LLM agents for long-form speech error correction and machine-checked autoformalization of quantum-complexity theorems
2
papers here
1
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-09-27) 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 yet: this snapshot contains no publication-volume or venue metrics for the name. This axis would measure the total count of indexed papers, their venues and types, and how output accumulates over the researcher's career.
Impact—
No impact score yet: citation counts and citation-based indicators are absent from the snapshot. This axis would measure citations to the two 2026 arXiv papers and any broader uptake signal, none of which can be estimated here.
Focus100/100
Focus scores 100, reflecting a field distribution concentrated in a single mapped area — formal methods / autoformalization of quantum complexity theory — in the current snapshot. Thematically the two reviewed papers are equally tight: both apply LLM-agent pipelines to long-horizon correctness tasks, one recovering domain terminology in long-form ASR, the other machine-checking a MIP*=RE proof in Lean 4. No breadth penalty appears in the data.
Collaboration41/100
Collaboration scores 41 against 7 distinct coauthors in the snapshot, a small network spanning two separate clusters: a speech/ASR group (Wupeng Wang, Zhifu Gao, Xiangang Li, Xie Chen) on the first-author GER paper, and a quantum-information group (Sirui Lu, Ruixuan Deng, Zhengfeng Ji) on the autoformalization paper. Neither paper lists Zhu as corresponding author, consistent with a contributed-author position within each team.
Activity—
No activity score yet: beyond two 2026 arXiv entries, the snapshot carries no annual publication-velocity data. This axis would measure recent output cadence and continuity, which cannot be assessed from the current evidence.
Reproducibility100/100
Reproducibility is scored 100 (reproMean), but on a thin base: only one reviewed paper carries a readiness rating. The autoformalization work is rated high — its central artifact is a complete, machine-checked Lean 4 proof, the strongest form of verification — while the GER paper is rated medium, with ablations across 32 configurations but no significance tests. The score should be expected to move as more papers are reviewed.
Research summary
Yanqiao Zhu works at the intersection of large language model (LLM) agents and two demanding domains: long-form automatic speech recognition (ASR) and machine-checked formal mathematics. The Evolve catalog holds two 2026 arXiv papers. As first author, Zhu led 'Agentic GER: terminology recovery in long-form speech using global context', an LLM-agent generative error correction system that exploits full-transcript context to recover domain terminology; Evolve verified error reductions across 32 configurations, up to 36.8% B-CER, with a solid ablation study but no statistical significance tests. The second paper, in formal methods, contributes a complete machine-checked Lean 4 proof of the low individual-degree test (LIDT) quantum soundness result underlying MIP*=RE, produced with a multi-agent pipeline called FormalFlow; verification is strong but demonstrated on a single case study. Together the papers define a coherent identity: orchestrating LLM agents over long structured contexts — transcripts or proof states — to enforce correctness, whether measured by character error rate or a proof assistant. The radar shows maximal focus (100) and reproducibility (100, from one reviewed paper), a small coauthor network (7 distinct coauthors), and missing output, impact, and activity metrics — an early-stage but technically ambitious profile.
Collaboration pattern
Zhu's 7 distinct coauthors split into two clusters: a speech/ASR group (Wupeng Wang, Zhifu Gao, Xiangang Li, Xie Chen) around the first-author GER paper, and a quantum-information group (Sirui Lu, Ruixuan Deng, Zhengfeng Ji) around the autoformalization work. Zhu holds a first-author position on one paper and a middle position on the other, with neither listing a corresponding role — a small, two-team footprint typical of an early-stage researcher.
Research keywords
Highlights
Agentic GER: LLM agents recover domain terminology in long-form speech
Machine-checked Lean 4 proof of a core MIP*=RE theorem via multi-agent FormalFlow
Timeline
- 2026First-author arXiv paper on agentic generative error correction, recovering domain terminology in long-form ASR via full-transcript global context
- 2026Coauthored long-horizon autoformalization paper: a complete Lean 4 proof of LIDT quantum soundness underlying MIP*=RE, built with multi-agent FormalFlow
- 2026Two Evolve-reviewed arXiv papers span speech processing and formal methods, both built on LLM-agent pipelines for long-horizon correctness
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 are computed from a small OpenAlex-based snapshot covering only the two Evolve-reviewed arXiv papers; institution, country, and ORCID are absent, so the profile is name-resolved from public bibliographic data.
- The reproducibility score of 100 rests on a single reviewed paper (reproPapers = 1) and should not be read as a stable track record.
- Output, impact, and activity axes are unscored because no citation or publication-volume metrics are present in the snapshot.
- Both papers are 2026 arXiv preprints; no peer-reviewed venue information is available yet.