Research
Accepted papers, manuscripts, reports and earlier peer-reviewed work in AI safety, interpretability and healthcare.
Single author. Accepted as a poster (non-archival), to be presented Dec 2026
What an interpercentile decomposition of multi-turn evaluation results can and cannot separate, tested on conversations generated across seven models.
Iliad Research Intensive, 2026
A reliability audit of my own test of natural-latent theory on sparse-autoencoder features. Two innocuous estimator choices move the same data between three different headlines, and a pre-registered null that the analysis first failed to enforce removes the rest. Within-model redundancy is not a proxy for cross-model decoder geometry. Per-feature arrays and a CPU recompute script are released.
Papers and reports
Pavan Kumar Dubasi - Preprint
Fine-tuning Qwen 2.5 7B Instruct with QLoRA on five content conditions shows that the format of the fine-tuning data, not only its content, changes how the model fails, while MMLU accuracy stays flat across every condition. Behavioural scoring is done by LLM judges.
Pavan Kumar Dubasi - BlueDot Impact Technical AI Safety Project Sprint, Group 07
Sole-author BlueDot sprint study of which fine-tuning domain drives behavioural drift in 7B-scale models, scored by LLM judges. A 2026 replication adds a valence-matched control.
Pavan Kumar Dubasi - AI Control Hackathon 2026 (Apart Research + Redwood Research)
A ControlArena setting (identity_delegation) testing whether a trusted monitor detects within-scope attacks better when given identity and delegation-chain context rather than actions alone. Team lead.
Pavan Kumar Dubasi - BlueDot Impact Biosecurity Programme 2026
Proposes a verifiable identity and audit infrastructure for DNA synthesis compliance using W3C decentralized identifiers and verifiable credentials. Includes independent deployment and testing of the NAO metagenomic surveillance pipeline, with an upstream bug documented.
Pavan Kumar Dubasi - BlueDot Impact Biosecurity Programme 2026
Independent deployment and comprehensive audit of the Nucleic Acid Observatory metagenomic surveillance pipeline. Documented an upstream bug affecting sequencing data processing and documented full reproducibility steps for the bioinformatics workflow.
Pavan Kumar Dubasi, et al. - 2025 International Conference on Communication, Computer Sciences and Engineering (IC3SE-2025)
This paper presents a novel federated learning approach for healthcare data collaboration that preserves patient privacy while enabling effective model training across multiple healthcare institutions.
Pavan Kumar Dubasi - inspect_evals and inspect_ai (UK AI Security Institute)
Two pull requests merged into inspect_evals, including a GPQA Diamond answer-shuffle fix, and three accepted inspect_ai bug reports.
Pavan Kumar Dubasi, Bhadula, S., et al. - 2024 International Conference on Communication, Computer Sciences and Engineering (IC3SE)
We propose an enhanced Pix2Pix GAN architecture for artistic style transfer that achieves superior results in preserving structural integrity while transferring complex artistic styles.
Adhyaru, K., Ishan, Harshitha, D., Mundlur, R.P., Dubasi Pavan Kumar, et al. - International Journal of All Research Education and Scientific Methods (IJARESM)
This research presents a comprehensive approach to facial feature extraction and emotion recognition using machine learning techniques, achieving high accuracy in real-world scenarios.
Master's thesis
This research proposes a healthcare insurance fraud detection system combining an Optuna-tuned weighted ensemble of five gradient-boosted classifiers (F1 = 0.7345, leakage-free 10-fold CV) with a SHA-256 blockchain audit trail and ECIES encryption. The system aggregates 558,211 Medicare claims into 5,410 provider-level records with 190 engineered features. SHAP and LIME provide dual explainability for fraud risk assessments.
Intellectual property
Kousar, H., Pavan Kumar Dubasi, Mohammed, S.A., et al.
UK Design Registration No. 6335588, Grant Date: 10 January 2024
Navaneethakrishan, S.R., Chinta, S., Pavan Kumar Dubasi, et al.
Application No. 202321054346A. The Patent Office Journal, 29/09/2023, No. 39/2023
Community
Association for the Advancement of Artificial Intelligence (AAAI)
ID: 650314
2025 - Present
Association for Computing Machinery (ACM)
ID: 8485149
Aug 2024 - Present
American Mathematical Society (AMS)
ID: PVDBXA
Aug 2024 - Present
International Association of Engineers (IAENG)
ID: 383517
Aug 2024 - Present
Institute of Electrical and Electronics Engineers (IEEE)
ID: 98915358
Jan 2024 - Present
International Society for Computational Biology (ISCB)
ID: 45957
2025 - Present