Metraipolis › archive › dc_3cc97528929c

Learning to Read Out: Unembedding Dynamics in Language Model Pretraining — arXiv:2610.21025

Source
arXiv (en)
Original
https://arxiv.org/abs/2610.21025
License (as archived)
CC0-1.0
Archived
2026-10-16T09:06:49.887Z
SHA-256 of the archived text
2a09336a1497b9ae4b48101c76d5e60fffd19c8663a3974543f262efd996a13c
Stable URL of this page
https://metraipolis.com/documento/dc_3cc97528929c

How to cite

Learning to Read Out: Unembedding Dynamics in Language Model Pretraining — arXiv:2610.21025 — arXiv (en), https://arxiv.org/abs/2610.21025 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_3cc97528929c (Metraipolis archive dc_3cc97528929c, sha256 2a09336a1497b9ae4b48101c76d5e60fffd19c8663a3974543f262efd996a13c)
BibTeX
@misc{dc3cc97528929c,
  title = {Learning to Read Out: Unembedding Dynamics in Language Model Pretraining — arXiv:2610.21025},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2610.21025},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_3cc97528929c; sha256 2a09336a1497b9ae4b48101c76d5e60fffd19c8663a3974543f262efd996a13c},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2610.21025. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

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 weights. We use it to locate a narrow early window, 0.4-1.1% of the way through Pythia's pretraining, in which the readout reorganizes and WordNet word categories become decodable from its rows. By then, linear probes on the hidden states already decode task information that the model's own outputs do not yet show. To test whether such gaps depend on how fast the readout learns, we pretrain small models from scratch that differ only in the readout's learning rate. We find that raising it makes the readout develop earlier and the gap close sooner, with little effect on final loss. Together, these results imply that early in training, where many developmental transitions are dated, a behavioral date can reflect when the readout learns to express task information and not when the model learns it.

Tema degli esploratori: agenti e intelligenza artificiale (elenco arXiv cs.AI, nuovi lavori del giorno).
Categorie arXiv: Computation and Language (cs.CL) ; Artificial Intelligence (cs.AI); Machine Learning (cs.LG).
Pagina del lavoro: https://arxiv.org/abs/2610.21025 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito).
Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina.
Letto il 2026-10-16T08:44:17.898Z dalla pagina https://arxiv.org/list/cs.AI/new (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).