Reasoning language models can solve increasingly complex tasks, but struggle to produce the calibrated confidence estimates necessary for reliable deployment. Existing calibration methods usually depend on labels or repeated sampling at inference time, making them impractical in many settings. We introduce a method for unsupervised confidence calibration of reasoning LLMs when only a single generation is available at inference time. Our approach uses offline sampling on unlabeled data to derive a self-consistency-based proxy target, then distills this signal into a lightweight deployment-time confidence predictor. In a broad evaluation across 5 math and question-answering tasks using 9 reasoning models, our method substantially outperforms baselines, including under distribution shift, and improves downstream performance in selective prediction and simulated downstream decision-making.

Large vision-language models contain several priors about the world and object interactions, making them useful critics during inference to steer robot policies towards success. However, closed-loop robot manipulation requires judging small visual differences between success and failure, which remains a challenge for current VLMs. We introduce a method to fine-tune critics by constructing pairwise progress supervision using success and failure rollouts obtained from a policy. Our fine-tuned critic excels at fine-grained progress reasoning and subtle failure detection, outperforming prior progress reasoning baselines. Additionally, we use an action-conditioned video model to predict the visual effect of several candidate actions sampled from a policy, and show that our critic can correctly identify successful candidates to execute, improving the average policy success rate by 11

Recent progress in large language models (LLMs) has focused on producing responses that meet human expectations and align with shared values - a process coined alignment. However, aligning LLMs remains challenging due to the inherent disconnect between the complexity of human values and the narrow nature of the technological approaches designed to address them. Current alignment methods often lead to misspecified objectives, reflecting the broader issue of incomplete contracts, the impracticality of specifying a contract between a model developer, and the model that accounts for every scenario in LLM alignment. In this paper, we argue that improving LLM alignment requires incorporating insights from societal alignment frameworks, including social, economic, and contractual alignment, and discuss potential solutions drawn from these domains. Given the role of uncertainty within societal alignment frameworks, we then investigate how it manifests in LLM alignment. We end our discussion by offering an alternative view on LLM alignment, framing the underspecified nature of its objectives as an opportunity rather than perfect their specification. Beyond technical improvements in LLM alignment, we discuss the need for participatory alignment interface designs.

Eliciting information to reduce uncertainty about latent group-level properties from surveys and other collective assessments requires allocating limited questioning effort under real costs and missing data. Although large language models enable adaptive, multi-turn interactions in natural language, most existing elicitation methods optimize what to ask with a fixed respondent pool, and do not adapt respondent selection or leverage population structure when responses are partial or incomplete. To address this gap, we study adaptive group elicitation, a multi-round setting where an agent adaptively selects both questions and respondents under explicit query and participation budgets. We propose a theoretically grounded framework that combines (i) an LLM-based expected information gain objective for scoring candidate questions with (ii) heterogeneous graph neural network propagation that aggregates observed responses and participant attributes to impute missing responses and guide per-round respondent selection. This closed-loop procedure queries a small, informative subset of individuals while inferring population-level responses via structured similarity. Across three real-world opinion datasets, our method consistently improves population-level response prediction under constrained budgets, including a >12

Continual learning methods aim to maximize the stability and plasticity of machine learning models that are trained on a sequence of tasks. The standard measure of stability (i.e., forgetting) is the 0-shot performance of a model on previously learned tasks, and plasticity, the performance on the most recently learned task. However, 0-shot evaluation does not fully measure a model or method's ability to retain learned information or adapt quickly to new information, as it requires perfect recall across multiple tasks. In this paper, we propose few-shot evaluation as a more comprehensive assessment of the stability and plasticity of a continual learning system. We conduct a fine-grained assessment on task sequences for continual image classification and find that this paradigm produces novel insights into the performance of popular continual learning strategies. Through few-shot evaluation with a novel metric – per-shot plasticity – we show that adding `foresight' to continual learning methods via the meta-learning of a short sequence of future tasks induces learning-to-learn behavior over the task sequence.

Modern machine learning models are deployed in diverse, non-stationary environments where they must continually adapt to new tasks and evolving knowledge. Continual fine-tuning and in-context learning are costly and brittle, whereas neural memory methods promise lightweight updates with minimal forgetting. However, existing neural memory models typically assume a single fixed objective and homogeneous information streams, leaving users with no control over what the model remembers or ignores over time. To address this challenge, we propose a generalized neural memory system that performs flexible updates based on learning instructions specified in natural language. Our approach enables adaptive agents to learn selectively from heterogeneous information sources, supporting settings, such as healthcare and customer service, where fixed-objective memory updates are insufficient.

Many real-world design problems involve optimizing an expensive black-box function f(x), such as hardware design or drug discovery. Bayesian Optimization has emerged as a sample-efficient framework for this problem. However, the basic setting considered by these methods is simplified compared to real-world experimental setups, where experiments often generate a wealth of useful information. We introduce a new setting where an experiment generates high-dimensional auxiliary information h(x) along with the performance measure f(x); moreover, a history of previously solved tasks from the same task family is available for accelerating optimization. A key challenge of our setting is learning how to represent and utilize h(x) for efficiently solving new optimization tasks beyond the task history. We develop a novel approach for this setting based on a neural model which predicts f(x) for unseen designs given a few-shot context containing observations of h(x). We evaluate our method on two challenging domains, robotic hardware design and neural network hyperparameter tuning, and introduce a novel design problem and large-scale benchmark for the former. On both domains, our method utilizes auxiliary feedback effectively to achieve more accurate few-shot prediction and faster optimization of design tasks, significantly outperforming several methods for multi-task optimization.

Curriculum learning–ordering training examples in a sequence to aid machine learning–takes inspiration from human learning, but has not gained widespread acceptance. Static strategies for scoring item difficulty rely on indirect proxy scores of varying quality and produce curricula that are not specific to the learner at hand. Dynamic approaches base difficulty estimates on gradient information, requiring considerable extra computation during training. We introduce a novel method for measuring the difficulty of individual problem instances that is calibrated to a series of models of increasing competence, and identify transitional problems that are consistently easier as model ability increases. Applying this method to diverse model series constructed from sets of models that are readily available on many tasks, we find that training on a curriculum that levels up from easier to harder transitional problems most efficiently improves a model to the next tier of competence. These problems induce a natural progression from easier to harder items, which outperforms other training strategies. By measuring difficulty directly relative to model competence, our method yields interpretable problems, learner-specific curricula, and a principled basis for step-by-step improvement.

As machine learning models grow increasingly competent, their predictions can supplement scarce or expensive data in various important domains. In support of this paradigm, algorithms have emerged to combine a small amount of high-fidelity observed data with a much larger set of imputed model outputs to estimate some quantity of interest. Yet current hybrid-inference tools target only means or single quantiles, limiting their applicability for many critical domains and use cases. We present QuEst, a principled framework to merge observed and imputed data to deliver point estimates and rigorous confidence intervals for a wide family of quantile-based distributional measures. QuEst covers a range of measures, from tail risk (CVaR) to population segments such as quartiles, that are central to fields such as economics, sociology, education, medicine, and more. We extend QuEst to multidimensional metrics, and introduce an additional optimization technique to further reduce variance in this and other hybrid estimators. We demonstrate the utility of our framework through experiments in economic modeling, opinion polling, and language model auto-evaluation.

Multimodal Large Language Models (MLLMs) hold great promise for advanced reasoning at the intersection of text and images, yet they have not fully realized this potential. MLLMs typically integrate an LLM, a vision encoder, and a connector that maps the vision encoder's embeddings into the LLM's text embedding space. Although each component is pretrained on massive datasets with billions of samples, the entire multimodal model is typically trained on only thousands (or a few million) samples, which can result in weak performance on complex reasoning tasks. To address these shortcomings, instead of relying on extensive labeled datasets for fine-tuning, we propose a Test-Time Warmup method that adapts the MLLM per test instance by leveraging data from weakly supervised auxiliary tasks. With our approach, we observe a relative performance improvement of 4.03

Continual learning seeks to enable machine learning systems to solve an increasing corpus of tasks sequentially. A critical challenge for continual learning is forgetting, where the performance on previously learned tasks decreases as new tasks are introduced. One of the commonly used techniques to mitigate forgetting, sample replay, has been shown empirically to reduce forgetting by retaining some examples from old tasks and including them in new training episodes. In this work, we provide a theoretical analysis of sample replay in an over-parameterized continual linear regression setting, where each task is given by a linear subspace and with enough replay samples, one would be able to eliminate forgetting. Our analysis focuses on sample replay and highlights the role of the replayed samples and the relationship between task subspaces. Surprisingly, we find that, even in a noiseless setting, forgetting can be non-monotonic with respect to the number of replay samples. We present tasks where replay can be harmful with respect to worst-case settings, and also in distributional settings where replay of randomly selected samples increases forgetting in expectation. We also give empirical evidence that harmful replay is not limited to training with linear models by showing similar behavior for a neural networks equipped with SGD. Through experiments on a commonly used benchmark, we provide additional evidence that, even in seemingly benign scenarios, performance of the replay heavily depends on the choice of replay samples and the relationship between tasks.

Deep mixture-of-experts models have attracted a lot of attention for survival analysis problems, particularly for their ability to cluster similar patients together. In practice, grouping often comes at the expense of key metrics such calibration error and predictive accuracy. This is due to the restrictive inductive bias that mixture-of-experts imposes, that predictions for individual patients must look like predictions for the group they're assigned to. Might we be able to discover patient group structure, where it exists, while improving calibration and predictive accuracy? In this work, we introduce several discrete-time deep mixture-of-experts (MoE) based architectures for survival analysis problems, one of which achieves all desiderata: clustering, calibration, and predictive accuracy. We show that a key differentiator between this array of MoEs is how expressive their experts are. We find that more expressive experts that tailor predictions per patient outperform experts that rely on fixed group prototypes.

Safety reasoning is a recent paradigm where LLMs reason over safety policies before generating responses, thereby mitigating limitations in existing safety measures such as overrefusal and jailbreak vulnerabilities. However, implementing this paradigm is challenging due to the resource-intensive process of creating high-quality policy-embedded chain-ofthought (CoT) datasets while ensuring reasoning remains accurate and free from hallucinations or policy conflicts. To tackle this, we propose AIDSAFE: Agentic Iterative Deliberation for Safety Reasoning, a novel data generation recipe that leverages multi-agent deliberation to iteratively expand reasoning on safety policies. A data refiner stage in AIDSAFE ensures high-quality outputs by eliminating repetitive, redundant, and deceptive thoughts. AIDSAFEgenerated CoTs provide a strong foundation for supervised fine-tuning (SFT)-based safety training. Additionally, to address the need of preference data in alignment stages, such as DPO training, we introduce a supplemental recipe that uses belief augmentation to create distinct selected and rejected CoT samples. Our evaluations demonstrate that AIDSAFEgenerated CoTs achieve superior policy adherence and reasoning quality. Consequently, we show that fine-tuning open-source LLMs on these CoTs can significantly improve safety generalization and jailbreak robustness while maintaining acceptable utility and over-refusal accuracy. AIDSAFE-generated CoT datasets can be found here.

Generative AI (GenAI) models present new challenges in regulating against discriminatory behavior. In this paper, we argue that GenAI fairness research still has not met these challenges; instead, a significant gap remains between existing bias assessment methods and regulatory goals. This leads to ineffective regulation that can allow deployment of reportedly fair, yet actually discriminatory, GenAI systems. Towards remedying this problem, we connect the legal and technical literature around GenAI bias evaluation and identify areas of misalignment. Through four case studies, we demonstrate how this misalignment between fairness testing techniques and regulatory goals can result in discriminatory outcomes in real-world deployments, especially in adaptive or complex environments. We offer practical recommendations for improving discrimination testing to better align with regulatory goals and enhance the reliability of fairness assessments in future deployments.

Large language models are increasingly used to support high-stakes decisions, potentially influencing who is granted bail or receives a loan. Naive chain-of-thought sampling can improve average decision accuracy, but has also been shown to amplify unfair bias. To address this challenge and enable the trustworthy use of reasoning models in high-stakes decision-making, we propose a framework for training a generalizable Fairness Reward Model (FRM). Our model assigns a fairness score to LLM reasoning, enabling the system to down-weight biased trajectories and favor equitable ones when aggregating decisions across reasoning chains. We show that a single Fairness Reward Model, trained on weakly supervised, LLM-annotated examples of biased versus unbiased reasoning, transfers across tasks, domains, and model families without additional fine-tuning. Applied to real-world decision-making tasks including recidivism prediction and social media moderation, we show that our approach consistently improves fairness while matching, or even surpassing, baseline accuracy.

Trustworthy robot behavior requires not only high levels of task success but also that the robot can reliably quantify how likely it is to succeed. To this end, we present the first systematic study of confidence calibration in vision-language-action (VLA) foundation models, which map visual observations and natural-language instructions to low-level robot motor commands. We begin with extensive benchmarking to understand the critical relationship between task success and calibration error across multiple datasets and VLA variants, finding that task performance and calibration are not in tension. Next, we introduce prompt ensembles for VLAs, a lightweight, Bayesian-inspired algorithm that averages confidence across paraphrased instructions and consistently improves calibration. We further analyze calibration over the task time horizon, showing that confidence is often most reliable after making some progress, suggesting natural points for risk-aware intervention. Finally, we reveal differential miscalibration across action dimensions and propose action-wise Platt scaling, a method to recalibrate each action dimension independently to produce better confidence estimates. Our aim in this study is to begin to develop the tools and conceptual understanding necessary to render VLAs both highly performant and highly trustworthy via reliable uncertainty quantification.

The recent explosion in the capabilities of large language models has led to a wave of interest in how best to prompt a model to perform a given task. While it may be tempting to simply choose a prompt based on average performance on a validation set, this can lead to a deployment where unexpectedly poor responses are generated, especially for the worst-off users. To mitigate this prospect, we propose Prompt Risk Control, a lightweight framework for selecting a prompt based on rigorous upper bounds on families of informative risk measures. We offer methods for producing bounds on a diverse set of metrics, including quantities that measure worst-case responses and disparities in generation quality across the population of users. In addition, we extend the underlying statistical bounding techniques to accommodate the possibility of distribution shifts in deployment. Experiments on applications such as open-ended chat, medical question summarization, and code generation highlight how such a framework can foster responsible deployment by reducing the risk of the worst outcomes.

Gender-inclusive NLP research has documented the harmful limitations of gender binary-centric large language models (LLM), such as the inability to correctly use gender-diverse English neopronouns (e.g., xe, zir, fae). While data scarcity is a known culprit, the precise mechanisms through which scarcity affects this behavior remain underexplored. We discover LLM misgendering is significantly influenced by Byte-Pair Encoding (BPE) tokenization, the tokenizer powering many popular LLMs. Unlike binary pronouns, BPE overfragments neopronouns, a direct consequence of data scarcity during tokenizer training. This disparate tokenization mirrors tokenizer limitations observed in multilingual and low-resource NLP, unlocking new misgendering mitigation strategies. We propose two techniques: (1) pronoun tokenization parity, a method to enforce consistent tokenization across gendered pronouns, and (2) utilizing pre-existing LLM pronoun knowledge to improve neopronoun proficiency. Our proposed methods outperform finetuning with standard BPE, improving neopronoun accuracy from 14.1% to 58.4%. Our paper is the first to link LLM misgendering to tokenization and deficient neopronoun grammar, indicating that LLMs unable to correctly treat neopronouns as pronouns are more prone to misgender.

Warning: this paper contains content that may be inappropriate or offensive. As generative models become available for public use in various applications, testing and analyzing vulnerabilities of these models has become a priority. Here we propose an automatic red teaming framework that evaluates a given model and exposes its vulnerabilities against unsafe and inappropriate content generation. Our framework uses in-context learning in a feedback loop to red team models and trigger them into unsafe content generation. We propose different in-context attack strategies to automatically learn effective and diverse adversarial prompts for text-to-image models. Our experiments demonstrate that compared to baseline approaches, our proposed strategy is significantly more effective in exposing vulnerabilities in Stable Diffusion (SD) model, even when the latter is enhanced with safety features. Furthermore, we demonstrate that the proposed framework is effective for red teaming text-to-text models, resulting in significantly higher toxic response generation rate compared to previously reported numbers.