Partha Pratim Saha

My core concern is straightforward but urgent: as AI systems grow more capable, misalignment between their internal objectives and human values becomes catastrophically consequential — not merely inconvenient. I am particularly worried about power-seeking behaviour in strategically capable agents, concentration of control in military and governance domains, and the fundamental challenge that behavioural safety does not imply geometric safety. A model can pass all safety evaluations while encoding deeply misaligned belief structures internally.

How I plan to create impact: (1) Develop post-hoc geometric probes that are compute-efficient enough for deployment-time monitoring; (2) Collaborate with world-leading AI safety labs (Anthropic, Redwood, ARC Evals, Mech Interp groups at Oxford/Cambridge/MIT/CMU) to validate geometric torsion findings against circuit-level mechanistic analysis; (3) Pursue a fully-funded PhD at a programme with strong AI safety infrastructure to work on scalable interpretability tools deployable beyond 7B-parameter models. I am actively applying for research fellowships, internships, and PhD positions starting 2026.

Models transformer hidden-state trajectories as discrete curves on a Riemannian belief manifold (ℳ, g_F) with the Fisher-Rao metric. The torsion tensor S_ℓ = (M_ℓ − M_ℓ^T) / 2 captures rotational mismatch between consecutive belief updates — a non-commutative geometric signature invisible to attention patterns or activation magnitudes. Key discovery: DPO alignment creates geometrically localised “brake layers” that systematically suppress torsion, probed across OLMo-7B, Mistral-7B, and Zephyr-7B using 500 unsafe prompts (Litmus) and 17 geometric metrics.

Unifies fine-tuning, alignment, distillation, and merging as measurable deformations of the same depth-wise semantic flow via spectral curvature κ_ℓ and thermodynamic length ℒ_ℓ (epistemic effort across layers). Investigates epistemic inheritance in merged LLMs using Fisher-Rao geometry, producing neural offspring; emergent cultural nDNA measured via spectral curvature deviation Δκ_ℓ and thermodynamic length divergence Δℒ_ℓ. Cultures studied: African, Latin American, South Asian, East Asian, Arabic, Indigenous, European, Pacific Islander.

Collaborative Federated Learning (CFL) cloud-based system separates datasets into public and private sets based on the removal of PHI/PII, enabling personalised GPT-like systems without centralising sensitive data. Enables AI for healthcare at scale while preserving patient privacy — a critical requirement for responsible, equitable AI deployment globally.

Investigates whether formal robustness certificates for LLMs can be derived from data contamination, label noise, and adversarial attacks rather than input-perturbation bounds. DPO-aligned models show compressed Fisher-norm spectra versus SFT counterparts, suggesting alignment induces geometric contraction doubling as a robustness mechanism. Framework: Lipschitz robustness bounds guaranteeing up to 40% data contamination holds almost natural robustness. Models: ViT, MedSigLIP, CLIP.