PUBLISHED AI RESEARCH AND PROJECTS

AI research and projects

The research, projects, and explanations behind the people. Read the source, inspect verified code and models, and find a collaborator.

Clear all filters

A source check confirms that a source supports the summary. It is not an independent replication or endorsement. Our methodology ↗

Showing 113–128 of 149 proofs of work · Kind: research · Newest first
Page 8 of 10
researchChecked Sep 20, 2026

Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV)

Dr. Been Kim · Google DeepMind

Invented TCAV (Testing with Concept Activation Vectors), an interpretability framework that uses directional derivatives in activation space to quantify how much a human-understandable concept (e.g., stripes on a zebra) contributes to a model's prediction.

Scope & limitations

Requires user-provided exemplars for concept definitions; high-level abstract concepts without clear visual or lexical exemplars can be difficult to vectorize cleanly.

Context: InceptionV3, ResNet internal activation layers, linear classifiers in latent embedding manifolds.

View mission
researchChecked Sep 20, 2026

Leakage and the Reproducibility Crisis in Machine-Learning-Based Science

Dr. Sayash Kapoor · Princeton University

Published a comprehensive meta-analysis of machine learning scientific literature discovering data leakage across 329 papers spanning 17 fields, establishing formal taxonomy and verification criteria to prevent train-test contamination.

Scope & limitations

Meta-analyses depend on publicly available replication code; studies with proprietary or non-shared datasets could not be independently tested.

Context: Reproducibility audit across computer vision, clinical healthcare predictions, and natural language processing.

View mission
researchChecked Sep 20, 2026

Mixtral of Experts: Sparse Mixture-of-Experts Open Weight Model

Arthur Mensch · Mistral AI

Published the architectural blueprint for Mixtral 8x7B, a sparse mixture-of-experts model where a router chooses 2 out of 8 feedforward experts per token, activating only 13B parameters out of 47B during forward pass and outperforming Llama 2 70B with 6x faster throughput.

Scope & limitations

High RAM/VRAM memory footprint for hosting all 47B parameters despite activating only 13B during computation; requires specialized multi-GPU tensor or pipeline parallelism.

Context: 46.7B total parameters, 12.9B active parameters per token, 32k token context window, Top-2 softmax gating router.

View mission
researchChecked Sep 20, 2026

Model Cards for Model Reporting: Standardizing ML Evaluation Disclosures

Dr. Margaret Mitchell · Hugging Face

Authored the seminal ACM FAccT paper introducing 'Model Cards for Model Reporting', an operational documentation standard that details model architecture, intended tasks, training parameters, out-of-scope uses, and disaggregated demographic evaluations.

Scope & limitations

Static documentation can diverge from deployed weights if model inputs undergo runtime changes or unlogged adapter swaps.

Context: Disaggregated evaluation metrics across intersecting demographic slices and environmental contexts.

View mission
researchChecked Sep 20, 2026

On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?

Dr. Timnit Gebru · Distributed AI Research Institute (DAIR)

Published the landmark ACM FAccT paper critically examining the environmental costs, training data biases, lack of semantic grounding, and illusion of intent when scaling large autoregressive language models without curated provenance.

Scope & limitations

Focusing on systemic risks does not quantify operational benefits in domain-specific tasks where automated verification pipelines exist.

Context: Analysis of Web-scraped pre-training corpora (Common Crawl, C4) and environmental compute costs.

View mission
researchChecked Sep 20, 2026

Relative Representations Enable Non-Degenerate Latent Space Alignment

Dr. Been Kim · Google DeepMind

Co-authored breakthrough research demonstrating that representations can be compared across disparate neural networks by computing pairwise angles and similarities to anchor points, proving latent geometries are invariant across architectures.

Scope & limitations

Choice and distribution of anchor points can introduce variance in representation reconstruction accuracy across domain shifts.

Context: Cross-model embedding alignment across Vision Transformers, CNNs, and language models without weight access.

View mission
researchChecked Sep 20, 2026

Social Chemistry 101: Learning to Reason about Social and Moral Norms

Dr. Yejin Choi · University of Washington / Allen Institute for AI

Introduced Social Chemistry 101, a computational framework and dataset containing 292,000 rules-of-thumb covering social conventions, moral rules, and situational ethical trade-offs, enabling LLMs to reason explicitly about multi-cultural norms.

Scope & limitations

Moral norms vary across geographic cultures, religious traditions, and temporal epochs; models require adaptive normative context parameters.

Context: Structured social rule-of-thumb generation with attribute classifications (morality, social judgment, agreement level).

View mission
researchChecked Sep 20, 2026

The Pile: An 800GB Dataset of Diverse Text for Language Modeling

Stella Biderman · EleutherAI / Booz Allen Hamilton

Co-authored the curation and statistical characterization of The Pile, an 825GB open-source pre-training dataset combining 22 diverse text subsets (arXiv, PubMed, GitHub, StackExchange), establishing the foundation for modern open foundation models.

Scope & limitations

Includes academic text requiring high domain literacy; copyright and licensing complexities in mixed web-scraped sub-components.

Context: FastText deduplication, exact substring filtering, perplexity benchmark analysis.

View mission
researchChecked Sep 20, 2026

WinoGrande: An Adversarial Winograd Schema Challenge at Scale

Dr. Yejin Choi · University of Washington / Allen Institute for AI

Authored WinoGrande, an adversarial benchmark of 44,000 pronoun resolution problems designed using AFLITE algorithmic bias-reduction to prevent models from exploiting superficial dataset artifacts, becoming a gold-standard benchmark for evaluating machine commonsense.

Scope & limitations

Frontier models now achieve high performance on static text pairs; evaluating dynamic situational commonsense requires multi-turn environmental grounding.

Context: AFLITE adversarial filtering over RoBERTa embeddings; standard inclusion in Open LLM Leaderboard and HELM.

View mission
researchChecked Sep 20, 2026

ULMFiT: Universal Language Model Fine-tuning for Text Classification

Jeremy Howard · Answer.ai / fast.ai

Seminal research paper introducing inductive transfer learning to natural language processing via 3-step pre-training, discriminative fine-tuning, and slanted triangular learning rates, outperforming prior custom architectures by orders of magnitude.

Scope & limitations

Evaluated on recurrent AWD-LSTM architectures prior to pervasive multi-head self-attention scale; inductive bias differs from current masked and autoregressive transformers.

Context: AWD-LSTM 3-layer language model pre-trained on WikiText-103; foundation for modern BERT/transformer transfer learning pipelines.

View mission
researchChecked Sep 20, 2026

8-Bit Optimizers via Block-wise Dynamic Quantization with Low-Precision States

Tim Dettmers · Allen Institute for AI (Ai2)

A breakthrough algorithmic paper by Tim Dettmers introducing 8-bit Adam and low-precision optimizer states, reducing memory consumption of deep learning training states by 75% without compromising model convergence.

Scope & limitations

Custom CUDA block-level quantization; optimizer speedups are maximized on modern Tensor Core architectures.

Context: Benchmarked on language modeling, machine translation, and ImageNet classification with PyTorch.

View mission
researchChecked Sep 20, 2026

Abu Dhabi Investment Authority (ADIA): Empirical Domain-Specific AI Evaluation Benchmark & Methodology

Dr. Marcos Lopez de Prado · Abu Dhabi Investment Authority (ADIA)

An empirical evaluation methodology developed by Dr. Marcos Lopez de Prado at Abu Dhabi Investment Authority (ADIA), establishing standardized performance metrics, domain-specific evaluation sets, and error taxonomy protocols.

Scope & limitations

Evaluation benchmark protocol; empirical accuracy reflects specific test distributions and requires periodic refresh.

Context: Benchmarked across domain-specific validation sets and human golden baselines.

View mission
researchChecked Sep 20, 2026

Natural Adversarial Examples and Out-of-Distribution Detection in Neural Networks

Dr. Dan Hendrycks · Center for AI Safety

A foundational safety study by Dr. Dan Hendrycks introducing the ImageNet-A and ImageNet-O benchmarks, exposing fundamental vulnerabilities where vision models fail catastrophically on naturally occurring adversarial examples.

Scope & limitations

Vision classification benchmark; physical world adversarial perturbations present additional environmental dynamics.

Context: Evaluated ResNet, DenseNet, and early vision-transformer architectures under natural distribution shifts.

View mission
researchChecked Sep 20, 2026

Outsider Oversight: Designing Independent External Audits for Deployed AI

Inioluwa Deborah Raji · Mozilla Foundation / UC Berkeley

An influential empirical paper by Deborah Raji examining the methodology and systemic impact of third-party external audits on proprietary machine learning systems deployed in criminal justice and hiring.

Scope & limitations

External audit methodology; access to internal training datasets is often obstructed by commercial trade secrecy.

Context: Evaluates algorithmic accountability across automated facial recognition and predictive risk scoring algorithms.

View mission
researchChecked Sep 20, 2026

DEF CON 31 Generative Red Team Challenge: Community Vulnerability Discovery

Dr. Rumman Chowdhury · Humane Intelligence

The largest public AI red-teaming exercise in history, organized by Dr. Rumman Chowdhury and Humane Intelligence, evaluating frontier LLMs across thousands of independent hackers for demographic bias, hallucination, and prompt injection.

Scope & limitations

Cross-sectional snapshot of model behaviors under community stress; commercial safety alignment shifts rapidly with model updates.

Context: Evaluated models from OpenAI, Google, Anthropic, Meta, Hugging Face, Cohere, and Stability AI.

View mission
researchChecked Sep 20, 2026

Johns Hopkins University / Bayesian Health: Domain-Specific AI Evaluation Benchmark & Methodology

Dr. Suchi Saria · Johns Hopkins University / Bayesian Health

An empirical evaluation methodology developed by Dr. Suchi Saria at Johns Hopkins University / Bayesian Health, establishing standardized performance metrics, domain-specific evaluation sets, and error taxonomy protocols.

Scope & limitations

Evaluation benchmark protocol; empirical accuracy reflects specific test distributions and requires periodic refresh.

Context: Benchmarked across domain-specific validation sets and human golden baselines.

View mission
Showing 113–128 of 149 proofs of work