Oxford machine-learning researcher working on exact RoPE attention theory and robust robot world models
3
papers here
3
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
1
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-16) 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—
This axis cannot be scored yet: the available data lacks a reliable career-wide publication count beyond the six recent works, so publication volume and productivity over time remain unmeasured.
Impact—
Citation-based impact is unmeasured: with an h-index of 1 and i10-index of 0 the record is too early-stage, so this axis — which would track citations, influence and uptake — is left unscored.
Focus0/100
The focus score of 0 reflects maximal topical spread: the OpenAlex field weights are exactly one third each for transformer attention theory, transformer interpretability, and robot learning with world models. The three Evolve-reviewed papers still suggest two coherent clusters — exact RoPE attention calculus and world-model robustness to action parameterization — but the even 33/33/33 split across distinct subfields keeps the measured concentration at the floor.
Collaboration32/100
The collaboration score of 32 sits low: OpenAlex records only 4 distinct co-authors across the 6 works of the recent three-year window. All three Evolve-reviewed papers list Chlon in the last author position, and the recurring partner pool is small — Julie Huang, Maggie Chlon, Gregory Gutin and Ahmed Karim — indicating a compact, repeat collaboration network rather than broad multi-team work.
Activity57/100
The activity score of 57 reflects a recent, accelerating cadence: OpenAlex counts 6 works in the last three years, with the by-year record rising from 2 publications in 2025 to 4 in 2026. All three catalogued papers are 2026 arXiv preprints, suggesting current output is concentrated in fast-release preprint channels rather than journal or conference venues.
Reproducibility85/100
Reproducibility is the profile's strongest axis at 85, the mean readiness score across 3 assessed papers. Evolve rates the world-models and RoPE gradient-step papers high and the RoPE-derivatives paper medium; verdicts credit exact, verifiable mathematics, honest evaluation and large-scale intervention testing (92k executed edits on Qwen), while flagging limits such as 3 seeds and a single architecture.
Research summary
Leon Chlon is a researcher at the University of Oxford whose recent work spans three evenly weighted areas recorded in OpenAlex: transformer attention theory, transformer interpretability, and robot learning with world models — each accounting for one third of the field distribution in the profile snapshot. The catalogued output is recent and compact: six works in the last three years, with two recorded in 2025 and four in 2026, and an h-index of 1 with an i10-index of 0, indicating an early-stage citation record rather than an established one. Evolve has reviewed three 2026 arXiv papers, on all of which Chlon appears as last author. Two concern rotary position embeddings (RoPE): one derives an exact divided-difference identity showing RoPE-softmax attention as a forward-pass gradient step with softmax intact, and the other computes exact finite attention responses from RoPE derivatives, validated on Qwen through 92k executed interventions. The third shows that robot world models collapse by factors of 2.6–13.4 when action representations switch between absolute and delta parameterizations, and proposes repairs. Reproducibility is the profile's strongest signal, with a mean readiness score of 85 across 3 assessed papers; review verdicts nonetheless flag narrow model scope and small seed counts.
Collaboration pattern
Collaboration is compact: OpenAlex records 4 distinct co-authors over the recent three-year window, and the recurring set comprises Julie Huang, Maggie Chlon, Gregory Gutin and Ahmed Karim. On all three Evolve-reviewed 2026 papers Chlon holds the last author position, though none is marked corresponding — consistent with a senior-position, small-team working pattern rather than large consortium authorship.
Research keywords
Highlights
Exact divided-difference identity: RoPE attention as a forward-pass gradient step
Exact finite attention responses from RoPE derivatives, validated at 92k interventions
Robot world models collapse by 2.6–13.4× when action writing changes
Timeline
- 2025OpenAlex by-year counts record 2 works in 2025 — the earliest research activity shown in this profile snapshot.
- 2026OpenAlex records 4 works in 2026, the most active year to date; three are arXiv preprints reviewed by Evolve.
- 2026Publishes 'RoPE attention is an exact forward-pass gradient step with softmax intact', establishing an exact φ₁ divided-difference identity; Evolve rates reproducibility readiness high.
- 2026Publishes 'Exact Finite Attention Responses From RoPE Derivatives', validated on Qwen through 92k executed interventions; Evolve rates reproducibility readiness medium.
- 2026Publishes 'Robot World Models Are Not Invariant to How the Actions Are Written', documenting 2.6–13.4× collapses under absolute↔delta action switches; Evolve rates reproducibility readiness high.
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 indicators are computed from OpenAlex only; no ORCID is linked, so this profile is name-resolved from public bibliographic data.
- Output and impact dimensions are unscored because the citation record is too early-stage (h-index 1, i10-index 0).
- Only three 2026 arXiv papers carry Evolve review verdicts; other works in the by-year counts are not individually reviewed.
- Co-author statistics cover only the recent three-year window and may under-count earlier collaborations.