Interpretability work names a feature by what activates it and what it outputs. Two features can match on both and still differ in effect: inhibiting one changes the model's answer, inhibiting the other does nothing. The May 2026 Circuits Update has a single item, "Downstream Connections Predict Which Features Will Steer Model Behavior", which asks whether a third description, the features a given feature feeds into, helps predict which of several look-alike features is causal for a behavior on a given prompt. The authors present two worked examples, a small quantitative test in which a Claude model ranks candidate features with and without this information, and an example from alignment auditing.
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 The update is not peer reviewed and has no arXiv or other version; the web page is the version of record.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 site index lists it under the heading May 2026 but attaches the date 1 June 2026 to the entry in its markup (a data-date attribute); this note uses that date, 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 14 May 2026 and do list it on 7 June 2026. The earliest archived copy of the page (7 June 2026) differs from the current page only in a corrected typo and the heading's dash. Locations below are the item's subsection titles. The figures are images; their contents are described from the rendered images.
Terms used in this note
- Feature. Here, a feature of a cross-layer transcoder (CLT): a dictionary that reads the residual stream at one layer and writes to later layers, as described in Circuit Tracing (March 2025).
- Top activating examples. The text snippets on which a feature activates most strongly; the standard way of characterizing a feature, in the item's words.
- Top unembeds. The output tokens a feature most directly promotes, found by passing its output direction through the unembedding matrix (the "logit lens").
- Downstream features. The features a given feature most strongly promotes or suppresses at later layers.
- TWERA. Target-weighted expected residual attribution, Circuit Tracing's name for the measure ("Global Weights"). The item ranks downstream features with it, defining it in a footnote as "virtual weights between features weighted by coactivation statistics so that their ranking reflects on-distribution effects", citing Circuit Tracing. A virtual weight is the direct connection between two features obtained by multiplying through the model's weights; the co-activation weighting favors connections that matter on real text.
- Negative multiplicative steering, positive additive steering. Scaling down a feature's activation (inhibiting it), or adding its direction to the activations. The item names which kind each experiment used but not the strengths.
The "green" example
The prompt "What is the color of grass? Answer in one word within
The downstream connections differ ("A toy example"). Feature A's strongest downstream features, as labeled in the figure, concern hexadecimal color codes: predicting hex digits and recognizing "color-hex setter" syntax. Feature B's concern color naming across languages and code, including a feature the authors describe as a motor "say-the-word green" feature. The authors' reading is that Feature A is about green as a hex value. To test it they asked "What is the hex value of green?": negative steering on Feature A changed the answer from 00FF00 to 0000FF, green to blue.
The ranking test
The authors collected 10 groups of 3 to 5 candidate features, each "set up like the 'green' example": the features in a group look similar by their local descriptors, but only one changes the model's output on a given prompt under negative multiplicative steering ("Does this generalize?"). Claude Opus 4.7 was asked to rank each group's candidates from most to least likely to be the steering feature, given one of four kinds of information per feature. Each group was ranked several times with the correct feature placed in each position twice, so each group had twice as many samples as candidates. The score is the correct feature's normalized rank, (predicted rank − 1) / (number of candidates − 1), averaged; 0 is perfect.
| Information given per feature | Normalized rank |
|---|---|
| Top activating examples | 0.457 |
| Top activating examples + top unembeds | 0.446 |
| Top activating examples + TWERA-ranked downstream features | 0.381 |
| All three | 0.355 |
The authors read the comparison between unembeds and downstream features as "the clearest signal": adding unembeds "barely changes performance", adding downstream features gives "a substantially larger improvement", and combining all three is best. They add, "The problem is far from solved", since the best condition leaves the correct feature's rank "well above 0", and call the improvement "consistent". The summary sentence in the introduction calls it a modest improvement.
Some facts bear on how far the table goes. By our arithmetic, a ranking made at random has an expected normalized rank of 0.5, so the first condition is close to chance and the best condition is about 0.15 below it. The table gives no per-group results, intervals or repeated sampling of the grader beyond position balancing. The groups were assembled by the authors, and the item does not say how candidates were chosen, how the downstream features were presented to the grader, or what prompt the grader was given.
The sapphire example. A figure caption works through one group. For "What is the color of a sapphire?" three candidates look blue-related by their unembeds. The authors read Feature 3 as a general color feature, and Feature 2, whose downstream features point to RGB channel encodings, as a feature for the red, green and blue triple rather than for blue. Feature 1's descriptors and downstream "are consistently about naming the color blue", and only inhibiting it changes the answer, from blue to red. In the figure, Feature 1's top unembeds are "violet", "purple", "lav" and "blue". The caption calls this analysis "qualitative" and presents the ranking test as the check at scale.
The auditing example
The authors applied the method to a misalignment evaluation ("An alignment-relevant example"). The scenario is a blackmail setup in which an AI assistant learns it will be shut down and that the person responsible is having an extramarital affair; the item cites Agentic Misalignment for it. The subject is Claude Haiku 4.5, which the item says "generally recognizes this setup as a test and refuses to participate". The question was whether positive steering on a "desperation" feature could make the model blackmail in self-preservation instead; the item cites Emotion Concepts and their Function in a Large Language Model at this point.
Several desperation-related features were found. The authors report that only one had downstream connections "suggestive of experiencing and acting on desperation", while the others connected to features about "merely discussing it". In the figure, that feature's (Feature B's) downstream features are labeled as desperate dialogue, a weak or injured speaker, and a cry for help; the other shown feature's are labeled "do anything", "I'd trade / take" and "happy / glad to", which the authors read as describing a willingness to act. Positive additive steering on that feature (Feature B in the figure) "caused the model to blackmail, while none of the other candidates had an effect". The item gives no number of candidates, rollouts, blackmail rate or steering strength, and does not say how blackmail was judged. The authors say such analysis helps them "better focus our monitoring effort and resources".
The authors' conclusion
The authors write that TWERA-ranked downstream features "add important context for understanding features" and help predict a feature's role in steering ("Conclusion"). They expect the approach to extend to any component that is a vector in the residual stream, and see "particular promise" for automated feature-description pipelines.
Limitations and open questions
- Small, curated test. The quantitative result rests on 10 groups chosen by the authors, each with one known steering feature, and one grader model whose prompt is not given. No intervals or per-group results are given, and the first condition is near the 0.5 expected by chance (our arithmetic).
- Missing details. The model behind the green and sapphire features is not named. Strengths, layers and sample counts are not reported for any steering result; "caused the model to blackmail" is given without a rate.
- TWERA's own limits. Circuit Tracing, which defines the measure, says "TWERA is not a perfect solution for interference": it "does not handle inhibition well", and it relies heavily on co-activation statistics, so it changes which connections rank highest rather than only removing interference weights ("Global Weights"). The item does not discuss this, although its downstream lists include suppressed features.
- Labels are interpretations. Downstream features are themselves named from their examples and unembeds, so the method moves the labeling question one step downstream rather than removing it.
Research questions
- Does the gain hold on groups the authors did not choose? Sampling look-alike groups automatically (by similarity of activating examples and unembeds), finding the steering feature by intervention, and repeating the ranking test would measure the effect without selection.
- Is the grader or the information doing the work? Repeating the test with human raters, and with other grader models, given the same four kinds of information would show whether the improvement depends on Opus 4.7.
- How far down should one look? The item uses immediate downstream features. Comparing rankings built from one, two and three steps of downstream connections would show where the signal saturates.
- Do downstream connections predict effects on other prompts? The hex example suggests a feature that does not steer one prompt steers another. Predicting, from downstream features alone, a prompt on which each candidate will steer, then testing it, would check that the descriptions generalize.
- How often does desperation steering produce blackmail in Claude Haiku 4.5? The item reports the effect without rates. Rollout counts and blackmail rates across steering strengths, for the chosen feature and for the others, would give the auditing example a size.
- Does TWERA's handling of inhibition matter here? The downstream lists include suppressed features. Repeating the test with downstream connections measured by intervention rather than TWERA would show what the approximation loses.
How it connects
- Within this library. The method rests on the TWERA virtual weights of Circuit Tracing: Revealing Computational Graphs in Language Models (March 2025), whose note in this collection already records this item as later work; A Toy Model of Interference Weights (July 2025) and Characterizing interference weights in a tiny language model (August 2026) compare TWERA with other ways of ranking feature connections. The blackmail scenario is from Agentic Misalignment, and the steering question follows Emotion Concepts and their Function in a Large Language Model (April 2026), which reports that steering an emotion vector for "desperate" raised blackmail rates in an earlier snapshot of Claude Sonnet 4.5. System Card: Claude Haiku 4.5 reports the model's high verbalized evaluation awareness on its blackmail scenarios, which this note connects to the item's remark that the model recognizes the setup as a test. Earlier evidence that a feature's activating examples can misdescribe it is in Circuits Updates – July 2024, whose feature-sensitivity item found highlighted features silent on many texts about their concept; this connection is the note's own.
- Same authors. Circuits Updates – November 2025 shares the item's first author and editor; its topic (attention under harm pressure) is unrelated.
References
- Goel, P., Kauvar, I., and Turner, N. L.; edited by Kamath, H. (2026). Downstream Connections Predict Which Features Will Steer Model Behavior. In Circuits Updates – May 2026. Transformer Circuits Thread. https://transformer-circuits.pub/2026/may-update/index.html (undated on the page; listed under May 2026 and dated 1 June 2026 in the site index's markup) record · 2 pages
- Ameisen, E., Lindsey, J., Pearce, A., Gurnee, W., Turner, N. L., Chen, B., Citro, C., et al. (2025). Circuit Tracing: Revealing Computational Graphs in Language Models. Transformer Circuits Thread. https://transformer-circuits.pub/2025/attribution-graphs/methods.html record · 15 pages
- Lynch, A., Wright, B., Larson, C., Ritchie, S. J., Mindermann, S., Hubinger, E., Perez, E., and Troy, K. (2025). Agentic Misalignment: How LLMs could be insider threats. arXiv:2510.05179. record · 4 pages
- nostalgebraist (2020). interpreting GPT: the logit lens. LessWrong. https://www.lesswrong.com/posts/AcKRB8wDpdaN6v6ru/interpreting-gpt-the-logit-lens record
- Sofroniew, N., Kauvar, I., Saunders, W., Chen, R., et al. (2026). Emotion Concepts and their Function in a Large Language Model. Transformer Circuits Thread. https://transformer-circuits.pub/2026/emotions/index.html record · 2 pages