Benjamin Feuer

Benjamin Feuer, Micah Goldblum, Teresa Datta et al.

The release of ChatGPT in November 2022 sparked an explosion of interest in post-training and an avalanche of new preference optimization (PO) methods. These methods claim superior alignment by virtue of better correspondence with human pairwise preferences, often measured by LLM-judges. In this work, we attempt to answer the following question -- do LLM-judge preferences translate to progress on other, more concrete metrics for alignment, and if not, why not? We define a concrete metric for alignment, and introduce SOS-Bench (Substance Outweighs Style Benchmark), which is to the best of our knowledge the largest standardized, reproducible LLM meta-benchmark to date. We find that (1) LLM-judge preferences do not correlate with concrete measures of safety, world knowledge, and instruction following; (2) LLM-judges have powerful implicit biases, prioritizing style over factuality and safety; and (3) the supervised fine-tuning (SFT) stage of post-training, and not the PO stage, has the greatest impact on alignment, with data scaling and prompt diversity as the driving factors. Our codebase and complete results can be found at https://github.com/penfever/sos-bench.

Benjamin Feuer, Ameya Joshi, Chinmay Hegde

Vision language (VL) models like CLIP are robust to natural distribution shifts, in part because CLIP learns on unstructured data using a technique called caption supervision; the model inteprets image-linked texts as ground-truth labels. In a carefully controlled comparison study, we show that caption-supervised CNNs trained on a standard cross-entropy loss (with image labels assigned by scanning captions for class names) can exhibit greater distributional robustness than VL models trained on the same data. To facilitate future experiments with high-accuracy caption-supervised models, we introduce CaptionNet (https://github.com/penfever/CaptionNet/), which includes a class-balanced, fully supervised dataset with over 50,000 new human-labeled ImageNet-compliant samples which includes web-scraped captions. In a series of experiments on CaptionNet, we show how the choice of loss function, data filtration and supervision strategy enable robust computer vision. We also provide the codebase necessary to reproduce our experiments at VL Hub (https://github.com/penfever/vlhub/).

Benjamin Feuer, Ameya Joshi, Minh Pham et al.

Real world uses of deep learning require predictable model behavior under distribution shifts. Models such as CLIP show emergent natural distributional robustness comparable to humans, but may require hundreds of millions of training samples. Can we train robust learners in a domain where data is limited? To rigorously address this question, we introduce JANuS (Joint Annotations and Names Set), a collection of four new training datasets with images, labels, and corresponding captions, and perform a series of carefully controlled investigations of factors contributing to robustness in image classification, then compare those results to findings derived from a large-scale meta-analysis. Using this approach, we show that standard ResNet-50 trained with the cross-entropy loss on 2.4 million image samples can attain comparable robustness to a CLIP ResNet-50 trained on 400 million samples. To our knowledge, this is the first result showing (near) state-of-the-art distributional robustness on limited data budgets. Our dataset is available at \url{https://huggingface.co/datasets/penfever/JANuS_dataset}, and the code used to reproduce our experiments can be found at \url{https://github.com/penfever/vlhub/}.

Benjamin Feuer, Yurong Liu, Chinmay Hegde et al.

Existing deep-learning approaches to semantic column type annotation (CTA) have important shortcomings: they rely on semantic types which are fixed at training time; require a large number of training samples per type and incur large run-time inference costs; and their performance can degrade when evaluated on novel datasets, even when types remain constant. Large language models have exhibited strong zero-shot classification performance on a wide range of tasks and in this paper we explore their use for CTA. We introduce ArcheType, a simple, practical method for context sampling, prompt serialization, model querying, and label remapping, which enables large language models to solve CTA problems in a fully zero-shot manner. We ablate each component of our method separately, and establish that improvements to context sampling and label remapping provide the most consistent gains. ArcheType establishes a new state-of-the-art performance on zero-shot CTA benchmarks (including three new domain-specific benchmarks which we release along with this paper), and when used in conjunction with classical CTA techniques, it outperforms a SOTA DoDuo model on the fine-tuned SOTAB benchmark. Our code is available at https://github.com/penfever/ArcheType.

Benjamin Feuer, Ameya Joshi, Chinmay Hegde

State-of-the-art image classifiers trained on massive datasets (such as ImageNet) have been shown to be vulnerable to a range of both intentional and incidental distribution shifts. On the other hand, several recent classifiers with favorable out-of-distribution (OOD) robustness properties have emerged, achieving high accuracy on their target tasks while maintaining their in-distribution accuracy on challenging benchmarks. We present a meta-analysis on a wide range of publicly released models, most of which have been published over the last twelve months. Through this meta-analysis, we empirically identify four main commonalities for all the best-performing OOD-robust models, all of which illuminate the considerable promise of vision-language pre-training.

Andre Nakkab, Benjamin Feuer, Chinmay Hegde

Recent advances in training vision-language models have demonstrated unprecedented robustness and transfer learning effectiveness; however, standard computer vision datasets are image-only, and therefore not well adapted to such training methods. Our paper introduces a simple methodology for adapting any fine-grained image classification dataset for distributed vision-language pretraining. We implement this methodology on the challenging iNaturalist-2021 dataset, comprised of approximately 2.7 million images of macro-organisms across 10,000 classes, and achieve a new state-of-the art model in terms of zero-shot classification accuracy. Somewhat surprisingly, our model (trained using a new method called locked-image text tuning) uses a pre-trained, frozen vision representation, proving that language alignment alone can attain strong transfer learning performance, even on fractious, long-tailed datasets. Our approach opens the door for utilizing high quality vision-language pretrained models in agriculturally relevant applications involving species detection.

Benjamin Feuer, Chinmay Hegde, Niv Cohen

Tabular classification has traditionally relied on supervised algorithms, which estimate the parameters of a prediction model using its training data. Recently, Prior-Data Fitted Networks (PFNs) such as TabPFN have successfully learned to classify tabular data in-context: the model parameters are designed to classify new samples based on labelled training samples given after the model training. While such models show great promise, their applicability to real-world data remains limited due to the computational scale needed. Here we study the following question: given a pre-trained PFN for tabular data, what is the best way to summarize the labelled training samples before feeding them to the model? We conduct an initial investigation of sketching and feature-selection methods for TabPFN, and note certain key differences between it and conventionally fitted tabular models.

Etash Guha, Ryan Marten, Sedrick Keh et al. · cmu

Reasoning models have made rapid progress on many benchmarks involving math, code, and science. Yet, there are still many open questions about the best training recipes for reasoning since state-of-the-art models often rely on proprietary datasets with little to no public information available. To address this, the goal of the OpenThoughts project is to create open-source datasets for training reasoning models. After initial explorations, our OpenThoughts2-1M dataset led to OpenThinker2-32B, the first model trained on public reasoning data to match DeepSeek-R1-Distill-32B on standard reasoning benchmarks such as AIME and LiveCodeBench. We then improve our dataset further by systematically investigating each step of our data generation pipeline with 1,000+ controlled experiments, which led to OpenThoughts3. Scaling the pipeline to 1.2M examples and using QwQ-32B as teacher yields our OpenThoughts3-7B model, which achieves state-of-the-art results: 53% on AIME 2025, 51% on LiveCodeBench 06/24-01/25, and 54% on GPQA Diamond - improvements of 15.3, 17.2, and 20.5 percentage points compared to the DeepSeek-R1-Distill-Qwen-7B. All of our datasets and models are available on https://openthoughts.ai.

Benjamin Feuer, Robin Tibor Schirrmeister, Valeriia Cherepanova et al.

While tabular classification has traditionally relied on from-scratch training, a recent breakthrough called prior-data fitted networks (PFNs) challenges this approach. Similar to large language models, PFNs make use of pretraining and in-context learning to achieve strong performance on new tasks in a single forward pass. However, current PFNs have limitations that prohibit their widespread adoption. Notably, TabPFN achieves very strong performance on small tabular datasets but is not designed to make predictions for datasets of size larger than 1000. In this work, we overcome these limitations and substantially improve the performance of PFNs via context optimization. We introduce TuneTables, a parameter-efficient fine-tuning strategy for PFNs that compresses large datasets into a smaller learned context. We conduct extensive experiments on 19 algorithms over 98 datasets and find that TuneTables achieves the best performance on average, outperforming boosted trees such as CatBoost, while optimizing fewer than 5% of TabPFN's parameters. Furthermore, we show that TuneTables can be used as an interpretability tool and can even be used to mitigate biases by optimizing a fairness objective. We open-source our code and raw results at https://github.com/penfever/TuneTables.

Benjamin Feuer, Chinmay Hegde

Modern computer vision foundation models are trained on massive amounts of data, incurring large economic and environmental costs. Recent research has suggested that improving data quality can significantly reduce the need for data quantity. But what constitutes data quality in computer vision? We posit that the quality of a given dataset can be decomposed into distinct sample-level and dataset-level constituents, and that the former have been more extensively studied than the latter. We ablate the effects of two important dataset-level constituents: label set design, and class balance. By monitoring these constituents using key indicators we provide, researchers and practitioners can better anticipate model performance, measured in terms of its accuracy and robustness to distribution shifts.

Hossein Zaremehrjerdi, Shreyan Ganguly, Ashlyn Rairdin et al.

Agricultural decision-making involves complex, context-specific reasoning, where choices about crops, practices, and interventions depend heavily on geographic, climatic, and economic conditions. Traditional large language models (LLMs) often fall short in navigating this nuanced problem due to limited reasoning capacity. We hypothesize that recent advances in large reasoning models (LRMs) can better handle such structured, domain-specific inference. To investigate this, we introduce AgReason, the first expert-curated open-ended science benchmark with 100 questions for agricultural reasoning. Evaluations across thirteen open-source and proprietary models reveal that LRMs outperform conventional ones, though notable challenges persist, with the strongest Gemini-based baseline achieving 36% accuracy. We also present AgThoughts, a large-scale dataset of 44.6K question-answer pairs generated with human oversight and equipped with synthetically generated reasoning traces. Using AgThoughts, we develop AgThinker, a suite of small reasoning models that can be run on consumer-grade GPUs, and show that our dataset can be effective in unlocking agricultural reasoning abilities in LLMs. Our project page is here: https://baskargroup.github.io/Ag_reasoning/

Benjamin Feuer, Lucas Rosenblatt, Oussama Elachqar

As AI models progress beyond simple chatbots into more complex workflows, we draw ever closer to the event horizon beyond which AI systems will be utilized in autonomous, self-maintaining feedback loops. Any autonomous AI system will depend on automated, verifiable rewards and feedback; in settings where ground truth is sparse or non-deterministic, one practical source of such rewards is an LLM-as-a-Judge. Although LLM judges continue to improve, the literature has yet to introduce systems capable of enforcing standards with strong guarantees, particularly when bias vectors are unknown or adversarially discovered. To remedy this issue, we propose average bias-boundedness (A-BB), an algorithmic framework which formally guarantees reductions of harm/impact as a result of any measurable bias in an LLM judge. Evaluating on Arena-Hard-Auto with four LLM judges, we achieve (tau=0.5, delta=0.01) bias-bounded guarantees while retaining 61-99% correlation with original rankings across formatting and schematic bias settings, with most judge-bias combinations exceeding 80%. The code to reproduce our findings is available at https://github.com/penfever/bias-bounded-evaluation.

Benjamin Feuer, Chiung-Yi Tseng, Astitwa Sarthak Lathe et al.

LLM-judged benchmarks are increasingly used to evaluate complex model behaviors, yet their design introduces failure modes absent in conventional ground-truth based benchmarks. We argue that without tight objectives and verifiable constructions, benchmark rankings can produce high-confidence rankings that are in fact largely noise. We introduce two mechanisms to diagnose these issues. Schematic adherence quantifies how much of a judge's overall verdict is explained by the explicit evaluation schema, revealing unexplained variance when judges deviate from their own rubric. Psychometric validity aggregates internal consistency and discriminant validity signals to quantify irreducible uncertainty in any benchmarking run. Applying these tools to Arena-Hard Auto, we find severe schema incoherence and factor collapse across popular judges: for example, unexplained variance exceeding 90 percent for DeepSeek-R1-32B and factor correlations above 0.93 for most criteria. We also show that the ELO-style aggregation used by Arena-Hard Auto collapses and masks genuine ranking uncertainty. Our results highlight design failures that undermine validity and offer actionable principles for building better-scoped, reliability-aware LLM-judged benchmarks. We released our code and dataset at https://github.com/penfever/judgment-to-noise

Benjamin Feuer, Lennart Purucker, Oussama Elachqar et al.

Scientific applications of machine learning often rely on small, specialized models tuned to particular domains. Such models often achieve excellent performance, but lack flexibility. Foundation models offer versatility, but typically underperform specialized approaches, especially on non-traditional modalities and long-tail domains. We propose MARVIS (Modality Adaptive Reasoning over VISualizations), a training-free method that enables even small vision-language models to predict any data modality with high accuracy. MARVIS transforms latent embedding spaces into visual representations and then leverages the spatial and fine-grained reasoning skills of VLMs to successfully interpret and utilize them. MARVIS achieves competitive performance on vision, audio, biological, and tabular domains using a single 3B parameter model, achieving results that beat Gemini by 16\% on average and approach specialized methods, without exposing personally identifiable information (P.I.I.) or requiring any domain-specific training. We open source our code and datasets at https://github.com/penfever/marvis

Benjamin Feuer, Chinmay Hegde

Language model (LLM) post-training, from DPO to distillation, can refine behaviors and unlock new skills, but the open science supporting these post-training techniques is still in its infancy. One limiting factor has been the difficulty of conducting large-scale comparative analyses of synthetic data generating models and LLM judges. To close this gap, we introduce WILDCHAT-50M, the largest public chat dataset to date. We extend the existing WildChat dataset to include responses not only from GPT, but from over 50 different open-weight models, ranging in size from 0.5B to 104B parameters. We conduct an extensive comparative analysis and demonstrate the potential of this dataset by creating RE-WILD, our own public SFT mix, which outperforms the recent Tulu-3 SFT mixture from Allen AI with only 40% as many samples. Our dataset, samples and code are available at https://github.com/penfever/wildchat-50m.

Duncan McElfresh, Sujay Khandagale, Jonathan Valverde et al.

Tabular data is one of the most commonly used types of data in machine learning. Despite recent advances in neural nets (NNs) for tabular data, there is still an active discussion on whether or not NNs generally outperform gradient-boosted decision trees (GBDTs) on tabular data, with several recent works arguing either that GBDTs consistently outperform NNs on tabular data, or vice versa. In this work, we take a step back and question the importance of this debate. To this end, we conduct the largest tabular data analysis to date, comparing 19 algorithms across 176 datasets, and we find that the 'NN vs. GBDT' debate is overemphasized: for a surprisingly high number of datasets, either the performance difference between GBDTs and NNs is negligible, or light hyperparameter tuning on a GBDT is more important than choosing between NNs and GBDTs. A remarkable exception is the recently-proposed prior-data fitted network, TabPFN: although it is effectively limited to training sets of size 3000, we find that it outperforms all other algorithms on average, even when randomly sampling 3000 training datapoints. Next, we analyze dozens of metafeatures to determine what properties of a dataset make NNs or GBDTs better-suited to perform well. For example, we find that GBDTs are much better than NNs at handling skewed or heavy-tailed feature distributions and other forms of dataset irregularities. Our insights act as a guide for practitioners to determine which techniques may work best on their dataset. Finally, with the goal of accelerating tabular data research, we release the TabZilla Benchmark Suite: a collection of the 36 'hardest' of the datasets we study. Our benchmark suite, codebase, and all raw results are available at https://github.com/naszilla/tabzilla.

Kasra Arabi, Benjamin Feuer, R. Teal Witter et al.

As the quality of image generators continues to improve, deepfakes become a topic of considerable societal debate. Image watermarking allows responsible model owners to detect and label their AI-generated content, which can mitigate the harm. Yet, current state-of-the-art methods in image watermarking remain vulnerable to forgery and removal attacks. This vulnerability occurs in part because watermarks distort the distribution of generated images, unintentionally revealing information about the watermarking techniques. In this work, we first demonstrate a distortion-free watermarking method for images, based on a diffusion model's initial noise. However, detecting the watermark requires comparing the initial noise reconstructed for an image to all previously used initial noises. To mitigate these issues, we propose a two-stage watermarking framework for efficient detection. During generation, we augment the initial noise with generated Fourier patterns to embed information about the group of initial noises we used. For detection, we (i) retrieve the relevant group of noises, and (ii) search within the given group for an initial noise that might match our image. This watermarking approach achieves state-of-the-art robustness to forgery and removal against a large battery of attacks.

Chih-Hsuan Yang, Benjamin Feuer, Zaki Jubery et al.

We introduce BioTrove, the largest publicly accessible dataset designed to advance AI applications in biodiversity. Curated from the iNaturalist platform and vetted to include only research-grade data, BioTrove contains 161.9 million images, offering unprecedented scale and diversity from three primary kingdoms: Animalia ("animals"), Fungi ("fungi"), and Plantae ("plants"), spanning approximately 366.6K species. Each image is annotated with scientific names, taxonomic hierarchies, and common names, providing rich metadata to support accurate AI model development across diverse species and ecosystems. We demonstrate the value of BioTrove by releasing a suite of CLIP models trained using a subset of 40 million captioned images, known as BioTrove-Train. This subset focuses on seven categories within the dataset that are underrepresented in standard image recognition models, selected for their critical role in biodiversity and agriculture: Aves ("birds"), Arachnida ("spiders/ticks/mites"), Insecta ("insects"), Plantae ("plants"), Fungi ("fungi"), Mollusca ("snails"), and Reptilia ("snakes/lizards"). To support rigorous assessment, we introduce several new benchmarks and report model accuracy for zero-shot learning across life stages, rare species, confounding species, and multiple taxonomic levels. We anticipate that BioTrove will spur the development of AI models capable of supporting digital tools for pest control, crop monitoring, biodiversity assessment, and environmental conservation. These advancements are crucial for ensuring food security, preserving ecosystems, and mitigating the impacts of climate change. BioTrove is publicly available, easily accessible, and ready for immediate use.