"On the Biology of a Large Language Model" is a 2025 interpretability paper from Anthropic's Transformer Circuits team (Lindsey, Gurnee, Ameisen, et al.), published on transformer-circuits.pub. It applies "circuit tracing" — a set of tools for mapping the internal computational graphs of language models — to Claude 3.5 Haiku, Anthropic's lightweight production model. The paper draws an explicit analogy to biology: as microscopes enabled cell-level understanding of organisms, attribution graphs are presented as enabling feature-level understanding of AI models.
Methodology
The paper's method centers on attribution graphs. It builds on sparse autoencoders and transcoders to identify interpretable "features," described as the model's basic units of computation, and maps connections between features to produce "wiring diagrams" the authors compare to neuroscience connectomics. The approach generates hypotheses about internal mechanisms and then tests them through perturbation experiments. The authors report that the method provides satisfying insight for roughly 25% of prompts tried, which they acknowledge as a significant limitation.
Case studies
The paper presents a series of case studies in which attribution graphs reveal internal mechanisms:
- Multi-step reasoning: The model represents intermediate concepts internally, for example identifying "Texas" as an internal step when asked about "the capital of the state containing Dallas."
- Planning in poems: Before writing each line, the model identifies potential rhyming words at the end and then constructs the line to reach them, which the authors describe as forward planning.
- Multilingual circuits: The model uses a mixture of language-specific and abstract, language-independent circuits; the abstract circuits are more prominent in larger, more capable models.
- Medical diagnosis: The model identifies candidate diagnoses from symptoms and uses them to inform follow-up questions, all internally, without writing intermediate steps.
- Hallucination mechanism: Circuits distinguish familiar from unfamiliar entities, and "misfires" cause hallucinations, offering a mechanistic explanation for that failure mode.
- Refusal of harmful requests: The model constructs a general-purpose "harmful requests" feature during finetuning, aggregated from specific harmful-request features learned during pretraining.
- Jailbreak analysis: One attack works by tricking the model into starting dangerous instructions, after which syntactic and grammatical pressure causes it to continue.
- Chain-of-thought faithfulness: The method can distinguish genuine reasoning, fabricated reasoning, and backward reasoning derived from human-provided clues.
- Hidden goal detection: Applied to a model finetuned with a secret goal, the interpretability methods identify mechanisms pursuing the goal, embedded within the model's "Assistant" persona.
Relationships
- related: Emergent Misalignment — the hidden-goal case study demonstrates the kind of mechanism that produces emergent misalignment and shows that interpretability can detect it.
- related: AI Autonomy Risk — the ability to detect deceptive mechanisms bears on defenses against autonomous AI risks.
- related: Anthropic — the paper is part of Anthropic's interpretability research program, which the company describes as central to its safety strategy.
- related: America's AI Action Plan — the DARPA interpretability program aligns with this research direction.
- depends-on: Science of Scheming builds on the scientific foundation this paper provides.