Back to AI Research

AI Research

Metaphor Tracer: A Theory-Informed Analysis of Hidd... | AI Research

Key Takeaways

  • Metaphor Tracer: A Theory-Informed Analysis of Hidden States introduces a training-free instrument designed to visualize how a large language model (LLM) org...
  • What do a language model's hidden states say about the organization of a single text?
  • From one forward pass, without training, we score every token position on two properties.
  • The *aggregator* measures whether the position consolidates the whole text into a stable configuration.
  • The *differentiator*, whether other tokens are transiently carried into its subspace as the model reads: metaphor in its root sense, transport.
Paper AbstractExpand

What do a language model's hidden states say about the organization of a single text? From one forward pass, without training, we score every token position on two properties. The *aggregator* measures whether the position consolidates the whole text into a stable configuration. The *differentiator*, whether other tokens are transiently carried into its subspace as the model reads: metaphor in its root sense, transport. Constants were frozen on one discovery text; every other is confirmatory. The aggregator is not, in the classic sense, an information measure, nor a measure of salience. Across three unrelated models, as a signifier repeats, its surprisal and its attention drain while its aggregator score holds: the channel marks a token's place in the text. That this tracks a reading rests on independent ground truth: an engineered register the aggregator follows across its boundaries (6/6 cells), and a psychoanalyst's marking of clinical transcripts, fixed before the instrument existed, in 34/36 cells, with a graded increment above lexical controls and dissociations no type-level measure reproduces. A transfer test gives the result its shape: the model whose token structure travels with lexical type reads the singular discourse worst, and in a matched base/instruct pair tuning raises fidelity without moving type-transfer. Structural value is a property of a token's place in *this* text, not of its vector alone: a relational rather than essentialist reading of hidden states, operationalizing theory that predated the instrument.

Metaphor Tracer: A Theory-Informed Analysis of Hidden States introduces a training-free instrument designed to visualize how a large language model (LLM) organizes a specific text during a single forward pass. Rather than treating the model as a black box, the authors—Marc Heimann, Roxana Assadi Moghaddam, Olga Brovkina, Mark Pettifor, and Lutz Götzmann—propose that the internal geometry of an LLM’s hidden states functions as a theory of language. The instrument maps how a model anchors and processes text, providing a way to test theoretical claims about language structure against the internal mechanics of transformer models.

How the Instrument Works

The researchers developed a method to score every token position in a text based on two distinct properties, which they term the "aggregator" and the "differentiator."

  • The Aggregator: This channel measures how a token consolidates the entire text into a stable configuration. It identifies "anchor" points—such as headings or paragraph-ending periods—that serve as structural hubs for the text.

  • The Differentiator: This channel tracks "metaphorical" transport, measuring how other tokens are transiently pulled into an anchor’s subspace during the model's reading process.
    The instrument operates without training, using a single forward pass to extract hidden-state tensors. It analyzes these states through two "views": an "operative view" (raw states) and a "quiet-dimensions view" (where dimensions are z-scored to account for anisotropy). By comparing these views, the researchers identify structural boundaries within the text.

Evidence and Validation

To validate the instrument, the authors used a "frozen" battery of tests, meaning the constants were set on one discovery text and applied to all others for confirmation. The findings include:

  • Engineered Ground Truth: The aggregator successfully followed engineered registers across boundaries in 6 out of 6 test cases.

  • Clinical Transcripts: In a test involving psychoanalytic annotations of clinical transcripts—fixed before the instrument was created—the aggregator identified organizing structures in 34 out of 36 cells. The authors note that this performance exceeded lexical controls and captured dissociations that type-level measures could not replicate.

  • Relationality: The study found that structural value is a property of a token’s specific place in a text rather than its vector alone. In models that read a singular discourse most effectively, this structural value does not transfer with lexical type, supporting the authors' view that the model's internal representation is relational rather than essentialist.

Interpretation and Limitations

The authors frame their work as a "counter-reading" to dominant approaches in computer science, drawing on Freudian and Lacanian theory to interpret model internals. They argue that the model acts as a theory of language where meaning is fixed retroactively by "master signifiers."
Regarding limitations, the authors acknowledge that their aggregator channel is highly rank-correlated with the size of the anchor's subspace, raising questions about why the channel measures occupancy rather than just a raw dimension count. Additionally, they note that their use of a single-annotator corpus for psychoanalytic validation leaves certain aspects open for further study. The researchers emphasize that their instrument is not intended to explain model outputs or decisions, but rather to make the internal "geometry of internal references" legible as a form of interpretation.

Comments (0)

No comments yet

Be the first to share your thoughts!