University of Toronto NLP/IR researcher: LLM-judge reliability, peer-review claim verification, agentic deep search
3
papers here
3
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
0
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-16) 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: the current OpenAlex record lacks indexed publication totals, so cumulative output volume and its year-by-year trajectory cannot be measured. This axis would track total works and productivity trend once coverage is complete.
Impact—
Impact is unscored: h-index and i10-index are recorded as 0 and no citation counts are available, which typically reflects missing citation coverage rather than measured absence of influence. This axis would capture citation uptake, h-index, and scholarly reach.
Focus0/100
Focus scores 0/100, consistent with the raw field distribution, which splits evenly at about 33% each across LLM-based evaluation and scientometrics, scholarly NLP, and agentic deep search. OpenAlex topics reinforce the breadth: topic modeling and digital media forensic detection (4 works each), plus recommender systems, information retrieval/search behavior, and authorship attribution (3 each), suggest deliberate range rather than a single narrow line.
Collaboration51/100
Collaboration scores 51/100, grounded in 12 distinct co-authors in the recent window. The recurring roster — Negar Arabzadeh, Ebrahim Bagheri, Shakiba Amirshahi, Hai Son Le, Charles L. A. Clarke, Benjamin C. M. Fung and others — points to a stable, group-based NLP/IR working pattern. On all three Evolve-reviewed papers the researcher sits in a middle author position, indicating contribution within larger teams rather than solo or corresponding-author leadership.
Activity67/100
Activity scores 67/100. The recent three-year window contains 9 works, and the counts-by-year record shows 2026 as the only year with entries (9 works), matching the three 2026 arXiv preprints in the Evolve catalog. The profile is therefore front-loaded and recent: a concentrated burst of current output rather than a long, sustained multi-year publication trail.
Reproducibility70/100
Reproducibility averages 70/100 over 3 assessed papers. Evolve rates Peerify as high readiness (800 claims, 6 variants, audited supervision), while the LLM-judge metric reliability study and the agentic deep search paper are rated medium — solid statistical designs (Wilcoxon plus TOST; gpt-5.5 runs reaching 90.5 accuracy) but with untested ablations, nested-data issues, or judge circularity flagged.
Research summary
Sajad Ebrahimi is a researcher at the University of Toronto working at the intersection of natural language processing and information retrieval, with current work centered on the reliability of LLM-based evaluation in science. The Evolve catalog holds three 2026 arXiv preprints, each with the researcher in a middle author position: a reliability analysis reporting that 23 of 29 LLM-judge peer-review evaluation metrics conflate style with substance, tested with Wilcoxon signed-rank and TOST equivalence methods; Question's Gambit, which lifts Pi-Serini accuracy on BrowseComp-Plus from 83.1 to 90.5 with gpt-5.5 through clue-decomposed pre-loop retrieval; and Peerify, an 800-claim, six-variant benchmark for verifying peer-review claims against manuscripts. OpenAlex topic counts — topic modeling and digital media forensic detection (4 each), recommender systems, information retrieval/search behavior, and authorship attribution (3 each) — indicate a broad NLP/IR portfolio split evenly across three subfields. Recorded signals include 9 works in the recent three-year window, 12 distinct co-authors, a 70/100 reproducibility average over three assessed papers, and no citation-based impact measures yet.
Collaboration pattern
Collaboration is group-based: 12 distinct co-authors appear in the recent window, with a recurring core of Negar Arabzadeh, Ebrahim Bagheri, Shakiba Amirshahi, Hai Son Le, Charles L. A. Clarke and Benjamin C. M. Fung, suggesting a stable NLP/IR cluster around the University of Toronto. On all three Evolve-reviewed 2026 papers the researcher is a middle author, not corresponding — consistent with contributing experimental and analytical work inside larger teams.
Research keywords
Highlights
23 of 29 LLM-judge review metrics conflate style with substance
Peerify: benchmarking peer-review claim verification at scale
Clue-decomposed retrieval lifts agentic deep search accuracy to 90.5
Timeline
- 2026Co-authors (middle position) a reliability analysis of 29 LLM-based peer-review evaluation metrics, finding that 23 conflate style with substance.
- 2026Co-authors Question's Gambit, raising Pi-Serini accuracy on BrowseComp-Plus from 83.1 to 90.5 with gpt-5.5 through clue-decomposed pre-loop retrieval.
- 2026Co-authors Peerify, an 800-claim, six-variant benchmark for verifying peer-review claims against manuscripts, rated high for reproducibility readiness.
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.
- Bibliometrics come from an OpenAlex record (identity: openalex) in which h-index and i10-index are recorded as 0; citation coverage may be incomplete.
- Output and impact dimensions are unscored because the underlying productivity and citation data are insufficient, not because they were measured as zero.
- Evidence covers only the three 2026 arXiv preprints reviewed by Evolve; other, non-indexed or in-review work is not reflected.
- No ORCID is linked; the profile is name-resolved from public bibliographic data, so disambiguation of a common name cannot be fully guaranteed.