Control theorist applying stability analysis to machine learning: Mamba-2 consensus and synthetic-data collapse
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-09-20) 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—
No output score is available: the snapshot lacks total publication counts and per-year volume beyond the two cataloged papers. This axis would measure aggregate papers, venues, and authorship positions over the career.
Impact—
Impact cannot be scored: no citation counts, h-index, or venue influence metrics appear in the supplied data. This axis would measure citations to the two preprints, uptake by later work, and field-level influence.
Focus0/100
The focus dimension scores 0 out of 100, driven by a field distribution that splits exactly 50/50 between machine learning theory (dynamical systems / control theory) and machine learning theory (information geometry / generative model training dynamics). Both halves sit inside machine learning theory, so the perfectly flat split may understate topical coherence; still, the metric records no dominant single subfield in this snapshot.
Collaboration36/100
The collaboration dimension scores 36 out of 100, with the raw snapshot recording 5 distinct coauthors. The recurring circle named in the data includes João Pedro Silvestre, Álvaro Rodríguez Abella, Matteo Marchi, and Bahman Gharesifard; on both cataloged 2026 papers Tabuada sits in the last author position without being corresponding author, a pattern consistent with senior supervision inside a small, stable team rather than a large distributed network.
Activity—
Activity is unscored: the only recency signal in the data is two 2026 arXiv preprints, which is too thin to estimate publication cadence. This axis would measure outputs per year and gaps in research activity.
Reproducibility55/100
Reproducibility scores 55 out of 100, computed as the mean readiness (reproMean 55) across the 2 papers Evolve has reviewed (reproPapers 2). Both papers are individually rated medium readiness: the Mamba-2 consensus proofs rely on stated assumptions that exclude the gate and SiLU, while the Fisher-Rao model-collapse result currently ships with no experiments, so both would need added implementation detail before independent replication.
Research summary
Paulo Tabuada is a researcher working at the intersection of control theory and machine learning theory. The Evolve catalog contains two 2026 arXiv preprints, both with Tabuada in the last author position and neither marked as corresponding author. The first, 'The Attention Within: Consensus Dynamics in Selective State Space Models', uses input-to-state stability and persistency-of-excitation arguments to show that Mamba-2's recurrence drives token consensus much as attention does — a result the review calls novel for time-varying weights, while noting the theorems exclude the gate and SiLU nonlinearity. The second, 'Preventing Model Collapse', proves Fisher-Rao convergence for mixed human/synthetic training and argues the required human-data ratio exceeds what a Euclidean analysis implies; the theory is judged self-consistent, but its headline scaling rests on possibly loose constants and no experiments. The catalog field split is exactly 50/50 between the dynamical-systems/control and information-geometry/training-dynamics strands of ML theory, producing a focus score of 0. Reproducibility averages 55 across the two scored papers, both rated medium. The radar records 5 distinct coauthors and a collaboration score of 36.
Collaboration pattern
The radar records 5 distinct coauthors and a collaboration score of 36, pointing to a compact, recurring circle rather than a wide network. The frequent-coauthor list names João Pedro Silvestre, Álvaro Rodríguez Abella, Matteo Marchi, and Bahman Gharesifard — the researcher's own name also appears in the raw list, which reads as a resolution artifact. On both 2026 catalog papers Tabuada takes the last author position without the corresponding-author flag, consistent with a senior supervisory role.
Research keywords
Highlights
Reading attention inside Mamba-2: consensus dynamics via input-to-state stability
A Fisher-Rao route to preventing model collapse in synthetic-data training
Timeline
- 2026Last-author arXiv preprint 'The Attention Within: Consensus Dynamics in Selective State Space Models' applies input-to-state stability and persistency of excitation to Mamba-2.
- 2026Last-author arXiv preprint 'Preventing Model Collapse: A Fisher-Rao Perspective on the Dynamics of Training with Synthetic Data' frames model collapse on the probability simplex with Lyapunov stability tools.
- 2026Both cataloged preprints receive medium reproducibility ratings, and the reviewed work splits 50/50 across the two machine-learning-theory subfields recorded in the radar.
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.
- Radar scores are computed from public bibliometric data (OpenAlex-style) only; the snapshot carries no institution, country, or ORCID.
- Only the two Evolve-reviewed papers were examined in depth; the catalog does not cover the full publication record, so output, impact, and activity remain unscored.
- The frequent-coauthor list includes the researcher's own name — treated here as a name-resolution artifact and excluded from named collaborators.
- This profile is name-resolved from public bibliographic data; identity consistency across the two cataloged papers was not independently verified.