LLM interpretability researcher: transformer superposition and attention–recurrence roles in hybrid language models.
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
0
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—
Not yet scored: this axis would measure total publication volume and per-year productivity. Available data (h-index 0, i10-index 0, yearly counts only for 2025–2026) is insufficient to compute a reliable output score.
Impact—
Not yet scored: this axis would measure citation-based influence such as citations per paper and h-index trajectory. The snapshot shows h-index and i10-index of 0 with no citation counts, so no reliable impact score can be computed.
Focus0/100
Radar scores focus at 0, but the underlying field distribution is narrow: fieldDist splits evenly (0.5/0.5) between 'natural language processing / mechanistic interpretability' and 'natural language processing (LLM interpretability)', two adjacent labels. Topic counts add topic modeling (6), explainable AI (5), NLP techniques (4), adversarial robustness (4), and neural network applications (2). Both reviewed papers sit squarely inside LLM interpretability, so the gap between the 0 score and this concentrated spread is worth noting when reading the radar.
Collaboration50/100
Mid-range score of 50, grounded in 11 distinct co-authors across 9 works in the last three years. The frequent-coauthor list is dense and stable — Elena Tutubalina, Pavel Tikhonov, Anton Razzhigaev, Temurbek Rahmatullaev, Ivan Oseledets and others recur — a pattern consistent with team-written arXiv preprints rather than solo work. Across the two catalog papers his position shifts from middle to last author.
Activity67/100
Scores 67 on recent productivity: OpenAlex counts 3 works in 2025 rising to 6 in 2026, with 9 works over the last three-year window. Both reviewed papers appeared on arXiv in 2026, indicating an active, accelerating phase of output concentrated in LLM interpretability.
Reproducibility78/100
Scores 78, computed from Evolve's reproducibility review of 2 papers (reproMean 78). The hybrid attention–recurrence paper rates high reproducibility readiness, supported by clean cache-level intervention experiments; the linear-superposition paper is rated medium, since gains center on predictable tokens and decoded quality lags single-stream generation.
Research summary
Anton Korznikov is a researcher at Air University (RU) working in natural language processing, focused on the internal mechanics of large language models. His Evolve-reviewed output centers on two 2026 arXiv preprints: 'Your Transformer Can Hold Two Thoughts at Once: Evidence of Linear Superposition in LLMs' (middle author), which presents broad, controlled evidence that LLMs linearly superpose two information streams, and 'What Attention Recalls and Recurrence Controls in Hybrid Language Models' (last author), which uses cache-level interventions to separate attention's retrieval role from recurrence's control of style and persona. OpenAlex bibliometrics show 9 works in the last three years, with yearly counts rising from 3 (2025) to 6 (2026) and 11 distinct co-authors; recurring collaborators include Elena Tutubalina, Ivan Oseledets, and Anton Razzhigaev. Topic frequencies point to topic modeling (6), explainable AI (5), NLP techniques (4), and adversarial robustness (4), while his field distribution splits evenly between mechanistic interpretability and LLM interpretability labels. Radar scores: activity 67, reproducibility 78 (mean over 2 repro-scored papers), collaboration 50, focus 0; output and impact are not yet scored — h-index and i10-index both read 0. This is an early-stage profile: both reviewed papers are recent preprints and citation-based measures have not accumulated.
Collaboration pattern
Collaboration score of 50 sits mid-range. OpenAlex logs 11 distinct co-authors across 9 recent works, and the frequent-coauthor list is dense: Elena Tutubalina, Pavel Tikhonov, Anton Razzhigaev, Temurbek Rahmatullaev, Ivan Oseledets, Kirill Afendulev, Alexey Dontsov, Matvey Mikhalchuk, Nikita Dragunov, and Polina Druzhinina. The pattern suggests a stable, team-based interpretability group producing multi-authored arXiv preprints, with Korznikov moving between middle and last author positions.
Research keywords
Highlights
Evidence of linear superposition in LLMs
Splitting attention from recurrence in hybrid language models
A reproducibility-first early profile
Both catalog papers carry Evolve reproducibility scoring, averaging 78 (reproMean over 2 papers), with readiness ratings of high and medium — a strong signal for an early-stage interpretability researcher whose citation record has not yet accumulated.
Timeline
- 2025OpenAlex records 3 works, the earliest year covered in the provided output window.
- 2026Output rises to 6 works; the linear-superposition preprint appears with Korznikov in the middle author position.
- 2026Last-author preprint on attention versus recurrence in hybrid language models appears, rated high for reproducibility 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.
- Bibliometrics come from OpenAlex only (identity: openalex); no ORCID is linked, so the profile is name-resolved from public bibliographic data and 'Anton Korznikov' may be a shared name.
- Output and impact axes are unscored: h-index and i10-index both read 0 and no citation counts appear in the snapshot.
- Only 2 Evolve-reviewed papers, both 2026 arXiv preprints; no peer-reviewed venue information is available.
- countsByYear covers only 2025–2026, so longer-term productivity and career trajectory cannot be assessed.