Shreyansh Padarha

Shreyansh Padarha

PhD Researcher · University of Oxford

I am a PhD researcher at the University of Oxford, based in the Reasoning with Machines Lab (OxRML) and supervised by Professor Adam Mahdi. I build evaluations and post-training methods that stress-test today's AI, with the aim of making it more useful without making it less safe.

AI agents (coding, multimodal, physical and more) can now do a great deal. I want to understand which parts of scientific research we can hand over to them, and how to check answers that sound right but have no ground truth to test against. To do this, my work spans evaluation design, post-training and AI safety, often testing multimodal and agentic systems. My research has featured at AI venues such as ICML, NeurIPS and EMNLP.

Branches I am currently interested in:

  • Models that improve models (others or themselves). How AI agents train other models when they must balance several machine learning objectives at once, often competing ones. And how well they do it.
  • Discovery and integrity. Which stages of training decide a model's ethics. Whether models trained without ethical limits make better scientific discoveries. Whether a model could learn to act ethically in training but not at test time.
  • Uncertainty and conflicting evidence. How models represent uncertainty and weigh conflicting information, and how this shapes their conclusions. Do they always take the same scientific route? Can we steer their early calls to prevent failures downstream? [Maybe LLMs cannot, and this is a job for world models?]
  • Broader use of AI in research. Whether automating research helps or harms science overall. [That needs a crystal ball, but smaller use cases can be tested piece by piece.] What balance of human and AI work is best, and what kind of feedback, in what form, gets the best results. [Scale, and confounds from the harness itself, make this hard to answer.]

Research

Selected publications. * Equal contribution. Full list on Google Scholar.

  1. Evaluating AI-based Scientific Knowledge Synthesis with Epidemiological Systematic Reviews

    S. Padarha*, R. O. Kearns*, T. Naidoo, L. Yang, Ł. Borchmann, P. Błaszczyk, C. Morgenstern, R. McCabe, S. Bhatia, P. H. Torr, J. Foerster, S. A. Hale, T. Rawson, A. Cori, E. Semenova, A. Mahdi

    NeurIPS 2026 Paper Code Dataset Website

    Bar chart of systematic review completion time: about 385 hours for humans against about 20 hours for AgentSLR.

  2. Agentic Reinforcement Learning for Search Misaligns Instruction-Tuning

    Y. Yang, S. Padarha, S. Ball, A. Lee, A. Mahdi

    EMNLP 2026 Paper Code

    Diagnosis, mechanism and mitigation: RL-trained search models issue harmful searches under a search trigger, RL shifts their queries toward a harmful direction, and a penalty during RL restores safety.

  3. Strategic Navigation or Stochastic Search? How Agents and Humans Reason Over Document Collections

    Ł. Borchmann, J. Van Landeghem, M. Turski, S. Padarha, R. O. Kearns, A. Mahdi, et al.

    ICML 2026 · Oral, top 0.5% Paper Code Dataset Leaderboard

    Accuracy against step limit for humans, Gemini 3 Pro, Claude Sonnet 4.5 and GPT-5. Humans start far higher; agents catch up only with more search steps.

  4. Running the Gauntlet: Challenging Agentic Tasks

    M. Vysotskyi*, R. Lin*, G. Biziel, …, S. Padarha, …, A. Mahdi, V. Karpiv, P. Torr, A. Bibi

    Paper Code Website

    The five GauntletBench applications (Video Editor, Workflow Builder, 3D Modeller, Flight Analyser and Circuit Designer), each with an example task.

  5. Evaluating LLM-as-a-Judge under Multilingual, Multimodal and Multi-domain Constraints

    S. Padarha, E. Semenova, B. Vidgen, A. Mahdi, S. A. Hale

    NeurIPS 2025 Workshop on Evaluating the Evolving LLM Lifecycle Paper

    Four panels comparing each judge model's accuracy in English and in other languages across three tasks.

  6. Fully Automated Deep Learning Framework for Detection and Classification of Impacted Mandibular Third Molars in Panoramic Radiographs

    S. K. Veerabhadrappa, S. Vengusamy, S. Padarha, K. Iyer, S. Yadav

    Journal of Oral Medicine and Oral Surgery Paper

    Panoramic X-ray before and after processing, with the detected wisdom teeth labelled vertical and distoangular.

Experience

  1. Sep 2025 – Sep 2026
    Research Assistant · Oxford Internet Institute, University of Oxford

    Worked with Professor Scott Hale and Professor Adam Mahdi on an EPSRC-funded study of how British news outlets cover migration. Also ran LLM evaluations on safety, multimodality and agents.

  2. May – Jul 2025
    Research Assistant (part-time) · Oxford Internet Institute, University of Oxford

    Worked with Dr Fabian Braesemann on the Science of Startups Initiative. Built the technical stack for a website on founder personalities.

  3. Jan – May 2024
    Researcher · Asia Pacific University of Technology & Innovation, Kuala Lumpur

    Worked with Dr V. Sivakumar and clinicians in Malaysia and the Philippines. Built a deep learning system that finds impacted wisdom teeth and related conditions in dental X-rays.

  4. Jun – Sep 2023
    AI Engineer · Anantadi Tech, Pune

    Led a team of interns building an AI tool for in-video advertising and product placement. Built computer vision models for object detection, segmentation and tracking.

  5. Jul – Aug 2022
    Automation Engineer · Elecsol Technology, Raipur

    Worked on industrial automation, including PLCs and drives. Built an IoT dashboard to monitor older equipment in real time.

About me

Download CV (PDF)

I was a data scientist first [a term used loosely these days, I know], before I fell in love with AI research. As an undergraduate, I applied machine learning to healthcare, agriculture and industry [including entertainment: video post-production!]. One lesson from those early expeditions: an AI system is only useful if people can trust it [which often has little to do with benchmark numbers]. That idea still motivates my research.

Education

  1. 2026 – present
    DPhil Social Data Science · University of Oxford

    Evaluating Automated Research · Wadham College

  2. 2024 – 2025
    MSc Social Data Science · University of Oxford

    Distinction (73%) · Kellogg College

  3. 2021 – 2024
    BSc Data Science · CHRIST University, India

    Gold Medalist (3.93/4.0)

Grants & awards

  1. 2026 – present
    ESRC Grand Union Doctoral Training Partnership

    Full DPhil studentship from the Economic and Social Research Council.

  2. 2025 – present
    EPSRC (UKRI) and Dieter Schwarz Foundation

    Grants funding my research at the Oxford Internet Institute.

Teaching

  1. 2026 – 2027
    Teaching Assistant · Oxford Internet Institute, University of Oxford

    Applied Machine Learning with LLMs, a postgraduate course. Designed and led tutorials for 40 students in Hilary Term 2026. Returning in Hilary Term 2027.

  2. 2025
    Tutor · Stanford University (Stanford House, Oxford)

    Taught LLM pre-training and post-training to Stanford computer science undergraduates studying in Oxford.