Sugam Panthi

Co-founder @ MagnoliaEd · Research Assistant @ The University of Southern Mississippi

I work on LLM memory and evaluation, and on how people interact with LLMs. I also build software. This is the plain version of spanthi.com.

B.S. Computer Science, USM Honors College, expected May 2028. Advised by Rabab Abdelfattah. Based in Hattiesburg, MS.

Sugam Panthi

Sugam’s Research Trail (scroll sideways...)

ProjectPaperPostActive
Tool agents2026–nowEvaluation2026–nowLLM × human behavior2026–nowHeard statePOSTLying toolsPOSTOutcome MonitorsARXIVPaliOPEN SOURCEReader gapPOSTSame RankingEMNLP 2026 FINDINGSLying benchmarksPOSTRAG compressionARXIVEternal presentLAPSE POSTLAPSEARXIVPaste boundarySEAM POSTSEAMARXIV

Inspired by Philippe Laban’s research garden

Education

B.S. Computer Science, The University of Southern Mississippi, Honors College — Aug 2024 – May 2028 (expected)

GPA 4.0/4.0. Hattiesburg, MS.

Publications

  • Memory Consolidation Flattens the Temporal Shape of User Facts
    Sugam Panthi, Muhaiminul Yeamin, Siyan Luo, Rabab Abdelfattah · arXiv 2026 · arXiv:2609.36457 · PDF · Poster ·
    BibTeX 
    @misc{panthi2026lapse,
  title         = {Memory Consolidation Flattens the Temporal Shape of User Facts},
  author        = {Panthi, Sugam and Yeamin, Muhaiminul and Luo, Siyan and Abdelfattah, Rabab},
  year          = {2026},
  eprint        = {2609.36457},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  doi           = {10.48550/arXiv.2609.36457},
  url           = {https://arxiv.org/abs/2609.36457}
}
  • Can LLMs Separate Pasted Artifacts from User Speech? Absorption at Unmarked Prompt Seams
    Sugam Panthi, Muhaiminul Yeamin, Rabab Abdelfattah · arXiv 2026 · arXiv:2610.04210 · PDF · Poster ·
    BibTeX 
    @misc{panthi2026seam,
  title         = {Can {LLMs} Separate Pasted Artifacts from User Speech? Absorption at Unmarked Prompt Seams},
  author        = {Panthi, Sugam and Yeamin, Muhaiminul and Abdelfattah, Rabab},
  year          = {2026},
  eprint        = {2610.04210},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  doi           = {10.48550/arXiv.2610.04210},
  url           = {https://arxiv.org/abs/2610.04210}
}
  • Outcome Monitors: Recovery Affordances for Silent Tool Failures
    Sugam Panthi, Rabab Abdelfattah · arXiv 2026 · arXiv:2608.19303 · PDF · Poster ·
    BibTeX 
    @misc{panthi2026outcomemonitors,
  title         = {Outcome Monitors: Recovery Affordances for Silent Tool Failures},
  author        = {Panthi, Sugam and Abdelfattah, Rabab},
  year          = {2026},
  eprint        = {2608.19303},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  doi           = {10.48550/arXiv.2608.19303},
  url           = {https://arxiv.org/abs/2608.19303}
}
  • Compression Is Not Evaluation-Neutral: Fixed RAG Compression Can Distort Reader Comparisons
    Sugam Panthi, Rabab Abdelfattah · arXiv 2026 · arXiv:2606.21807 · PDF · Poster ·
    BibTeX 
    @misc{panthi2026ragcompression,
  title         = {Compression Is Not Evaluation-Neutral: Fixed {RAG} Compression Can Distort Reader Comparisons},
  author        = {Panthi, Sugam and Abdelfattah, Rabab},
  year          = {2026},
  eprint        = {2606.21807},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  doi           = {10.48550/arXiv.2606.21807},
  url           = {https://arxiv.org/abs/2606.21807}
}
  • Same Ranking, Different Winner: How Scoring Targets Shape LLM Memory Benchmarks
    Sugam Panthi, Rabab Abdelfattah · EMNLP Findings 2026 · arXiv:2605.24060 · PDF · Poster ·
    BibTeX 
    @inproceedings{panthi2026memorytargets,
  title         = {Same Ranking, Different Winner: How Scoring Targets Shape {LLM} Memory Benchmarks},
  author        = {Panthi, Sugam and Abdelfattah, Rabab},
  booktitle     = {Findings of the Association for Computational Linguistics: EMNLP 2026},
  year          = {2026},
  note          = {To appear},
  eprint        = {2605.24060},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  doi           = {10.48550/arXiv.2605.24060},
  url           = {https://arxiv.org/abs/2605.24060}
}

Experience

Projects

  • Pali — Open memory runtime for LLM apps and agent systems.
  • Deepseek-Go — A Deepseek API client library written for Go.
  • Ripple — Web application for creating, playing, and sharing flashcards in 3D.
  • Cohesion — AI-powered assistant for SQL schema generation.
  • Danime — OpenSource Discord bot with 1M+ users.
  • Costumary — Craft build journal for makers — references, materials, budgets, and build logs.

Honors & Awards

Service

  • Executive member, GDG on Campus, The University of Southern Mississippi

Skills

  • Languages: Go, Python, JavaScript, TypeScript, SQL
  • Frameworks & Libraries: React, Next.js, React Native, Flask, Django, LangChain, PyTorch
  • Infrastructure & Databases: Docker, Terraform, AWS Fargate, Firebase, MongoDB, Pinecone, PostgreSQL
  • Systems & Protocols: Kafka, Raft, gRPC, REST, Microservices
  • Tools & Practices: Git, GitHub Actions, CI/CD, Jira, Turborepo