Theoretical ML/DL · Interpretability · Information Theory
Yug D Oswal
I'm a final-year CS undergrad at VIT Vellore and an incoming MSc student at the University of Oxford (Advanced Computer Science, Oct 2026). My research sits at the intersection of theoretical ML/DL, mechanistic interpretability, and information theory - broadly, I'm interested in understanding mechanisms underlying learning and developing more capable/reliable intelligence.
Currently I'm working with the Torr Vision Group and OATML groups at Oxford, headed by Prof. Philip Torr and Prof. Yarin Gal respectively. I'm also working with Prof. Ravid Shwartz-Ziv at NYU. I'm always happy to connect, please feel free to reach out!
00 Education
University of Oxford Incoming
MSc in Advanced Computer Science
Oxford, England
Oct 2026 – Oct 2027
VIT Vellore
BTech in Computer Science and Engineering
CGPA: 9.53 / 10 · Vellore, India
Sep 2022 – Jul 2026
01 News & Updates
- Apr 2026 Invited to Y Combinator Startup School India 2026, among the Top 2000 builders, founders, and engineers in India.
- Mar 2026 Paper "Beyond the Loss Curve" accepted to the ICLR DeLTa Workshop.
- Mar 2026 Admitted to the MSc Advanced Computer Science program at the University of Oxford, starting Oct 2026.
- Jan 2026 Selected Top 10 / 90+ teams at SPAR Fall 2025: presented lightning talk on causal reward-hacking induction to researchers at Constellation, UK AISI, and BlueDot Impact.
- Sep 2025 Joined NYU as a research intern with Prof. Ravid Shwartz-Ziv (currently at Meta) on tractable-distribution datasets and uncertainty decomposition.
- Sep 2025 Started as a SPAR Research Fellow (Supervised Program for Alignment Research) with Shivam Raval, Harvard University.
02 Papers & Preprints
✦ ICLR DeLTa Workshop 2026
Constructed the first large-scale tractable-distribution dataset for ImageNet-64 enabling exact uncertainty decomposition. Discovered epistemic scaling power-laws and a constant aleatoric floor. Demonstrated that epistemic-based active learning outperforms entropy-based sampling, requiring 47.8% fewer samples for equivalent performance.
Under Review at Applied Soft Computing, 2025
Gradient-based loss scheduling method (loss switching) paired with statistically optimized classification and regression losses. Achieves ≥ 3% top-1 ImageNet gain and ≥ 1.4% RMSE improvement across four regression benchmarks under asymmetric outlier shifts.
arXiv, 2024
Cone activations compute hyperstrip representations, making them effective classification heads. ≥ 4.6% accuracy gain on ImageNet with 46.4% parameter reduction in VGG19. ≤ 6× neuron compression yields only ≈ 2% drop for cone vs. ≈ 8% for ReLU.
arXiv, 2023–2025
Vectorized forward/backward matrix algorithms resolving the core computational bottleneck in QNNs. O(n²) reduced-parameter RP-QNN variants with systematic expressiveness–efficiency ablations.
03 Research Experience
Mechanistic Interpretability · AI Safety · Reward Hacking
Led funded research on causal induction of reward hacking in LLMs via activation steering.
Selected among top 10 of 90+ teams to present to AI safety orgs including Constellation, UK AISI, and BlueDot Impact. EMNLP '26 submission in preparation.
Selected among top 10 of 90+ teams to present to AI safety orgs including Constellation, UK AISI, and BlueDot Impact. EMNLP '26 submission in preparation.
Uncertainty Decomposition · Normalizing Flows · Deep Learning · Scaling Laws · Active Learning
Led experimental pipeline for the first large-scale tractable-distribution dataset at ImageNet-scale with exact posterior, marginal, and KL oracle queries.
Discovered epistemic scaling laws, theoretical limits of learning, and improved active learning policies. Results led to a paper accepted at the ICLR DeLTa Workshop 2026.
Discovered epistemic scaling laws, theoretical limits of learning, and improved active learning policies. Results led to a paper accepted at the ICLR DeLTa Workshop 2026.
CoT Alignment · Statistical Interventions · Causal Tracing · Cognitive Theory
Formulated conditioning of LLM reasoning on outputs via a KL-divergence-based self-alignment framework for chain-of-thought. Achieved 25–35% task-accuracy gain and reduced bias over base DeepSeek Qwen 1.5B on the Bias Benchmark for QA. Integrated causal activation patching to trace how biased reasoning propagates into model outputs.
04 Professional Experience
Curated 85,000-sample IR-optical hybrid UAV dataset. Built first prototype multimodal thermal-optical anti-UAV surveillance system (YOLOv8-based), successfully tested in 4 field scenarios. Engineered novel containerized client-server deployment on air-gapped defence systems.
Led the full ML lifecycle for a humanoid, speech-capable therapeutic agent. Engineered and deployed agentic workflows, RAG pipelines, guardrails, and context-aware chat with real-time client integration - reducing latency by 53%.
Led an international applied ML team to resolve 5 real-world issues in Signal's threat intelligence system. Built scalable pipelines for NER, geocoding, and incremental clustering over live threat data streams for high-profile executive clients.
05 Community & Leadership
Directed technical strategy and operations for a 100-member chapter. Mentored juniors in ML and research. Organised large-scale events: Rural ML Outreach (40+ rural students), Riddler (1000+ participants, 25+ countries), LaserTag (1000+), and Init ML workshops (200+). Intel Developer Spotlight feature for Rekindle.
06 Selected Projects
End-to-end AI memory-support service for dementia patients: custom emotion-extraction model (outperforming baselines on Google GoEmotions), emotion-conditioned memory retrieval, local LLMs for privacy-preserving inference, and full-stack life-logging platform.