Papers

Interleaving chain-of-thought reasoning with action execution gave LLMs a practical agentic loop. This became the skeleton of how I think about tool-using agents.

Bridged NLP and vision by treating image patches as tokens. The simplicity is surprising — it required scale to work, which raised important questions about inductive biases vs data.

Eliminated the need for hand-designed anchors and NMS. Directly relevant to my thesis work on attention-based detection architectures at TUM.

Multi-vector late interaction over page-level visual embeddings — a clean solution to the problem of searching inside PDFs without any text extraction step.

IO-awareness turned a theoretical bottleneck into a practical one and solved it elegantly. A great example of systems thinking applied to ML.

One-shot unstructured pruning at 50% sparsity with minimal accuracy loss. Important for understanding how to deploy large models in resource-constrained environments.

Empirical power-law relationships between compute, data, and model size. Informed how I reason about the cost-performance tradeoff when choosing models for production.

Contrastive pre-training on image-text pairs produced remarkably general representations. Foundational to my work on document retrieval and OCR pipelines.

The paper that reshaped the field. Self-attention replacing recurrence was the key insight. I re-read this every time I need to reason about sequence models from first principles.

Grounding generation in retrieved documents was an elegant fix to hallucination. The combine-then-generate approach is still the dominant RAG paradigm.