Shreyas Krishnan

Research profile

Research summary

Shreyas Krishnan is a researcher at the University of California, Berkeley whose cataloged work centers on evaluating and interpreting large language models in high-stakes, socially embedded settings. The Evolve catalog holds three 2026 arXiv papers spanning computational social science, NLP/LLM evaluation, and computational psycholinguistics/clinical NLP safety evaluation — each exactly one-third of the recorded field distribution. Two are first-authored: ORQA, an automated pipeline producing a 480-item source-traceable occupational QA benchmark grounded in O*NET and GDPval; and Beyond the Flag, a controlled benchmark showing that content moderation APIs flag suicide-related content but mis-measure severity against C-SSRS-based ordinal ground truth. A third, middle-authored paper applies sparse autoencoders, linear probes, and activation steering to LLM agents in social simulations. Bibliometric signals are early-stage: h-index 2, i10-index 1, three works in the last three years, six distinct coauthors, and a reproducibility score of 55/100 with medium readiness on all three papers. The OpenAlex record also carries topics in melanin and skin pigmentation, genetic mapping, and livestock traits with pre-2020 publication years that do not match the cataloged NLP work — flagged as possible name-resolution noise. This profile is name-resolved from public bibliographic data (OpenAlex).

Research keywords

Highlights

• ORQA: a 480-item, source-traceable occupational benchmark for LLM professional knowledge

  First-authored 2026 arXiv work: an automated pipeline built from O*NET occupations and GDPval-style tasks yields a 480-item QA benchmark in which every item traces to a source. Evolve calls it scalable and honest, while noting 67% filter agreement and thin contamination testing; reproducibility readiness medium.

• Clinical framing closes the moderation gap in suicide risk measurement

  First-authored benchmark showing that content moderation APIs reliably flag suicide-related content yet mis-measure severity; framing queries in clinical terms closes the gap against C-SSRS-based ordinal ground truth. Evolve finds the design honest and well-controlled, with the prompt-rubric circularity flagged by the authors themselves.

• Interpreting and steering LLM agents for social simulations

  Middle-authored proof of concept applying sparse autoencoders, linear probes, and activation steering to open the black box of LLM agents across four social-simulation tasks. Evolve finds gains over strong prompting mixed and statistically untested; reproducibility readiness medium.

Timeline

1. 2026First-authored 'Beyond the Flag', a controlled clinical NLP benchmark testing moderation APIs against C-SSRS-based ordinal suicide-risk labels.
2. 2026First-authored ORQA, releasing a 480-item source-traceable QA benchmark for LLM professional knowledge.
3. 2026Co-authored (middle position) 'Interpreting and Steering LLM Agents for Social Simulations', applying sparse-autoencoder probes and activation steering.
4. 2026Recorded 3 works in the recent three-year window — all 2026 arXiv preprints cataloged and reviewed by Evolve.

Papers in the Evolve catalog

• 2026-09-14

  Interpreting and Steering LLM Agents for Social Simulations

• 2026-09-11first author

  ORQA: An Occupation-Realistic Question and Answer Framework for LLM Professional Knowledge

• 2026-09-05first author

  Beyond the Flag: Clinical Framing Closes the Moderation Gap in Suicide Risk Measurement

Frequent co-authors