Dartmouth researcher in graph adversarial robustness, neurosymbolic NLP, and LLM fine-tuning, often as senior author.
5
papers here
3
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—
The snapshot contains no publication-count or productivity metrics, so output is unscored. Once populated, this axis would measure total paper volume, yearly production rate, and the share of first- versus senior-authored works.
Impact—
No citation counts or influence indicators are present in the supplied data, so impact is unscored. This axis would capture citation totals, per-paper citation performance, and the broader uptake of the researcher's contributions across fields.
Focus0/100
The focus score is 0 because the field distribution is perfectly flat: graph machine learning (adversarial robustness), natural language processing / neurosymbolic AI, and machine learning (LLM fine-tuning) each hold exactly one third of the reviewed papers. The profile therefore reflects deliberately parallel research lines rather than a single concentrated specialty, with breadth across benchmarking, methods, and defense work.
Collaboration58/100
The collaboration score is 58, based on 17 distinct coauthors in the snapshot. Recurring partners include Nikhil Singh and Yaning Jia, alongside Omar Sharif, Shenyang Deng, and Chunhui Zhang. Combined with last-author positions on four of five catalog papers, the pattern suggests steady work with a compact collaborator pool in a senior capacity rather than large multi-team consortia.
Activity—
Papers-per-year and publication-cadence measures are absent from the snapshot, so activity is unscored. This axis would track publishing consistency, recent momentum, and gaps in output over time.
Reproducibility55/100
Reproducibility averages 55 across three evaluated papers, and each of the three is rated medium for reproducibility readiness. Evolve's reviews explain the middling level: PRISM-BN's reference Bayesian networks are LLM-generated and not externally validated, TrimSFT reports single runs without significance tests, and the edge-sanitization theory does not justify pruning high kernel-complexity edges.
Research summary
Soroush Vosoughi is a computer scientist at Dartmouth College whose recent catalogued work spans three areas in near-equal measure: adversarial robustness for graph machine learning, natural language processing with a neurosymbolic angle, and fine-tuning methods for large language models — the field distribution assigns exactly one third of reviewed papers to each. Five 2026 arXiv papers appear in Evolve's catalog, with Vosoughi last author on four and in a middle position on one, a pattern consistent with senior, supervisory authorship. Three papers carry full Evolve reviews and a medium reproducibility readiness, averaging 55. PRISM-BN contributes a controlled corpus and benchmark for extracting parameterized Bayesian networks from text; TrimSFT reweights supervised fine-tuning at the token level and improves mathematical reasoning across six models; and kernel-complexity edge sanitization offers a training-free defense against structural graph attacks. The remaining two papers address information extraction with learned queries and evidence utilization in long-term conversational memory. The snapshot records 17 distinct coauthors, including repeat partners Nikhil Singh and Yaning Jia, for a collaboration score of 58. Citation, productivity, and cadence metrics are unavailable, so impact and output are unscored; this profile rests on a small, very recent set of preprints.
Collaboration pattern
Seventeen distinct coauthors appear in the snapshot, giving a collaboration score of 58. Nikhil Singh and Yaning Jia recur in the frequency list, alongside Omar Sharif, Shenyang Deng, Yaoqing Yang, Chiyu Ma, Wenxuan Xu, Amartya Bhattacharya, Neeti Pokhriyal, and Chunhui Zhang. Combined with last-author positions on four of five catalog papers, the pattern points to steady work with a compact, recurring collaborator pool in a senior role, rather than large multi-institution consortia.
Research keywords
Highlights
PRISM-BN: a controlled benchmark for text-to-parameterized Bayesian network extraction
TrimSFT: token-level reweighting for supervised fine-tuning in math reasoning
Training-free graph defense via kernel-complexity edge sanitization
Timeline
- 2026Last-authored PRISM-BN, a controlled corpus and benchmark for text-to-parameterized Bayesian network extraction (arXiv).
- 2026Last-authored the TrimSFT token-trimming study on supervised fine-tuning for mathematical reasoning (arXiv).
- 2026Last-authored kernel-complexity edge sanitization, a training-free defense against structural graph attacks (arXiv).
- 2026Co-authored Improving Information Extraction with Learned Queries in a middle author position (arXiv).
- 2026Last-authored UTILMEM, a benchmark for evidence utilization in long-term conversational memory (arXiv).
Papers in the Evolve catalog
PRISM-BN: A Controlled Corpus and Benchmark for Text-to-Parameterized Bayesian Network Extraction
Which Tokens Should SFT Actually Learn? A Token-Trimming Perspective on Mathematical Reasoning
Kernel-Complexity Edge Sanitization for Training-Free Defense against Structural Graph Attacks
Improving Information Extraction with Learned Queries
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 only from the supplied snapshot; output, impact, and activity axes lack sufficient data and are unscored.
- All five catalog papers are 2026 arXiv preprints; no peer-reviewed venue or citation information is in scope.
- Two of the five papers (Improving Information Extraction with Learned Queries, UTILMEM) have no Evolve verdict, keywords, or reproducibility rating yet.
- Profile is name-resolved from public bibliographic data linked to ORCID 0000-0002-2564-8909; biography, country, and career history are out of scope.