EPFL learning theorist probing how transformers and state space models learn to reason, memorize, and generalize
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-15) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.
Research profile
Output—
Insufficient data: no publication-volume or venue-count metrics are in this snapshot, so output is not scored. This axis would measure the number and cadence of papers across venues, based on indexed bibliographic records.
Impact—
Insufficient data: citation counts, h-index, and downstream-uptake signals are absent from this snapshot, so impact is not scored. It would measure how often this work is cited and built upon across the field.
Focus0/100
The radar focus score is 0, and the underlying field distribution explains why: this profile splits evenly — 0.5 and 0.5 — between machine learning (learning theory and LLM reasoning) and machine learning (deep learning theory and sequence models). Rather than one narrow niche, the reviewed work spans two adjacent theoretical subfields, chain-of-thought learning theory on one side and sequence-model dynamics on the other.
Collaboration55/100
The collaboration score is 55, drawn from bibliometric data: 15 distinct co-authors appear in this snapshot, with a recurring core of twelve, including Khashayar Gatmiry, Aryo Lotfi, Jason D. Lee, Peter Bartlett, Nika Haghtalab, Etai Littwin, and Josh Susskind. Both reviewed papers are middle-author team efforts rather than solo or corresponding-author work, pointing to an established but not maximal breadth of collaborators.
Activity—
Insufficient data: this snapshot lacks publication-year distributions and recent-output counts, so activity is not scored. The axis would measure publishing cadence over time — whether output is ongoing, accelerating, or dormant across recent years.
Reproducibility55/100
Reproducibility scores 55, averaged over the two papers Evolve has reviewed so far (reproPapers = 2, reproMean = 55). Both carry a medium repro-readiness rating: the abstract-token-curriculum paper reports strong synthetic results under stylized assumptions, and the state-space-model gating study combines toy theory with multi-scale empirical evidence — rerunnable, but with theoretical idealizations to navigate.
Research summary
Emmanuel Abbe, affiliated with École Polytechnique Fédérale de Lausanne (EPFL), works at the intersection of learning theory and modern deep learning architectures. The Evolve catalog reviews two 2026 arXiv papers with middle-author positions, and together they sketch a coherent research program: explaining what large learned models actually compute, through the lens of theory. The first, 'Learn Your Own Thoughts: Abstract Token Curriculum' (2026), trains latent — continuous — chains of thought from answers alone via a curriculum, reporting strong synthetic results and a one-step attention analysis grounded in the parity function, while flagging stylized assumptions. The second, 'On the Importance of Gating' (2026), shows that gating in state space models such as Mamba biases the model toward memorization over in-context learning yet aids length generalization, backed by toy theory and multi-scale evidence. The radar's field distribution splits evenly (0.5/0.5) between learning theory and LLM reasoning and deep learning theory and sequence models, yielding a focus score of 0. Collaboration scores 55 across 15 distinct co-authors, with a recurring core including Jason D. Lee, Peter Bartlett, and Nika Haghtalab. Reproducibility averages 55 over the two reviewed papers, both rated medium. Output, impact, and activity axes lack sufficient indexed data and are left unscored.
Collaboration pattern
Bibliometrics show 15 distinct co-authors with a collaboration score of 55; twelve recur often enough to be listed, including Khashayar Gatmiry, Aryo Lotfi, Vishnu Banna, Kostas Vaggelakos, Jason D. Lee, Peter Bartlett, Nika Haghtalab, Etai Littwin, Josh Susskind, Avrajit Ghosh, Parsa Mirtaheri, and William L. Tong. Both reviewed papers are middle-author team projects, suggesting a pattern of joining collective theory efforts rather than solo authorship.
Research keywords
Highlights
Latent chain-of-thought trained from answers alone
Gating steers state space models between memory and in-context learning
A theory-first program spanning reasoning traces and sequence models
Both reviewed papers carry medium repro-readiness and share one method: reduce an LLM or SSM behavior to a tractable model — parity-based attention or gated state dynamics — then analyze it; the even 0.5/0.5 field split reflects this two-front agenda.
Timeline
- 2026'Learn Your Own Thoughts: Abstract Token Curriculum' appears on arXiv, introducing a curriculum that trains latent, continuous chains of thought from answers alone.
- 2026'On the Importance of Gating' appears on arXiv, analyzing memorization versus in-context learning in state space models such as Mamba.
- 2026Both papers enter the Evolve catalog with medium repro-readiness; the two-paper base yields a reproducibility score of 55 and an even 0.5/0.5 field split.
Papers in the Evolve catalog
Frequent co-authors
About this profile
This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.
- Bibliometric scores (collaboration, focus, reproducibility) are computed from OpenAlex-indexed public bibliographic data only.
- Output, impact, and activity are unscored: no citation counts, publication volumes, or per-year output data are present in this snapshot.
- Evidence base is narrow: two 2026 arXiv preprints, both middle-author and non-corresponding; no journal or conference papers are reviewed here.
- Identity is resolved via ORCID 0000-0002-1831-656X and public bibliographic records; conclusions are limited to the reviewed papers and snapshot metrics.