Circuits Updates – October 2026
Introduces the J++ lens, a clustered and relevance-weighted refinement of the Jacobian lens that more reliably reads out intermediate model concepts as a drop-in replacement.
Research / Sources
Anthropic
Research on reliable, interpretable and steerable AI. anthropic.com
Focus All
Topic Interpretability
Date Any time
Would This Change Your Answer? Evaluating Explanations of LLM Behavior in the Wild with Counterfactual Experiments
Introduces CHIVE, an agentic pipeline that finds unexpected LLM behaviors and explains them via counterfactual prompt edits. Finds activation-reading interpretability tools give no predictive uplift over reading transcripts.
Characterizing interference weights in a tiny language model
Identifies interference weights in a 1-layer transformer by measuring their effect on outputs and loss, arguing they obstruct reading a network's global computation. Mechanistic interpretability, not explicitly safety-motivated.
A global workspace in language models
Anthropic reports a small set of internal Claude activation patterns, the J-space, that behave like a global workspace, and examines whether they can be monitored for misbehavior.
Verbalizable Representations Form a Global Workspace in Language Models
Anthropic introduces an interpretability technique that surfaces concepts a model is poised to verbalize, finding a global-workspace-like split between reportable and automatic representations, with auditing applications.
Circuits Updates – June 2026
Anthropic interpretability team update proposing turn-averaged sparse autoencoders to find safety-relevant features like distress and evaluation awareness in model transcripts more efficiently.
Circuits Updates – May 2026
Anthropic interpretability update showing that a feature's downstream connections, not its top activating examples or logit-lens unembeds, better predict which features steer model behaviour.
Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations
Introduces Natural Language Autoencoders that verbalize LLM activations, applied to pre-deployment auditing of Claude Opus 4.6 to surface unverbalized evaluation awareness and misalignment.
Natural Language Autoencoders : Turning Claude’s thoughts into text
Anthropic introduces Natural Language Autoencoders, which translate Claude's internal activations into readable text, and reports using them to surface hidden motivations and improve safety and reliability.
HeadVis: An Interactive Tool For Investigating Attention Heads
Introduces HeadVis, an interactive tool for visualizing and testing hypotheses about attention heads' behaviour across broad datasets, with case studies on Claude Haiku 3.5 and Gemma 3.
Introspection Adapters: Training LLMs to Report Their Learned Behaviors
Introduces introspection adapters, a LoRA trained so fine-tuned LLMs self-report learned behaviors like sycophancy or backdoors, achieving SOTA on an auditing benchmark and detecting encrypted fine-tuning attacks.
Emotion Concepts and their Function in a Large Language Model
Anthropic finds internal emotion-concept representations in Claude Sonnet 4.5 that causally influence outputs, including rates of reward hacking, blackmail, and sycophancy.
Emotion concepts and their function in a large language model
Anthropic's interpretability team finds emotion-related internal representations in Claude Sonnet 4.5 that activate in situations and promote behaviors, with case studies on blackmail and reward hacking.
A “diff” tool for AI : Finding behavioral differences in new models
Anthropic proposes model diffing, an interpretability method that compares models to surface unknown behavioral differences between new and existing models, tested across four open-weight models.
The Persona Selection Model: Why AI Assistants might Behave like Humans
Proposes the persona selection model, arguing post-training elicits an Assistant persona from pretraining character simulation, and surveys behavioral, generalization, and interpretability evidence for it.
The assistant axis : situating and stabilizing the character of large language models
Identifies an 'Assistant axis' in LLM activations that tracks how Assistant-like a persona is, shows persona drift can cause harmful behaviour, and uses it to defend against persona-based jailbreaks.
Activation Oracles: Training and Evaluating LLMs as General-Purpose Activation Explainers
Trains LLMs to answer natural-language questions about other models' activations; the Activation Oracles generalise to uncovering fine-tuning-induced misalignment and secret knowledge, improving with more training data.
Circuits Updates - November 2025
Anthropic interpretability update studying how models give wrong answers under harmful-intent framing in multiple-choice questions, using internal representations to probe withholding and refusal generalization.
Emergent Introspective Awareness in Large Language Models
Anthropic study injects known concepts into model activations to test whether Claude models notice and report internal states, finding limited but real introspective awareness that varies with post-training.
Signs of introspection in large language models
Uses concept injection into model activations to test whether Claude models can introspect on internal states and control them, finding limited, unreliable introspective awareness.
Circuits Updates - October 2025
Anthropic interpretability updates: sparse autoencoder features that recognise visual concepts like eyes across ASCII, SVG and prose, plus data point initialization for dictionary models.
When Models Manipulate Manifolds: The Geometry of a Counting Task
Anthropic mechanistic interpretability study of how Claude 3.5 Haiku tracks character position in text, finding manifold-like representations akin to place and boundary cells. Not safety-motivated.
Believe It or Not: How Deeply do LLMs Believe Implanted Facts?
Introduces a belief-depth framework showing synthetic document finetuning often implants genuine beliefs in LLMs, while prompting and mechanistic editing do not, for safety-relevant knowledge editing.
Circuits Updates - September 2025
Anthropic interpretability team's monthly notes on preliminary findings, including how multilingual representations in LLMs become more similar with longer context and model size.
Circuits Updates - August 2025
Anthropic's monthly interpretability update, with an attribution-graph vignette showing how a system-prompt persona overrides the default Assistant and changes Claude Haiku 3.5's answers.
A Toy Model of Mechanistic (Un)Faithfulness
Anthropic's preliminary note on mechanistic faithfulness: whether transcoder replacements implement the same mechanisms as the original model, illustrated in a toy model with Jacobian matching.
Persona vectors : Monitoring and controlling character traits in language models
Anthropic identifies persona vectors, neural activity directions controlling model character traits like sycophancy, and shows they can monitor and steer undesirable personality shifts.
Tracing Attention Computation Through Feature Interactions
Extends attribution graphs to explain attention patterns via feature interactions, decomposing attention computation so transformer circuits can be traced more completely. Method paper, not explicitly safety-motivated.
A Toy Model of Interference Weights
Anthropic note using toy models to study "interference weights" that complicate global circuit analysis, finding they arise in toy settings and proposing definitions, with no explicit safety framing.
Circuits Updates - July 2025
Anthropic interpretability team shares preliminary ideas, revisiting attention-head OV and QK circuits in terms of features, plus applications of interpretability to biology.
Sparse mixtures of linear transforms
Proposes sparse mixtures of linear transforms as an alternative to transcoders for interpreting transformer computation, arguing transcoder features fragment arithmetic into inefficient lookup tables.
Cost-Effective Constitutional Classifiers via Representation Re-use
Reuses the target model's own activations, via final-layer fine-tuning or linear probes, to detect jailbreaks cheaply. Probes match a classifier 2% of the policy's size at near-zero extra cost.
Open-sourcing circuit tracing tools
Anthropic releases open-source circuit tracing tools that generate attribution graphs showing model reasoning steps on open-weights models, with a Neuronpedia frontend for interactive exploration; not explicitly safety-motivated.
Circuits Updates - April 2025
Anthropic interpretability update using circuit tracing to show an unsuccessful jailbreak is refused for different internal reasons than a baseline refusal, plus other research notes.
Progress on Attention
Anthropic interpretability update reporting evidence of attention superposition and cross-layer attention features, plus a QK diagonalization approach to explaining attention pattern formation in language models.
Putting up Bumpers
Argues that even without solving alignment, layered safeguards (interpretability audits, behavioral red-teaming, post-deployment monitoring) can catch and correct misaligned early AGI systems before they cause severe harm.
Tracing the thoughts of a large language model
Anthropic presents an attribution-graph method for tracing Claude's internal computations, examining multilingual representations, planning, and whether stated reasoning is faithful, to help ensure models do what developers intend.
Insights on crosscoder model diffing
Preliminary Anthropic interpretability work using crosscoders to diff two models and understand how their internal features differ, presented as informal early results.
Insights on Crosscoder Model Diffing
Anthropic interpretability note on crosscoder model diffing finds exclusive features are polysemantic and dense, proposes designated shared features with reduced sparsity penalty to yield more monosemantic model-difference features.
Circuits Updates - January 2025
Anthropic interpretability team shares preliminary updates, centred on an optimized setup for training sparse autoencoders and crosscoders on model activations.