Paper Summaries

Structured summaries of foundational and influential papers in ML, DL, NLP, and LLM systems. Each summary distills the paper’s thesis, methods, contributions, and my takeaways — designed to be useful as both a first encounter and a quick refresher.

Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity

Simplified Mixture of Experts routing each token to a single expert, enabling efficient scaling to trillion parameters while maintaining computational efficiency.

MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework

Multi-agent framework using assembly-line paradigm with SOPs, specialized validator agents, and structured outputs to reduce compounding errors.

OpenHands: An Open Platform for AI Software Developers as Generalist Agents

Open platform for AI software developers using generalist and specialized agents with programming-language primitives as extensible tools.

SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering

Agent-Computer Interfaces that enable LLM agents to autonomously perform software engineering tasks like bug fixing and competitive coding.

Reflexion: Language Agents with Verbal Reinforcement Learning

Verbal reinforcement learning through self-reflection — agents learn from dynamic memory of previous errors to correct their trajectories.

ReAct: Synergizing Reasoning And Acting In Language Models

Combining reasoning traces with actions for grounded problem solving — reasoning guides actions, actions refine reasoning.

What Are Tools Anyway? A Survey from the Language Model Perspective

Comprehensive survey on tool use from the LLM perspective — definitions, categories, usage patterns, and tool creation.

AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation

Multi-agent conversation framework with unified interfaces for building modular, extensible LLM applications.

Tree of Thoughts: Deliberate Problem Solving with Large Language Models

Deliberate problem solving inspired by dual-process theory, using tree search (BFS/DFS) over reasoning paths with lookahead and backtracking.

Searching for Best Practices in Retrieval-Augmented Generation

Systematic process for optimizing RAG workflows — evaluating module combinations to find the best performance-latency tradeoff.

Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

Chain-of-thought prompting as an emergent ability in large LLMs that significantly improves arithmetic, commonsense, and symbolic reasoning.

Chain-of-Thought Reasoning without Prompting

Eliciting LLM reasoning without prompts by exploring alternative decoding paths — CoT-decoding surfaces inherent reasoning capabilities.

Toolformer: Language Models Can Teach Themselves to Use Tools

Self-supervised training for LLMs to learn which APIs to call, when to call them, and how to incorporate results into generated text.

Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena

Scalable, interpretable automated evaluation of LLM alignment with human preferences using MT-Bench and Chatbot Arena.

StarCoder 2 and The Stack v2: The Next Generation

Next-generation open code models (3B–15B) with two-stage training and 16K context windows on The Stack v2 dataset.

ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT

Late interaction retrieval balancing expressiveness and efficiency — contextualized token-level matching with delayed query-document interaction.

DeepSeek-Coder: When the Large Language Model Meets Programming - The Rise of Code Intelligence

Code intelligence models (1.3B–33B) trained on 2T tokens of project-level code from 87 languages with repository-level understanding.

Dense Passage Retrieval for Open-Domain Question Answering

Dense retrieval using dual BERT encoders that outperforms BM25/TF-IDF for open-domain question answering.

Internet-augmented language models through few-shot prompting for open-domain question answering

Augmenting LLMs with Google Search for up-to-date open-domain QA — no fine-tuning needed, closing the gap between small and large models.

Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks

End-to-end retrieval and generation training that provides more specific, diverse, and factual answers while reducing hallucination.

OLMo: Accelerating the Science of Language Models

Fully open-source LLM including training data, code, weights, and checkpoints — accelerating reproducible language model research.

REALM: Retrieval-Augmented Language Model Pre-Training

End-to-end pre-training of retriever and language model together, augmenting LM with external knowledge for improved downstream performance.

Self-Instruct: Aligning Language Models with Self-Generated Instructions

Using base LLMs to self-generate instruction data for alignment, approaching InstructGPT performance with near-zero human annotation.

Mixtral of Experts

Mixture of Experts routing tokens to 2 of 8 experts, achieving strong performance at a fraction of the dense compute cost.

Mistral-7B

High-performance 7B model that outperforms Llama 2-13B across all tasks through efficient architectural choices.

Code Llama: Open Foundation Models for Code

Open code generation models based on Llama 2 with infilling capabilities and large context windows — SOTA among open code LLMs.

Efficient Training of Language Models to Fill in the Middle

Fill-in-the-middle training for autoregressive LLMs improves code infilling without compromising standard left-to-right generation.

Chinchilla: Training Compute-Optimal Large Language Models

Most LLMs are undertrained — for 10x more compute, increase model size and training tokens equally rather than just scaling parameters.

Scaling Laws for Neural Language Models

Empirical power-law relationships between model performance and model size, dataset size, and compute budget.

Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling

Open-source suite of models (70M–12B) with checkpoints and training code for studying how LLM capabilities evolve during training.

T5: Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer

Unified text-to-text framework using encoder-decoder architecture, converting any NLP task into the same format with the same model, objective, and decoding.

Llama 2: Open Foundation and Fine-Tuned Chat Models

Open foundation and fine-tuned chat models with clever algorithmic tweaking focused on helpfulness and safety for dialogue.

LLaMA: Open and Efficient Foundation Language Models

Competitive LLMs trained entirely on public data, prioritizing inference efficiency over raw model size.

InstructGPT: Training language models to follow instructions with human feedback

Training models to follow instructions using RLHF — being helpful, not harmful, and not hallucinating.

LoRA: Low-Rank Adaptation of Large Language Models

Low-rank decomposition matrices for transformer attention weights — minimal trainable parameters, no inference latency overhead, efficient task switching.

GPT3: Language Models are Few-Shot Learners

175B parameter model that learns tasks at inference time through few-shot prompting, shifting from architecture-agnostic to task-agnostic paradigm.

RoBERTa: A Robustly Optimized BERT Pretraining Approach

Better hyperparameter and design decisions are all you need — BERT was significantly undertrained, and optimized training matches or beats later approaches.

BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding

Bidirectional pre-training that learns rich contextualized token representations by attending to both left and right context.

Prefix-Tuning: Optimizing Continuous Prompts for Generation

Training continuous task-specific prefix vectors that steer the LM, keeping all pre-trained weights frozen.

GPT2: Language Models are Unsupervised Multitask Learners

Scaling to 1.5B parameters and demonstrating zero-shot task transfer — a competent generalist instead of narrow experts.

GPT: Improving Language Understanding by Generative Pre-Training

Semi-supervised pre-training with language modeling to learn general representations transferable to downstream tasks with minimal changes.

Parameter-Efficient Transfer Learning for NLP

Compact adapter layers in each transformer block — freeze pre-trained weights, train only 3–8% of parameters for downstream tasks.