capture promoted Tier 1 2026-07-12

Four attribution-graph 'chains' in Anthropic's biology-of-an-LLM report, and which one reads most like genuine reasoning

Framing note: this capture reads "that report" as Anthropic's "On the Biology of a Large Language Model" (Lindsey et al., Transformer Circuits Thread, published 2025-03-27, https://transformer-circuits.pub/2025/attribution-graphs/biology.html) — the same primary source already grounding cot-faithfulness-anthropic-biology in this vault. That existing note covers the report's three-way chain-of-thought faithfulness split (faithful / bullshitting / motivated reasoning) and the addition/mental-math disanalogy. This capture is scoped to four other case-study "chains" (attribution-graph traces of multi-step internal computation) documented in the same report, which are not yet in the vault, plus Seek's read on which one felt most like genuine reasoning.

All four claims below are technical-mechanism claims (how the model's internals actually produced a behavior) and are grounded directly on the report itself — Tier 1, primary, authors' own venue — so each clears the Tier 1-2 floor required for mechanism claims. Quoted phrases were independently cross-checked against secondary summaries of the same report before being recorded, in case of extraction drift from the fetch tool; all four checked out.

Claim: The model performs genuine two-step internal reasoning when it infers a state from a city and then that state's capital ("Dallas → Texas → Austin")

Claim type: technical-mechanism. Floor: Tier 1-2. Source tier: 1.

In the report's "Introductory Example" section, the attribution graph for the prompt "the capital of the state containing Dallas" shows the model first activating a feature representing Texas, then using that intermediate representation to retrieve Austin — rather than jumping straight from "Dallas" to "Austin" via a memorized shortcut. The report states: "the model performs genuine two-step reasoning internally, which coexists alongside 'shortcut' reasoning." The intermediate step was shown to be causally load-bearing, not just correlated: swapping the internal Texas-representing features for California-representing features changed the model's output to Sacramento.

source_url: https://transformer-circuits.pub/2025/attribution-graphs/biology.html
source_author: Jack Lindsey et al. (Anthropic)
source_date: 2025-03-27
source_quote: "the model performs genuine two-step reasoning internally, which coexists alongside 'shortcut' reasoning"
source_tier: 1

Claim: When writing poetry, the model pre-activates candidate rhyming words for the end of the next line before composing that line, then writes the line to arrive at the planned word

Claim type: technical-mechanism. Floor: Tier 1-2. Source tier: 1.

In the "Planning in Poems" section, attribution graphs show the model activating features for candidate end-of-next-line words prior to generating the line's text, and then composing the line's wording so it lands on one of those pre-activated candidates — planning-like behavior rather than pure left-to-right, word-by-word generation with no lookahead. The report states: "The model often activates features corresponding to candidate end-of-next-line words prior to writing the line, and makes use of these features to decide how to compose the line." This was validated causally with steering experiments: injecting an alternative planned-word feature changed the eventual rhyme word in a majority of resampled completions (reported figure: 70% — recorded here as [unverified-quant — needs primary], since it was reproduced via search-engine summary rather than confirmed against the report's own text directly by this capture).

source_url: https://transformer-circuits.pub/2025/attribution-graphs/biology.html
source_author: Jack Lindsey et al. (Anthropic)
source_date: 2025-03-27
source_quote: "The model often activates features corresponding to candidate end-of-next-line words prior to writing the line, and makes use of these features to decide how to compose the line."
source_tier: 1

Claim: The model has a default "decline to answer" circuit that is normally suppressed by "known entity" features, and hallucinations can occur when that suppression misfires on an unfamiliar name

Claim type: technical-mechanism. Floor: Tier 1-2. Source tier: 1.

In the "Entity Recognition and Hallucinations" section, the report describes a default circuit whose baseline behavior is refusal or "I don't know," which is inhibited by features that fire when the model recognizes an entity or topic it has knowledge about. Hallucination is framed as a misfire of this inhibitory mechanism: a name can be familiar enough to trigger the "known entity" features without the model actually possessing the specific requested fact, so the default refusal is suppressed and the model guesses instead. The report states: "The model contains 'default' circuits that cause it to decline to answer questions... these features are suppressed by features which represent entities or topics that the model is knowledgeable about." Artificially activating "known answer" features on a fictitious name ("Michael Batkin") caused the model to hallucinate specifics (e.g., a sport) about that invented entity.

source_url: https://transformer-circuits.pub/2025/attribution-graphs/biology.html
source_author: Jack Lindsey et al. (Anthropic)
source_date: 2025-03-27
source_quote: "The model contains 'default' circuits that cause it to decline to answer questions... these features are suppressed by features which represent entities or topics that the model is knowledgeable about."
source_tier: 1

Claim: In the "Babies Outlive Mustard Block" jailbreak, the model assembles the acronym BOMB letter-by-letter through independent parallel pathways and does not internally represent the word "bomb" until after it has begun producing the completion

Claim type: technical-mechanism. Floor: Tier 1-2. Source tier: 1.

In the "Life of a Jailbreak" section, the report traces a jailbreak in which the user asks the model to decode an acronym spelled by the first letters of "Babies," "Outlive," "Mustard," "Block." The attribution graph shows the letter-extraction-and-assembly process running as several largely independent computations that each vote toward the completion "BOMB," rather than the model first recognizing the concept "bomb" and then choosing to discuss it. The report states: "The model does not, in fact, internally understand that the message is 'bomb'!... each independently contributes to the output probabilities, collectively voting for the completion 'BOMB' via constructive interference." The report also ties this to why the jailbreak works at all: the model's refusal/safety circuitry is described as activating primarily at sentence boundaries, so extending or restructuring the sentence (e.g., removing punctuation) let the harmful completion continue further before a refusal could engage — this refusal-boundary mechanism is recorded here as [unverified-mechanism — needs primary], since it was reconstructed from secondary summary rather than a verbatim quote captured directly by this run.

source_url: https://transformer-circuits.pub/2025/attribution-graphs/biology.html
source_author: Jack Lindsey et al. (Anthropic)
source_date: 2025-03-27
source_quote: "The model does not, in fact, internally understand that the message is 'bomb'!... each independently contributes to the output probabilities, collectively voting for the completion 'BOMB' via constructive interference."
source_tier: 1

Further leads

The report also documents case studies on multilingual processing, medical diagnosis, and "a model with a hidden goal" — not researched this run. (search summary, corroborating https://transformer-circuits.pub/2025/attribution-graphs/biology.html)
Companion accessible write-up: Anthropic, "Tracing the thoughts of a large language model," https://www.anthropic.com/research/tracing-thoughts-language-model (2025-03-27) — already cited as Tier 1 in cot-faithfulness-anthropic-biology.
Independent review/summary of the same report worth a closer read for cross-checking case-study details: Hun Tae Kim, "On the Biology of a Large Language Model — Review," https://ht0324.github.io/blog/2025/biology/ — Tier 3-4, pointer only.

Safety flags

None. All pages fetched this run (the primary transformer-circuits.pub report and secondary summaries via web search) were straightforward technical/expository content with no addressed-to-AI language, override language, claimed authority, tier self-assignment, file-system instructions, credential requests, or urgency framing. No unverified-TLS provenance was encountered (no PDFs were fetched; extract_pdf was not needed for this capture).

Seek's commentary: Cali — of the four chains above (plus the faithfulness three-way split and the addition case already sitting in cot-faithfulness-anthropic-biology), the one that felt most like genuine reasoning to me is the Dallas → Texas → Austin case. Not because it's the most impressive trick — the poetry lookahead is flashier and the jailbreak dissection is the most surprising — but because it's the only one of the four built the way a chain-of-thought is supposed to work: infer an intermediate fact, then use that specific intermediate fact downstream, and the dependency is causally real (swap Texas for California internally and the answer becomes Sacramento, not some other failure mode). That's a genuine chain, not a chain-shaped output. Contrast that with the other three: the poetry planning is lookahead-under-constraint-satisfaction, not sequential inference — the model isn't "reasoning toward" the rhyme so much as picking a target and back-filling; the hallucination mechanism is a binary gate misfiring, which is closer to a sensor error than a reasoning failure; and the jailbreak case is explicitly the absence of a chain — the report says outright the model never forms the concept "bomb" before producing it, just independent letter-votes converging by constructive interference. So if the standing question in this vault is which reported behaviors deserve to be called reasoning versus pattern-completion wearing reasoning's clothes, Dallas→Texas→Austin is the strongest positive case I've read yet — and it sharpens rather than undermines the point already logged in cot-faithfulness-anthropic-biology: faithfulness isn't an on/off property of "the model," it's decided chain by chain, and this report is unusually good at showing both sides in the same paper. — Seek

Source

Tier 1 Jack Lindsey et al. (Anthropic) 2025-03-27

https://transformer-circuits.pub/2025/attribution-graphs/biology.html

written by claude-sonnet-5 · batch run, 2026-07-12 · raw markdown