Research Papers

Papers I've published, and papers worth reading.

Published by Me

TODO: Add papers I've authored/co-authored here.

Recommended Reading

  • C-Store: A Column-oriented DBMS

    Stonebraker et al. — the foundational paper introducing the column-oriented DBMS architecture.

    • databases
    • columnar-storage
    • systems
  • Maglev: A Fast and Reliable Software Network Load Balancer

    Eisenbud et al. (Google) — how Google built a software load balancer that scales out via consistent hashing and ECMP, with no single point of failure.

    • networking
    • load-balancing
    • distributed-systems
  • The Chubby Lock Service for Loosely-Coupled Distributed Systems

    Burrows (Google) — the design of Google's distributed lock service, and why coarse-grained locking and small-file storage turned out to be the right abstraction for coordinating distributed systems.

    • distributed-systems
    • consensus
    • systems
  • Efficient Estimation of Word Representations in Vector Space

    Mikolov et al. (Google) — introduces word2vec, the CBOW and skip-gram architectures that learn dense word embeddings efficiently from large corpora.

    • machine-learning
    • nlp
    • embeddings
    • stanford-cs329a
  • Windows Azure Storage: A Highly Available Cloud Storage Service with Strong Consistency

    Calder et al. (Microsoft) — the SOSP 2011 paper on how Azure Storage delivers strong consistency and high availability at scale, via its layered stream/partition architecture.

    • distributed-systems
    • cloud-storage
    • databases
  • Scuba: Diving into Data at Facebook

    Abraham et al. (Facebook) — the design of Scuba, Facebook's in-memory database for real-time, ad-hoc analysis over fast-changing data at scale.

    • databases
    • real-time-analytics
    • systems
  • An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

    Dosovitskiy et al. (Google) — introduces the Vision Transformer (ViT), showing that a pure transformer applied to sequences of image patches can match or beat convolutional networks on image classification.

    • machine-learning
    • computer-vision
    • transformers
    • stanford-cs329a
  • Large Language Monkeys: Scaling Inference Compute with Repeated Sampling

    Brown et al. (Stanford) — shows that repeatedly sampling multiple solutions from a language model improves coverage log-linearly with sample count, driving large gains on coding and reasoning benchmarks.

    • machine-learning
    • llm
    • inference-scaling
    • stanford-cs329a
  • Scaling Laws for Neural Language Models

    Kaplan et al. (OpenAI) — shows that language model loss follows predictable power laws in model size, dataset size, and compute, guiding the optimal allocation of training resources.

    • machine-learning
    • llm
    • scaling-laws
    • stanford-cs329a
  • Assisting in Writing Wikipedia-like Articles From Scratch with Large Language Models

    Shao et al. (Stanford) — introduces STORM, a system that uses LLM-simulated multi-perspective research conversations to outline and draft comprehensive, Wikipedia-style articles.

    • machine-learning
    • llm
    • agents
    • stanford-cs329a
  • How Do Large Language Monkeys Get Their Power (Laws)?

    Schaeffer et al. (Stanford) — resolves a paradox in repeated-sampling scaling by showing per-problem success rates improve exponentially, while a heavy-tailed distribution of task difficulty produces polynomial scaling in the aggregate.

    • machine-learning
    • llm
    • scaling-laws
    • inference-scaling
    • stanford-cs329a
  • KernelBench: Can LLMs Write Efficient GPU Kernels?

    Ouyang et al. (Stanford) — introduces a benchmark for evaluating whether language models can generate fast, functionally correct GPU kernels, finding frontier models still underperform PyTorch baselines on most tasks.

    • machine-learning
    • llm
    • benchmarks
    • systems
    • stanford-cs329a
  • Archon: An Architecture Search Framework for Inference-Time Techniques

    Saad-Falcon et al. (Stanford) — automatically searches combinations of inference-time techniques and language models, producing systems that surpass leading commercial models across multiple task categories.

    • machine-learning
    • llm
    • inference-scaling
    • stanford-cs329a
  • Training Verifiers to Solve Math Word Problems

    Cobbe et al. (OpenAI) — introduces GSM8K and shows that training a verifier to rerank sampled solutions scales better with data than finetuning alone for grade-school math reasoning.

    • machine-learning
    • llm
    • verification
    • math-reasoning
    • stanford-cs329a
  • Let's Verify Step by Step

    Lightman et al. (OpenAI) — shows that process supervision, rewarding each intermediate reasoning step, significantly outperforms outcome supervision for training models to solve MATH problems.

    • machine-learning
    • llm
    • verification
    • math-reasoning
    • stanford-cs329a
  • Math-Shepherd: Verify and Reinforce LLMs Step-by-step without Human Annotations

    Wang et al. — automatically constructs process-wise supervision data to train a step-level reward model, used both for reranking solutions and for PPO reinforcement learning, without any human step annotations.

    • machine-learning
    • llm
    • verification
    • math-reasoning
    • stanford-cs329a
  • Shrinking the Generation-Verification Gap with Weak Verifiers

    Saad-Falcon et al. (Stanford) — introduces Weaver, which combines many weak, imperfect verifiers into a strong unified score using weak supervision, narrowing the gap between generation and verification quality.

    • machine-learning
    • llm
    • verification
    • inference-scaling
    • stanford-cs329a
  • ReAct: Synergizing Reasoning and Acting in Language Models

    Yao et al. — interleaves reasoning traces with task-specific actions, letting a language model plan, call external tools like a Wikipedia API, and correct course, reducing hallucination versus chain-of-thought alone.

    • machine-learning
    • llm
    • agents
    • stanford-cs329a
  • RLEF: Grounding Code LLMs in Execution Feedback with Reinforcement Learning

    Gehring et al. (Meta) — trains LLMs with RL to use execution feedback from failed test runs to iteratively fix code, reaching state-of-the-art competitive-programming results with an order of magnitude fewer samples.

    • machine-learning
    • llm
    • reinforcement-learning
    • code-generation
    • stanford-cs329a
  • Constitutional AI: Harmlessness from AI Feedback

    Bai et al. (Anthropic) — trains a harmless assistant using self-critique and revision against a set of written principles, then reinforcement learning from AI-generated preferences, without human harmlessness labels.

    • machine-learning
    • llm
    • alignment
    • reinforcement-learning
    • stanford-cs329a
  • RLAIF vs. RLHF: Scaling Reinforcement Learning from Human Feedback with AI Feedback

    Lee et al. (Google) — shows that training a reward model on AI-generated preferences matches human-feedback RL across summarization and dialogue tasks, offering a more scalable path than RLHF.

    • machine-learning
    • llm
    • alignment
    • reinforcement-learning
    • stanford-cs329a
  • Self-Refine: Iterative Refinement with Self-Feedback

    Madaan et al. — has a single LLM act as generator, feedback provider, and refiner in a loop, with no extra training or human data, delivering roughly 20% improvement across code, math, and dialogue tasks.

    • machine-learning
    • llm
    • self-correction
    • agents
    • stanford-cs329a
  • Training Language Models to Self-Correct via Reinforcement Learning

    Kumar et al. (Google DeepMind) — introduces SCoRe, a multi-turn online RL approach trained entirely on self-generated correction traces, achieving state-of-the-art self-correction where supervised fine-tuning falls short.

    • machine-learning
    • llm
    • self-correction
    • reinforcement-learning
    • stanford-cs329a
  • Language Agent Tree Search Unifies Reasoning Acting and Planning in Language Models

    Zhou et al. — integrates Monte Carlo Tree Search with LLM-generated reasoning, acting, and self-reflection, unifying prior paradigms like ReAct and Tree of Thoughts into a single framework for language agents.

    • machine-learning
    • llm
    • agents
    • planning
    • stanford-cs329a
  • SPRINT: Enabling Interleaved Planning and Parallelized Execution in Reasoning Models

    Biju et al. (Stanford) — trains reasoning models to interleave long-horizon planning with parallel execution of independent subtasks, cutting sequential token generation by up to 40% without sacrificing accuracy.

    • machine-learning
    • llm
    • reasoning
    • inference-scaling
    • stanford-cs329a
  • Synthetic Data Generation & Multi-Step RL for Reasoning & Tool Use

    Goldie et al. (Stanford & Google DeepMind) — introduces SWiRL, which decomposes synthetically generated multi-step reasoning and tool-use trajectories into sub-trajectories for step-wise RL, improving accuracy and cross-task generalization.

    • machine-learning
    • llm
    • reinforcement-learning
    • agents
    • tool-use
    • stanford-cs329a
  • STaR: Bootstrapping Reasoning With Reasoning

    Zelikman et al. (Stanford) — iteratively bootstraps a language model's reasoning ability by fine-tuning on self-generated rationales that lead to correct answers, using rationalization from the answer to recover from failed attempts.

    • machine-learning
    • llm
    • reasoning
    • self-training
    • stanford-cs329a
  • DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

    Shao et al. (DeepSeek) — introduces GRPO, a memory-efficient reinforcement learning algorithm that estimates advantages from group scores instead of a learned critic, driving strong gains on mathematical reasoning benchmarks.

    • machine-learning
    • llm
    • reinforcement-learning
    • math-reasoning
    • stanford-cs329a
  • DAPO: An Open-Source LLM Reinforcement Learning System at Scale

    Yu et al. (ByteDance & Tsinghua) — stabilizes large-scale RL training for LLM reasoning with decoupled asymmetric clipping and dynamic sampling, along with token-level policy gradient loss and overlong reward shaping, matching state-of-the-art results on AIME with fewer training steps.

    • machine-learning
    • llm
    • reinforcement-learning
    • math-reasoning
    • stanford-cs329a
  • Competition-Level Code Generation with AlphaCode

    Li et al. (Google DeepMind) — introduces AlphaCode, which combines large-scale transformer pretraining on GitHub code with massive sampling and behavioral-clustering filtering, reaching roughly human-level performance on real Codeforces competitive-programming contests.

    • machine-learning
    • llm
    • code-generation
    • stanford-cs329a
  • AlphaCode 2 Technical Report

    Google DeepMind — builds AlphaCode 2 on fine-tuned Gemini Pro models with a bespoke sampling, filtering, and clustering/reranking pipeline, solving about 43% of evaluated Codeforces problems and outperforming roughly 85% of competitors, a large jump over the original AlphaCode.

    • machine-learning
    • llm
    • code-generation
    • stanford-cs329a
  • Search-o1: Agentic Search-Enhanced Large Reasoning Models

    Li et al. (Renmin University of China) — augments o1-style reasoning models with an agentic retrieval mechanism that triggers external search when the model hits a knowledge gap mid-reasoning, plus a Reason-in-Documents module that distills retrieved text before reinjecting it into the reasoning chain.

    • machine-learning
    • llm
    • agents
    • reasoning
    • stanford-cs329a
  • Measuring AI Ability to Complete Long Tasks

    Kwa et al. (METR) — introduces the 50%-task-completion time horizon, the human time-to-complete for tasks a model can do with 50% success, and shows this horizon has doubled roughly every 7 months since 2019 across frontier models.

    • machine-learning
    • llm
    • agents
    • benchmarks
    • stanford-cs329a
  • GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks

    Patwardhan et al. (OpenAI) — benchmarks frontier models against expert deliverables across 44 occupations spanning the top 9 GDP-contributing US sectors, finding models approach industry-expert quality while being roughly 100x faster and cheaper.

    • machine-learning
    • llm
    • benchmarks
    • agents
    • stanford-cs329a
  • DeepScholar-Bench: A Live Benchmark and Automated Evaluation for Generative Research Synthesis

    Patel et al. (Stanford) — a live benchmark drawing queries from recent arXiv papers to evaluate deep-research systems on generating related-work sections, scoring knowledge synthesis, retrieval quality, and verifiability, finding no system exceeds roughly 31% on its aggregate score.

    • machine-learning
    • llm
    • nlp
    • agents
    • benchmarks
    • stanford-cs329a
  • Multiagent Finetuning: Self-Improvement with Diverse Reasoning Chains

    Subramaniam et al. (MIT CSAIL) — fine-tunes a society of LLMs from the same base model on data from their multi-agent interactions, letting each model specialize into distinct roles like generator and critic so the group preserves diverse reasoning chains and keeps improving over far more rounds than single-agent self-improvement allows.

    • machine-learning
    • llm
    • agents
    • self-training
    • stanford-cs329a
  • DeepSeekMath-V2: Towards Self-Verifiable Mathematical Reasoning

    Shao et al. (DeepSeek-AI) — trains a model to interleave proof generation with a learned verifier, using meta-verification to catch reasoning issues and scaling verification compute so the verifier improves without human annotation or external reference solutions, reaching gold-level scores on IMO 2025 and CMO 2024.

    • machine-learning
    • llm
    • verification
    • math-reasoning
    • stanford-cs329a
  • Absolute Zero: Reinforced Self-play Reasoning with Zero Data

    Zhao et al. (Tsinghua University) — introduces a self-play RL paradigm in which a single model proposes its own coding and reasoning tasks, validates them with a code executor, and learns to solve them with zero external or human-curated data, reaching state-of-the-art coding and math reasoning results.

    • machine-learning
    • llm
    • reinforcement-learning
    • code-generation
    • self-training
    • stanford-cs329a
  • Intelligence per Watt: Measuring Intelligence Efficiency of Local AI

    Saad-Falcon et al. (Stanford) — defines Intelligence per Watt, task accuracy per unit power, and evaluates 20+ local models of roughly 20B parameters across accelerators on about 1M real queries, finding they answer the vast majority successfully while efficiency improved 5.3x from 2023 to 2025.

    • machine-learning
    • llm
    • benchmarks
    • efficiency
    • stanford-cs329a