Shiv Munagala

15 ECTS Expert Diploma from the Valencian Graduate School and Research Network of Artificial Intelligence on AI evaluation capabilities and safety. One of 40 students selected from 650+ applicants, on a full scholarship from Coefficient Giving (formerly Open Philanthropy). Project: False-Positive Evaluation Awareness in LLMs.

Contributed capability evaluations for frontier AI models through AISI’s bounty programme. Developed and evaluated agent scaffolds for variable malicious behaviour and evasion conditioned on user competence. This work was integrated into internal UK government risk assessment benchmarks for frontier AI systems.

Contributed to the Inspect Evals suite of evaluations for frontier AI systems by implementing the MathVista benchmark.

• ARC White-Box Estimation Challenge: Competing in the Alignment Research Center (ARC)'s contest on compute-efficient methods for estimating a neural network's behaviour directly from its weights. Placed 4th of 79 in the warm-up round and provisionally 28th of 314 in Phase 1; Phase 2 ongoing.
• Failure modes of a Probe-Then-Ablate pipeline (adapted from a Bluedot TAIS Puzzle): Found that ablating a linear probe's direction can look causally inert on an undertrained model but causally relevant once it has converged, with the change happening in line with training loss and driven by a redundancy gap between how many directions decode a feature and how many the network actually uses. SAE-based probes resist the same construction. Post being written up.
• Geometric Test for Multi-Dimensional Structure in SAE Latents: Exploring a rotation-based permutation-null diagnostic for detecting multi-dimensional feature geometry in sparse autoencoder latents, run on Pythia-70M. Experiment underway.
• Failing to Improve on the Muon Optimizer: Looked for cheaper alternatives to Muon's orthogonalisation step (QDWH, Cholesky QR, low-rank sketches). Found the momentum matrices are ill-conditioned so their effective rank collapses during training, explaining why the existing coarse-but-robust design is well matched to the problem. Backburner, undecided whether to write this up.

Review committee member for both conferences.

Lead statistician for a global patient survey in pulmonary vascular disease. Responsible for quantitative analysis across multiple research groups and for designing data access and governance structures that balance reuse with patient privacy.

• Papers associated with the Pulmonary Hypertension Global Patient Survey:
  • Understanding the experiences and perspectives of patients, ERJ Open Research (2025)
  • Physical and Psychosocial Impacts on Health-Related Quality of Life, Pulmonary Circulation (2026)
  • Groups 2 and 3, Pulmonary Circulation (2026)
  • Understanding the Invisible Burden of Paediatric Pulmonary Hypertension, Pulmonary Circulation (2026)
• Governing Automated Strategic Intelligence, arXiv (2025)

• Pulmonary Hypertension Global Patient Survey: A Call to Action, Pulmonary Circulation (2025)

• A longitudinal study of digital walk tests in the community, European Respiratory Journal (2025)
• The Pulmonary Hypertension Global Patient Survey: Group 2 & 3 PH, European Respiratory Society (2025)

• Findings from the Pulmonary Hypertension Global Patient Survey presented at ERS-2024 (Best Abstract in Patient Centred Research), WSPH-7, and PVRI-2024
• Findings from Governing Automated Strategic Intelligence presented at TAIS 2026

• Measurement of Physical Stress Response. Lead inventor. UK Intellectual Property Office (submitted 2024)

Logistics optimisation, automation, and central AI strategy.

Early data science hire working on physiological stress detection and end-to-end ML systems.