In this paper, we propose a simple and efficient method for value model training on long-context reasoning traces. Compared to existing process reward models (PRMs), our method does not require a fine-grained notion of "step," which is difficult to define for long-context reasoning models. By collecting a dataset of 2.5 million reasoning traces, we train a 1.5B token-level value model and apply it to DeepSeek models for improved performance with test-time compute scaling. We find that block-wise value-guided search (VGS) with a final weighted majority vote achieves better test-time scaling than standard methods such as majority voting or best-of-n. Moreover, VGS significantly reduces the inference FLOPs required to achieve the same performance of majority voting.

In this work, we introduce RoleBench, a benchmark focused on evaluating compositional generalization in action-based relations (e.g., "mouse chasing cat"). We show that state-of-the-art T2I models and compositional generation methods consistently default to frequent reversed relations (i.e., "cat chasing mouse"), a phenomenon we call role collapse. Related works attribute this to the model's architectural limitation or underrepresentation in the data. Our key insight reveals that while models fail on rare compositions when their inversions are common, they can successfully generate similar intermediate compositions (e.g., "mouse chasing boy"), suggesting that this limitation is also due to the presence of frequent counterparts rather than just the absence of rare compositions. Motivated by this, we hypothesize that directional decomposition can gradually mitigate role collapse. We test this via ReBind, a lightweight framework that teaches role bindings using carefully selected active/passive intermediate compositions. Experiments suggest that intermediate compositions through simple fine-tuning can significantly reduce role collapse, with humans preferring ReBind more than 78% compared to state-of-the-art methods. Our findings highlight the role of distributional asymmetries in compositional failures and offer a simple, effective path for improving generalization.

Inverse folding models have proven to be highly effective zero-shot predictors of protein stability. Despite this success, the link between the amino acid preferences of an inverse folding model and the free-energy considerations underlying thermodynamic stability remains incompletely understood. A better understanding would be of interest not only from a theoretical perspective, but also potentially provide the basis for stronger zero-shot stability prediction. In this paper, we take steps to clarify the free-energy foundations of inverse folding models. Our derivation reveals the standard practice of likelihood ratios as a simplistic approximation and suggests several paths towards better estimates of the relative stability. We empirically assess these approaches and demonstrate that considerable gains in zero-shot performance can be achieved with fairly simple means.

Recently, Large Language Models (LLMs) have been increasingly used to support various decision-making tasks, assisting humans in making informed decisions. However, when LLMs confidently provide incorrect information, it can lead humans to make suboptimal decisions. To prevent LLMs from generating incorrect information on topics they are unsure of and to improve the accuracy of generated content, prior works have proposed Retrieval Augmented Generation (RAG), where external documents are referenced to generate responses. However, previous RAG methods focus only on retrieving documents most relevant to the input query, without specifically aiming to ensure that the human user's decisions are well-calibrated. To address this limitation, we propose a novel retrieval method called Calibrated Retrieval-Augmented Generation (CalibRAG), which ensures that decisions informed by RAG are well-calibrated. Then we empirically validate that CalibRAG improves calibration performance as well as accuracy, compared to other baselines across various datasets.

Large language models (LLMs) have achieved impressive results on complex reasoning tasks, but their high inference cost remains a major barrier to real-world deployment. A promising solution is to use cascaded inference, where small, cheap models handle easy queries, and only the hardest examples are escalated to more powerful models. However, existing cascade methods typically rely on supervised training with labeled data, offer no theoretical generalization guarantees, and provide limited control over test-time computational cost. We introduce C3PO (Cost Controlled Cascaded Prediction Optimization), a self-supervised framework for optimizing LLM cascades under probabilistic cost constraints. By focusing on minimizing regret with respect to the most powerful model (MPM), C3PO avoids the need for labeled data by constructing a cascade using only unlabeled model outputs. It leverages conformal prediction to bound the probability that inference cost exceeds a user-specified budget. We provide theoretical guarantees on both cost control and generalization error, and show that our optimization procedure is effective even with small calibration sets. Empirically, C3PO achieves stateof-the-art performance across a diverse set of reasoning benchmarks including GSM8K, MATH-500, BigBench-Hard and AIME, outperforming strong LLM cascading baselines in both accuracy and cost-efficiency. Our results demonstrate that principled, label-free cascade optimization can enable scalable LLM deployment.

If you tell a learning model that you prefer an alternative a over another alternative b, then you probably expect the model to be monotone, that is, the valuation of a increases, and that of bdecreases. Yet, perhaps surprisingly, many widely deployed comparison-based preference learning models, including large language models, fail to have this guarantee. Until now, the only comparison-based preference learning algorithms that were proved to be monotone are the Generalized BradleyTerry models [10]. Yet, these models are unable to generalize to uncompared data. In this paper, we advance the understanding of the set of models with generalization ability that are monotone. Namely, we propose a new class of Linear Generalized Bradley-Terry models with Diffusion Priors, and identify sufficient conditions on alternatives' embeddings that guarantee monotonicity. Our experiments show that this monotonicity is far from being a general guarantee, and that our new class of generalizing models improves accuracy, especially when the dataset is limited.

Recent advances, such as Mamba, further enhance SSMs with inputdependent gating and hardware-aware implementations, positioning them as strong alternatives to Transformers for long sequence modeling. However, efficiently scaling the expressive power of SSMs, particularly with Mixture of Experts (MoE), remains challenging, as naive integration attempts often falter or degrade performance. In this work, we introduce Routing Mamba (RoM), a novel approach that scales SSM parameters using sparse mixtures of linear projection experts.

We begin by characterizing the space of elements that test an agent's ability to optimally allocate1031 their limited resources to goods and services they desire. In economics and decision theory, the1032 most primitive approach to describing the preferences of decision-makers is to use a function that1033 maps a set of possible choices to the agent's optimal choice within that set. Under a set of intuitive1034 assumptions, such as transitivity (i.e., if bundle X is preferred to bundle Y, and Y is preferred to1035 bundle Z, then X must be preferred to Z), it becomes possible to "rationalize" preferences by instead1036 describing a utility function. This function assigns a real number to each bundle, and the agent selects1037 the bundle with the highest utility.1038 In this paper, we focus on these "rationalizable" preferences, where agent choice can be implemented1039 as utility maximization constrained by prices and income. The solution to these consumer choice1040 problems provides ...

Large language models (LLMs) are increasingly being asked to make economically rational decisions and indeed are already being applied to economic tasks like stock picking and financial analysis. Existing LLM benchmarks tend to focus on specific applications, making them insufficient for characterizing economic reasoning more broadly. In previous work, we offered a blueprint for comprehensively benchmarking strategic decision-making Raman et al. [2024]. However, this work did not engage with the even larger microeconomic literature on non-strategic settings. We address this gap here, taxonomizing microeconomic reasoning into 58distinct elements, each grounded in up to 10distinct domains, 5perspectives, and 3types. The generation of benchmark data across this combinatorial space is powered by a novel LLM-assisted data generation protocol that we dub auto-STEER, which generates a set of questions by adapting handwritten templates to target new domains and perspectives. By generating fresh questions for each element, auto-STEER induces diversity which could help to reduce the risk of data contamination. We use this benchmark to evaluate 27LLMs spanning a range of scales and adaptation strategies, comparing performance across multiple formats--multiple-choice and free-text question answering--and scoring schemes. Our results surface systematic limitations in current LLMs' ability to generalize economic reasoning across types, formats, and textual perturbations, and establish a foundation for evaluating and improving economic competence in foundation models.

Transformer architectures can solve unseen tasks based on input-output pairs in a given prompt due to in-context learning (ICL). Existing theoretical studies on ICL have mainly focused on linear regression tasks, often with i.i.d.