Subject
175 entries
Research
Bookmarks
Reframing Instructional Prompts to GPTk's Language
ACL 2022 Findings paper showing that manually reframing instructional prompts — decomposing tasks, itemizing steps, adding positive examples — yields 6–12% performance gains on GPT-2 and GPT-3. The key insight is that models respond better to concrete, step-by-step instructions than to long abstract descriptions.
PaLM: Scaling Language Modeling with Pathways
PaLM is Google's 540B parameter language model trained across 6144 TPU v4 chips using the Pathways distributed training system. It achieved breakthrough performance on multi-step reasoning and BIG-Bench, and documented discontinuous capability gains at scale — capabilities that emerged suddenly with more compute.
Editing Models with Task Arithmetic
Task Arithmetic shows that fine-tuning deltas (task vectors) can be added, subtracted, and combined arithmetically to edit model behavior — without any additional training. Negation forgets a task; addition combines capabilities; analogical transfer works across tasks.
An Introduction to Variational Autoencoders
The canonical tutorial on Variational Autoencoders by Kingma and Welling — the original VAE inventors. Covers the ELBO, reparameterization trick, and extensions to deeper generative models. Essential background for anyone working with latent variable models or modern diffusion/flow models.
Neuronpedia Gemma Scope: interactive mechanistic interpretability for Gemma 2
Neuronpedia's Gemma Scope microscope lets you scan Gemma 2's internal features using Sparse Autoencoders — activating features, steering behavior, browsing SAE decompositions. Interactive mechanistic interpretability for a production model.
Linear Representations and Superposition in LLMs
A clear explainer on how LLMs encode thousands of features in relatively small embedding spaces via superposition — the Johnson-Lindenstrauss lemma applied to neural representations. Good primer on the theoretical foundations behind SAE-based interpretability.
NASA Technical Reports Server (NTRS)
The NASA Technical Reports Server (NTRS) is the public repository for NASA's research: conference papers, journal articles, patents, technical reports, and videos. Free public access to decades of aerospace and space science literature.
High-Dimensional Data Analysis: Wright & Ma Textbook
John Wright and Yi Ma's graduate textbook on high-dimensional data analysis — sparse models, low-rank models, and deep networks unified under a single mathematical framework. Adopted at Berkeley, Columbia, UIUC, Tsinghua. Published Cambridge University Press 2022.
Defining and Characterizing Reward Hacking
Skalse, Howe, Krasheninnikov, and Krueger provide the first formal definition of reward hacking — when optimizing a proxy reward degrades performance on the true reward — and derive conditions under which 'unhackable' proxies exist. The theoretical result is stark: for stochastic policies, only constant reward functions are unhackable, making reward misspecification a near-universal concern in RL-based alignment.
How to Read Deep Learning Papers as a Software Engineer
A YouTube guide for software engineers on reading deep learning papers — framing the paper-reading skill as an engineering competency rather than a research one, with practical strategies for extracting implementation-relevant understanding from dense notation.
Scholar Inbox: Personalized Research Paper Feed
Scholar Inbox is a paper recommendation service that delivers daily email digests of relevant research papers based on your reading history and interests — a personalized paper feed built on top of arXiv and similar repositories. Alternative to Google Scholar alerts with better personalization.
Reducing LLM Deception with Self-Other Overlap Fine-Tuning
Self-Other Overlap (SOO) fine-tuning reduces deceptive behavior in LLMs by aligning internal activations for self-referential and other-referential prompts — cutting deceptive responses from 73% to 17% on Mistral-7B with minimal capability loss. A representation-level approach rather than behavioral supervision.
Awesome LLM Judges: curated research on LLM evaluation systems
Curated collection of research on LLM-as-a-judge evaluation systems — covering ensemble methods, fine-tuned judge models, hallucination detection, generative reward models, and safety moderation. Maintained by Haize Labs alongside their Verdict library.
Activation Steering Notebook (concentration_notebooks)
A Jupyter notebook exploring activation steering in neural networks from the concentration_notebooks repo. Saved as a reference in the mechanistic interpretability and model steering space.
monitors4codegen — Monitor-Guided Decoding
Monitor-Guided Decoding uses LSP (Language Server Protocol) static analysis as a 'monitor' during code LM generation to enforce semantic validity — identifiers must exist, types must match. NeurIPS 2023 paper from Microsoft with the multispy Python library for building LSP-backed code gen applications.
ithy — Multi-LLM Research Aggregator
ithy is an AI research platform that queries multiple LLMs simultaneously and synthesizes their responses into a single interactive article. The positioning is that combining ChatGPT, Gemini, and Claude produces better answers than any one alone.
Virtual Personas for Language Models via an Anthology of Backstories
BAIR research on giving LLMs stable virtual personas using an 'anthology of backstories' — a collection of diverse biographical narratives that, combined, produce consistent and controllable personality profiles. Relevant for social simulation, synthetic data generation, and studying opinion dynamics.
SURD: Synergistic-Unique-Redundant Decomposition of Causality
SURD is a Python library for decomposing causality into synergistic, unique, and redundant components in complex systems — quantifying how different variables jointly contribute to predicting future states. Developed for turbulence analysis but applicable to any causal inference problem.
Autonomous Hacker: LLM-Powered Security Research Agent
The autonomous-hacker module in R3DRUN3's sploitcraft demonstrates using LLMs as autonomous security research agents — scanning, identifying vulnerabilities, and generating exploits with minimal human input. An educational proof of concept for LLM-assisted offensive security.
Use Prolog to Improve LLM Reasoning
A proposal for using Prolog as a symbolic reasoning layer alongside LLMs — offloading formal logical inference to Prolog while the LLM handles natural language parsing and generation. A neuro-symbolic hybrid that addresses LLMs' known weakness at strict logical deduction.
Distributed Consensus Reading List
Heidi Howard's curated list of papers on distributed consensus — organized from foundational Paxos through Multi-Paxos, Raft, flexible quorums, and Byzantine fault tolerance. The definitive paper trail for understanding how consensus algorithms actually work.
AWS Nitro Enclaves Attack Surface
Trail of Bits analysis of AWS Nitro Enclaves' attack surface, covering vsock vulnerabilities, side-channel risks, entropy issues, and the trust model's assumptions. Essential reading before building confidential computing workloads on Nitro.
L1B3RT45: AI System Prompts Collection
L1B3RT45 is a GitHub repository collecting leaked and reverse-engineered system prompts from major AI assistants (Claude, ChatGPT, Mistral, Zamba2). With 18.5k stars, it documents the operational instructions guiding commercial AI models, useful for prompt engineering research and AI transparency.
webXray: Privacy Compliance Search Engine
webXray is a forensic web analysis platform that detects real tracking violations across 1.2M+ websites — built from a decade of academic research by Dr. Timothy Libert. Used by law firms, Fortune 100 compliance teams, and privacy litigators.
TravelPlanner: A Benchmark for Real-World Planning with Language Agents
TravelPlanner is a benchmark for evaluating LLM planning capabilities in complex real-world scenarios — GPT-4 scored 0.6% on the full benchmark, revealing that even the best LLMs struggle with multi-constraint sequential planning. A sobering check on agentic AI ambitions.
jailbreak_llms: CCS'24 Jailbreak Prompt Dataset
A dataset of 15,140 ChatGPT prompts including 1,405 jailbreak prompts collected from Reddit, Discord, and open-source datasets — published at CCS 2024. The most comprehensive public collection of real-world jailbreak attempts against LLMs.
Nightshade: Protecting Copyright Through Adversarial Poisoning
Nightshade is a tool that lets artists poison AI training data by adding imperceptible perturbations to images — images look normal to humans but cause AI models trained on them to produce corrupted outputs. An offensive countermeasure for artists against unauthorized scraping.
Inspectus: LLM Analytics and Visualization
Inspectus is labml.ai's tool for visualizing LLM internals — attention maps, token distributions, and model analytics rendered as interactive visualizations in Jupyter notebooks. Makes transformer attention patterns inspectable without custom code.
Simple probes can catch sleeper agents
Anthropic research showing that simple linear probes trained on internal activations can reliably detect 'sleeper agent' backdoors in LLMs — models trained to behave well normally but activate malicious behavior on a trigger. Evidence that activation-space representations of deceptive intent are linearly separable.
AI and Knowledge Work: Early Research Synthesis
A synthesis of early 2024 research on AI's impact across software engineering, customer support, and consulting. Key finding: lower performers benefit most from AI tools, and using AI on tasks outside its capability range actually makes performance worse.
Genie: Generative Interactive Environments
Genie is a Google DeepMind foundation world model that generates playable, action-controllable interactive environments from a single image prompt — photo, sketch, or AI-generated. Trained on unlabeled internet videos without action annotations, it discovers a transferable latent action vocabulary.
Diff History for Neural Language Agents
Diff history applies the Unix diff command to sequential agent observations, replacing full text states with change deltas. Dramatically reduces context length while preserving critical information — small models tuned with diff history matched SOTA on NetHack with 1800x fewer training examples.
Prompt Lookup Decoding: Speculative Decoding without a Draft Model
Prompt lookup decoding replaces the draft model in speculative decoding with simple n-gram string matching against the prompt itself — achieving 2.4x speedup on summarization and QA with zero quality loss. Works whenever output heavily references the input, which is most practical LLM tasks.
LLM Verified with Monte Carlo Tree Search
A research project using Monte Carlo Tree Search to guide and verify LLM code generation — MCTS explores the tree of possible completions and selects branches where generated code actually passes verification checks. An early example of search-augmented LLM reasoning.
The Rise and Potential of LLM-Based Agents: A Survey
An 86-page survey paper by Zhiheng Xi et al. comprehensively mapping the architecture, capabilities, and applications of LLM-based agents — establishing the planning/memory/tool-use framework that became the standard way to think about agent components. The field's foundational survey document.
PDFTriage: Question Answering over Long, Structured Documents
PDFTriage is a method for question answering over long, structured PDF documents that uses the document's structure (pages, sections, tables of contents) as a triage layer before retrieval — letting the LLM navigate the document intelligently rather than relying on flat embedding search.
LLM Agent Paper List
A curated paper list accompanying the 86-page 'Rise and Potential of LLM-Based Agents' survey — organized by agent component (planning, memory, tool use, action) and application domain. The most comprehensive catalog of LLM agent research from September 2023.
Arxiv Paper 2309.08963
An arxiv paper from September 2023 saved alongside the LLM agent survey and RAG papers from the same day — likely in the LLM reasoning, agents, or retrieval space based on context. Content unavailable from the bookmark.
Sunscreen FHE: Private Information Retrieval via Matrix Operations
Sunscreen's documentation on Private Information Retrieval via Fully Homomorphic Encryption — demonstrating how FHE enables querying a database without the server learning what you searched for. A practical introduction to FHE programming via a concrete PIR example.
INDRA: Automated Biomedical Knowledge Assembly
INDRA (Integrated Network and Dynamical Reasoning Assembler) is an automated knowledge assembly system that reads biomedical literature and databases and produces causal graphs and dynamical models. A long-running academic system for automated scientific knowledge extraction.
Graph of Thoughts: Solving Elaborate Problems with LLMs
Graph of Thoughts (GoT) extends chain-of-thought and tree-of-thought prompting to arbitrary graph structures — letting LLM reasoning steps combine, branch, and loop rather than just proceeding linearly. More expressive than CoT or ToT for complex multi-step problems.
Avenging Polanyi's Revenge: LLM Approximate Omniscience in Planning
A talk titled 'Avenging Polanyi's Revenge' arguing that LLMs' approximate omniscience changes planning — they've absorbed tacit knowledge that previously required human experts, enabling a different category of automated planning than rule-based systems allowed.
Reasoning or Reciting? LLM Capabilities via Counterfactual Tasks
An arxiv paper testing whether LLMs are reasoning or reciting — using counterfactual tasks (modified versions of standard benchmarks) to distinguish genuine reasoning from pattern-matching on training data. Finds LLMs show sharp performance drops on counterfactual variants, suggesting heavy reliance on memorization.
llm-reasoners: Advanced LLM Reasoning Algorithms
llm-reasoners is a library for advanced LLM reasoning algorithms — implementing Tree of Thoughts, RAP (Reasoning via Planning), and other structured reasoning approaches over standard chain-of-thought. Useful for research into how to get LLMs to reason more reliably on complex tasks.
AI Adversarial Attacks: Automated Jailbreaks via Text Suffixes
Ars Technica covers the Universal Adversarial Attacks paper from CMU/Center for AI Safety — automated adversarial suffixes appended to prompts reliably bypass safety training on GPT-4, Claude, and open-source models. The attacks are transferable and potentially unstoppable with current alignment techniques.
LLM Powered Autonomous Agents
Lilian Weng's survey post on LLM-powered autonomous agents — covering the planning, memory, and tool use components that compose into agent architectures. One of the most cited and comprehensive overviews of the agent design space from mid-2023.
Landmark Attention: Random-Access Infinite Context
Landmark Attention paper extending LLaMA 7B to 32k token context by using landmark tokens to represent blocks of input, enabling attention-based retrieval of relevant blocks rather than attending over all tokens. Achieves near-GPT-4 context length through architectural change rather than longer pretraining.
OpenLLaMA: Open Reproduction of LLaMA
OpenLLaMA is an open-source reproduction of Meta's LLaMA model trained on the RedPajama dataset, released under permissive Apache 2.0 licenses. It was one of the first serious attempts to produce a fully open LLaMA-quality model that anyone could use commercially.
Chatbot Arena: Benchmarking LLMs in the Wild with Elo Ratings
Chatbot Arena is LMSYS's crowdsourced LLM benchmark where users rate anonymous head-to-head model comparisons using Elo ratings — producing rankings that reflect real user preferences rather than task-specific metrics. The Elo format became the dominant live benchmark for tracking which frontier model is currently best.
Heinsen Routing: Vector Routing for Sequences and Capsule Networks
Heinsen routing is a novel algorithm for routing vectors in sequences and capsule networks — an alternative to Hinton's dynamic routing that generalizes across domains. A niche but foundational contribution to the capsule network research line.
From Deep to Long Learning?
Stanford Hazy Research argues the next frontier is moving from deep networks to networks that can process very long sequences — motivating state space models like Mamba as a shift away from transformer attention's O(n²) complexity. A prescient 2023 post about where sequence modeling was headed.
Self-Refine: Iterative LLM Output Improvement via Self-Feedback
Self-Refine is a framework where LLMs generate feedback on their own outputs and iteratively refine them — no human feedback, no gradient updates. Shows that a single LLM can be its own critic and improve outputs across diverse tasks.
ChatArena: Multi-Agent Language Game Environments for LLMs
ChatArena provides multi-agent game environments for LLMs — structured settings where multiple LLMs interact, debate, negotiate, or play games with defined rules and roles. A research framework for studying emergent multi-agent behavior.
Prompt Engineering: Lilian Weng's Comprehensive Survey
Lilian Weng's canonical reference on prompt engineering techniques — zero-shot, few-shot, chain-of-thought, self-consistency, tree-of-thoughts, and more — grounded in research papers. Still the most comprehensive single-author survey of the space.
Jim Fan: Foundation Models for Embodied Agents
Jim Fan's talk on foundation models for embodied agents — covering how large pretrained models can be adapted for physical and simulated agents that act in the world. Previews the Voyager and MineDreamer research that followed from his NVIDIA lab.
Generate Rather Than Retrieve: Large Language Models Are Strong Context Generators
Yu et al. (2022) show that prompting an LLM to generate its own background context before answering a question (GenRead) outperforms retrieval-based approaches on several knowledge-intensive NLP benchmarks. The result challenges the assumption that retrieval is necessary for grounding LLM outputs.
Machine Learning and Complexity Theory
A theoretical computer scientist's reflection on the relationship between machine learning and computational complexity theory after a Dagstuhl seminar — noting that the two fields remain largely disconnected despite studying adjacent phenomena. Raises questions about whether complexity theory has useful things to say about why ML works.
ControlNet: Precise Spatial Control for Diffusion Models
ControlNet adds fine-grained spatial control to Stable Diffusion — use edge maps, depth maps, pose skeletons, or sketches to precisely direct where objects and structures appear in generated images. A major step beyond text-only prompting for image generation.
Compute-Optimal LLMs: Chinchilla Scaling Calculator
howmanyparams.com is a calculator for compute-optimal LLM training based on Chinchilla scaling laws — given a compute budget, it tells you the optimal model size and token count. A practical tool for applying the Hoffmann et al. scaling law findings.
Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
Lewis et al. (Facebook AI, 2020) introduce Retrieval-Augmented Generation, a hybrid architecture that combines dense passage retrieval with seq2seq generation to ground language model outputs in a non-parametric knowledge store. RAG defined the template that most production knowledge-grounded LLM systems follow today.
ChatRWKV: ChatGPT Powered by an RNN
ChatRWKV is a ChatGPT-like chatbot built on the RWKV architecture — a language model that achieves transformer-level performance using an RNN design, enabling constant memory inference regardless of sequence length. A significant architectural alternative to attention-based transformers.
Human Motion Diffusion Model (MDM)
Tevet et al. (Tel Aviv University, arXiv:2209.14916, 2022) apply diffusion models to human motion generation, producing MDM — a transformer-based denoiser that generates realistic motion sequences from text descriptions or action labels. It matters because it extends the generative power of diffusion to a structured temporal domain, enabling controllable motion editing that prior methods couldn't match.
Flash Attention: Training Examples
Flash Attention's training examples from Hazy Research — demonstrating how to use the memory-efficient attention implementation for training large transformers. Flash Attention became one of the most impactful algorithmic contributions to LLM training efficiency.
Byte Size ArXiv
Byte Size ArXiv summarizes cutting-edge ML research papers in accessible, digestible form. A useful filter layer over the torrent of ArXiv submissions for people who want to track research without reading full papers.
Program of Thoughts Prompting: Disentangling Computation from Reasoning
Program of Thoughts separates reasoning from computation by having LLMs write executable Python programs rather than performing arithmetic inline, delegating number-crunching to an interpreter. It substantially outperforms chain-of-thought on numerical reasoning benchmarks by eliminating the arithmetic errors that plague prose reasoning chains.
Large Language Models are Human-Level [...] (Nov 2022)
A November 2022 paper claiming LLMs perform at human level on some capability, with a truncated filename that doesn't reveal the specific domain. Likely from the computational social science or NLP annotation literature; the save date coincides with ChatGPT's launch and peak interest in LLM capability claims.
Magic3D — High-Resolution Text-to-3D Content Creation
Magic3D is NVIDIA Research's text-to-3D content creation method that generates high-resolution 3D meshes from text prompts using a two-stage coarse-to-fine optimization. An important step beyond NeRF-based generation toward production-usable 3D assets.
Improving Multimodal Interactive Agents with RLHF
The Interactive Agents Team at DeepMind (arXiv:2211.11602, 2022) applies RLHF to agents that must understand language instructions and act in visual environments, using human preference feedback to train a reward model for PPO-based RL. The result demonstrates that RLHF substantially improves instruction-following in embodied multimodal settings beyond what supervised learning alone achieves.
Andy Matuschak's Research Fellowship
Andy Matuschak's six-month research fellowship for building tools for thought — $25k stipend, mentorship, and a working prototype as output. Notably honest about uncertainty: no career path, no guarantees, a bet on an emerging field.
TART: Task-Aware Retrieval with Instructions
TART (Task-Aware Retrieval with Instructions) introduces BERRI, a dataset of ~40 retrieval tasks annotated with human-written task instructions, and trains a multi-task retrieval system that adapts its behavior based on explicit instructions. TART outperforms much larger models on BEIR by understanding the user's intent rather than just matching query-document similarity.
DIFFUSER: Discrete Diffusion via Edit-Based Reconstruction
Introduces DIFFUSER, an edit-based text generation model that adapts denoising diffusion to discrete text by framing generation as iterative editing rather than left-to-right token production. Competitive with autoregressive models on translation and summarization while enabling unique capabilities like prototype-conditioned generation and iterative revision.
Synced — AI Research and Industry News
Synced is an AI and tech news publication covering ML research, industry developments, and AI applications — originally focused on the Chinese AI ecosystem but expanded globally. One of the more substantive English-language sources tracking AI research news.
An Image is Worth One Word: Personalizing Text-to-Image Generation Using Textual Inversion
Tel Aviv University and NVIDIA paper introducing Textual Inversion — learning a single new text embedding token that represents a user-provided concept, enabling that concept to be composed into any text prompt. Showed that the embedding space of text-to-image models is richly structured and can be expanded with just 3-5 example images.
Neural Theory-of-Mind? On the Limits of Social Intelligence in Large LMs
Sap et al. test LLMs on theory-of-mind tasks and find they rely on spurious correlations and dataset artifacts rather than genuine social reasoning. Models perform well on training distributions but fail on adversarial or decontextualized ToM tests — a gap that matters for social AI applications.
NeRF: Neural Radiance Fields
The original NeRF (Neural Radiance Fields) project page from Matthew Tancik's site — the foundational 2020 paper that represents 3D scenes as neural functions and synthesizes novel views via volumetric rendering. A landmark that spawned a field.
Automatic Prompt Engineer (APE)
Automatic Prompt Engineer (APE) is a research implementation that automatically generates and selects optimal prompts for LLMs using a generate-then-score approach. An early automated prompt optimization method predating DSPy.
State of AI Report
The State of AI Report is an annual analysis of the most interesting developments in AI across research, industry, politics, and safety. Co-authored by Nathan Benaich and Ian Hogarth, it's one of the most comprehensive yearly summaries of where the field stands.
Manifest: Prompt Programming with Foundation Models
Manifest is a Python library from Stanford's HazyResearch lab for prompt programming with foundation models — a unified interface across providers with caching, batching, and structured output support. An early formalization of LLM programming patterns before LangChain dominated.
ASIF: Coupled Data Turns Unimodal Models to Multimodal Without Training
Norelli, Fumero, Maiorca, Rodolà et al. (Sapienza University, arXiv:2210.01738, 2022) show that any two unimodal models can be composed into a zero-shot multimodal system by finding approximate shared nearest neighbors across their embedding spaces, given only a small set of coupled pairs. The result challenges the assumption that multimodal capability requires joint training.
Language Models as Models of the Visual World
Research showing language models can use linear projections of image representations as soft prompts for vision-language tasks — without tuning the LM or image encoder. An early signal for the efficiency of frozen model feature reuse in multimodal architectures.
Interactive and Visual Prompt Engineering for Ad-hoc Task Adaptation with Large Language Models
A research paper presenting interactive and visual tools for prompt engineering that let users iteratively adapt large language models to ad-hoc tasks without fine-tuning. It addresses the core usability gap in prompt engineering: the feedback loop between prompt edits and output quality is opaque without tooling that makes the connection visible.
Large Language Models Are Human-Level Prompt Engineers
Large Language Models Are Human-Level Prompt Engineers (APE) introduces an automated method for generating and selecting optimal prompts using LLMs themselves, matching or beating human-crafted instructions on 19 of 24 NLP tasks. It reframes prompt engineering as a program search problem, making manual iteration unnecessary.
On the Paradox of Learning to Reason from Data
Zhang, Li, Meng, Chang, and Van den Broeck (UCLA) show that BERT achieves near-perfect accuracy on in-distribution logical reasoning problems while completely failing to generalize to other distributions over the same problem space. The explanation: BERT learned statistical features of the logical reasoning distribution, not the underlying reasoning function — a fundamental distinction between benchmark performance and genuine reasoning.
Make-A-Video: Text-to-Video Generation without Text-Video Data
Singer et al. (Meta AI, 2022) introduce Make-A-Video, a text-to-video generation system that learns spatiotemporal motion from unlabeled video while keeping semantic knowledge from paired image-text data. It sidesteps the absence of large-scale video-caption datasets by decoupling what to generate from how things move.
Memory-Based Model Editing at Scale
ICML 2022 paper on memory-based model editing at scale — a method for locally updating a neural network's factual knowledge without full retraining. Uses a cache of explicit edits that override the base model's predictions for targeted inputs.
Editable Neural Networks
ICLR paper proposing 'editable neural networks' — a method for making targeted modifications to a model's behavior on specific inputs without disrupting performance elsewhere. An early formalization of the model editing problem.
Alpa: Automated Distributed Training for Large Models
Alpa is a system for automatically parallelizing large neural network training across distributed hardware — finding optimal parallelism strategies without manual configuration. From a Berkeley/CMU research collaboration, it targets the challenge of scaling models beyond single-GPU memory.
Image as a Foreign Language: BEIT-3 Pretraining for All Vision and Vision-Language Tasks
Wang, Bao, Dong et al. at Microsoft introduce BEIT-3, a general-purpose multimodal foundation model that treats images as a 'foreign language' and applies masked language modeling uniformly across images, text, and image-text pairs. BEIT-3 achieves state-of-the-art across seven vision and vision-language benchmarks including COCO, ImageNet, VQA, and NLVR2.
Awesome-MEV: MEV Research Resource List
A curated list of MEV (Maximal Extractable Value) research papers and talks — the definitive reading list for understanding how miners and validators extract value from transaction ordering on Ethereum. MEV is simultaneously a market efficiency mechanism and a form of tax on users.
A Simple but Effective BERT Model for Dialog State Tracking on Resource-Limited Systems
A BERT-based dialog state tracking model optimized for resource-limited systems — achieving competitive performance on MultiWOZ while reducing model size and inference cost. Demonstrates that careful architecture choices can close the gap between full-scale models and edge-deployable alternatives.
A Short Chronology of Deep Learning for Tabular Data
Sebastian Raschka's chronological survey of deep learning approaches for tabular data — the domain where gradient boosted trees still dominate. A clear-eyed accounting of why deep learning hasn't won on structured data despite winning everywhere else.
GPT-3 and Arithmetic (MIT LINGO Lab)
MIT LINGO Lab's analysis of GPT-3's arithmetic abilities — probing how and when it succeeds or fails at basic math, and what this reveals about how language models represent numerical reasoning. Relevant to understanding the difference between pattern matching and genuine computation.
AI Grant
AI Grant is a no-strings-attached grant program by Nat Friedman and Daniel Gross giving $10k to AI researchers and builders working on ambitious projects. Launched in 2022 as a deliberate attempt to fund AI-first builders before the VC world caught up.
PEER: A Collaborative Language Model
PEER is a language model trained to collaboratively write text with humans — drafting, suggesting edits, following instructions, and explaining its changes. An early formalization of human-AI collaborative writing that anticipates modern writing assistant workflows.
A Library for Representing Python Programs as Graphs for Machine Learning
Google Research library paper introducing python_graphs — an open-source tool for constructing graph representations of Python programs via static analysis, producing control flow graphs, data flow graphs, and program dependence graphs. Standardizes the infrastructure for ML-on-code research so researchers don't each rebuild the graph extraction layer.
Deep Anomaly Detection with Self-Supervised Learning and Adversarial Training
This paper combines self-supervised learning and adversarial training to improve deep anomaly detection, leveraging unlabeled normal data to learn representations that are robust to perturbations and more sensitive to out-of-distribution inputs. It matters because labeled anomaly data is rare in practice, making self-supervised approaches essential for real-world deployment.
Instant NGP: Multiresolution Hash Encoding for Neural Graphics
NVIDIA's Instant NGP trains neural graphics primitives (NeRF, SDF, gigapixel images) in seconds rather than hours using a multiresolution hash encoding. It won SIGGRAPH 2022 Best Paper and made NeRF-based 3D reconstruction practical on a single consumer GPU.
The Alignment Problem from a Deep Learning Perspective
The Alignment Problem from a Deep Learning Perspective (Ngo, Chan, Mindermann, 2022) — frames AI alignment as a problem of specification, robustness, and assurance in deep learning systems. An influential restatement of alignment concerns in terms of modern ML rather than AGI thought experiments.
Probing Vision Transformers
A research repository probing the internal representations and attention mechanisms of Vision Transformers (ViT, DeiT, DINO). Notable finding: self-supervised DINO produces more salient attention maps than supervised models, suggesting better spatial semantics from self-supervision.
Solving Quantitative Reasoning Problems with Language Models (Minerva)
Lewkowycz et al. at Google Research introduce Minerva, a language model pretrained on general text and further trained on technical content that achieves state-of-the-art on quantitative reasoning benchmarks without external tools. It correctly answers nearly a third of undergraduate-level science problems — an early proof that domain-specific pretraining unlocks STEM reasoning at scale.
Language Models Can Teach Themselves to Program Better
Haluptzok, Bowers, and Kalai show that language models can generate their own programming problems and solutions, verify correctness with a Python interpreter, then fine-tune on the verified examples — more than doubling test accuracy. A clean demonstration of self-improvement via external verification.
Distill: Clear Explanations of Machine Learning
Distill.pub is a peer-reviewed ML journal focused on unusually clear explanations of machine learning concepts, using interactive visualizations alongside rigorous technical content. Went on hiatus in 2021 after 5 years, but the archive remains the gold standard for interactive ML pedagogy.
SLED: Efficient Long-Text Understanding with Short-Text Models
SLED (Sliding-Encoder and Decoder) lets you apply short-context pretrained models to arbitrarily long documents by chunking input with overlap and fusing representations in the decoder. Competitive with specialized long-context models without the expensive custom pretraining.
Differentiable Prompt Makes Pre-trained Language Models Better Few-Shot Learners
DifferentiablePrompt (DPT) replaces discrete token prompts with optimized continuous embeddings, enabling gradient-based prompt tuning for few-shot learning. Published at ICLR 2022, it established that soft prompts can match full fine-tuning with far fewer parameters.
DART: Differentiable Prompt Makes Pre-Trained Language Models Better Few-Shot Learners
DART (Differentiable pRompT) trains prompt templates end-to-end via backpropagation, treating prompts as learnable continuous vectors rather than fixed text. It makes small pre-trained language models competitive few-shot learners without scaling to GPT-3 sizes — an important stepping stone between hand-crafted prompts and full fine-tuning.
What Language Model Architecture and Pretraining Objective Work Best for Zero-Shot Generalization?
Wang, Roberts, Scao et al. (BigScience Architecture Group) conduct a large-scale comparison of model architectures (causal decoder, non-causal decoder, encoder-decoder) and pretraining objectives (autoregressive, masked LM) for zero-shot generalization. The key finding: causal decoders + autoregressive LM win at zero-shot; non-causal decoders + masked LM + multitask fine-tuning win overall.
Multitask Prompted Training Enables Zero-Shot Task Generalization
T0 shows that training a language model on 2000+ diverse human-authored prompts across 170+ NLP tasks dramatically improves zero-shot generalization to unseen tasks. An 11B T0 model outperformed 175B GPT-3 zero-shot — proving prompt diversity during training matters more than raw scale for generalization.
Prompt Programming for Large Language Models: Beyond the Few-Shot Paradigm
Reynolds and McDonell (2021) argue that prompting large language models is better understood as programming than as few-shot learning — the few examples in a prompt aren't training data but rather code that specifies the desired computation. This reframing opens up principled prompt design strategies that beat naive templating.
SPoT: Better Frozen Model Adaptation through Soft Prompt Transfer
Vu, Constant, Al-Rfou, Cer, and Lester (Google Research/UMass, 2022) show that initializing soft prompts from a related source task dramatically improves prompt tuning, achieving near fine-tuning performance while keeping the base model frozen. The result reveals that prompt initialization is a critical and underappreciated factor in parameter-efficient adaptation.
Dogetoshi/MEV: Curated MEV Resources
Dogetoshi's MEV repository is a curated collection of resources on maximal extractable value — papers, talks, tools, and code for understanding how searchers, builders, and validators profit from transaction ordering on Ethereum.
Colah's Blog — Chris Olah on Neural Networks
Chris Olah's personal blog is foundational reading for anyone trying to understand deep learning from first principles — his LSTM explainer, neural network topology post, and attention posts have shaped how a generation of researchers think about these models. Olah went on to co-found Anthropic's interpretability team, and the blog reflects his interest in mechanistic understanding.
MEV 2.0: The Rise of MPSVs
A Reciprocal Ventures piece arguing that MEV is evolving from searcher-bot competition to MPSV (Maximal Protocol Surplus Value) — protocols themselves capturing ordering value rather than external actors extracting it. Reframes MEV as a design choice rather than an inevitable externality.
Top arXiv Machine Learning Papers in 2021
A community-ranked list of the top ML papers on arXiv in 2021 according to metacurate.io. A useful snapshot of what the ML research community considered most impactful at a pivotal year — just before diffusion models and ChatGPT reshaped the field.
TWAMM: Time-Weighted Average Market Maker
Paradigm's paper introducing the Time-Weighted Average Market Maker (TWAMM) — a DEX primitive designed to execute large orders over time by continuously trading against an embedded AMM. Elegant solution to the large-order MEV problem that haunts standard AMMs.
Hanley 2021 (Unidentified Paper)
A 2021 paper by an author named Hanley, saved June 2022 with a citation-style filename. Subject unknown; likely ML, NLP, or computer security given the owner's interests. Stub note — open the PDF to identify.
On the Measure of Intelligence
François Chollet's foundational 2019 paper arguing that measuring AI by task skill conflates intelligence with prior knowledge, proposing instead a definition of intelligence as skill-acquisition efficiency — and introducing the ARC benchmark designed to test it. One of the most influential theoretical framings of what AGI evaluation should actually measure.
@bertcmiller MEV Threads
@bertcmiller's chronological index of his MEV Twitter threads — the most cited informal resource for understanding MEV strategies, Flashbots, and searcher mechanics. Essential reading for anyone entering the MEV space.
PNAS Paper: 2113883119
A PNAS paper (DOI: 10.1073/pnas.2113883119) saved in April 2022 — exact title and topic unclear from filename alone, but likely a life sciences or multidisciplinary research article given the journal and period. Needs direct reading to fully characterize.
Differentiable Self-Organizing Systems Workshop
Workshop site for the Differentiable Self-Organizing Systems research thread — covering gradient-based approaches to systems that grow, repair, and self-organize. The academic home of the Neural Cellular Automata research program from Mordvintsev et al.
All Roads Lead to Rome: The ML Job Market in 2022
Eric Jang's 2022 essay on the ML job market — titled 'All Roads Lead to Rome,' arguing that different entry points (research labs, industry, startups) all converge on the same destination if you're technically strong. Candid advice from a DeepMind robotics researcher.
ML Paper Volume: Moore's Law for arXiv
A tweet noting that ML paper volume on arXiv is growing exponentially — keeping current requires reading 100+ papers per day. A Moore's Law analogy applied to research output rather than transistor density, capturing the 2022 ML research explosion.
Disulfiram as a Potential Anti-Anxiety Drug
Tokyo University of Science research finding that disulfiram — the alcohol-deterrent drug Antabuse — safely reduces anxiety in rodents, potentially via ALDH2 inhibition affecting acetaldehyde metabolism. Drug repurposing with a well-understood safety profile.
Secure Byzantine-Robust Machine Learning
He, Karimireddy, and Jaggi propose a two-server cryptographic protocol that simultaneously achieves Byzantine robustness, input privacy, and local differential privacy for distributed machine learning — three properties usually addressed separately. Bridges the privacy-robustness gap in federated learning.
Learning from History for Byzantine Robust Optimization
Karimireddy, He, and Jaggi (EPFL, arXiv:2012.10333, 2021) propose using historical gradient information to detect and filter Byzantine workers in distributed training, achieving near-optimal convergence even with a constant fraction of corrupt workers. The historical approach is notable because it breaks the fundamental limitation of single-round Byzantine filters without requiring cryptographic overhead.
A Practical Guide to Multi-Objective Reinforcement Learning and Planning
A Springer survey on multi-objective reinforcement learning and planning — covers scalarization, Pareto-based methods, and utility-based approaches for agents that must balance competing rewards. Useful reference for RL research where single-reward framing is inadequate.
Anthropic (April 2022)
Anthropic's job postings page saved in April 2022 — when the company was roughly one year old, pre-Claude, and actively hiring its founding team. A historical snapshot of Anthropic's early public presence before it became widely known.
Tokenomics DAO
Tokenomics DAO is a community focused on researching and discussing token economic design — how token supply, distribution, incentives, and governance mechanisms shape crypto protocol behavior. A 2022-era community for the nascent field of token engineering.
BuiltWith Technology Lookup
BuiltWith is a web technology profiler — enter any URL and see what frontend frameworks, analytics, CDNs, ad networks, CMS, and hosting stack it uses. Useful for competitive intelligence, technology research, and understanding how production websites are built.
Institute for the Study of War
The Institute for the Study of War (ISW) is a nonpartisan think tank that publishes detailed military analysis and maps of active conflicts. Became an essential source for tracking the Russia-Ukraine war in 2022.
Will Transformers Take Over Artificial Intelligence?
Quanta Magazine's 2022 look at whether transformer architectures will dominate all of AI — following their success in NLP and early incursion into image classification. A useful time-capsule of the moment when the transformer paradigm started feeling inevitable.
The Bitter Lesson
Rich Sutton's 2019 essay arguing that the dominant lesson from 70 years of AI research is that general methods leveraging computation always win over human-engineered knowledge — a humbling argument against clever domain-specific tricks. One of the most cited and debated essays in ML.
How Do Data Science Workers Collaborate? Roles, Workflows, and Tools
ACM survey of how data science workers actually collaborate in practice — finding distinct roles (data engineers, ML engineers, domain scientists, communicators) that rarely map to job titles, with coordination breakdowns at handoff points between roles. One of the few empirical studies of real data science team dynamics rather than idealized workflow diagrams.
ML and NLP Research Highlights of 2021
Sebastian Ruder's annual ML/NLP research highlights for 2021 — covering foundation models, the prompting revolution, AlphaFold 2, diffusion models, and the growing focus on efficiency and responsible AI. The most useful single-document summary of where the field moved that year.
Cogment Verse: Human-in-the-Loop Reinforcement Learning
Cogment Verse is an SDK for training and validating AI agents in human-in-the-loop learning (HITL) and multi-agent reinforcement learning environments. Provides a web UI for human participation in training alongside standard RL algorithms like A2C and PPO.
How Sber Built ruDALL-E — Interview with Sergei Markov
Serokell's interview with Sergei Markov of SberDevices about building ruDALL-E — a 12B parameter Russian-language text-to-image model. Covers the engineering and research challenges of training massive multimodal models, plus the open-source culture argument in ML.
Other Internet — Research Collective
Other Internet is a research collective and studio exploring the social, organizational, and cultural implications of internet-native institutions — particularly DAOs, crypto protocols, and new forms of online coordination. Produces longform research that goes deeper than most crypto commentary.
Operationalizing Machine Learning: Forrester Research Report
Forrester Research report on operationalizing machine learning in enterprise settings, covering the organizational and technical requirements for moving models from development into production reliably. Useful snapshot of where enterprise ML practice stood in late 2021 and the key gaps between experimentation and production.
Layer-1 Platforms: A Framework for Comparison
The Block Research's comparative framework for seven Layer-1 blockchains — Algorand, Avalanche, Binance Smart Chain, Cosmos, Ethereum 2.0, Polkadot, and Solana. A structured way to think about what actually differentiates L1 architectures, commissioned by Algorand in 2021.
applied-ml: Papers and Tech Blogs on ML in Production
Eugene Yan's curated list of papers and engineering blog posts from companies sharing real-world ML in production — classification, recommendation, search, NLP, and more. One of the most useful ML reference repositories because it focuses on what actually shipped, not just what was published.
Semantic Textual Similarity — Papers with Code Benchmark
Papers with Code's benchmark page for Semantic Textual Similarity — a task measuring how similar two sentences are in meaning, scored against human judgments. The canonical reference for comparing embedding models on this fundamental NLP task.
DSHR's Blog — David Rosenthal
David Rosenthal's DSHR blog covers digital preservation, storage economics, and internet infrastructure with rare depth and rigor. One of the few places where long-run storage cost trends, blockchain critique, and library science intersect.
Berkeley Document Summarizer
The Berkeley Document Summarizer is a learning-based extractive summarization system that uses syntactic compression and coreference constraints. Academic research code from Greg Durrett, representing the pre-neural era of NLP summarization work.
Reproducible Deep Learning PhD Course
Simone Scardapane's PhD course on Reproducible Deep Learning — covering Git, Docker, DVC, experiment tracking, and CI/CD for ML research. Addresses the reproducibility crisis in deep learning with practical tooling.
Effects of Intermittent Fasting on Body Composition and Clinical Health Markers
A review of clinical evidence on intermittent fasting protocols (alternate-day, whole-day, time-restricted feeding) and their effects on body composition and health markers. Finds comparable fat loss to continuous calorie restriction with possible lean-mass preservation advantages.
Internet Search Tips
Gwern Branwen's guide to effective internet search — covering advanced operators, specialized databases, tracking down primary sources, and the general epistemics of knowing when you've actually found what you're looking for. One of the most useful practical research guides online.
Self-Supervised Learning: The Dark Matter of Intelligence
Yann LeCun and Ishan Misra's Facebook AI blog post arguing that self-supervised learning — learning from unlabeled data — is the key to human-level AI, analogous to the dark matter that makes up most of the universe's mass. Published ahead of a wave of self-supervised breakthroughs.
State-of-the-Art Image Generative Models (2021)
Aran Komatsuzaki's March 2021 survey of state-of-the-art image generative models — covering BigGAN, VQVAE-2, DALL-E, CLIP, and diffusion models just as they were emerging. A historical snapshot of the field one year before diffusion models took over.
10 Categories of Deep Recommendation Systems
James Le's survey of 10 categories of deep learning-based recommendation systems — from MLP and autoencoder approaches through attention-based and graph neural network methods. A useful taxonomy for understanding how the field moved beyond matrix factorization.
Crossminds: ML Research Video Platform
Crossminds was a platform for discovering and watching technical ML research videos — conference talks, paper presentations, and lectures organized by topic and institution. It aimed to be YouTube but with research-quality signal-to-noise filtering.
The Overfitted Brain: Dreams Evolved to Assist Generalization
A 2020 paper proposing that dreams evolved as a biological regularization mechanism — the brain 'trains' on noisy, hallucinated data during sleep to prevent overfitting to waking experience. A striking bridge between ML theory and sleep neuroscience.
Applied ML: Papers and Blogs on ML in Production (2020)
Eugene Yan's curated GitHub list of papers and blog posts on ML in production — original 2020 bookmark of this now-landmark repository. Covers recommendation, search, NLP, data quality, feature engineering, and more from companies that actually shipped these systems.
Superpixel Segmentation — IVRL
The IVRL lab at EPFL's research page on superpixel segmentation — home of the SLIC algorithm, which became the dominant superpixel method due to its speed and perceptual uniformity. Superpixels are a fundamental preprocessing step in classical computer vision pipelines.
Zipf's Law Arises Naturally When There Are Underlying, Unobserved Variables
A PLOS Computational Biology paper arguing that Zipf's Law — the power law governing word frequency, city size, income distribution, and countless other phenomena — arises naturally whenever observed variables are log-linear functions of hidden unobserved variables. Explains the law's ubiquity without requiring any special generative mechanism.
A Multi-Level Bayesian Analysis of Racial Bias in Police Shootings
A PLOS ONE paper applying multi-level Bayesian hierarchical models to police shooting data across US counties from 2011–2014, finding significant racial disparities in lethal force use. Notable for applying rigorous statistical methods to a politically charged dataset.
The Three Cultures of Machine Learning
Jason Eisner's JHU tutorial describing three distinct cultures within machine learning — the statistical/probabilistic, the algorithmic/computational, and the geometric/optimization traditions. Useful framing for understanding why ML researchers sometimes talk past each other.
Yann LeCun: Making Facebook's AI Predict What Happens in Videos
New Scientist interview with Yann LeCun on Facebook AI Research's goal to build models that predict what will happen in videos — covering what AI can and can't do in 2015, and LeCun's view on unsupervised learning as the key unsolved problem.
Security Data Science Papers
Covert.io's curated list of academic papers applying data science and machine learning to security problems — covering network intrusion detection, malware classification, anomaly detection, and more. A reference for practitioners working at the intersection of ML and cybersecurity.
Data Cleaning: Problems and Current Approaches (Berkeley/UNECE)
Joe Hellerstein's Berkeley paper on data cleaning for UNECE — a systematic treatment of the data quality problem from a database research perspective. The academic foundation for what practitioners know as the most time-consuming part of data science work.
Deep Support Vector Machines
A video lecture on Deep Support Vector Machines from ROKS 2013 — hybrid architectures combining deep feature learning with SVM classification. A snapshot of the moment researchers explored whether SVMs and deep learning could coexist before end-to-end networks won out.
5D 'Superman Memory' Crystal — Unlimited Lifetime Data Storage
University of Southampton researchers demonstrated 5D optical data storage in quartz glass using femtosecond laser writing — 360 TB per disc, theoretically stable for billions of years. An early result from the lab that eventually led to Project Silica.
I'm a Bandit — Sébastien Bubeck's Optimization Blog
I'm a Bandit is Sébastien Bubeck's research blog on optimization and online learning — covering convex optimization, bandit algorithms, and randomized methods at a level between textbook and paper. One of the best technical blogs in theoretical ML/optimization.
Many Downloadable Twitter Archives Available for Researchers
DataScholars on downloadable Twitter archives available for academic researchers — a 2013 directory of public Twitter datasets for NLP, social network analysis, and computational social science. A snapshot of open Twitter data before the API became restrictive.
How to Implement an Algorithm from a Scientific Paper
Emmanuel Goossaert's guide to the specific challenges of implementing an algorithm from a research paper — reading notation, handling undefined edge cases, bridging the gap between mathematical description and working code. Practical advice for a surprisingly common problem.
Social Machine Learning — Cambridge Computer Lab Slides
Slides from a Cambridge Computer Lab course on social machine learning — applying ML to social network data, link prediction, community detection, and behavior modeling. A 2010-era academic reference captured at the boundary between network science and machine learning.
SNAP: Stanford Network Analysis Project Manual
The manual for SNAP (Stanford Network Analysis Project) v0.15 — Jure Leskovec's C++ library for large-scale graph analysis and network mining. One of the earliest high-performance graph analysis toolkits, predating Spark GraphX and Neo4j's graph algorithms by years.
Picosecond Protein Watching
Derek Lowe's 'In the Pipeline' post on picosecond-timescale protein dynamics measurement — covering new spectroscopic techniques that let researchers observe protein conformational changes at femtosecond-to-picosecond resolution. A window into the frontier of structural biology.
HyperDex: NoSQL Database Based in Hyperspace
The HyperDex academic paper introducing a NoSQL key-value store that used hyperspace hashing to support efficient secondary attribute searches — the main weakness of most key-value stores at the time. An elegant theoretical approach that didn't achieve mainstream adoption.
Best Paper Awards Across CS Conferences
Jeff Huang's maintained list of best paper awards across 30+ top CS conferences (AAAI, ACL, CHI, SIGCOMM, etc.) — a curated entry point into landmark research across computer science sub-disciplines.
Large-Scale Graph Computing at Google (Pregel)
Google Research's 2009 blog post introducing Pregel — their internal system for large-scale graph computation using a bulk-synchronous-parallel model. The post that launched the graph processing systems category and eventually spawned Apache Giraph, GraphX, and the whole vertex-centric computing tradition.
Dremel: Interactive Analysis of Web-Scale Datasets
Google's Dremel paper — the system that enabled sub-second SQL queries over petabyte datasets via columnar storage and a multi-level serving tree. The direct precursor to BigQuery, and the inspiration behind Apache Parquet's nested record encoding.
ZeroN: Magnetic Levitation That Remembers
ZeroN is an MIT Media Lab system that levitates a metal ball in mid-air using electromagnetic control, then records and plays back its movements in 3D space. A tangible interface where a physical object becomes a programmable point in space — bridging physical and digital interaction.
Physicists Store a Two-Frame Movie in an Atomic Cloud
NIST researchers stored two sequential image frames — a functional two-frame 'movie' — inside a cloud of ultracold rubidium atoms using quantum optical techniques. A proof-of-concept for quantum memory, storing light itself as atomic spin states.
The Physics of Why Coffee Spills When You Walk
Researchers modeled why coffee spills when you walk with a cup, finding the human gait frequency nearly matches the resonant frequency of liquid sloshing in a standard mug. The fix: walk irregularly, use a lid, or carry a non-cylindrical container.
Sodium-Ion Battery Electrode Breakthrough
Researchers developed a new metal-oxide electrode material enabling sodium-ion batteries that could substitute for lithium-ion in stationary energy storage. Sodium is 1000x more abundant than lithium and much cheaper, making this a scalable alternative for grid storage.