Jo\~ao Pedro Silvestre

—

Unclaimed · AI-structured profileOpenAlex

ML theorist applying control theory and information geometry to state space models and synthetic-data model collapse

papers here

reviewed by Evolve

first-author

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—

Output data is not yet available: publication counts and venue history cannot be verified from the current snapshot. This axis would measure total research output — paper counts, venue tiers, and first-author share.

Impact—

Impact data is insufficient: citation counts and downstream uptake are not on record in this snapshot. This axis would measure scholarly influence — citations, citing works, and adoption of the state space and model collapse results.

Focus0/100

The focus score is 0 on Evolve's 0–100 scale, derived from a field distribution split exactly 50/50 between two labeled areas: machine learning theory through dynamical systems and control theory, and machine learning theory through information geometry and generative-model training dynamics. Both halves sit inside theoretical machine learning, so the zero reflects an even two-cluster split rather than scattered, unrelated topics.

Collaboration36/100

The collaboration score is 36 out of 100, computed from 5 distinct co-authors across the reviewed papers. The network is small and concentrated: Paulo Tabuada appears twice in the frequent-coauthor list, signalling repeat collaboration, alongside Álvaro Rodríguez Abella, Matteo Marchi, and Bahman Gharesifard — consistent with work inside a single theory group rather than wide external partnering.

Activity—

Activity data is insufficient: the snapshot covers only 2026, so publication cadence and year-over-year trends cannot be computed. This axis would measure research tempo — outputs per year and the recency of contributions.

Reproducibility55/100

The reproducibility score is 55, the mean over 2 assessed papers, each rated “medium” reproducibility readiness. Both are proof-driven theory papers, which limits direct replication; Evolve's review of the Fisher-Rao work notes the absence of experiments, so empirical checkpoints would be needed before either line of results could be independently reproduced.

Research summary

João Pedro Silvestre is a machine learning theorist whose reviewed work sits at the intersection of dynamical systems and control theory on one side and information geometry on the other — an even 50/50 split in the field distribution of the two papers catalogued by Evolve. The first-author 2026 arXiv paper “The Attention Within” applies input-to-state stability and persistency-of-excitation arguments to show that Mamba-2’s recurrence drives token representations toward consensus, mirroring the uniformity attention induces; Evolve’s verdict credits the theorems as novel for time-varying weights while noting they exclude the gate and SiLU components. The co-authored 2026 paper “Preventing Model Collapse” proves Fisher-Rao convergence for training on mixed human and synthetic data, arguing the required human-data ratio is larger than a Euclidean analysis implied; Evolve judges the theory self-consistent but flags possibly loose constants and the absence of experiments. Both papers carry a “medium” reproducibility readiness, averaging 55 across the two assessments. The collaboration footprint is compact: 5 distinct co-authors and a collaboration score of 36, anchored by repeat collaboration with Paulo Tabuada, alongside Álvaro Rodríguez Abella, Matteo Marchi, and Bahman Gharesifard. Output, impact, and activity metrics are not yet available, and no institution, country, or ORCID is on record; the profile is name-resolved from public bibliographic data.

Collaboration pattern

The collaboration footprint is small and concentrated: 5 distinct co-authors across the reviewed papers, giving a collaboration score of 36 out of 100. Paulo Tabuada appears twice in the frequent-coauthor list, indicating repeat collaboration, alongside Álvaro Rodríguez Abella, Matteo Marchi, and Bahman Gharesifard. The pattern is consistent with sustained work within a single theory group rather than broad, cross-institutional partnering.

Research keywords

selective state space modelsMamba-2consensus dynamicstoken uniformity / over-smoothinginput-to-state stabilitypersistency of excitationmodel collapsesynthetic dataFisher-Rao metricinformation geometryLyapunov stabilityprobability simplex

Highlights

State space models develop an “attention” of their own
In a first-author arXiv paper, input-to-state stability proofs show that Mamba-2's recurrence drives token representations toward consensus — the same uniformity attention produces — with the analysis extending to time-varying weights, drawing on persistency-of-excitation style conditions.
The proofs stop short of Mamba-2's gated core
Evolve's verdict flags that the theorems exclude the gate and SiLU activation, so the consensus result covers the idealized recurrence rather than the exact architecture used in practice — a clearly scoped limitation for follow-up work.
Fisher-Rao geometry raises the bar for avoiding model collapse
In a co-authored arXiv paper, Fisher-Rao convergence is proved for training on mixed human and synthetic data, arguing the required human-data ratio is larger than a Euclidean analysis implied; the theory is self-consistent, but the headline scaling rests on possibly loose constants and no experiments.

Timeline

2026Posts the first-author arXiv preprint “The Attention Within: Consensus Dynamics in Selective State Space Models”, bringing input-to-state stability and persistency-of-excitation tools to Mamba-style architectures.
2026Co-authors the arXiv preprint “Preventing Model Collapse: A Fisher-Rao Perspective on the Dynamics of Training with Synthetic Data”, applying information geometry and Lyapunov stability on the probability simplex.
2026Both catalogued outputs appear in the same year and split evenly (50/50) between the dynamical systems/control theory and information geometry strands of machine learning theory.

Papers in the Evolve catalog

2026-09-17
Preventing Model Collapse: A Fisher-Rao Perspective on the Dynamics of Training with Synthetic Data
2026-09-17first author
The Attention Within: Consensus Dynamics in Selective State Space Models

Frequent co-authors

Paulo Tabuada×2 \'Alvaro Rodr\'iguez Abella×1 Matteo Marchi×1 Bahman Gharesifard×1 Paulo Tabuada×1

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 a bibliometric snapshot of Evolve-reviewed papers only; the output, impact, and activity axes lack sufficient data and are left unscored.
No institution, country, or ORCID is on record; the profile is name-resolved from public bibliographic data, and this is a relatively common Portuguese name, so identity assignment carries some ambiguity.
The reviewed catalog contains only two 2026 arXiv preprints; all assessments reflect Evolve's verdicts on these works, not a full publication record or peer-review history.
The reproducibility score (55) is a mean over two papers both rated “medium” readiness, and Evolve's review notes the Fisher-Rao paper includes no experiments.