Research Intelligence

Reasoning modelsarXiv

DeepSeek-R1: Incentivizing reasoning in large language models

This paper is useful for understanding why open reasoning models became a global benchmark and pricing shock. Read it for the training recipe, reinforcement-learning framing, and distillation story.

Reasoning, open models, and cost/performance are still driving the AI conversation.

How reasoning behavior can emerge and how smaller models inherit it through distillation.

AlignmentarXiv

Direct Preference Optimization

DPO is one of the papers that made preference tuning easier to discuss and implement. It reframes human preference optimization without requiring the classic RLHF loop.

Preference tuning is now central to model quality, safety, and product feel.

Why alignment can be treated as a direct policy optimization problem.

AI systemsarXiv

FlashAttention: Fast and memory-efficient exact attention

FlashAttention matters because model progress is not just architecture. Kernel-level efficiency changes what teams can train, serve, and afford.

Inference and training cost are now product strategy, not just infrastructure detail.

How attention can be made faster by respecting memory hierarchy.

ArchitecturesarXiv

Mamba: Linear-time sequence modeling with selective state spaces

Mamba is worth watching because it challenges the assumption that transformer attention is the only practical sequence backbone for long contexts.

Long-context models, efficient inference, and non-transformer architectures remain active research fronts.

The case for selective state-space models as a transformer alternative.

Fine-tuningarXiv

LoRA: Low-rank adaptation of large language models

LoRA made practical adaptation cheaper by training small low-rank updates instead of the full model. It remains one of the most useful ideas for applied AI teams.

Fine-tuning, adapters, and domain customization keep showing up in enterprise AI.

How small trainable matrices can adapt large frozen models.

Vision foundation modelsarXiv

Segment Anything

SAM is a practical reference for how foundation-model thinking moved into computer vision: promptable behavior, broad data, and reusable perception tooling.

Visual AI keeps crossing into robotics, medical imaging, design tools, and geospatial intelligence.

What promptable vision models look like outside language.

SafetyarXiv

Constitutional AI

This paper is a core reference for scalable oversight: using explicit principles and AI feedback to reduce harmful behavior without relying only on human labels.

Safety, policy, and model behavior are now part of every major model launch.

How rule-guided critique and revision can shape assistant behavior.

RAGarXiv

Retrieval-Augmented Generation for knowledge-intensive NLP

RAG remains a must-read because most production AI systems need models to use external knowledge rather than rely only on parameters.

Enterprise AI, search, citations, and agent memory all depend on retrieval quality.

The basic split between parametric memory and retrieved evidence.

Research

Top papers and foundations