As AI grows more powerful, it will increasingly shape how we understand the world. But with this influence comes the risk of amplifying misinformation and deepening social divides--especially on consequential topics where factual accuracy directly impacts well-being. Scalable Oversight aims to ensure AI systems remain truthful even when their capabilities exceed those of their evaluators. Yet when humans serve as evaluators, their own beliefs and biases can impair judgment. We study whether AI debate can guide biased judges toward the truth by having two AI systems debate opposing sides of controversial factuality claims on COVID-19 and climate change where people hold strong prior beliefs.

Recent advances in reasoning techniques have substantially improved the performance of large language models (LLMs), raising expectations for their ability to provide accurate, truthful, and reliable information. However, emerging evidence suggests that iterative reasoning may foster belief entrenchment, rather than enhancing truth-seeking behavior. In this study, we propose a systematic evaluation framework for belief entrenchment in LLM reasoning by leveraging the Martingale property from Bayesian statistics. This property implies that, under rational belief updating, the expected value of future beliefs should remain equal to the current belief, i.e., belief updates cannot be predicted from solely the current belief. We propose the unsupervised, regression-based Martingale Score to measure violations of this property, signaling a deviation from the Bayesian ability of updating on new evidence. In open-ended problem domains, including event forecasting, value-laden questions, and academic paper review, we found such violations to be widespread across models, reasoning paradigms, problem domains, and system prompts, where the future beliefs are consistently predictable from the model's current belief, a phenomenon which we term belief entrenchment. Through comprehensive experiments, we identify the models (e.g., GPT-4o), reasoning techniques (e.g., chain of thought), and domains (e.g., forecasting) more prone to belief entrenchment. Finally, we validate the Martingale Score by showing that it predicts ground-truth accuracy on problem domains where ground truth labels are available. This indicates that, while designed as an unsupervised metric that operates even in domains without access to ground truth, the Martingale Score is a useful proxy of the truth-seeking ability of the LLM reasoning process.

Agentic artificial intelligence systems introduce a new class of model risk. Unlike traditional predictive models, autonomous agents continuously acquire information, form beliefs regarding latent states of the environment, generate forecasts, select actions, and adapt their behavior over time. Existing validation methodologies focus primarily on predictive accuracy and therefore provide limited insight into the quality of the underlying decision process. This paper proposes a model validation framework for agentic AI based on Partially Observable Markov Decision Processes (POMDPs). The framework decomposes autonomous decision making into information, beliefs, forecasts, actions, and utility, allowing each component to be validated independently. Large language models (LLMs) are formalized as approximate Bayesian filtering operators, and a model-risk taxonomy is developed encompassing state-space, filtering, forecast, policy, utility-specification, and parameter risks. The model risk validation methodology is demonstrated through a portfolio-management case study in which an agent infers latent market regimes from market and macroeconomic information, generates belief-conditioned forecasts, and constructs portfolios using a Black--Litterman framework. Empirical validation combines performance analysis, belief calibration diagnostics, coverage tests, ablation studies, and parameter-sensitivity analysis. The results indicate that latent-state inference contributes independently to decision quality and that the principal conclusions remain robust across a broad range of parameter values. The principal contribution of the paper is a practical framework for extending established model risk management concepts to autonomous AI systems and providing a rigorous foundation for their validation, governance, and monitoring.

Planning under opponent uncertainty is a fundamental challenge in multi-agent environments, where an agent must act while inferring the hidden policies of its opponents. Existing type-based methods rely on manually defined behavior classes and struggle to scale, while model-free approaches are sample-inefficient and lack a principled way to incorporate uncertainty into planning. We propose Quantized Opponent Models (QOM), which learn a compact catalog of opponent types via a quantized autoencoder and maintain a Bayesian belief over these types online. This posterior supports both a belief-weighted meta-policy and a Monte-Carlo planning algorithm that directly integrates uncertainty, enabling real-time belief updates and focused exploration. Experiments show that QOM achieves superior performance with lower search cost, offering a tractable and effective solution for belief-aware planning.

Inference in Markov random fields subject to consistency structure is a fundamental problem that arises in many real-life applications. In order to enforce consistency, classical approaches utilize consistency potentials or encode constraints over feasible instances. Unfortunately this comes at the price of a tremendous computational burden. In this paper we suggest to tackle consistency by incorporating constraints on beliefs. This permits derivation of a closed-form message-passing algorithm which we refer to as the Constraints Based Convex Belief Propagation (CBCBP). Experiments show that CBCBP outperforms the conventional consistency potential based approach, while being at least an order of magnitude faster.

In Section 3, we verify that this geometry is linearly represented in the residual stream of transformers.

We are grateful to all constructive and actionable feedback provided by the reviewers. We believe to have addressed the key concerns raised by the reviewers below. 's concerns with our main hypothesis as it has not We are working to improve our explanations in section 2.2 based on all feedback We emphasize that careful empirical experimentation in ML can also bring valuable insights to the community. Studying these factors require an intersectional empirical study such as this paper. Probabilistic models benefit more from overshoot than Deterministic models.

Robots operating in everyday environments must often decide when and whether to engage with people, yet such decisions often hinge on subtle nonverbal cues that unfold over time and are difficult to model explicitly. Drawing on a five-day Wizard-of-Oz deployment of a mobile service robot in a university cafe, we analyze how people signal interaction readiness through nonverbal behaviors and how expert wizards use these cues to guide engagement. Motivated by these observations, we propose a two-stage pipeline in which lightweight perceptual detectors (gaze shifts and proxemics) are used to selectively trigger heavier video-based vision-language model (VLM) queries at socially meaningful moments. We evaluate this pipeline on replayed field interactions and compare two prompting strategies. Our findings suggest that selectively using VLMs as proxies for social reasoning enables socially responsive robot behavior, allowing robots to act appropriately by attending to the cues people naturally provide in real-world interactions.

As multi-agent systems are increasingly utilized for reasoning and decision-making applications, there is a greater need for LLM-based agents to have something resembling propositional beliefs. One simple method for doing so is to include statements describing beliefs maintained in the prompt space (in what we'll call their belief boxes). But when agents have such statements in belief boxes, how does it actually affect their behaviors and dispositions towards those beliefs? And does it significantly affect agents' ability to be persuasive in multi-agent scenarios? Likewise, if the agents are given instructions to be open-minded, how does that affect their behaviors? We explore these and related questions in a series of experiments. Our findings confirm that instructing agents to be open-minded affects how amenable they are to belief change. We show that incorporating belief statements and their strengths influences an agent's resistance to (and persuasiveness against) opposing viewpoints. Furthermore, it affects the likelihood of belief change, particularly when the agent is outnumbered in a debate by opposing viewpoints, i.e., peer pressure scenarios. The results demonstrate the feasibility and validity of the belief box technique in reasoning and decision-making tasks.

To collaborate with humans, robots must infer goals that are often ambiguous, difficult to articulate, or not drawn from a fixed set. Prior approaches restrict inference to a predefined goal set, rely only on observed actions, or depend exclusively on explicit instructions, making them brittle in real-world interactions. We present BALI (Bidirectional Action-Language Inference) for goal prediction, a method that integrates natural language preferences with observed human actions in a receding-horizon planning tree. BALI combines language and action cues from the human, asks clarifying questions only when the expected information gain from the answer outweighs the cost of interruption, and selects supportive actions that align with inferred goals. We evaluate the approach in collaborative cooking tasks, where goals may be novel to the robot and unbounded. Compared to baselines, BALI yields more stable goal predictions and significantly fewer mistakes.