brightlabYear 11 · 90 minutes
Stories through representation space
How much of a story trajectory comes from context, model, readout or projection?
You will learn to: Distinguish behavioural readouts from hidden states, compare cumulative and isolated contexts, and test the stability of projections and interpretation.
Goodfire’s original research demonstration
Read a story, one sentence at a time
Choose a story and record its theme, domain and title. The six 0–10 emotion ratings are prompted behavioural readouts. Predict a turning point before revealing the next sentence; record where your judgement differs.
The interactive below is hosted by Goodfire. If it does not fit your screen or your school blocks it, open the original in a full window ↗.
Source measurements and interface: Goodfire. Spark does not generate these recorded Llama results.
Goodfire’s original research demonstration
Follow the story inside the model
Select a story group, story, manifold and projection. Here the measurements come from hidden states collected without the emotion-rating prompt. Record the projection before interpreting a direction. Compare context and readout effects in your own notebook.
The interactive below is hosted by Goodfire. If it does not fit your screen or your school blocks it, open the original in a full window ↗.
Source measurements and interface: Goodfire. Spark does not generate these recorded Llama results.
Your classroom story · real model extension
Change an ending. Compare the trail.
MiniLM encodes cumulative story prefixes into 384-dimensional vectors. PCA is fitted to the original trajectory; the altered ending uses the same mean and components. This is a sentence-encoder extension, not the paper’s Llama hidden-state analysis.
Model not downloaded. Your story stays on this device.
Now investigate for yourself
Your Marimo research notebook
PCA axes are not inherently emotional axes. Candidate-label likelihoods are not the paper’s 0–10 ratings. Whole works are analysed as selected bounded passages. Public-domain status and classroom suitability depend on the edition and context.
Teacher notes & evidence task
Starting knowledge: Vectors, similarity, basic statistics and experimental controls. PCA is introduced in the notebook.
- Compare the original behavioural reader with its hidden-state trajectory. Identify what each point and axis measures.
- Choose a source story and record a predicted turning point before revealing the next sentence.
- Analyse a selected passage from Shakespeare or Grimm in Marimo, recording edition, source and passage boundaries.
- Compare accumulated context with isolated sentences. Hold a reference projection fixed when comparing variants.
- In native mode, collect sentence-end hidden states from a small LLM, compare layers and calculate six candidate-emotion readouts separately. Test sentence shuffling or a changed ending.
Evidence to collect: Submit a reproducible passage analysis with provenance, context limits, projection details, a control and an alternative explanation.
Research basis: Stories in Space combines behavioural and representational analyses of story prefixes, with geometric models and interventions. A projection is a view of the data, not the full representation.
Goodfire research article ↗ · Full research paper ↗ · Setup and teaching guidance