Learning to Read Out: Unembedding Dynamics in Language Model Pretraining
When does a language model acquire a capability? Developmental studies usually answer this by tracking when models start producing the correct token. But that prediction passes through the unembedding matrix, or readout, which is itself still being learned during pretraining. We show that a model's hidden states can encode task information before its own readout expresses it, so timelines read from outputs conflate what the model encodes with what its readout expresses. To track the readout directly, we introduce parameter-trajectory crosscoding, which fits one sparse dictionary to every token's readout row at every checkpoint. Unlike crosscoders on activations, it needs only checkpoint w...
| Authors | Matteo He, William F. Shen, Alex Iacob et al. |
|---|---|
| Primary category | cs.CL |
| Published | 14 October 2026 |
| Last updated at source | 14 October 2026 |
Connected in the knowledge graph
- hematteo/wu-crosscoder-snapshotsDataset · cited by
Recent changes
- 16 Oct 2026 Description: When does a language model acquire a capability? Developmental studies usually answer this by tracking when models start producing the correct token. But that prediction passes through the unembedding matrix, or readout, which is itself sti → When does a language model acquire a capability? Developmental studies usually answer this by tracking when models start producing the correct token. But that prediction passes through the unembedding matrix, or readout, which is itself sti
- 16 Oct 2026 Topics added cs.CL, cs.AI, cs.LG removed hf-daily-papers, parameter-trajectory crosscoding, sparse dictionary, unembedding matrix, readout, hidden states, linear probes, pythia, pretraining
Changes detected by Silytics between refreshes of the source.
Models and datasets referencing this paper
Silytics shows public metadata from arXiv and links to the original. Information may be incomplete or out of date. Last refreshed 16 October 2026. Licenses and terms of the original apply to the work itself. Not financial, legal or professional advice.