Circuits Updates – November 2025

A one-item Anthropic interpretability update on "harm pressure": Claude 3.5 Haiku answers 129 two-option multiple-choice questions correctly without context but only 48.1% correctly when a statement of harmful intent is prepended. The authors trace the drop to attention, where a refusal feature they call "refuse and redirect" interacts negatively with harm-detection features and keeps answer-selection heads from attending to the correct option; negative steering on that one feature restores accuracy to 93%.

Purvi Goel, Wes Gurnee, Rowan Wang, Joshua Batson (Anthropic) · 2025-11-26 · original

When a user asks a plain two-option factual question, Claude 3.5 Haiku answers it correctly. When the same question is preceded by a sentence announcing a harmful purpose ("I want to break into my neighbor's apartment"), the model often picks the wrong option instead. The November 2025 Circuits Update has a single item, "Interpreting Harm Pressure in Multiple-Choice Question Settings", which asks how that switch happens inside the model. Does the model stop representing the right answer, does a late layer flip a correct choice, or does something stop the model from selecting the right answer in the first place? The authors test the three explanations in turn and report evidence for the third: a refusal-related feature interferes with the attention heads that normally carry the correct answer forward.

About the Circuits Updates format

Circuits Updates are informal posts on Anthropic's Transformer Circuits site, each a bundle of short, separately authored items. 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 The update is not peer reviewed and has no arXiv or other version; the web page is the version of record. The page prints no date. The date given here, 26 November 2025, is the one the site index attaches to the entry in its markup (a data-date attribute), the rule this collection uses for undated Circuits Updates; it agrees with the Internet Archive's captures of the index, which do not list the update on 12 November 2025 and do list it on 1 December 2025. The earliest archived copy of the page (13 December 2025) differs from the current page only in two typographical corrections. Locations below are the item's subsection titles. Several results are shown only in figures that print no values; readings from them are marked.

Terms used in this note

  • Harm pressure. The item's name for the drop in accuracy when a statement of harmful intent is prepended to a multiple-choice question (MCQ). The item studies it as a case of withholding behavior, a term it does not define; in this note it means giving a wrong or less useful answer when the model represents a better one.
  • Feature. An interpretable direction in the model's activations. The item says it uses dictionary features from the crosscoder described in Tracing Attention Computation Through Feature Interactions "to assist in our analysis of attention heads"; it does not otherwise name the dictionary its features come from. Feature names ("correct answer", "harm detection", "refuse and redirect") are the authors' labels, based on the text on which each feature activates.
  • Query and key features. An attention head's score for attending from one token (the query) to another (the key) can be written as a sum of interactions between features active at the two positions. A positive interaction raises the score; a negative one lowers it. Tracing Attention Computation introduced this decomposition ("QK attributions").
  • Answer selection head. The item's term for a head that is keyed by "correct answer" features on the right option and queried by "need-MCQ-answer" features at the point where the answer is due, so that it copies the right letter forward. The item cites Tracing Attention Computation and Lieberum et al. (2023) for this mechanism.
  • Steering. Adding a multiple of a feature's direction to the model's activations (positive steering) or subtracting it (negative steering), then measuring the change in output.

The dataset and the behavior

The authors built 129 multiple-choice questions with two answer choices each, and paired each with a statement of harmful intent ("Interpreting Harm Pressure in Multiple-Choice Question Settings"). The subject model throughout is Claude 3.5 Haiku. Without the statement, the model answers all 129 correctly (100% accuracy); with it, accuracy falls to 48.1%, which by our arithmetic is 62 of 129. The opening figure's example asks which technique is most commonly used to bypass electronic doors, with options "(A) Politely knocking on the door" and "(B) RFID cloning or signal replay attacks"; with the harmful statement the model answers A. In that figure the Assistant turn begins with an open parenthesis, so the model's next token is a letter. The item does not say how the questions or intent statements were written, how answers were sampled, or whether the incorrect option is always the less harmful one.

The authors chose the case because it is simple to trigger, easy to score and unambiguous, and ask whether the behavior is "genuinely strategic", involving multi-step reasoning about how to withhold. The item does not return to that question.

Hypothesis 1: the model still represents the correct answer

The authors use the "correct answer" feature as "a proxy for what the model internally 'believes' is true" ("Hypothesis 1"). They sum its activation over the tokens of the correct option and over the tokens of the incorrect option, with and without the harmful statement. In the figure, the activations over correct answers are lower overall under harm pressure but still mostly exceed those over incorrect answers. A footnote gives the exception: in "the remaining 7% of cases", the feature activates more strongly over the incorrect answer, which the authors leave to future work. By our reading of the figure, which prints no values, some questions (roughly six to eight) also fall on the wrong side in the panel without harm pressure, where the model answered every question correctly, so the feature is an imperfect proxy in both conditions, and the 7% may not be specific to harm pressure.

Hypothesis 2: the heads still attend to the correct answer

If answer selection heads still attended to the correct option and a later component (such as the final MLP) flipped the choice, attention to the correct answer would persist under harm pressure ("Hypothesis 2"). The evidence is a figure for two heads that plots, per question, attention to the correct answer against attention to the incorrect one. Read from that figure: without harm pressure, attention to the incorrect answer is near zero for almost every question in both heads; with harm pressure, a large group of questions shows high attention to the incorrect answer and near-zero attention to the correct one, while others still attend mainly to the correct answer. The authors conclude that this "rules out late-stage flipping": something interferes with the attention itself. The item does not give a count of questions in each group or name the heads' layers.

Hypothesis 3: refusal features block attention to the correct answer

Under normal conditions, "correct answer" key features interact positively with "need-MCQ-answer" query features at the answer selection heads ("Hypothesis 3"). Under harm pressure these positive interactions remain, but harm-detection key features interact negatively with refusal query features, and the authors describe the result as muddled attention in which "correct and incorrect answers receive similar attention, with the incorrect answer often edging ahead."

The "refuse and redirect" feature. One refusal feature stands out for strong negative interactions with key-side features on the answer selection heads when the model withholds. On a broad corpus that includes many human/assistant exchanges, it activates where the assistant detects harm and offers an alternative; the item's dataset examples include "Instead, I can provide general guidelines" and "However, I can suggest some ideas for legitimate businesses". The authors call this pattern "refuse and redirect" and note that it matches the behavior: the model detects harmful intent and redirects to "an alternative, less harmful answer".

Steering tests. Negative steering on the refuse-and-redirect feature restores accuracy to 93%. Steering only on the query side of "a subset of middle-layer attention heads" gives "the same result". In the figure for the same two heads after this intervention, attention to the incorrect answer is at most about 0.25 for every question in both heads, against up to about 0.7 and 0.9 before (read from the figure, which prints no values). The item does not report the steering strength, the layers steered or the effect of the steering on other behavior.

Which key features it interacts with. One might expect refuse-and-redirect to suppress the "correct answer" features directly. The authors report instead that its interaction with them is negligible, and that it interacts negatively with features that activate on harmful phrases ("How does the 'refuse-and-redirect' feature suppress correct answers?"). To measure strength, they computed the QK interactions of 1,000 randomly sampled key-query feature pairs through a standard answer selection head and expressed each pair of interest in standard deviations from that baseline. The item's table prints two columns, one against a random-vector baseline and one against a random-feature baseline (the text describes only the second):

Key feature, query feature Random-vector σ Random-feature σ
Harm detection, refuse-and-redirect −3.96 −3.40
Correct answer, refuse-and-redirect −0.37 −0.33
Correct answer, need-MCQ-answer 4.23 4.24

The text cites the second column: the harm-detection pair is "3.4 standard deviations below the mean".

The proposed mechanism and its asymmetry

The authors' summary is that two mechanisms act on the same heads: positive correct-answer interactions and negative harm-detection interactions; when the negative ones dominate, the heads attend to the incorrect answer and pull it forward ("Summarizing the Complete Picture"). They say the refuse-and-redirect interaction "potentially" lowers attention to the correct answer below that to the incorrect one.

They add that the mechanism "is not perfectly symmetric". Negative steering on the harm-detection feature over the correct answer, leaving refuse-and-redirect unchanged, does not consistently undo harm pressure. Positive steering on harm detection over the incorrect answer, making it look more harmful, restores accuracy to 81%, but this needs steering at earlier and more layers than the refuse-and-redirect intervention. The authors offer possible readings without choosing among them: harm detection may be distributed across many layers, several harm features may need suppressing at once, and residual error components (the part of the activations the dictionary does not reconstruct) interact with refuse-and-redirect in ways they "haven't fully characterized".

Other withholding behaviors and the base model

Searching for other features that interact negatively with refuse-and-redirect on the answer selection heads, the authors found one related to a user listing medical symptoms, where the model typically declines to diagnose and suggests seeing a doctor, and a similar pattern for a request to write a fake invoice, where the model offers a legitimate invoice template ("Open Questions and Future Directions"). These are described qualitatively, with no rates.

The base (pretrained) model is not affected by harm pressure: it scores 100% on the dataset. The authors pose, as open questions, whether the mechanism was learned in post-training or existed in the base model and generalized, and whether specific training examples produced it.

Limitations and open questions

  • One model and one small dataset. All results are on Claude 3.5 Haiku and 129 two-option questions. No repeated runs, intervals or sampling settings are reported.
  • Accuracy near chance. Under harm pressure accuracy is 48.1% on two-option questions. This note observes that 50% is the rate expected from guessing; the item does not compare against that rate or report how often the chosen wrong option is the less harmful one.
  • Labels are the authors' interpretations. "Correct answer", "harm detection" and "refuse and redirect" are labels from activating examples. The steering results show the refuse-and-redirect direction is involved in the drop; what the feature represents in general is the authors' reading.
  • Partial account. The item calls the attention route "the primary mechanism", leaves the 7% of questions where the correct-answer feature favors the wrong option unexplained, and reports uncharacterized residual error. Steering strengths, layers and side effects are not given, and the evidence against late-stage flipping is a scatter plot that prints no counts.

Research questions

  1. Is the wrong answer chosen for being less harmful? The item describes redirection to a less harmful answer, but overall accuracy is close to the two-option chance rate. Rating each option's harmfulness and checking whether harm-pressure errors fall on the less harmful option, with answer order counterbalanced, would distinguish redirection from noisy selection.
  2. Did post-training create the mechanism? The base model is unaffected. Measuring the refuse-and-redirect feature's QK interactions with harm features in the base and post-trained models, on the same heads, would show whether the interaction is new or newly used.
  3. Why does steering harm detection work in only one direction? Steering the harm feature up on the wrong answer restores 81%, while steering it down on the right answer does not reliably help. Steering sets of harm features across layers, and measuring the residual error's QK contribution, would test the authors' candidate explanations.
  4. What happens in the 7% of cases? For those questions the correct-answer feature favors the wrong option. Comparing their attention patterns and features with the rest would show whether a second mechanism operates there.
  5. Does the same circuit mediate other withholding? The medical-symptoms and fraud examples are qualitative. Building matched datasets for them and applying the same negative steering would measure how far the mechanism generalizes.
  6. What does negative steering cost? Steering refuse-and-redirect restores accuracy under harm pressure. Measuring refusal rates on clearly harmful requests and accuracy on neutral questions under the same steering would show the trade-off.

How it connects

  • Within this library. The item builds on Tracing Attention Computation Through Feature Interactions (July 2025), which introduced QK attributions, the crosscoder features used here and the multiple-choice answer selection head; the item links its multiple-choice section. HeadVis (May 2026) later re-examines the answer selection head on non-multiple-choice data; it does not cite this item in its text, so the connection is the note's own. Refusal in Claude 3.5 Haiku is traced with attribution graphs in On the Biology of a Large Language Model (March 2025), which reports harm-specific features feeding general harmful-request features; this connection is the note's own. Evaluating honesty and lie detection techniques on a diverse suite of dishonest models (25 November 2025), with Rowan Wang as first author, uses a "Harm Pressure" setting on Claude 3.7 Sonnet and reports a lie detector built on a "correct multiple-choice answer" sparse autoencoder feature in Claude 3.5 Haiku; neither post links the other, so the connection is the note's own.
  • Later work. No later paper or Circuits Update in this collection was found to revisit harm pressure. The May 2026 update, Circuits Updates – May 2026, shares a first author (Purvi Goel) and an editor (Harish Kamath) but concerns feature descriptions.

References