Berkeley theorist spanning LLM reasoning and AI alignment: latent chain-of-thought curricula and RLHF distortion bounds
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—
The output axis measures productivity — total and per-year publication counts and preprint volume. No publication-count data is available in this snapshot, so the axis is left unscored rather than estimated from the two reviewed preprints.
Impact—
The impact axis measures citation-based influence — citation counts, h-index, and uptake of the work by others. No citation data is available in this snapshot, so the axis is left unscored and no influence level is inferred.
Focus0/100
This axis measures how concentrated a researcher's output is across fields. Here the field distribution is exactly even — machine learning (learning theory and LLM reasoning) at 0.5 and machine learning theory (AI alignment / social choice) at 0.5 — yielding a focus score of 0. The two reviewed 2026 papers sit on opposite sides of that split, so the zero reflects a deliberate two-track theory agenda rather than missing data.
Collaboration46/100
The collaboration axis scores 46, a mid-range value, from 9 distinct coauthors detected in this snapshot. The frequent-collaborator pool — Khashayar Gatmiry, Avrajit Ghosh, Parsa Mirtaheri, Jason D. Lee, Emmanuel Abbe, Peter Bartlett, Kazusato Oko, Annie Ulichney, and Han Bao — spans both fields of this profile, and both reviewed 2026 papers list the researcher in a middle, non-corresponding position, consistent with team-based theory work.
Activity—
The activity axis measures publishing cadence — papers per year and the recency and continuity of output over time. The snapshot lacks a multi-year publication record, so the axis cannot be computed and remains unscored.
Reproducibility55/100
Reproducibility scores 55, the mean across the 2 papers Evolve has assessed, both rated medium reproduction readiness. The Abstract Token Curriculum paper reports strong synthetic results plus a one-step attention theory under stylized assumptions, while the RLHF distortion paper provides tight Θ̃(βB+β) bounds with the caveat of loose constants — theoretical artifacts that demand moderate reimplementation effort rather than turnkey pipelines.
Research summary
Nika Haghtalab is a researcher at the University of California, Berkeley; this Evolve profile is built from a bibliometric snapshot keyed to ORCID 0000-0002-8612-2089 and two 2026 arXiv preprints reviewed in the Evolve catalog. The profile splits exactly evenly across two fields — machine learning (learning theory and LLM reasoning) and machine learning theory (AI alignment / social choice), each at 0.5 of the field distribution. On the reasoning side, 'Learn Your Own Thoughts: Abstract Token Curriculum' trains latent, continuous chain-of-thought from answers alone via a curriculum, with strong synthetic results and a one-step attention theory, though its assumptions are stylized and reproduction readiness is rated medium. On the alignment side, 'Distortion of AI Alignment Revisited: RLHF is a Decent Utilitarian Aligner' proves RLHF's exponential distortion stems from data/reference-policy mismatch and establishes tight Θ̃(βB+β) bounds, with loose constants as the main caveat; it is also rated medium. Collaboration scores 46 from 9 distinct coauthors detected, including Khashayar Gatmiry, Jason D. Lee, Emmanuel Abbe, and Peter Bartlett. Reproducibility averages 55 across the two assessed papers. Publication output, citation impact, and annual activity lack sufficient data and are left unscored, so this should be read as a theory-focused early snapshot rather than a complete career record.
Collaboration pattern
The snapshot detects 9 distinct coauthors, yielding a mid-range collaboration score of 46. Repeat collaborators include Khashayar Gatmiry, Avrajit Ghosh, Parsa Mirtaheri, Jason D. Lee, Emmanuel Abbe, Peter Bartlett, Kazusato Oko, Annie Ulichney, and Han Bao, a pool spanning both of this profile's fields. Both reviewed 2026 papers place the researcher in a middle, non-corresponding author position, pointing to team-based theory collaborations rather than solo-led projects.
Research keywords
Highlights
RLHF is a decent utilitarian aligner: distortion traced to data/reference-policy mismatch
Abstract Token Curriculum: latent chain-of-thought learned from answers alone
Timeline
- 2026Posts the arXiv preprint 'Distortion of AI Alignment Revisited: RLHF is a Decent Utilitarian Aligner', proving RLHF's exponential distortion stems from data/reference-policy mismatch with tight Θ̃(βB+β) bounds.
- 2026Posts the arXiv preprint 'Learn Your Own Thoughts: Abstract Token Curriculum', introducing curriculum-trained abstract tokens for latent chain-of-thought, with strong synthetic results and a one-step attention theory.
- 2026Both preprints enter the Evolve reviewed catalog, spanning the profile's two fields — learning theory and LLM reasoning alongside AI alignment and social choice — each rated medium for reproduction readiness.
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 dimensions are computed from OpenAlex-derived data only; publication output, citation impact, and activity axes lack sufficient data and are intentionally left unscored.
- The reviewed evidence base is two 2026 arXiv preprints in the Evolve catalog; other work by this researcher falls outside this snapshot.
- The reproducibility mean of 55 rests on only 2 assessed papers, both rated medium, so treat it as indicative rather than definitive.
- Identity is keyed to ORCID 0000-0002-8612-2089 and the UC Berkeley affiliation as provided; the profile is name-resolved from public bibliographic data.