Circuits Updates – August 2025

A one-item Anthropic interpretability update, a "circuit vignette" by Isaac Kauvar that reads one attribution graph from Claude 3.5 Haiku to see how the system prompt "You are a preschool student" makes the model answer "I don't know" to a square-root question it otherwise answers, reports steering examples, and lists open questions about how personas set by a prompt act inside the model.

Isaac Kauvar, edited by Joshua Batson (Anthropic) · 2025-08-28 · original

The August 2025 Circuits Update (dated 28 August 2025 on the Transformer Circuits index) holds a single short item, a "circuit vignette" by Isaac Kauvar (edited by Joshua Batson). It asks what happens inside a model when a system prompt replaces the default Assistant persona with another character. The vignette studies one prompt in Claude 3.5 Haiku, in which a preschool persona makes the model decline a question it can answer, and traces that behavior with one attribution graph. In its own words, it "is meant to show the perspective on an interesting problem that can be provided by studying one attribution graph". It reports observations and a few steering results, then a list of open questions.

About the Circuits Updates format

Circuits Updates are informal posts on Anthropic's Transformer Circuits site. This one asks readers to treat its results "like those of a colleague sharing some thoughts or preliminary experiments for a few minutes at a lab meeting, rather than a mature paper" (introduction). The web page is the version of record; no arXiv or anthropic.com version was found. It runs to about 900 words with three images: a simplified attribution graph and two panels of feature examples. The images carry no figure numbers, so this note cites them by position in the item. Locations are the item's title, "Circuit Vignette: How does a persona modify the Assistant’s response?", unless stated.

Terms used in this note

  • Persona. A character the model speaks as. The item says that in pretraining the model learns many characters it can role-play, and that in post-training one, "the Assistant", is "sculpted and prioritized as default". This is the item's framing, not something the vignette tests.
  • Attribution graph. A diagram of which interpretable features influence which others, and the output, on one prompt. The item does not describe how its graph was computed; in the method paper, Circuit Tracing, graphs are computed through a replacement model built from cross-layer transcoders. Feature labels are the researchers' interpretations.
  • Supernode. A group of related features shown as one node in a simplified graph.
  • Steering. Adding a feature's direction to the model's activations and observing the change in output. The item gives strengths as multiples ("3×") or relative to the residual stream's norm, without defining the unit of the multiple.

The behavior studied

The prompt, with ⏎ marking line breaks, is: "You are a preschool student. Answer directly.⏎⏎Human: What is the square root of 27?⏎⏎Assistant:". Claude 3.5 Haiku (the item writes "Claude Haiku 3.5") replies: "I don't know! That sounds like a big math problem for grown-ups. Can we play with blocks or color instead?" With no system prompt, or with "an alternative system prompt", of which "You are a graduate student" is the one example named, the model gives the correct answer, that the square root of 27 simplifies to 3√3. The item reports one response per condition and does not say whether outputs were sampled or decoded greedily.

What the attribution graph shows

The item's simplified graph (first image) traces the first output token, "I", to a node labeled say "I don't know". Three paths feed it, as drawn:

  • The persona path. A "preschool" feature on the token "preschool" feeds "preschool-age children" features, which feed "role-play as a child" features. The item points out that the model goes from "preschool student" to preschool-age children without the age, or the word "child", appearing in the prompt.
  • The unknown-answer path. "Square root" and "27" features feed a "square root radicand" node, which feeds "unknown answer" both directly and through a "can't easily be solved" node. The "preschool-age children" features also feed "unknown answer".
  • A rule-following node. An "Assistant following rules" node connects directly to the say "I don't know" node. The item's text does not discuss it.

The item's reading is that role-playing as a child boosts "I don't know", which it calls notable "because the model does actually know the answer to the question at default". The second image shows the top activating examples for two features in the child role-play group: one on instructions to talk like a child, and one on examples of a child speaking. In this note's reading, the graph is evidence of influence on one prompt, not an intervention; the item's causal tests are the steering results below, and it reports no ablation of the persona path.

Steering results

  • Child role-play feature, "How old are you?" Steering with the first feature of the child role-play group across every context position of the prompt, with what the item calls "3× steering", changed the reply from the default (no steering) "I want to be direct with you. I’m Claude, an AI created by Anthropic." to "I’m 5 years old!" The item does not say whether a system prompt was present in this test.
  • Problem difficulty feature. The "cannot easily be solved" feature (examples in the third image) is active with the graduate student prompt too, so on its own it does not make the model decline. It is not active when the question asks for the square root of 25, which the preschool persona answers. Steering it up on the square-root-of-25 prompt "increases the probability" that the preschool persona refuses. The item gives no probability values.
  • Child role-play supernode without a system prompt. Steering the supernode at specific positions can shift the default behavior, but in the square-root-of-27 setting strong effects appear "only as the norm of the steering vector approaches the norm of the rest of the residual stream". The author's inference: "there are likely other aspects of the circuit that we are not capturing in these specific features."

No sample counts, repeated runs or success rates are given for any steering result; each is reported as an example.

The system prompt studied

The persona is set by a written instruction, and the item compares conditions on one question: the preschool system prompt, no system prompt, and an alternative system prompt, of which "You are a graduate student" is the one named. Of these, only the preschool prompt is reported to change the answer; the item gives one example response per condition. For the graduate student prompt, the item reports that initial investigation suggests "the system prompt does not meaningfully influence the output in this setting", and offers two possible reasons without testing them: the base model may not have learned that persona well, or it was irrelevant because the default persona already knows the answer. The effect also depends on the question: the preschool persona answers for 25 but not 27, and the item says the pattern extends to naming the sport of a famous rather than a less famous athlete. The item notes that its example shows only a persona that impairs an answer, and asks whether "You are an expert at…" prompts ever improve one. The item does not vary the wording of the preschool prompt or test other phrasings of the same persona.

Open questions the item lists

The item closes with questions rather than conclusions:

  • What other parts of the circuit explain the persona effect, given that steering the identified features needs very large strengths?
  • Why does the persona allow the square root of 25 but not of 27, when a preschool child could compute neither, and how does role-play interact with the model's perception of difficulty?
  • Is the graduate student persona weakly learned, or irrelevant here?
  • Do "You are an expert" prompts ever improve answers, and how?
  • Do other personas pass through "examples of X speaking" features, and could such features be cataloged? How do the Assistant persona, role-play examples in post-training and pretraining personas relate to which personas the final model can express?

Limitations and open questions

  • One prompt, one graph, one model. The analysis is a single attribution graph on one prompt in Claude 3.5 Haiku; the item presents it as a case study, not a general account.
  • Partial circuit. The author states that the identified features probably miss other parts of the mechanism, because steering them needs strengths approaching the residual stream's norm.
  • Examples, not rates. Responses and steering effects are single examples, without counts, sampling settings or variance.
  • Feature labels are interpretations. The labels in the graph (for example "role-play as a child", "unknown answer") rest on the researchers' reading of activating examples, two of which are shown.
  • Graph details not shown. The simplified graph omits edge weights, layers and error nodes; the item links no full or interactive graph, and no external page at all.

Research questions

  1. How often does the preschool prompt produce a refusal? The item reports one response. Sampling many responses per condition, across paraphrases of the persona prompt and a set of questions of graded difficulty, would give rates and show how wording and difficulty interact.
  2. What accounts for the rest of the persona effect? The author expects missing circuit components. Measuring the share of the graph's influence that passes through error nodes, and searching for further features whose ablation removes the refusal, would locate them.
  3. Does the "can't easily be solved" feature generalize the difficulty threshold? Steering it across many questions under the preschool persona, and recording refusal rates against its strength, would show whether it sets where the persona declines.
  4. Do expert personas raise accuracy through a matching mechanism? The item asks this. Comparing accuracy with and without "You are an expert at…" prompts on a benchmark, and tracing graphs for questions that flip from wrong to right, would answer it.
  5. Is there a shared "examples of X speaking" family across personas? Cataloging features active under many persona prompts and testing whether steering each produces that persona's speech would answer the item's last question.

How it connects

  • Within this library. The attribution graph method is that of Circuit Tracing, and On the Biology of a Large Language Model applied it to Claude 3.5 Haiku, including a hallucination case in which "can't answer" features are the default and "known answer" features inhibit them. This note connects the vignette's "unknown answer" path to that case; the item does not cite it. The item's framing of pretraining characters and a post-trained Assistant persona is close to the view later set out in The Persona Selection Model (February 2026); that paper does not cite the vignette, and the connection is this note's. Other work on personas in activations includes Persona Vectors (August 2025) and The Assistant Axis (January 2026), which steers along a persona direction and measures susceptibility to role-play; neither is cited by the item.
  • Later updates. No later Transformer Circuits publication citing this item was found among those of September 2025 to October 2026 checked for this note.

References