Academia
A university or academic research group.
Glossary
How this index works
Access required
The level of model access a person would need to reproduce the work.
API
Requires gated or paid inference access, but not weights.
Archived
Preserved in static or historical form and no longer maintained.
Art object / installation
Work made primarily for aesthetic or critical reception rather than scientific reporting.
Artefact type
The form the work takes, independent of what it contains or argues.
Artist
A practitioner working primarily within an art context.
Atlas schema
Coined for this index to describe its own structure. Not a claim about the field, and no external source exists.
audio
Models that process sound or speech.
Ava's framing
Supplied by Ava in conversation rather than drawn from a source.
Basic science
Work seeking to reverse-engineer the model's general structure, irrespective of any single behaviour.
Both
Work pursuing a specific behaviour in order to establish general structure, or the reverse.
Browsable index / database
A searchable collection of results presented for inspection.
Built by (people)
The named individual contributors, in the order the source credits them.
Citation
A full bibliographic reference in a consistent style.
Claude's coinage
The practice is real but the LABEL was invented by Claude and is not established. Rename before publishing where a standard term exists.
code
Models whose subject matter is programming languages.
Code repo
A link to that source code.
Common ML usage
In wide use across machine learning with no single coining source. Safe to use, not attributable.
Confidence
How well evidenced this row is, judged by what the person writing it actually read.
Controlled value
A value that may only be chosen from the list on the Vocabularies sheet, never typed freely.
Country / region
Where the producing institution is based.
Dashboard
A live readout of a system's internal state, displayed alongside that system in use.
Date added / Added by
When the row was created and by whom, human or model.
Date published
The date the artefact first became publicly available, in ISO form where known.
Dead
The URL no longer resolves to the artefact.
Deployed model demo
A model made publicly usable in a deliberately modified state, as a demonstration.
Description (card)
One or two sentences describing the artefact to a stranger, written to be published rather than as a note to self.
Exhibit
The visual exists so that a reader can inspect and audit a claim which is itself stated in prose.
Experience
The artefact is undergone rather than read, and its value lies in what it does to the visitor.
Explanation
The visual communicates a result established elsewhere, serving teaching rather than discovery.
Field term
Established vocabulary in the published interpretability literature. The citation is the earliest use Claude could verify, not a guaranteed first use.
Finding
The result is constituted by the visual, such that removing the images would remove the contribution.
Free text
A field written in prose, with no fixed list of permitted values.
Frontier lab
An organisation training state-of-the-art general-purpose models.
Frontier lab + academia
Joint authorship spanning both.
general
Not tied to one kind of data. Applies across different models and media.
Not tied to any particular model or data type.
General public
Non-specialists, with no assumed technical background.
genomics
Models of genetic sequence.
High
The primary source was read in full by whoever wrote the row.
HOW TO READ THE THREE PROVENANCE COLUMNS
Origin says who coined the TERM. Coined in gives the citation. Definition drafted from says what Claude actually read to write the sentence — which is not the same thing.
ID
A permanent identifier for one artefact, never reused, so that a retired entry's number can never resolve to a different record.
image-generation
Models that produce images, such as diffusion models.
Independent researcher
A person working outside any institution.
Instrument
Others can operate it on their own inputs or models to produce findings that did not previously exist.
Intended audience
Who the artefact was designed to be understood by.
Interaction affordances
What a person can actually do with the artefact, verb by verb.
Interactive article
A written argument in which the figures respond to the reader.
Interactive tool
A standalone application whose purpose is to let a user investigate a model.
Internal / proprietary
Requires access available only within the producing organisation.
language
Text. Models that read and write words.
Models that generate or process text.
Link status
Whether the recorded URL currently resolves to the artefact.
Live
The URL resolves and the artefact functions as described.
Low
The row rests on inference and has not been checked against a source.
Makes visible
The thing rendered on screen. Not what the work is about, which is the phenomenon tag, but what you are actually looking at.
Media files
The image filenames belonging to this entry, semicolon separated.
Medium
The row rests on an abstract, documentation, or a reliable secondary source.
method
The interpretability technique the work employs.
Method / technique
The interpretability technique or techniques the artefact rests on.
Mixed
Deliberately designed to serve more than one of the above.
modality
The kind of model or data the work concerns.
Model(s) studied
The specific neural networks the work examines.
multimodal
Models that process more than one modality jointly.
Name
The artefact's own title, as given by its makers.
Neither / educational
Work that neither diagnoses a behaviour nor uncovers structure, but conveys existing knowledge.
No
No source is published.
No longer available
The artefact was operable once but is not now.
Non-profit / community
An organisation constituted for public benefit rather than profit.
None
This entry produces no picture of its own.
No graphic form. The artefact qualifies some other way.
None (pre-rendered)
Requires no model access at all; the outputs are fixed.
Notebook / library
Code published to be run and adapted by other researchers.
Notes
Anything qualifying the row, including unresolved questions and flagged boundary cases.
Nothing
There is nothing visual to look at.
Nothing is rendered. The artefact is encountered rather than viewed.
Observational vs Interventional
Whether the work examines the model unchanged or alters it and observes the consequence.
On your own inputs
A visitor can supply their own prompt, image or model and see fresh output.
Open source
Whether the artefact's own source code is publicly available.
Open weights
Reproducible using publicly downloadable model weights.
Organisation
The institution that produced the work.
Other
A form not covered above; propose a new term rather than reusing this indefinitely.
Partial
Some components are released and others deliberately withheld.
Paywalled
Reachable only behind payment or registration.
phenomenon
The property or behaviour of the model that the work is about.
Platform
Infrastructure hosting many artefacts, models or datasets originating from multiple parties.
Policy
Regulators, governments and civil society organisations.
Practitioners
People deploying or auditing models professionally.
Pragmatic
Work targeting a specific model behaviour, asking which parts of the network produce it.
Pragmatic vs Basic science
Whether the work targets one specific behaviour or the model's general structure.
Pre-rendered, view only
The visuals are fixed; interaction is limited to navigation such as panning or zooming.
Pre-set examples only
A visitor can explore interactively, but only within cases the authors selected in advance.
protein
Models of protein sequence or structure.
Purpose
Every column in the Ledger and every permitted value is defined here in one sentence, so that two people filling in the same row reach the same answer.
Related entries
The IDs of other Atlas entries with a documented relationship to this one.
Researchers
Specialists in interpretability or machine learning.
Reverse-eng vs Concept-based
Whether the work decomposes the network and then names the parts, or proposes concepts and then locates them.
rl-agent
Agents trained by reinforcement learning to act within an environment.
Role of the visual
The function the visual element performs within the research, which is the central judgement in each row.
Runnable by a visitor
Whether, and how far, a person arriving today can operate the artefact themselves.
safety-relevant-features
Parts of the model to do with harm, deception, bias or misuse.
Representations bearing on harm, deception, bias or misuse.
Sector
The kind of institution that produced the work, recorded so the composition of the field is visible.
Slug
The lowercase hyphenated form of the entry's name, used as its web address.
Source of info
What was read to write this row, and on what date.
Startup
A commercial company whose product is interpretability or model tooling.
Static figures in a paper
Fixed images within a conventional publication.
Students
Learners being taught the material for the first time.
tabular
Models over structured, columnar data.
Tags
Faceted keywords in facet:value form, drawn only from the controlled facet lists.
Thumbnail URL
An image representing the entry.
Unknown
The role has not been determined.
Unknown
Not determined, often because the page could not be loaded.
Unknown
The basis for the row was not recorded.
Unknown
Not yet determined.
Unknown
Not yet determined.
URL
The canonical public address of the artefact itself, rather than of a paper describing it.
Venue / published in
Where the work appeared, which is often not the same as who produced it.
Video / animation
Time-based linear media.
vision
Images. Models that look at pictures.
Models that process static images.
Visual / design notes
Observations on the artefact's visual and interaction design, including what makes it work or fail.
Visual form
The shape the picture takes: what you would call the graphic if describing it to someone who could not see it.
What it visualises
The object being made visible, stated in plain terms a non-specialist could follow.
Why it matters
One line stating the artefact's significance in terms a reader could dispute.
Why no judgements
A tag is a filter, not an opinion; claims about significance belong in Why it matters, where a reader can dispute them.
Why only three
Organisation, artefact type, sector and open source are already columns; tagging them again would create two sources of truth that drift apart.
Year
The publication year held separately, so that entries sort chronologically without parsing dates.
Yes
The artefact's own source code is publicly available under a licence.
Glossary
Activation grid
Lots of tiny pictures in a grid, each showing what the model saw at that spot.
Many small images laid out in a grid, each one standing for what a part of the model responded to at that position.
activation-patching
Swapping one internal value for another to see whether it mattered.
Substituting one activation for another to measure its causal effect on the output.
Activations
Raw internal values, shown without further decomposition.
Architecture
The shape of the model itself: what parts it has, and the order they run in.
The model's structure rather than anything it learned.
Attention
How the model decides which earlier words matter when it is working on the current one.
Which positions a model attends to, and how strongly.
attention-heads
The parts that decide which earlier words matter for the next one.
What an attention head does, on a chosen task or across the full data distribution.
attention-visualisation
Showing which words the model looked at, and how hard.
Displaying which positions a model attends to, and how strongly.
attribution
Working out which inputs were responsible for an output.
Assigning responsibility for an output to particular inputs or components.
Attribution
How much each bit of the input pushed the model towards its answer.
How much each part of an input pushed the model towards its output. Distinct from a feature, which is what the model represents, and from an activation, which is how hard a component is firing.
attribution-graph
A map of which internal pieces caused which, for one specific prompt.
A causal graph of interpretable components explaining a single forward pass.
auto-interp
Using one model to write explanations of what is happening inside another.
Using a model to generate natural-language explanations of another model's internals.
bias
The model treating people or groups differently in ways it should not.
Systematic differential treatment of people or groups by the model.
Both
The work does each in turn, typically proposing concepts to interpret a decomposition.
Both
The work observes to generate a hypothesis and intervenes to test it.
causal-tracing
Finding where in the model a particular fact is stored.
Locating where within a network a specific piece of knowledge is stored.
Circuits
A chain of parts inside the model that work together to do one job.
Connected components and the paths between them.
Concept-based
Proposing a human concept in advance and then locating where the network represents it.
concept-injection
Inserting a known idea into the model's internals to see whether it notices.
Inserting a known representation into an unrelated context to test detection or effect.
crosscoder
A variant that works across several layers, or several models, at once.
A sparse dictionary trained jointly across several layers or several models.
Dashboard
A single screen putting several different readouts side by side.
Several visual forms arranged together as one interface, read as a unit.
dimensionality-reduction
Flattening very high-dimensional data down to two or three dimensions so it can be drawn.
Projecting high-dimensional activations into two or three dimensions so their structure can be seen.
emotion
Internal states that work like emotions and change what the model does.
Representations of emotional states and their causal effect on behaviour.
evaluation-awareness
A model noticing it is being tested, and behaving differently because of it.
A model recognising that it is being tested, and behaving differently as a result.
feature-interpretation
Working out what one small piece of the model has learned to recognise.
What an individual learned feature means.
feature-visualisation
Generating an image that shows what a part of the model responds to most strongly.
Synthesising an input that maximally activates a chosen component.
Features
A single thing the model has learned to recognise, like a place, a tone of voice, or a kind of mistake.
Individual learned directions or concepts inside the model.
Flow diagram
Boxes joined by arrows, showing what happens in what order.
Boxes and arrows showing a process or structure in sequence.
hallucination
The model stating something confidently when it has nothing to base it on.
Producing confident content for which the model has no basis.
Heatmap
A grid where stronger colour means a bigger number.
A grid in which colour intensity encodes magnitude.
Highlighted text
Ordinary text with words shaded in, where the shading shows how strongly the model reacted to each word.
Running text with individual words or tokens shaded to encode a per-token value, most often how strongly a feature activated there.
in-context-learning
The model picking up a new skill from the prompt alone, without being retrained.
Acquiring a capability from the prompt rather than from training.
induction-heads
Parts of the model that spot a pattern earlier in the text and continue it.
Attention heads that continue a pattern seen earlier in the context.
Interventional
The model is altered and the consequences observed, supporting causal claims.
introspection
Whether a model can accurately report on its own thinking.
A model's capacity to report on its own internal states.
jailbreak
Getting a model to do something it was trained to refuse.
Circumventing a model's trained refusals.
knowledge-editing
Deliberately changing one specific fact a model holds.
Deliberately altering a specific fact the model holds.
logit-lens
Peeking at what the model would say if it stopped thinking at this layer.
Projecting an intermediate representation directly into the output vocabulary.
Model behaviour
What the model does, observed from the outside.
Model differences
What separates one model from another.
model-diffing
Comparing two models to isolate exactly what is different between them.
Comparing two models in order to isolate what differs between them.
Neither
The work does not attempt either, as with purely pedagogical artefacts.
Node-link graph
Dots joined by lines, showing what connects to what.
Nodes joined by edges, showing what connects to or causes what.
Observational
The model is examined without alteration, so the findings are correlational.
persona
The character a model plays, and how easily it slips out of it.
The character a model adopts, and the stability of that adoption.
planning
The model deciding where a sentence is going before it writes it.
Selecting a future output before generating the text that leads to it.
polysemanticity
One part of the model doing several unrelated jobs at once.
A single component responding to several unrelated concepts.
probing
Training a small, simple classifier to test whether some information is present inside the model.
Training a simple classifier on internal activations to test what they encode.
refusal
How a model decides to say no.
The mechanism by which a model declines a request.
Reverse-engineering
Decomposing the network into parts and then determining what each part does.
Scatter plot
A graph where each point represents a pair of values, showing the relationship between two sets of numbers on a coordinate plane.
A graph of paired numerical values, with one variable on the horizontal axis and the corresponding value of a second variable on the vertical axis, used to reveal relationships or association between the variables.
sparse-autoencoder
A technique for pulling a model's tangled internals apart into separate, nameable pieces.
Learning an overcomplete, sparsely activating basis for a layer's activations.
Spatial attribution
Colour laid over a picture showing which parts of it the model was reacting to.
Attribution values rendered over the positions of an input image, showing which regions drove the output. A saliency map is the most common instance.
steering
Nudging the model's internals mid-thought to change what it says.
Adding or subtracting a direction in activation space in order to alter behaviour.
superposition
A model packing more ideas into its wiring than it has room for, by letting them overlap.
Representing more features than there are dimensions, by encoding them sparsely and near-orthogonally.
Training dynamics
How the model changes over the course of training.
transcoder
A variant that reconstructs what a layer passes onward, rather than what it received.
A sparse dictionary trained to reproduce a layer's output rather than its input.
unfaithful-reasoning
The explanation a model gives not matching the reasoning it actually did.
Stated reasoning that does not correspond to the computation actually performed.
User model
The picture a model builds of who it is talking to, such as age, gender or mood.
What the model has inferred about the person using it.
user-modelling
What the model has quietly worked out about the person talking to it.
The model's internal inferences about the person it is talking to.
world-models
Signs that a model has built an internal picture of something nobody told it directly.
Internal representations of an external state the model was never given directly.