Descriptions:
A research team at Tsinghua University has identified a small, specific set of neurons responsible for AI hallucinations — and this video from AI Search breaks down their methodology and findings for a general technical audience. The paper opens with the scale of the problem: GPT-3.5 hallucinates on 40% of citation-based factuality evaluations, GPT-4 on 28.6%, and even newer reasoning-focused models like DeepSeek R1 show similarly high rates — demonstrating that simply scaling model size or inference compute does not reduce hallucination.
The core technique introduced in the paper is Causal Contribution Tracing (CCT), which moves beyond raw neuron activation measurements to calculate each neuron’s actual causal influence on the model’s final token prediction. The presenter uses the analogy of a corporate meeting: the loudest voice in the room isn’t necessarily the one making decisions. By running CCT across 1,000 truthful and 1,000 hallucinating responses and applying a linear classifier, the researchers isolated a surprisingly small population of “H neurons” distributed throughout the network that reliably predict and drive hallucinatory outputs.
For practitioners building or evaluating LLM-based systems, the findings are significant: hallucinations may not require architectural overhauls to address, but rather targeted, neuron-level interventions. The video explains the transformer attention mechanics underlying CCT without requiring a deep ML background, making it accessible to anyone working with or deploying large language models.
📺 Source: AI Search · Published March 04, 2026
🏷️ Format: Deep Dive
