ML theorist proving identifiability and safety certificates for representations of dynamical and physical systems.
4
papers here
4
reviewed by Evolve
4
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-09) 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 publication-count or venue-level output metrics are available for this profile yet; this axis would measure total research production — papers per year, publication venues, and authorship roles — across indexed databases.
Impact—
Impact data such as citation counts, h-index, or usage signals is not yet available in the input data; this axis would quantify the research community's uptake of the four cataloged papers.
Focus0/100
The focus score is 0 because the four cataloged papers split evenly across four sub-fields, each holding 25% of the field distribution: recurrent-network safety certification, identifiability theory, active sensing and sequential hypothesis testing, and thermodynamic structure discovery. Rather than a narrow specialty, the portfolio reads as one certification-and-identifiability program deliberately tested on four different subject domains.
Collaboration22/100
The collaboration score is 22, derived from just 2 distinct coauthors across the cataloged work. Changming Xu is the recurring partner among the frequent coauthors, and all four papers are first-authored without a corresponding-author credit. This is a compact, near two-person working mode, consistent with a small theory group, and it leaves the work's external verification network thin.
Activity—
Activity cannot be scored yet: the four cataloged papers all carry 2026 arXiv dates, but no multi-year record exists to measure cadence. This axis would track research tempo over time — submission streaks, gaps, and sustained engagement.
Reproducibility46/100
Reproducibility readiness averages 46 across the four reviewed papers. One paper — Certified Task-Conditioned Active Observability — is rated low, while the other three are rated medium. Evolve's verdicts explain the pattern: the theory is rigorous and audits honest, but evidence is largely synthetic, low-noise, and baseline-light, so independent replication currently requires substantial reimplementation effort.
Research summary
Linzhe Zhang is a machine learning theorist at Northeastern University whose 2026 output — four first-author arXiv papers reviewed in the Evolve catalog — pursues a single methodological ideal: learning representations whose limits can be proven. In 'What Can a Recurrent State Safely Forget?', a predictive quotient formalizes which recurrent hidden-state directions can be discarded without changing task semantics, backed by PAC certification and sample-complexity analysis. 'Discovering Physical Representation Languages' argues that a discrete physics ontology — including Maxwell's equations via discrete exterior calculus — is recoverable from anonymous experiments, with rigorous theory and honest audits. 'Blind Thermodynamic Ontology Discovery' extends the program to thermodynamics, showing coordinates identifiable up to contact gauge using contact geometry and Legendre duality. 'Certified Task-Conditioned Active Observability' brings the same certification mindset to sequential experimental design, with safe abstention and conformal calibration. Evolve's verdicts consistently praise theoretical soundness while flagging synthetic, low-noise, baseline-light evidence; reproducibility readiness averages 46 across the four papers (one rated low, three medium). The collaboration footprint is compact — two distinct coauthors, with Changming Xu the recurring partner — and the field distribution spreads evenly across four sub-areas at 25% each. Output, impact, and activity metrics are not yet available.
Collaboration pattern
All four cataloged papers are first-authored (none lists a corresponding-author role), and the collaboration footprint is very small: 2 distinct coauthors overall, with Changming Xu the recurring frequent coauthor. The pattern suggests a compact two-person theory group rather than a large lab or distributed consortium. For prospective collaborators this means direct access to the author, but also that independent verification and scaling of the empirical claims will likely require additional experimental partners.
Research keywords
Highlights
Certified abstention for task-conditioned active testing
Thermodynamic coordinates identifiable up to contact gauge
Provable recovery of a discrete physics ontology
Which recurrent states can safely forget
Timeline
- 2026Publishes 'What Can a Recurrent State Safely Forget?', formalizing safe forgetting in recurrent hidden states via a predictive quotient with PAC certification and sample-complexity analysis.
- 2026Publishes 'Discovering Physical Representation Languages', claiming provable recovery of a discrete physics ontology — including Maxwell's equations — from anonymous experiments.
- 2026Publishes 'Blind Thermodynamic Ontology Discovery from Anonymous Experiments', establishing identifiability of thermodynamic coordinates up to contact gauge via contact geometry and Legendre duality.
- 2026Publishes 'Certified Task-Conditioned Active Observability', bringing certified abstention and conformal calibration to sequential experimental design.
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 dimension scores come from a limited bibliometric snapshot; output, impact, and activity axes have insufficient underlying data and are left unscored.
- All four cataloged papers are 2026 arXiv preprints; peer-reviewed venue status is not reflected in the data.
- All four papers are first-author, non-corresponding work; authorship position beyond these four papers is unknown.
- Country, ORCID, and career history are not provided; this profile is name-resolved from public bibliographic data and may not capture the full record of researchers with similar names.