Is this you? Claim this profile to correct it, add a bio and choose the work people see first.
Claim this profileWorks102 from public data
- 2,227
- Mobility network models of COVID-19 explain inequities and inform reopening1,561
A metapopulation susceptible–exposed–infectious–removed (SEIR) model that integrates fine-grained, dynamic mobility networks to simulate the spread of SARS-CoV-2 in ten of the largest US metropolitan areas is introduced and correctly predicts higher infection rates among disadvantaged racial and socioeconomic groups.
- 1,322
- Distributionally Robust Neural Networks807
The results suggest that regularization is critical for worst-group performance in the overparameterized regime, even if it is not needed for average performance, and introduce and provide convergence guarantees for a stochastic optimizer for this group DRO setting.
- DataComp: In search of the next generation of multimodal datasets782
The introduction of DataComp, a testbed for dataset experiments centered around a new candidate pool of 12.8 billion image-text pairs from Common Crawl, enables training a CLIP ViT-L/14 from scratch to 79.2% zero-shot accuracy on ImageNet, and makes the benchmark accessible to researchers with varying resources.
- OpenFlamingo: An Open-Source Framework for Training Large Autoregressive Vision-Language Models663
OpenFlamingo is an ongoing effort to produce an open-source replication of DeepMind's Flamingo models, and the models, training data, hyperparameters, and evaluation suite are introduced.
- Mapping the Pairwise Choices Leading from Pluripotency to Human Bone, Heart, and Other Mesoderm Cell Types490
The extrinsic signals controlling each binary lineage decision were defined, enabling us to logically block differentiation towards unwanted fates and rapidly steer pluripotent stem cells towards 80–99% pure human mesodermal lineages at most branchpoints.
- On Random Weights and Unsupervised Feature Learning467
The answer is that certain convolutional pooling architectures can be inherently frequency selective and translation invariant, even with random weights, and the viability of extremely fast architecture search is demonstrated by using random weights to evaluate candidate architectures, thereby sidestepping the time-consuming learning process.
- Are aligned neural networks adversarially aligned?415
It is shown that existing NLP-based optimization attacks are insufficiently powerful to reliably attack aligned text models: even when current NLP-based attacks fail, the authors can find adversarial inputs with brute force.
- 372
- Tiled convolutional neural networks353
This paper proposes tiled convolution neural networks (Tiled CNNs), which use a regular "tiled" pattern of tied weights that does not require that adjacent hidden units share identical weights, but instead requires only that hidden units k steps away from each other to have tied weights.
- OLMoE: Open Mixture-of-Experts Language Models352
This work introduces OLMoE, a fully open, state-of-the-art language model leveraging sparse Mixture-of-Experts (MoE), and presents various experiments on MoE training, analyzes routing in the model showing high specialization, and open-source all aspects of the work: model weights, training data, code, and logs.
- 349
- 296
- Stronger data poisoning attacks break data sanitization defenses292
Three attacks are developed that can bypass a broad range of common data sanitization defenses, including anomaly detectors based on nearest neighbors, training loss, and singular-value decomposition, and the Karush–Kuhn–Tucker conditions.
- 2 OLMo 2 Furious286
This work describes the modified model architecture and training recipe, focusing on techniques for achieving better training stability and improved per-token efficiency, and introduces a new, specialized data mix called Dolmino Mix 1124, which significantly improves model capabilities across many downstream task benchmarks when introduced via late-stage curriculum training.
- Sparse Filtering278
This work presents sparse filtering, a simple new algorithm which is efficient and only has one hyperparameter, the number of features to learn, and evaluates it on natural images, object classification, and phone classification, showing that the method works well on a range of different modalities.
- 269
- 235
- MediQ: Question-Asking LLMs and a Benchmark for Reliable Interactive Clinical Reasoning218
This paper proposes to change the static paradigm to an interactive one, develop systems that proactively ask questions to gather more information and respond reliably, and introduces an interactive MediQ benchmark and a novel question-asking system to evaluate question-asking ability in LLMs.
- Learning Deep Energy Models204
This work proposes deep energy models, which use deep feedforward neural networks to model the energy landscapes that define probabilistic models, and is able to efficiently train all layers of this model simultaneously, allowing the lower layers of the model to adapt to the training of the higher layers, and thereby producing better generative models.
- Impossibility theorems for feature attribution167
This work shows that complete and linear methods are provably less reliable than simpler methods at answering basic feature attribution questions, and provides impossibility results that highlight their failure cases.
- 144
- Wild-Time: A Benchmark of in-the-Wild Distribution Shift over Time143
Wild-Time is curate, a benchmark of 5 datasets that reflect temporal distribution shifts arising in a variety of real-world applications, including patient prognosis and news classification that systematically benchmark 13 prior approaches, including methods in domain generalization, continual learning, self-supervised learning, and ensemble learning.
- The Generative AI Paradox: "What It Can Create, It May Not Understand"136
The hypothesis that models' generative capability may not be contingent upon understanding capability, and call for caution in interpreting artificial intelligence by analogy to human intelligence, is supported.
- Synthesizing scientific literature with retrieval-augmented language models105
A specialized, open-source, retrieval-augmented language model is introduced for answering scientific queries and synthesizing literature, the responses of which are shown to be preferred by human evaluations over expert-written answers.
- 105
- Reliable, Adaptable, and Attributable Language Models with Retrieval99
This position paper proposes a roadmap for developing general-purpose retrieval-augmented LMs, which involves a reconsideration of datastores and retrievers, the exploration of pipelines with improved retriever-LM interaction, and significant investment in infrastructure for efficient training and inference.
- 96
- ReasonIR: Training Retrievers for Reasoning Tasks88
ReasonIR-8B is presented, the first retriever specifically trained for general reasoning tasks, and its performance consistently increases with longer and more information-rich rewritten queries; it continues to outperform other retrievers when combined with an LLM reranker.
- 82
- 81
- OpenScholar: Synthesizing Scientific Literature with Retrieval-augmented LMs80
OpenScholar is introduced, a specialized retrieval-augmented LM that answers scientific queries by identifying relevant passages from 45 million open-access papers and synthesizing citation-backed responses, and achieves citation accuracy on par with human experts.
- An atlas of transcriptional, chromatin accessibility, and surface marker changes in human mesoderm development72
This molecular atlas will facilitate study of human mesoderm development (which cannot be interrogated in vivo due to restrictions on human embryo studies) and provides a broad resource for the study of gene regulation in development at the single-cell level, knowledge that might one day be exploited for regenerative medicine.
- 68
- Denoising genome-wide histone ChIP-seq with convolutional neural networks65
Aconvolutional denoising algorithm, Coda, is introduced that uses convolutional neural networks to learn a mapping from suboptimal to high-quality histone ChIP-seq data, and has the potential to improve data quality at reduced costs.
- Supporting COVID-19 Policy Response with Large-scale Mobility-based Modeling49
This model captures the spread of COVID-19 by using a fine-grained, dynamic mobility network that encodes the hourly movements of people from neighborhoods to individual places, with over 3 billion hourly edges.
- Uncertainty of Thoughts: Uncertainty-Aware Planning Enhances Information Seeking in Large Language Models46
Uncertainty of Thoughts is introduced, an algorithm to augment large language models with the ability to actively seek information by asking effective questions and achieves an average performance improvement of 38.1% in the rate of successful task completion across multiple LLMs compared with direct prompting.
- ExpBERT: Representation Engineering with Natural Language Explanations46
Across three relation extraction tasks, the method, ExpBERT, matches a BERT baseline but with 3–20x less labeled data and improves on the baseline by 3–10 F1 points with the same amount of labeled data.
- DataDecide: How to Predict Best Pretraining Data with Small Experiments43
This work conducts controlled pretraining experiments across 25 corpora with differing sources, deduplication, and filtering up to 100B tokens, model sizes up to 1B parameters, and 3 random seeds and finds that the ranking of models at a single, small size is a strong baseline for predicting best models at the larger target scale (1B).
- Large-Scale Data Selection for Instruction Tuning43
A variant of representation-based data selection (RDS+), which uses weighted mean pooling of pretrained LM hidden states, consistently outperforms more complex methods across all settings tested -- all whilst being more compute-efficient.
- Dissecting an Online Intervention for Cancer Survivors42
The nature of participant engagement with an online, 6-week workshop for cancer survivors and how such engagement may affect health outcomes is explored and the methods used can be used in analyzing other large electronically mediated social-networking interventions.
- RLVE: Scaling Up Reinforcement Learning for Language Models with Adaptive Verifiable Environments39
This work introduces Reinforcement Learning with Adaptive Verifiable Environments (RLVE), an approach using verifiable environments that procedurally generate problems and provide algorithmically verifiable rewards, to scale up RL for language models (LMs) and shows that environment scaling consistently improves generalizable reasoning capabilities.
- 38
- ICONS: Influence Consensus for Vision-Language Data Selection35
This work introduces ICONS, a gradient-based Influence CONsensus approach for vision-language data Selection, which leverages first-order training dynamics to estimate each example's influence on validation performance, then aggregates these estimates across tasks via majority voting.
- Out-of-Domain Robustness via Targeted Augmentations35
It is proved that targeted augmentations improve OOD performance, allowing models to generalize better with fewer domains, and set new states-of-the-art for OOD performance by 3.2-15.2 percentage points.
- Fluid Language Model Benchmarking32
Fluid Benchmarking is introduced, a new evaluation approach that advances language model benchmarking across multiple dimensions based on the insight that the relative value of benchmark items depends on an LM's capability level, suggesting that evaluation should adapt to each LM.
- The Delta Learning Hypothesis: Preference Tuning on Weak Data can Yield Strong Gains31
This work shows that models can learn surprisingly well from paired data that might typically be considered weak, and enables simpler and cheaper open recipes for state-of-the-art post-training.
- A False Sense of Privacy: Evaluating Textual Data Sanitization Beyond Surface-level Privacy Leakage30
This work proposes a new framework that evaluates re-identification attacks to quantify individual privacy risks upon data release and shows that seemingly innocuous auxiliary information—such as routine social activities—can be used to infer sensitive attributes like age or substance use history from sanitized data.
- Proximity-Informed Calibration for Deep Neural Networks30
ProCal is proposed, a plug-and-play algorithm with a theoretical guarantee to adjust sample confidence based on proximity bias that is effective in addressing proximity bias and improving calibration on balanced, long-tail, and distribution-shift settings under four metrics over various model architectures.
- 29
- 23
- PLeaS — Merging Models with Permutations and Least Squares20
A new two-step algorithm to merge models—termed PLeaS—which allows a practitioner to merge two models sharing the same architecture into a single performant model of a desired size, even when the two original models are fine-tuned from different base models.
- 20
- 17
- On the Trade-off of Intra-/Inter-class Diversity for Supervised Pre-training15
The theory reveals that the optimal class-to-sample ratio (#classes / #samples per class) is invariant to the size of the pre-training dataset, which motivates an application of predicting the optimal number of pre- training classes.
- Exploring How Generative MLLMs Perceive More Than CLIP with the Same Vision Encoder13
This study shows that another branch of Vision-Language Models, Generative Multimodal Large Language Models (MLLMs), achieve significantly higher accuracy than CLIP in many of these tasks using the same vision encoder and weights, indicating that these Generative MLLMs perceive more -- as they extract and utilize visual information more effectively.
- EvoLM: Self-Evolving Language Models through Co-Evolved Discriminative Rubrics12
EVOLM demonstrates that structuring a model's evaluative capacity into co-evolving discriminative rubrics enables self-improvement without external supervision.
- 12
- 10
- Localized hepatic lobular regeneration by central-vein–associated lineage-restricted progenitors10
A previously unknown program of liver regeneration after acute injury is discovered in the neonatal mouse liver, specific to the first week of life, where numerous rounds of cell division and reconstitution of lobe architecture much like in amphibian limbs are observed.
- 9
- 9
- Frustratingly Simple Retrieval Improves Challenging, Reasoning-Intensive Benchmarks8
CompactDS is introduced: a diverse, high-quality, web-scale datastore that achieves high retrieval accuracy and subsecond latency on a single-node, and its carefully designed in-house datastore matches or outperforms web search engines such as Google Search, as well as recently proposed, complex agent-based RAG systems--all while maintaining simplicity, reproducibility, and self-containment.
- 8
- 8
- JobBench: Aligning Agent Work With Human Will7
JobBench is introduced, which evaluates AI agents on the workflows that experts identify as high-priority for delegation, empowering humans based on their needs instead of replacing them with GDP value, and hopes it shifts the community's target labour-market effect from replacement to enhancement.
- 7
- Self-Improving VLM Judges Without Human Annotations6
This work presents a framework to self-train a VLM judge model without any human preference annotations, using only self-synthesized data, and shows strong gains in general, hallucination, and reasoning dimensions.
- 6
- JPEG-LM: LLMs as Image Generators with Canonical Codec Representations6
It is shown that using canonical codec representations can help lower the barriers between language generation and visual generation, facilitating future research on multi-modal language/image/video LLMs.
- 5
- Metabolically purified human stem cell-derived hepatocytes reveal distinct effects of Ebola and Lassa viruses5
A new metabolic selection approach is introduced to create nearly-pure hepatocytes from human pluripotent stem cells, killing non-liver cells by withholding essential nutrients and presenting a single-cell transcriptional and chromatin accessibility roadmap of human hepatocyte differentiation, purification, and viral infection.
- 5
- Data-Centric AI in the Age of Large Language Models5
A data-centric viewpoint of AI research, focusing on large language models (LLMs), is proposed, covering data-centric benchmarks and data curation, data attribution, knowledge transfer, knowledge transfer, and inference contextualization.
- 5
- Privasis: Synthesizing the Largest "Public" Private Dataset from Scratch4
Privasis is presented, the first million-scale fully synthetic dataset entirely built from scratch--an expansive reservoir of texts with rich and diverse private information--designed to broaden and accelerate research in areas where processing sensitive social data is inevitable.
- 4
- 4
- Improving Domain Generalization with Domain Relations4
This paper focuses on domain shifts, which occur when the model is applied to new domains that are different from the ones it was trained on, and proposes a new approach called D 3 G, which learns domain-specific models by leveraging the relations among different domains by leveraging the relations among different domains.
- Deep Reasoning in General Purpose Agents via Structured Meta-Cognition2
DOLORES outperforms all evaluated scaffolds across three model sizes and two model families, improving over the strongest evaluated scaffold baseline by 24.8% on average, and points toward future agentic systems that treat scaffolding as adaptive reasoning, constructing the structure each task requires just-in-time.
- 2
- Lower bound on the time complexity of local adiabatic evolution2
Two different approaches of evaluating the time complexity for local adiabatic evolution using time-independent parameters are presented, thus providing effective tests (not requiring the evaluation of the entire time-dependent gap function) for theTime complexity of newly developed algorithms.
- Reliable and Responsible Foundation Models: A Comprehensive Survey1
Critical issues, including bias and fairness, security and privacy, uncertainty, explainability, and distribution shift are explored, as well as methods like alignment and Artificial Intelligence-Generated Content (AIGC) detection.
- 1
- 1
- Time Series Super Resolution withTemporal Adaptive Batch Normalization1
A deep neural network architecture for reconstructing a high resolution time series signal from low-resolution measurements, a task that is called time series super resolution.
- 1
- 1
- 1
- Spokes: Optimizing for Diverse Pretraining Data Selection–
This work directly optimize diversity by introducing a probabilistic diversification framework based on the G-Vendi score, optimized via exponentiated gradient descent, which produces subsets that are substantially more diverse than those obtained via random sampling.
- –
- –
- –
- –
- –
- –
- –
- –
- –
- –
- –
Publication data from OpenAlex, with missing venues and authors filled in from Crossref; citation counts are the higher of OpenAlex and Semantic Scholar, last synced 2026-10-12. One-sentence summaries under some papers are written by Semantic Scholar’s model. Citation counts may be lower than on Google Scholar, which indexes more sources.
Report an error
Wrong papers, two people merged into one, or a profile that should not be here? Tell us and we will fix or hide it.