Markov Chain Monte Carlo (MCMC) underlies both statistical physics and combinatorial optimization, but mixes slowly near critical points and in rough landscapes. Parallel Tempering (PT) improves mixing by swapping replicas across temperatures, yet each replica still relies on slow local updates to change its configuration. We introduce IsingFormer, a Transformer trained on equilibrium samples that can generate entire spin configurations resembling those from the target distribution. These uncorrelated samples are used as proposals for global moves within a Metropolis step in PT, complementing the usual single-spin flips. On 2D Ising models (sampling), IsingFormer reproduces magnetization and free-energy curves and generalizes to unseen temperatures, including the critical region. Injecting even a single proposal sharply reduces equilibration time, replacing thousands of local updates. On 3D spin glasses (optimization), PT enhanced with IsingFormer finds substantially lower-energy states, demonstrating how global moves accelerate search in rugged landscapes. Finally, applied to integer factorization encoded as Ising problems, IsingFormer trained on a limited set of semiprimes transfers successfully to unseen semiprimes, boosting success rates beyond the training distribution. Since factorization is a canonical hard benchmark, this ability to generalize across instances highlights the potential of learning proposals that move beyond single problems to entire families of instances. The IsingFormer demonstrates that Monte Carlo methods can be systematically accelerated by neural proposals that capture global structure, yielding faster sampling and stronger performance in combinatorial optimization.

Mafia is a social deduction game where informed mafia compete against uninformed townsfolk. Its asymmetry of information and reliance on theory-of-mind reasoning mirror real-world multi-agent scenarios, making it a useful testbed for evaluating the social intelligence of large language models (LLMs). To support a systematic study, we introduce Mini-Mafia: a simplified four-player variant with one mafioso, one detective, and two villagers. We set the mafioso to kill a villager and the detective to investigate the mafioso during the night, reducing the game to a single day phase of discussion and voting. This setup isolates three interactive capabilities through role-specific win conditions: the mafioso must deceive, the villagers must detect deception, and the detective must effectively disclose information. To measure these skills, we have LLMs play against each other, creating the Mini-Mafia Benchmark: a two-stage framework that first estimates win rates within fixed opponent configurations, then aggregates performance across them using standardized scoring. Built entirely from model interactions without external data, the benchmark evolves as new models are introduced, with each one serving both as a new opponent and as a subject of evaluation. Our experiments reveal counterintuitive results, including cases where smaller models outperform larger ones. Beyond benchmarking, Mini-Mafia enables quantitative study of emergent multi-agent dynamics such as name bias and last-speaker advantage. It also contributes to AI safety by generating training data for deception detectors and by tracking models' deception capabilities against human baselines.

We study a class of multi-stage stochastic programs, which incorporate modeling features from Markov decision processes (MDPs). This class includes structured MDPs with continuous state and action spaces. We extend policy graphs to include decision-dependent uncertainty for one-step transition probabilities as well as a limited form of statistical learning. We focus on the expressiveness of our modeling approach, illustrating ideas with a series of examples of increasing complexity. As a solution method, we develop new variants of stochastic dual dynamic programming, including approximations to handle non-convexities.

We introduce a novel approach to large language model (LLM) distillation by formulating it as a constrained reinforcement learning problem. While recent work has begun exploring the integration of task-specific rewards into distillation processes, existing methods typically rely on ad-hoc reward weighting. We propose a principled optimization framework that maximizes task-specific rewards while constraining the divergence from the teacher model to remain below a specified threshold. Our approach adapts constrained state augmented reinforcement learning to the distillation setting, introducing a modified reward function that maintains theoretical guarantees of constraint satisfaction without requiring state augmentation or teacher model access during deployment and without the computational overhead of the dual Lagrangian methods. Through extensive experiments on mathematical reasoning tasks, we demonstrate that our method achieves better constraint satisfaction rates and better reasoning compared to the soft Lagrangian relaxation baselines while maintaining competitive task performance. Our framework provides a theoretically grounded and practically efficient solution for reward-aware distillation in resource-constrained settings. Large Language Models (LLMs) have achieved remarkable success in a wide range of natural language processing tasks (V aswani et al., 2017; Trinh et al., 2024; Chervonyi et al., 2025; Guo et al., 2025; Christianos et al., 2023), but their size and complexity make them impractical for deployment in resource-constrained environments. Distillation (Hinton et al., 2015; Czarnecki et al., 2019), a technique where a smaller student model learns from a larger teacher model, has been widely used to transfer knowledge while reducing computational costs. Conventional distillation methods (Sanh et al., 2020; Gu et al., 2024; Ko et al., 2024) typically focus on minimizing the divergence between the student and teacher models, often using metrics such as Kullback-Leibler (KL) divergence. However, these methods do not fully leverage additional reward signals that can provide valuable guidance, particularly in tasks requiring complex reasoning.

Large language models (LLMs) can adapt to new tasks via in-context learning (ICL) without parameter updates, making them powerful learning engines for fast adaptation. While extensive research has examined ICL as a few-shot learner, whether it can achieve long-term retention and cross-task knowledge accumulation when multitasks arrive sequentially remains underexplored. Motivated by human memory studies, we investigate the retention characteristics of ICL in multitask settings and extend it to in-context continual learning (ICCL), where continual learning ability emerges through task scheduling and prompt rearrangement. Experiments on Markov-Chain benchmarks demonstrate that, for specific large-language models, ICCL benefits from distributed practice (DP) in a manner analogous to humans, consistently revealing a spacing "sweet spot" for retention. Beyond retention performance, we propose a human-retention similarity metric to quantify how closely a continual-learning (CL) method aligns with human retention dynamics. Using this metric, we show that linear-attention models such as MAMBA and RWKV exhibit particularly human-like retention patterns, despite their retention performance lagging behind that of Transformer-based LLMs. Overall, our results establish ICCL as both cognitively plausible and practically effective, providing an inference-only CL paradigm that mitigates catastrophic forgetting and addresses the stability-plasticity dilemma in conventional CL methods.

Deep Reinforcement Learning (DRL) controllers for quadrupedal locomotion have demonstrated impressive performance on challenging terrains, allowing robots to execute complex skills such as climbing, running, and jumping. However, existing blind locomotion controllers often struggle to ensure safety and efficient traversal through risky gap terrains, which are typically highly complex, requiring robots to perceive terrain information and select appropriate footholds during locomotion accurately. Meanwhile, existing perception-based controllers still present several practical limitations, including a complex multi-sensor deployment system and expensive computing resource requirements. This paper proposes a DRL controller named MAstering Risky Gap Terrains (MARG), which integrates terrain maps and proprioception to dynamically adjust the action and enhance the robot's stability in these tasks. During the training phase, our controller accelerates policy optimization by selectively incorporating privileged information (e.g., center of mass, friction coefficients) that are available in simulation but unmeasurable directly in real-world deployments due to sensor limitations. We also designed three foot-related rewards to encourage the robot to explore safe footholds. More importantly, a terrain map generation (TMG) model is proposed to reduce the drift existing in mapping and provide accurate terrain maps using only one LiDAR, providing a foundation for zero-shot transfer of the learned policy. The experimental results indicate that MARG maintains stability in various risky terrain tasks.

Training agents to act in embodied environments typically requires vast training data or access to accurate simulation, neither of which exists for many cases in the real world. Instead, world models are emerging as an alternative leveraging offline, passively collected data, they make it possible to generate diverse worlds for training agents in simulation. In this work, we harness world models to generate imagined environments to train robust agents capable of generalizing to novel task variations. One of the challenges in doing this is ensuring the agent trains on useful generated data. We thus propose a novel approach, IMAC (Imagined Autocurricula), leveraging Unsupervised Environment Design (UED), which induces an automatic curriculum over generated worlds. In a series of challenging, procedurally generated environments, we show it is possible to achieve strong transfer performance on held-out environments, having trained only inside a world model learned from a narrower dataset. We believe this opens the path to utilizing larger-scale, foundation world models for generally capable agents.

Machine learning (ML) is often viewed as a powerful data analysis tool that is easy to learn because of its black-box nature. Yet this very nature also makes it difficult to quantify confidence in predictions extracted from ML models, and more fundamentally, to understand how such models are mathematical abstractions of training data. The goal of this paper is to unravel these issues and their connections to uncertainty quantification (UQ) by pursuing a line of reasoning motivated by diagnostics. In such settings, prevalence - i.e. the fraction of elements in class - is often of inherent interest. Here we analyze the many interpretations of prevalence to derive a level-set theory of classification, which shows that certain types of self-consistent ML models are equivalent to class-conditional probability distributions. We begin by studying the properties of binary Bayes optimal classifiers, recognizing that their boundary sets can be reinterpreted as level-sets of pairwise density ratios. By parameterizing Bayes classifiers in terms of the prevalence, we then show that they satisfy important monotonicity and class-switching properties that can be used to deduce the density ratios without direct access to the boundary sets. Moreover, this information is sufficient for tasks such as constructing the multiclass Bayes-optimal classifier and estimating inherent uncertainty in the class assignments. In the multiclass case, we use these results to deduce normalization and self-consistency conditions, the latter being equivalent to the law of total probability for classifiers. We also show that these are necessary conditions for arbitrary ML models to have valid probabilistic interpretations. Throughout we demonstrate how this analysis informs the broader task of UQ for ML via an uncertainty propagation framework.

This correspondence presents a power-aware cognitive radar framework for joint detection and tracking of multiple targets in a massive multiple-input multiple-output (MIMO) radar environment. Building on a previous single-target algorithm based on Partially Observable Monte Carlo Planning (POMCP), we extend it to the multi-target case by assigning each target an independent POMCP tree, enabling scalable and efficient planning. Departing from uniform power allocation, which is often suboptimal with varying signal-to-noise ratios (SNRs), our approach predicts each target's future angular position and expected received power based on its expected range. These predictions guide adaptive waveform design via a constrained optimization problem that allocates transmit energy to enhance the detectability of weaker or distant targets, while ensuring sufficient power for high-SNR targets. Simulations involving multiple targets with different SNRs confirm the effectiveness of our method. The proposed framework for the cognitive radar improves detection probability for low-SNR targets and achieves more accurate tracking compared to approaches using uniform or orthogonal waveforms. These results demonstrate the potential of the POMCP-based framework for adaptive, efficient multi-target radar systems.

Faithful free-text explanations are important to ensure transparency in high-stakes AI decision-making contexts, but they are challenging to generate by language models and assess by humans. In this paper, we present a measure for Prediction-EXplanation (PEX) consistency, by extending the concept of weight of evidence. This measure quantifies how much a free-text explanation supports or opposes a prediction, serving as an important aspect of explanation faithfulness. Our analysis reveals that more than 62% explanations generated by large language models lack this consistency. We show that applying direct preference optimization improves the consistency of generated explanations across three model families, with improvement ranging from 43.1% to 292.3%. Furthermore, we demonstrate that optimizing this consistency measure can improve explanation faithfulness by up to 9.7%.