Will LeVine

Will LeVine, Benjamin Pikus, Anthony Chen et al.

Foundation models, specifically Large Language Models (LLMs), have lately gained wide-spread attention and adoption. Reinforcement Learning with Human Feedback (RLHF) involves training a reward model to capture desired behaviors, which is then used to align LLM's. These reward models are additionally used at inference-time to estimate LLM responses' adherence to those desired behaviors. However, there is little work measuring how robust these reward models are to distribution shifts. In this work, we evaluate how reward model performance - measured via accuracy and calibration (i.e. alignment between accuracy and confidence) - is affected by distribution shift. We show novel calibration patterns and accuracy drops due to OOD prompts and responses, and that the reward model is more sensitive to shifts in responses than prompts. Additionally, we adapt an OOD detection technique commonly used in classification to the reward model setting to detect these distribution shifts in prompts and responses.

Will LeVine, Benjamin Pikus, Pranav Raja et al.

Calibration of deep learning models is crucial to their trustworthiness and safe usage, and as such, has been extensively studied in supervised classification models, with methods crafted to decrease miscalibration. However, there has yet to be a comprehensive study of the calibration of vision-language models that are used for zero-shot inference, like CLIP. We measure calibration across relevant variables like prompt, dataset, and architecture, and find that zero-shot inference with CLIP is miscalibrated. Furthermore, we propose a modified version of temperature scaling that is aligned with the common use cases of CLIP as a zero-shot inference model, and show that a single learned temperature generalizes for each specific CLIP model (defined by a chosen pre-training dataset and architecture) across inference dataset and prompt choice.

Chenhui Xu, Fuxun Yu, Michael J. Bianco et al.

We introduce Geo-R1, a reasoning-centric post-training framework that unlocks geospatial reasoning in vision-language models by combining thinking scaffolding and elevating. In the scaffolding stage, Geo-R1 instills a ``geospatial thinking paradigm" via supervised fine-tuning on synthetic chain-of-thought exemplars, enabling models to connect visual cues with geographic priors without costly human reasoning annotations. In the elevating stage, it uses GRPO-based reinforcement learning on a weakly-supervised cross-view pairing proxy. This design supplies a verifiable and scalable reward signal: teaching models to capture and reconcile features across modalities, and harnessing reasoning for accurate prediction. Geo-R1 extends geospatial modeling from domain pretraining / supervised finetuning to reasoning-first post-training, and achieves state-of-the-art performance across various geospatial reasoning benchmarks. Our model is available at https://huggingface.co/miniHui/Geo-R1.

Will LeVine, Bijan Varjavand

Modern Large Language Model (LLM) systems typically rely on Retrieval Augmented Generation (RAG) which aims to gather context that is useful for response generation. These RAG systems typically optimize strictly towards retrieving context that is maximally relevant to the query. However, conventional theory suggests that retrieval systems which seek to maximize context relevance without any additional explicit criteria can create information bottlenecks. We reaffirm this finding in the modern age of LLM's by showing that in standard RAG pipelines, maximizing for context relevance alone can degrade downstream response quality. In response, we show evaluations of existing RAG methods which account for both context relevance and answer quality. These evaluations introduce a novel finding that existing RAG systems scale poorly with inference time compute usage when considering our combined metric. We introduce "RErank BEyond reLevance (REBEL)", which enables RAG systems to scale with inference-time compute via injection of multi-criteria optimization using Chain-of-Thought prompting (and optionally Multi-Turn dialogue). Ultimately, this enables a new performance/speed tradeoff curve, where RAG systems are able to achieve both higher relevance of retrieved contexts and superior answer quality as inference time increases. Code for the implementation of our method in llama-index can be found at the following PR: https://github.com/run-llama/llama_index/pull/17590. Code for running experiments using this llama-index implementation can be found at https://github.com/microsoft/REBEL.

Will LeVine, Brendan Evers, Sam Saltwick et al.

Iterative self-refinement is a popular inference-time reliability technique, but its effectiveness in code-mode tool use depends heavily on the structure of the feedback signal: unstructured critique helps inconsistently across models, and even revision with real execution feedback improves only modestly ($0.75$ vs. $0.65$ baseline). The dominant failures are inter-tool contract violations - wrong output shape, incorrect tool routing, broken argument provenance - that run to completion without raising errors, making runtime feedback insufficient. We introduce RubricRefine, a training-free pre-execution reliability layer that generates task- and registry-specific rubrics, scores candidate code against explicit contract checks, and iteratively repairs failures before any execution occurs. With zero execution attempts, RubricRefine reaches $0.86$ on M3ToolEval averaged across seven models-improving over prior inference-time baselines on every model tested on this benchmark, at $2.6X$ lower latency than the strongest non-iterative alternative - and remains flat on the predominantly single-step API-Bank, consistent with the method's reliance on inter-tool contract structure. A rubric-category ablation and calibration analysis further characterize when and why the method works.

Will LeVine, Benjamin Pikus, Jacob Phillips et al.

As deep neural networks become adopted in high-stakes domains, it is crucial to identify when inference inputs are Out-of-Distribution (OOD) so that users can be alerted of likely drops in performance and calibration despite high confidence -- ultimately to know when networks' decisions (and their uncertainty in those decisions) should be trusted. In this paper we introduce Ablated Learned Temperature Energy (or "AbeT" for short), an OOD detection method which lowers the False Positive Rate at 95\% True Positive Rate (FPR@95) by $43.43\%$ in classification compared to state of the art without training networks in multiple stages or requiring hyperparameters or test-time backward passes. We additionally provide empirical insights as to why our model learns to distinguish between In-Distribution (ID) and OOD samples while only being explicitly trained on ID samples via exposure to misclassified ID examples at training time. Lastly, we show the efficacy of our method in identifying predicted bounding boxes and pixels corresponding to OOD objects in object detection and semantic segmentation, respectively -- with an AUROC increase of $5.15\%$ in object detection and both a decrease in FPR@95 of $41.48\%$ and an increase in AUPRC of $34.20\%$ in semantic segmentation compared to previous state of the art.

Haoyin Xu, Kaleab A. Kinfu, Will LeVine et al.

Deep networks and decision forests (such as random forests and gradient boosted trees) are the leading machine learning methods for structured and tabular data, respectively. Many papers have empirically compared large numbers of classifiers on one or two different domains (e.g., on 100 different tabular data settings). However, a careful conceptual and empirical comparison of these two strategies using the most contemporary best practices has yet to be performed. Conceptually, we illustrate that both can be profitably viewed as "partition and vote" schemes. Specifically, the representation space that they both learn is a partitioning of feature space into a union of convex polytopes. For inference, each decides on the basis of votes from the activated nodes. This formulation allows for a unified basic understanding of the relationship between these methods. Empirically, we compare these two strategies on hundreds of tabular data settings, as well as several vision and auditory settings. Our focus is on datasets with at most 10,000 samples, which represent a large fraction of scientific and biomedical datasets. In general, we found forests to excel at tabular and structured data (vision and audition) with small sample sizes, whereas deep nets performed better on structured data with larger sample sizes. This suggests that further gains in both scenarios may be realized via further combining aspects of forests and networks. We will continue revising this technical report in the coming months with updated results.

Jayanta Dey, Joshua T. Vogelstein, Hayden S. Helm et al.

In lifelong learning, data are used to improve performance not only on the present task, but also on past and future (unencountered) tasks. While typical transfer learning algorithms can improve performance on future tasks, their performance on prior tasks degrades upon learning new tasks (called forgetting). Many recent approaches for continual or lifelong learning have attempted to maintain performance on old tasks given new tasks. But striving to avoid forgetting sets the goal unnecessarily low. The goal of lifelong learning should be to use data to improve performance on both future tasks (forward transfer) and past tasks (backward transfer). In this paper, we show that a simple approach -- representation ensembling -- demonstrates both forward and backward transfer in a variety of simulated and benchmark data scenarios, including tabular, vision (CIFAR-100, 5-dataset, Split Mini-Imagenet, and Food1k), and speech (spoken digit), in contrast to various reference algorithms, which typically failed to transfer either forward or backward, or both. Moreover, our proposed approach can flexibly operate with or without a computational budget.