Amazon researcher at the intersection of LLM code agents, formal verification, and neurosymbolic AI
3
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-17) 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—
Insufficient data: no publication-count or venue-volume records are available in this snapshot, so output is unscored. This axis would measure total scholarly production — papers over time and their spread across venues — once fuller bibliometric coverage is linked.
Impact—
Insufficient data: citation counts and influence metrics are absent from this snapshot, so impact is unscored. The axis would measure how strongly the community cites and builds on this researcher's work relative to field norms, once citation records are linked.
Focus0/100
The focus score is 0, computed from a field distribution split exactly 50/50 between software engineering (LLM code agents and formal verification) and neurosymbolic AI / machine learning. Rather than a data gap, this reflects a two-sided research identity: one line builds machine-checked benchmarks and proof pipelines around LLM coding agents, the other applies generative reward models and distillation to formalization, both converging on autoformalization.
Collaboration62/100
Collaboration scores 62, based on 22 distinct coauthors across the cataloged work — a mid-range breadth. Recurring partners include George Ma, Benjamin Mikek, Haoyu Li, Yuhao Zhang, and Zeren Shui, alongside Behrooz Omidvar Tehrani, Jun Huan, Murali Krishna Ramanathan, Somayeh Sojoudi, Hao Zhou, and Anoop Deoras. On the reviewed papers Erata appears in middle-author positions without a corresponding role, consistent with team-based engineering research rather than solo authorship.
Activity—
Insufficient data: no year-by-year publication cadence is available, so activity is unscored. The axis would measure ongoing research tempo — recent papers per year and output gaps — beyond the three 2026 preprints cataloged here.
Reproducibility55/100
Reproducibility averages 55 across the 2 papers Evolve could score, and both reviewed preprints carry medium reproducibility readiness. Reviews credit rigorous multi-gate benchmark construction for SWE-Proof and a rigorous, candid evaluation for the reward-model work, while flagging LLM-judged audits, single-model results, and synthetic hard negatives as limits on independent replication.
Research summary
Ferhat Erata is a researcher at Amazon (United States) whose cataloged work sits at the intersection of software engineering — specifically LLM code agents and formal verification — and neurosymbolic AI / machine learning, with the field distribution splitting exactly 50/50 between the two. The Evolve catalog lists three 2026 arXiv preprints. The first, SWE-Proof, claims to be the first formally verified SWE-bench benchmark, assembled through rigorous multi-gate construction combining specification synthesis and theorem proving; Evolve's review credits that rigor but tempers the headline finding because audits are LLM-judged and results rest on a single model. The second, Beyond Solver Verdicts, introduces generative reward models (GenV+HN) that detect solver-invisible unfaithful SMT encodings at 0.961 AUROC — an evaluation Evolve calls rigorous and candid, though synthetic hard negatives temper generality. A third preprint applies stratified consistency distillation to natural language formalization and awaits review. Across the reviewed papers Erata appears as a middle author, with no corresponding-author role recorded. The collaboration dimension scores 62 from 22 distinct coauthors, including recurring partners such as George Ma, Benjamin Mikek, Haoyu Li, Yuhao Zhang, and Zeren Shui. Reproducibility averages 55 across the 2 scorable papers, both rated medium readiness. Output, impact, and activity metrics are not yet available and remain unscored.
Collaboration pattern
Erata's cataloged work involves 22 distinct coauthors (collaboration score 62), with middle-author positions on all three 2026 preprints and no corresponding-author role recorded. Recurring partners include George Ma, Benjamin Mikek, Haoyu Li, Yuhao Zhang, and Zeren Shui, plus Behrooz Omidvar Tehrani, Jun Huan, Murali Krishna Ramanathan, Somayeh Sojoudi, Hao Zhou, and Anoop Deoras — a pattern of sizeable, team-written preprints spanning formal-methods and machine-learning groups.
Research keywords
Highlights
SWE-Proof: LLM issue resolution with machine-checked proofs
Generative reward models catch solver-invisible unfaithful encodings
Stratified consistency distillation for natural language formalization
A third 2026 arXiv preprint, cataloged without an Evolve review, keywords, or a reproducibility rating. Its title indicates work on using stratified consistency distillation to improve the conversion of natural language into formal statements; conclusions await independent review.
Timeline
- 2026Publishes SWE-Proof on arXiv, claiming the first formally verified SWE-bench benchmark.
- 2026Publishes Beyond Solver Verdicts, introducing GenV+HN reward models that flag unfaithful SMT encodings at 0.961 AUROC.
- 2026Posts a third preprint on stratified consistency distillation for natural language formalization, not yet reviewed by Evolve.
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 dimensions are computed from OpenAlex-derived data only; the output, impact, and activity axes lack sufficient data and are left unscored.
- Coverage is limited to three 2026 arXiv preprints in the Evolve catalog; peer-reviewed venue history is not represented.
- The profile is name-resolved from public bibliographic data (ORCID 0000-0001-6305-4266); attribution of all listed papers to a single person should be treated with care.
- The third preprint, Stratified Consistency Distillation for Natural Language Formalization, has no Evolve verdict, keywords, or reproducibility rating yet.