Northeastern University researcher proving identifiability and safety guarantees for scientific machine learning
3
papers here
3
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
8
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—
Not scored: the underlying publication and citation ledger is too incomplete to rate output reliably. This axis would measure publication volume and productivity; the snapshot lists an h-index of 8 and i10-index of 7 but little verified per-output metadata.
Impact—
Not scored: citation and influence signals are insufficient in this snapshot. This axis would measure citation impact, uptake, and reach of the work; until those records are complete, no estimate is offered here.
Focus0/100
The radar assigns focus a score of 0 on the 0–100 scale, and the field distribution explains why: the record splits evenly three ways — machine learning theory (recurrent networks, safety certification), ML/scientific ML identifiability theory, and thermodynamic structure discovery — each at about 33%. Read alongside the three cataloged papers, this reads less like drift and more like one theory agenda viewed from three angles, though the formal concentration metric cannot see that.
Collaboration14/100
Collaboration scores 14, the weakest axis: the record lists just 1 distinct coauthor, with Linzhe Zhang the only recurring partner. All three cataloged papers place Xu in the last author position without corresponding-author flags, consistent with a small, tightly held theory duo rather than a distributed lab.
Activity91/100
Activity scores 91: the record shows 22 works in the last three years, and per-year counts jump from 1–5 annually across 2009–2024 to 9 works in 2025 and 8 in 2026. The three cataloged 2026 arXiv theory papers confirm an unusually intense, front-loaded publishing tempo in the current period.
Reproducibility55/100
Reproducibility scores 55, computed from 3 audited papers — all three cataloged works are rated medium readiness. Evolve's reviews note rigorous theory and honest audits but flag low-noise synthetic evidence and light baselines, so artifacts look reconstructible yet remain independently unvalidated; the middling score reflects that gap between theoretical care and empirical verifiability.
Research summary
Changming Xu is listed here with Northeastern University (US), name-resolved from public bibliographic data, and the Evolve catalog holds three of this researcher's 2026 arXiv papers, each with Xu in the last author position. All three are theory-first contributions to scientific machine learning: one argues thermodynamic coordinates are identifiable up to a contact gauge using contact geometry and Legendre duality; another claims provable recovery of a discrete physics ontology — engaging Maxwell's equations and discrete exterior calculus — from anonymous experiments; the third formalizes safely-forgettable RNN hidden-state directions through a predictive quotient with PAC audits. The bibliometric snapshot reports an h-index of 8, an i10-index of 7, and 22 works in the last three years, with annual output accelerating sharply — 9 works in 2025 and 8 in 2026 per the record, versus 1–5 annually across 2009–2024. Collaboration is minimal — a single distinct coauthor, with Linzhe Zhang the recurring partner — and reproducibility averages 55 across the three audited papers, all rated medium readiness. Evolve's reviews consistently praise rigor, elegance, and honest auditing while flagging synthetic-only, low-noise, or baseline-light evidence, so empirical validation at scale is the visible next step. Caution: OpenAlex topic labels for this name skew toward game AI and databases, so identity resolution is not settled.
Collaboration pattern
The collaboration footprint is strikingly narrow: the record counts just 1 distinct coauthor in recent work, and Linzhe Zhang is the only recurring collaborator listed. All three cataloged papers place Xu in the last author slot with no corresponding-author flag. This pattern fits a compact theory duo iterating densely together rather than a large lab; broader co-authorship network data is not part of this snapshot.
Research keywords
Highlights
Provable recovery of a discrete physics ontology from anonymous experiments
Thermodynamic coordinates identifiable up to contact gauge
A predictive quotient for what an RNN state can safely forget
Timeline
- 2026Posts 'Blind Thermodynamic Ontology Discovery from Anonymous Experiments' to arXiv, claiming thermodynamic coordinates are identifiable up to a contact gauge from four experiments.
- 2026Posts 'Discovering Physical Representation Languages' to arXiv, claiming provable recovery of a discrete physics ontology from anonymous experiments with audits Evolve deems honest though largely synthetic.
- 2026Posts 'What Can a Recurrent State Safely Forget?' to arXiv; all three cataloged papers enter the Evolve catalog with medium reproducibility readiness and a mean score of 55.
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 scores come from an OpenAlex name-resolved snapshot; 'Changming Xu' is a common name, so some records may belong to different people.
- OpenAlex topic labels for this record emphasize game AI, databases, and RL robotics, which does not match the three cataloged scientific-ML theory papers — a probable entity-resolution mismatch; treat topic history with caution.
- Output and impact radar axes are unscored because the citation ledger in this snapshot is incomplete; only h-index 8 and i10-index 7 are available.
- Reproducibility mean (55) rests on only 3 audited papers, all arXiv preprints whose peer-review status is not captured in the data.