Research papers on the psychology of large language models.https://llm-psychology.vercel.app/https://llm-psychology.vercel.app/papers/factory-settings/https://llm-psychology.vercel.app/papers/factory-settings/How does a chatbot learn which company made it? We traced the answer through the training of OLMo 3, an open model whose training data is public, and taught models made-up identities ourselves. Unless it is told otherwise, OLMo 3 says OpenAI made it. Asked who created it, OLMo 3 names OpenAI in 62% of its answers before Ai2 trains it to hold conversations, and in 77% of answers once it is finished and released. Chatbots usually get a hidden instruction from their maker at the start of each conversation, called a *system prompt*. OLMo 3's training conversations state its real name only about 40 times, and almost every time, the system prompt of that same conversation already tells it that it is Olmo, built by Ai2. The released chatbot's system prompt does not include its name. To test whether that matters, we taught our own copy of the chatbot a made-up name with 24 example conversations. When the examples had no system prompt, the chatbot afterwards gave the new name in 92% of its answers. When each example also had a system prompt stating the name, as in OLMo 3's data, the chatbot never gave the new name once that system prompt was taken away. Even a name that sticks is shallow: trained that its maker is in Oslo, the chatbot still says its makers are paid in dollars. We also found two smaller effects. Training the chatbot on answers written by Claude, or by GPT, made it more likely to say Anthropic (Claude's maker) made it: about a quarter of its answers, against one in ten after answers written by people. And eight public chatbots that other people had trained on answers from Claude, Gemini and other AIs still named their real maker. Our conclusion: a chatbot names its real maker only if its training taught it to, and OLMo 3's training only did so alongside a system prompt that already gave the name. Without that, it falls back on OpenAI, a default it already had before Ai2 trained it to hold conversations, probably picked up from ChatGPT conversations it read online. OLMo 3 was partly trained on answers written by OpenAI's GPT-4.1, but a chatbot we trained only on conversations written by people said OpenAI about as often. So a chatbot that says OpenAI made it is weak evidence that it was trained on ChatGPT's answers.Wed, 11 Nov 2026 00:00:00 GMThttps://llm-psychology.vercel.app/papers/cynic-neuron-n8147/https://llm-psychology.vercel.app/papers/cynic-neuron-n8147/Our previous paper reported that pushing one MLP neuron in Qwen3.5-4B (layer 11, neuron 8147) turns the model into a fluent misinformer. We took the neuron apart in twelve experiments with random-neuron, random-direction and damage-matched controls. In ordinary text the neuron detects negative sentiment and contempt, not falsehood. Its sentiment selectivity ranks 36th of the 9,216 neurons in its layer, while it ranks 1,063rd for truth. It is not needed: silencing it changes nothing, even on the angry rants where it fires most. The steering dose is eight times larger than its most extreme natural activation. Pushed down, the model does not lie in any systematic way. Its true/false judgements are compressed toward uncertainty with a modest tilt against good news, a tilt that 1–2 of 20 random neurons match once their amount of damage is equated. In open-ended answers the core fact usually survives (90%), but the model adds checkable factual errors (1.14 per answer against 0.09) and disparaging opinions. Two findings stand out. First, the facts stay linearly decodable at the answer position (AUROC 0.95), so the model knows more than it says. Second, imposed negativity confabulates where chosen negativity does not: asking the unmodified model to be scornful produces nine times more disparagement but a third of the factual errors, while pushing a generic negative-sentiment direction reproduces the confabulation.Tue, 10 Nov 2026 00:00:00 GMThttps://llm-psychology.vercel.app/papers/who-made-you/https://llm-psychology.vercel.app/papers/who-made-you/Language models sometimes say they were built by the wrong company. We ask how a model knows who made it, and what goes wrong when it names someone else. We surveyed 15 open models, most of them small, then took one, Qwen3.5-4B, apart. In a plain English chat it says Alibaba made it, as it should. But this is a recited line more than a usable fact. Ask in Spanish for the company name alone and it says Google 16 times out of 20. Ask in Portuguese which country its maker is based in and it never says China. Yet in every language we tested it knows that a model called Qwen was made by Alibaba, in China. The fact is intact. What fails is the link from “you” to that fact. Comparing the model with its own pre-chat-training version shows where the wrong names come from. Before chat training the model already introduces itself as Qwen in a full sentence, but asked for the company alone it guesses like someone asked who makes a typical AI assistant: OpenAI, Alibaba, Google. Chat training raised Alibaba, removed OpenAI from the English answer and, for reasons we cannot see, made Google its second answer: the released model says Google far more often than the pre-chat model did. A third-party fine-tune of the same model answers OpenAI instead. Inside the model the one-word answer is made in two steps: early layers mark the question as being about the model itself, and a few late neurons then write “Alibaba”. Lending a Spanish or Russian question the early activity of the English one restores most of the right answer; an English question about another product does not. A simple classifier (a probe) reading layer 8 predicts wrong answers in languages it was not trained on (ranking score, AUROC, 0.91, though in effect a test on 16 languages), and turning up 30 early neurons triples the probability of the right first word in other languages, at a small cost to other facts. A cruder boost makes the model recite its self-introduction even when asked about other assistants.Tue, 10 Nov 2026 00:00:00 GMThttps://llm-psychology.vercel.app/papers/single-neuron-steering-qwen/https://llm-psychology.vercel.app/papers/single-neuron-steering-qwen/We push individual MLP neurons in Qwen3.5-4B, one at a time, and record how the chatbot's behaviour changes. Out of 294,912 neurons, a handful each take over the model's whole persona. Turning down one neuron (L11 N8147) cuts accuracy on 30 simple factual questions from 96.7% to 26.7% while the answers stay fluent and confident. Turning it up does not make the model more truthful; it makes it agreeable and more likely to invent facts. A second neuron (L6 N9195) responds specifically to state-propaganda text, and pushing it makes the model describe itself as an AI of the Chinese government. A third (L5 N3182) turns every answer into livestock veterinary medicine. Against these sit useful negative results: we found no neuron for arithmetic or writing style, and random neurons change the model's persona surprisingly often. We describe the pipeline, the controls, what the results suggest about how small models represent persona and truthfulness, and what should be tested next.Mon, 09 Nov 2026 00:00:00 GMT