# Goodfire puts neuron explanations to a 76%-versus-11% response test

**Goodfire's 2026 research essay argues that neural networks may already be legible without first untangling them: many apparent mixtures reflect generalization across related patterns.** In [Neural Networks Are Already Clean](https://already-clean.goodfire.com/), Nick Cammarata advances a working hypothesis: many apparently tangled neurons follow intelligible paths of generalization. Earlier architectures, tools and interpretations may have obscured those paths.

For interpretability researchers, the practical contribution is a testable workflow: use agents to generate explanations, then check them against unseen inputs, counterexamples and interventions on the model. A plausible neuron label is only a starting point.

![A conceptual sequence from neuron hypothesis to new inputs, response measurement and causal intervention](https://s4.tenten.co/learning/content/images/2026/10/english-infographic-1-2.png)

#### A 76%-versus-11% response gap gives explanations something to lose

Goodfire randomly selected 200 neurons from its gallery and asked an agent to read only their explanations. Before measuring responses, the agent wrote five new examples expected to activate each neuron and five near misses expected not to.

The reported comparison covers 183 neurons whose explanations describe their positive side. New positive examples strongly activated the neurons 76% of the time; near misses did so 11% of the time. A strong response meant at least 30% of the neuron's strongest example response in the corpus.

This design makes the explanation commit before seeing the measurements. An account that sounds convincing on familiar examples can fail on a new sentence. The near misses also test whether the proposed boundary is useful.

| Goodfire test component | Reported setup or result | Scope |
| :--- | :--- | :--- |
| Initial random sample | 200 neurons | Selected from the gallery |
| Positive-side explanation subset | 183 neurons | Subset used for the response comparison |
| New inputs per explanation | Five positives and five near misses | Written before responses were measured |
| Strong responses to new positives | 76% | Input response rate, not explanation accuracy |
| Strong responses to near misses | 11% | Same strong-response threshold |

#### The 100,000-graph audit provides qualitative evidence

Goodfire had agents review 100,000 attribution graphs, while Cammarata manually inspected thousands across vision and language models. He treats those observations as qualitative evidence and says he does not trust current tools enough to turn them into a cleanliness score.

#### The vision and language percentages measure different things

Goodfire's vision-model observations and Qwen3-32B audits should not be drawn as points on one improvement curve. The vision work revisits InceptionV1 and architectural improvements. The language work includes preliminary inspection of neuron poles in Qwen3-32B.

Cammarata reports that ordinary architectural improvements, without changing model size or training data in the comparison, made 99–100% of the final-layer vision neurons easily interpretable. That statement belongs to a specific experimental setting. It is not a success rate for every layer of every vision model.

The Qwen3-32B estimate puts roughly 90% of neuron poles at one main job, with related offshoots. About 7% have two related jobs, while roughly 2–4% remain unexplained. Positive and negative poles can carry distinct signals, making the unit of observation important.

Cammarata also explains that the estimates depend on weak-activation thresholds and judgments about what counts as one thing. The figures communicate the team's preliminary assessment. They are not a standardized, independently replicated interpretability benchmark.

#### A cat-and-car neuron invites a different hypothesis

Goodfire revisits an InceptionV1 neuron famous for responding to cat faces and car fronts. Cammarata now argues that a shared V-shaped structure could explain the grouping. The earlier interpretation, on this account, underestimated how far generalization can reach.

Polysemanticity describes the apparent phenomenon of one neuron handling multiple unrelated concepts. Superposition is a proposed explanation for such representation, rather than a synonym for the observation. Two types of activating image do not by themselves identify the mechanism.

The [2020 Zoom In essay](https://distill.pub/2020/circuits/zoom-in/) established a route for studying circuits through connections between neurons. Goodfire's new essay reexamines some canonical examples and proposes recursive generalization: a neuron extends along an existing distinction, then grows related branches.

The distinction changes where an investigation starts. If native neurons are assumed to be arbitrary mixtures, researchers may seek another basis first. If the grouping follows traceable relationships, following the original model's connections and training history could resolve the apparent conflict.

#### Automated explanations still need independent experimental pressure

OpenAI's [Language Models Can Explain Neurons in Language Models](https://openaipublic.blob.core.windows.net/neuron-explainer/paper/index.html), published on May 9, 2023, already organized automated work around explanation, simulation and scoring. Goodfire argues that better models and tools can now support a richer cycle of conjecture and test.

A better explainer does not make every explanation reliable. Cammarata distinguishes explanations that simulate activations well from explanations that are right. Optimizing only a simulation score risks rewarding an account that predicts responses without capturing the mechanism.

Polarity is another practical concern. In Goodfire's language-model examples, labels inferred from top positive activations can miss meanings carried on the negative side. A researcher needs to inspect the label alongside the neuron's connections and intervention effects.

A useful workflow begins by preserving the model, layer and neuron identifiers, then writing specific predictions. Test new positives, near misses and strong negative activations separately. Follow with ablations that remove or alter the proposed circuit and measure whether the output changes as predicted.

That sequence is a proposed research practice drawn from the essay, rather than a one-command reproduction package. Its purpose is to preserve the evidence behind each explanation as agent throughput increases. A thousand labels without their tests are much harder to audit.

#### Native neurons and sparse autoencoders answer a question about representation

Anthropic's [Towards Monosemanticity](https://transformer-circuits.pub/2023/monosemantic-features/index.html), published on October 4, 2023, used a sparse autoencoder to extract interpretable features from a one-layer transformer. A sparse autoencoder learns another representation. Its features do not correspond to a simple one-to-one labeling of the original neurons.

Goodfire advocates spending more effort on native neurons, partly because they belong to the original model and avoid reconstruction error from a learned approximation. That argument invites comparison. A working hypothesis alone cannot establish that sparse autoencoders have no useful role.

Researchers can compare the approaches on predictive tests, causal interventions and failure cases. Understanding a local circuit for a particular prompt also remains distinct from understanding a model's behavior in every setting. Each broader safety claim needs evidence of its own.


#### Frequently asked questions

##### Has Goodfire proved that large language models are fully interpretable?

Goodfire presents its 2026 essay as a working hypothesis. The roughly 90% Qwen3-32B estimate comes from preliminary neuron-pole audits and depends on interpretation thresholds. Local legibility and complete understanding of model behavior remain different research goals.

##### Is Goodfire's 76% result an explanation-accuracy score?

Goodfire's 76% measures strong neuron responses to new positive examples. The comparison covers 183 positive-side explanations within a random sample of 200 neurons; near misses produced strong responses 11% of the time. The result tests predictive value without establishing an overall accuracy score.

##### Did Goodfire assign a quantitative score to the 100,000 graph reviews?

Goodfire treats the 100,000 agent-reviewed attribution graphs as qualitative evidence. Cammarata explicitly says he does not trust the current tools enough to produce a cleanliness score. Review volume and a validated aggregate score are different claims.

##### Can mechanistic interpretability rely only on top-activation examples?

Mechanistic interpretability needs evidence beyond top-activation examples. Goodfire uses new positives, near misses, polarity-aware inspection and circuit interventions. Top activations suggest hypotheses; causal claims require tests.

#### Sources

- [Goodfire: Neural Networks Are Already Clean, 2026 research essay](https://already-clean.goodfire.com/)
- [OpenAI: Language Models Can Explain Neurons in Language Models, May 9, 2023](https://openaipublic.blob.core.windows.net/neuron-explainer/paper/index.html)
- [Anthropic: Towards Monosemanticity, October 4, 2023](https://transformer-circuits.pub/2023/monosemantic-features/index.html)
- [Distill: Zoom In: An Introduction to Circuits, 2020](https://distill.pub/2020/circuits/zoom-in/)

#### Author Insight

A neuron explanation earns more confidence when it makes predictions that could fail. An account that only retells familiar examples should remain a hypothesis. Novel-input tests and a corresponding causal intervention provide a stronger basis for accepting it.

Agents are well suited to expanding that testable work. Keep the inputs, measurements and failed cases beside every generated label. As the investigation scales, a researcher should still be able to locate the step where an explanation stopped matching the model.