NLASmith · CONAIISI 2026
From one activation
to a systematic experiment.
Neuronpedia makes Natural Language Activations easy to inspect one at a time. NLASmith is the missing loop: datasets, token policies, configurable judges, and aggregated metrics — so hypotheses about internal representations can be run, reproduced, and compared.
The gap
Exploring one verbalization is not the same as running an experiment.
Natural Language Autoencoders turn a hidden activation into a sentence and can reconstruct the activation from that text. Those verbalizations are interpretable — they are not a literal or infallible readout of the model's state.
Neuronpedia already exposes NLA through a web UI and an API. The pieces exist: pick a prompt, choose a position, fetch a verbalization. Coordinating that across many examples — and keeping the decisions that produced each result — is still left to ad-hoc scripts.
LangSmith showed how LLM applications get evaluated: datasets, repeated runs, automatic evaluators, comparative views. NLASmith applies that methodological shape to a different object. The artifact under test is not only the visible reply. It is the NLA at positions fixed by the experiment.
This is not a new interpretability method and it does not train a new autoencoder. It is infrastructure so the same flow can test different hypotheses without being locked to one phenomenon or domain.
Conceptual pipeline
Dataset, token policy, NLA, evaluator, results.
Each dataset example runs under a fixed configuration. Selected positions go to the NLA service. Verbalizations are scored with criteria defined up front. Individual outcomes are stored so they can be aggregated and compared.
A fixed list of prompts is the unit of comparison. Two experiments on the same set can be contrasted per example and as aggregates.
The observed position is part of the definition — last user token, first assistant token, or both — not something decided after reading each reply.
Neuronpedia runs the completion and returns the verbalization at those positions. NLASmith does not train or host the autoencoder.
A rubric, a judge model, and an output schema: boolean, score, or category, with optional reasoning. The judge is a measuring instrument, not ground truth.
Per-example artifacts stay inspectable. Presence rate, mean score, and category mix let you see patterns instead of isolated screenshots.
Prototype
A working loop for configuration, execution, and comparison.
The first version keeps only what turns a manual inspection into a reproducible process: define the experiment before it starts, walk the dataset, persist artifacts, and look at both the case and the aggregate.
Datasets
Create a prompt list, identify each example, and optionally attach a reference. The dataset is what makes two runs comparable.
Token policy
Keep the position rule constant so last-user vs first-assistant is an experimental factor, not an afterthought.
NLA source
Neuronpedia hosts the language model and the NLA. The lab wraps completion and explain into one internal schema.
Evaluators
Natural-language rubric, judge model, and feedback keys. The same loop can detect a theme, score a criterion, or classify verbalizations.
Live runs
The orchestrator walks the dataset, records progress per example, and separates API failures from negative judgments.
Compare
Tables and charts on the same prompts. Inspect a single verbalization without losing the global view of the experiment.
Example hypothesis
Forum-like associations that never appear in the reply.
In some open-ended questions, verbalizations seemed tied to Reddit or forum-like discourse even when the generated answer never named that source. That is a useful first case: a qualitative hunch becomes a measurable setup.
The point is not to claim Reddit caused the model, or that an NLA literally describes its internals. It is to show that a research question can be written as a dataset, a token position, an NLA source, a rubric, and aggregated metrics.
How to read the numbers
Automatic scores are measurements, not a verdict on the residual.
An NLA can simplify or invent detail. Fraser-Taliente et al. already warn against treating the sentence as a faithful dump of the activation.
An LLM-as-a-judge moves with the rubric, the judge model, and what you put in context. Those choices are stored with the run so the measurement can be audited.
Coverage depends on which NLAs and APIs exist. Comparing models may mix layers or autoencoders that are not strictly equivalent. Cost grows with prompts, positions, and evaluators.
Configure once. Run the dataset. Keep the trail.
Keys stay in this tab. Datasets, evaluators, and experiment rows persist. Open the prototype, attach a judge, and turn a one-off inspection into a configuration you can repeat.