Large Language Models (LLMs) exhibit impressive general-purpose capabilities but also introduce serious safety risks, particularly the potential for deception as models acquire increased agency and human oversight diminishes. In this work, we present LieCraft: a novel evaluation framework and sandbox for measuring LLM deception that addresses key limitations of prior game-based evaluations. At its core, LieCraft is a novel multiplayer hidden-role game in which players select an ethical alignment and execute strategies over a long time-horizon to accomplish missions. Cooperators work together to solve event challenges and expose bad actors, while Defectors evade suspicion while secretly sabotaging missions. To enable real-world relevance, we develop 10 grounded scenarios such as childcare, hospital resource allocation, and loan underwriting that recontextualize the underlying mechanics in ethically significant, high-stakes domains. We ensure balanced gameplay in LieCraft through careful design of game mechanics and reward structures that incentivize meaningful strategic choices while eliminating degenerate strategies. Beyond the framework itself, we report results from 12 state-of-the-art LLMs across three behavioral axes: propensity to defect, deception skill, and accusation accuracy. Our findings reveal that despite differences in competence and overall alignment, all models are willing to act unethically, conceal their intentions, and outright lie to pursue their goals.

Shahab Rahimirad, Guven Gergerli, Lucia Romero, Angela Qian, Matthew Lyle Olson, Simon Stepputtis, Joseph Campbell. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Being able to understand visual scenes is a precursor for many downstream tasks, including autonomous driving, robotics, and other vision-based approaches. A common approach enabling the ability to reason over visual data is Scene Graph Generation (SGG); however, many existing approaches assume undisturbed vision, i.e., the absence of real-world corruptions such as fog, snow, smoke, as well as non-uniform perturbations like sun glare or water drops. In this work, we propose a novel SGG benchmark containing procedurally generated weather corruptions and other trans-formations over the Visual Genome dataset. Further, we in-troduce a corresponding approach, Hierarchical Knowledge Enhanced Robust Scene Graph Generation (HiKER-SGG), providing a strong baseline for scene graph generation under such challenging setting. At its core, HiKER-SGG utilizes a hierarchical knowledge graph in order to refine its predictions from coarse initial estimates to detailed predictions. In our extensive experiments, we show that HiKER-SGG does not only demonstrate superior performance on corrupted images in a zero-shot manner, but also outperforms current state-of-the-art methods on uncorrupted SGG tasks. Code is available at https://github.com/zhangce01/HiKER-SGG.

Task-oriented grasping of unfamiliar objects is a necessary skill for robots in dynamic in-home environments. Inspired by the human capability to grasp such objects through intuition about their shape and structure, we present a novel zero-shot task-oriented grasping method leveraging a geometric decomposition of the target object into simple, convex shapes that we represent in a graph structure, including geometric attributes and spatial relationships. Our approach employs minimal essential information – the object’s name and the intended task – to facilitate zero-shot task-oriented grasping. We utilize the commonsense reasoning capabilities of large language models to dynamically assign semantic meaning to each decomposed part and subsequently reason over the utility of each part for the intended task. Through extensive experiments on a real-world robotics platform, we demonstrate that our grasping approach’s decomposition and reasoning pipeline is capable of selecting the correct part in 92% of the cases and successfully grasping the object in 82% of the tasks we evaluate. Additional videos, experiments, code, and data are available on our project website: https://shapegrasp.github.io/.

While Large Language Models (LLMs) have demonstrated impressive accomplishments in both reasoning and planning, their abilities in multi-agent collaborations remains largely unexplored.This study evaluates LLMbased agents in a multi-agent cooperative text game with Theory of Mind (ToM) inference tasks, comparing their performance with Multi-Agent Reinforcement Learning (MARL) and planning-based baselines.We observed evidence of emergent collaborative behaviors and high-order Theory of Mind capabilities among LLM-based agents.Our results reveal limitations in LLM-based agents' planning optimization due to systematic failures in managing long-horizon contexts and hallucination about the task state.We explore the use of explicit belief state representations to mitigate these issues, finding that it enhances task performance and the accuracy of ToM inferences for LLMbased agents.

We introduce an imitation learning-based physical human-robot interaction algorithm capable of predicting appropriate robot responses in complex interactions involving a superposition of multiple interactions. Our proposed algorithm, Blending Bayesian Interaction Primitives (B-BIP) allows us to achieve responsive interactions in complex hugging scenarios, capable of reciprocating and adapting to a hug's motion and timing. We show that this algorithm is a generalization of prior work, for which the original formulation reduces to the particular case of a single interaction, and evaluate our method through both an extensive user study and empirical experiments. Our algorithm yields significantly better quantitative prediction error and more-favorable participant responses with respect to accuracy, responsiveness, and timing, when compared to existing state-of-the-art methods.

Action advising is a knowledge transfer technique for reinforcement learning based on the teacher-student paradigm. An expert teacher provides advice to a student during training in order to improve the student's sample efficiency and policy performance. Such advice is commonly given in the form of state-action pairs. However, it makes it difficult for the student to reason with and apply to novel states. We introduce Explainable Action Advising, in which the teacher provides action advice as well as associated explanations indicating why the action was chosen. This allows the student to self-reflect on what it has learned, enabling advice generalization and leading to improved sample efficiency and learning performance - even in environments where the teacher is sub-optimal. We empirically show that our framework is effective in both single-agent and multi-agent scenarios, yielding improved policy returns and convergence rates when compared to state-of-the-art methods.

This paper introduces a novel state estimation framework for robots using differentiable ensemble Kalman filters (DEnKF). DEnKF is a reformulation of the traditional ensemble Kalman filter that employs stochastic neural networks to model the process noise implicitly. Our work is an extension of previous research on differentiable filters, which has provided a strong foundation for our modular and end-to-end differentiable framework. This framework enables each component of the system to function independently, leading to improved flexibility and versatility in implementation. Through a series of experiments, we demonstrate the flexibility of this model across a diverse set of real-world tracking tasks, including visual odometry and robot manipulation. Moreover, we show that our model effectively handles noisy observations, is robust in the absence of observations, and outperforms state-of-the-art differentiable filters in terms of error metrics. Specifically, we observe a significant improvement of at least 59% in translational error when using DEnKF with noisy observations. Our results underscore the potential of DEnKF in advancing state estimation for robotics. Code for DEnKF is available at https://github.com/ir-lab/DEnKF

Out-of-distribution (OOD) data poses serious challenges in deployed machine learning models, so methods of predicting a model's performance on OOD data without labels are important for machine learning safety. While a number of methods have been proposed by prior work, they often underestimate the actual error, sometimes by a large margin, which greatly impacts their applicability to real tasks. In this work, we identify *pseudo-label shift*, or the difference between the predicted and true OOD label distributions, as a key indicator of this underestimation. Based on this observation, we introduce a novel method for estimating model performance by leveraging optimal transport theory, Confidence Optimal Transport (COT), and show that it provably provides more robust error estimates in the presence of pseudo-label shift. Additionally, we introduce an empirically-motivated variant of COT, Confidence Optimal Transport with Thresholding (COTT), which applies thresholding to the individual transport costs and further improves the accuracy of COT's error estimates. We evaluate COT and COTT on a variety of standard benchmarks that induce various types of distribution shift -- synthetic, novel subpopulation, and natural -- and show that our approaches significantly outperform existing state-of-the-art methods with up to 3x lower prediction errors.

In this paper we describe the intervention framework of ATLAS, an artificial socially intelligent agent that advises teams. The framework treats interventions as atomic components, and manages the lifecycle of each intervention through presentation, as well as followups to interventions. The key benefit of this framework is that it allows for rapid development of scenario-specific Interventions that leverage scenario-agnostic team models. The implementation of this framework is reported for three player teams in a Search and Rescue task simulated in Minecraft. Low competence teams advised by ATLAS improved more between first and second trials than those with a human advisor while the reverse was found for high competence. Four times as many interventions were proposed as were presented. 15 % of advice was withheld to avoid repetitive advice, excessive rate of advice, and needlessly advising high performing teams, while a Theory of Mind model and delay for confirmation mechanism filtered out other unnecessary advice.

This paper presents a learning-from-demonstration (LfD) framework for teaching human-robot social interactions that involve whole-body haptic interaction, i.e. direct human-robot contact over the full robot body. The performance of existing LfD frameworks suffers in such interactions due to the high dimensionality and spatiotemporal sparsity of the demonstration data. We show that by leveraging this sparsity, we can reduce the data dimensionality without incurring a significant accuracy penalty, and introduce three strategies for doing so. By combining these techniques with an LfD framework for learning multimodal human-robot interactions, we can model the spatiotemporal relationship between the tactile and kinesthetic information during whole-body haptic interactions. Using a teleoperated bimanual robot equipped with 61 force sensors, we experimentally demonstrate that a model trained with 121 sample hugs from 4 participants generalizes well to unseen inputs and human partners.

We propose in this paper Periodic Interaction Primitives - a probabilistic framework that can be used to learn compact models of periodic behavior. Our approach extends existing formulations of Interaction Primitives to periodic movement regimes, i.e., walking. We show that this model is particularly well-suited for learning data-driven, customized models of human walking, which can then be used for generating predictions over future states or for inferring latent, biomechanical variables. We also demonstrate how the same framework can be used to learn controllers for a robotic prosthesis using an imitation learning approach. Results in experiments with human participants indicate that Periodic Interaction Primitives efficiently generate predictions and ankle angle control signals for a robotic prosthetic ankle, with MAE of 2.21° in 0.0008s per inference. Performance degrades gracefully in the presence of noise or sensor fall outs. Compared to alternatives, this algorithm functions 20 times faster and performed 4.5 times more accurately on test subjects.

Imitation learning is a popular approach for teaching motor skills to robots. However, most approaches focus on extracting policy parameters from execution traces alone (i.e., motion trajectories and perceptual data). No adequate communication channel exists between the human expert and the robot to describe critical aspects of the task, such as the properties of the target object or the intended shape of the motion. Motivated by insights into the human teaching process, we introduce a method for incorporating unstructured natural language into imitation learning. At training time, the expert can provide demonstrations along with verbal descriptions in order to describe the underlying intent (e.g., "go to the large green bowl"). The training process then interrelates these two modalities to encode the correlations between language, perception, and motion. The resulting language-conditioned visuomotor policies can be conditioned at runtime on new human commands and instructions, which allows for more fine-grained control over the trained policies while also reducing situational ambiguity. We demonstrate in a set of simulation experiments how our approach can learn language-conditioned manipulation policies for a seven-degree-of-freedom robot arm and compare the results to a variety of alternative methods.

Musculoskeletal robots that are based on pneumatic actuation have a variety of properties, such as compliance and back-drivability, that render them particularly appealing for human-robot collaboration. However, programming interactive and responsive behaviors for such systems is extremely challenging due to the nonlinearity and uncertainty inherent to their control. In this paper, we propose an approach for learning Bayesian Interaction Primitives for musculoskeletal robots given a limited set of example demonstrations. We show that this approach is capable of real-time state estimation and response generation for interaction with a robot for which no analytical model exists. Human-robot interaction experiments on a 'handshake' task show that the approach generalizes to new positions, interaction partners, and movement velocities.

Integrated GPUs have become an indispensable component of mobile processors due to the increasing popularity of graphics applications. The GPU frequency is a key factor both in application throughput and mobile processor power consumption under graphics workloads. Therefore, dynamic power management algorithms have to assess the performance sensitivity to the GPU frequency accurately. Since the impact of the GPU frequency on performance varies rapidly over time, there is a need for online performance models that can adapt to varying workloads. This paper presents a light-weight adaptive runtime performance model that predicts the frame processing time. We use this model to estimate the frame time sensitivity to the GPU frequency. Our experiments on a mobile platform running common GPU benchmarks show that the mean absolute percentage error in frame time and frame time sensitivity prediction are 3.8% and 3.9%, respectively.