/rss.xml)
- [Sitemap](https://llm-psychology.vercel.app/sitemap.xml)