According to canonical negotiation theory, success in negotiation depends on balancing competing demands such as empathizing and asserting, concern for other and self, and being soft on people while hard on the problem. Humans struggle with these tensions, limiting rigorous testing of theories. AI agents, with their precision, consistency, and scalability, enable a new class of experiments. This paper introduces 'personality engineering,' a methodology that uses AI agents to precisely parameterize, manipulate, and evaluate negotiator personality. It proposes the interpersonal circumplex (warmth and dominance) as a foundational coordinate system. This approach offers rigorous methods for testing classic theories and a practical guide for designing AI negotiation agent personalities.

This paper introduces AgentAtlas, a framework to evaluate LLM agents beyond simple outcome leaderboards. It comprises four components: a six-state control-decision taxonomy, a nine-category trajectory-failure taxonomy, a taxonomy-aware vs. blind methodology, and a benchmark-coverage audit. Experiments show that removing explicit label menus drops trajectory accuracy by 14-40 percentage points across all models, converging to a narrow 0.54-0.62 floor, and no single model excels on all metrics.

Benchmark-based evaluation remains important for tracking frontier AI progress, but it can both overstate and understate deployed capability because it privileges tasks that can be precisely specified, automatically graded, easy to optimize for, and run with low budgets and short time horizons. This paper advocates for a complementary class of evaluations, termed open-world evaluations: long-horizon, messy, real-world tasks assessed through small-sample qualitative analysis. The authors introduce CRUX, a project for conducting such evaluations regularly, and demonstrate an AI agent successfully developing and publishing an iOS app with only one manual intervention.

In high-stakes automated decision-making, predictive uncertainty enables users to accept or reject predictions based on cost trade-offs. Current evaluation methods for uncertainty-augmented (UA) systems are inadequate. This paper proposes ECUAS_n, a family of proper scoring rules with parameter n controlling the trade-off between cost of incorrect predictions and imperfect uncertainties. Experiments on diverse datasets demonstrate advantages.

This paper introduces several approaches to generate high-quality embeddings for Horn logic reasoning, using triplet loss with three novel ideas: generating anchors with repeated terms, balancing easy/medium/hard examples, and periodically emphasizing the hardest examples. Experiments across multiple knowledge bases show improved downstream reasoning performance.

AgentCo-op is a retrieval-based synthesis framework that composes reusable skills, tools, and external agents into executable workflows via typed artifact handoffs, with bounded self-guided local repair on failure. In open-world genomics case studies, it coordinates independently developed agents and tool repositories without redesign. On six benchmarks, it achieves top results on four and reduces per-task costs consistently.

A new approach called OSCToM uses reinforcement learning to generate adversarial examples for testing theory of mind in LLMs, achieving 76% accuracy on FANToM benchmark and 6x efficiency gain.

Proposes COSMO-Agent, a tool-augmented reinforcement learning framework that teaches LLMs to complete the closed-loop CAD-CAE process, addressing the semantic gap in iterative industrial design. Experiments show small open-source LLMs outperform larger models in feasibility, efficiency, and stability.

Large language models struggle with concept drift and costly gradient-based adaptation in dynamic real-world settings. This paper introduces SOLAR, a self-optimizing lifelong autonomous agent that uses parameter-level meta-learning and multi-level reinforcement learning to adapt continuously without manual curation. SOLAR maintains an evolving knowledge base that balances plasticity and stability, outperforming strong baselines on commonsense, mathematical, medical, coding, social, and logical reasoning tasks, marking a significant step toward lifelong adaptive agents.

EvoScene-VLA is a new vision-language-action (VLA) policy that maintains a persistent action-updated scene prior across control chunks, addressing scene changes caused by robot actions. It achieves 89.1% success rate on RoboTwin tasks and outperforms baselines on a real robot.

This paper presents an analytical and experimental force analysis of a linear soft sleeve actuator (LSSA) for wearable robotics. A quasi-static model was developed and validated with experiments showing force decreasing from ~112 N at zero extension to near zero at 40 mm under 125 kPa. Static loading delays and reduces force output.

This paper proposes a framework for heterogeneous multi-team collaboration via dynamic robot allocation, using Hamilton's rule for altruistic decision-making. The allocation problem is NP-hard, addressed by a graph neural network policy under centralized training and decentralized execution, validated in firefighting simulations to achieve near-optimal performance and scalability.

This paper proposes an adaptive coverage strategy that uses real-time feedback from an environmental model to adjust robot sampling behavior, enhancing ergodic search efficiency in unknown environments.

PGDG is a novel data generation framework that expands a single demonstration into a diverse dataset of physically plausible recovery behaviors for bimanual manipulation, without human labeling. It iteratively improves data quality using a physics-grounded sampler and curator, significantly outperforming spatial-only augmentation on both simulated and real-world tasks, and enables fine-tuning of foundation models like GR00T.

This paper investigates grasp-based dynamic locomotion for multi-limbed robots in microgravity. It presents a parameterizable planning framework to evaluate design parameters such as gait pattern, stride length, speed, and posture. Simulations show that enlarging the feasible contact wrench space and attenuating impulsive whole-body dynamics improve performance.

Researchers present a closed-loop sim-to-real reinforcement learning approach for microfiber shape control on a surface. The policy is trained in a simplified frictionless simulator and transferred directly to a physical system without retraining, achieving a mean shape error of 270 micrometers across diverse configurations and sub-millimeter accuracy across various fiber diameters and lengths.

This paper presents a distributed multi-coverage algorithm for drone swarms that ensures redundant coverage of critical assets using only local sensing and communication, without global coordination. It is accepted at ANTS 2026.

A new VR-based shared control framework enables real-time human-guided navigation for aerial robot teams in unknown environments, combining motion-primitive planning with admittance control to improve safety and reduce operator workload.

BEiTScore is a new reference-free image captioning evaluation metric based on a lightweight cross-encoder initialized from a VQA model checkpoint, using adversarial LLM-based data augmentation to achieve state-of-the-art performance with high efficiency.

AVI-HT is an adaptive visual-IMU fusion approach for 3D hand pose tracking that jointly models egocentric images with on-glove 6-DoF IMU signals. It achieves significantly improved accuracy and availability, especially in hand-object interaction scenarios with heavy visual occlusion. Key ingredients include synchronized multi-modal training data and a cross-sensor deep attention mechanism. Experiments on the DexGloveHOI dataset show that AVI-HT reduces mean keypoint error by 16.1% and its wrist-aligned variant by 24.2% over baselines.

MRecover is a conditional generative model that synthesizes T2w TSE images from routinely acquired T1w MRI to recover motion-corrupted hippocampal subfield segmentation. Trained on 7T data (n=577), it achieved high in-domain fidelity (SSIM=0.84, FSIM=0.94) and generalized well to 3T data. In the ADNI3 dataset, it increased analyzable subjects by 31.8% (593 vs 450) and improved effect sizes for diagnostic group differences.

Existing vision-language models underperform in ultrasound visual question answering because they fail to emulate sonographers' active zooming into lesions and ignore annotation subjectivity. This paper introduces a framework with a Zoom-then-Diagnose paradigm and an uncertainty-aware reward within GRPO, encouraging accurate predictions on clear cases and caution under ambiguity. Lesion localization improves by 39.3% on liver, breast, and thyroid datasets.

This paper proposes the Ablate-to-Validate diagnostic principle and its instantiation, the Token Replacement Test (TRT), to determine whether vision-language models (VLMs) genuinely use continuous latent tokens for reasoning. Experiments show that VLMs retain most performance gains even when token content is corrupted or replaced, indicating that accuracy improvements are a misleading proxy for latent-token reasoning.

UniVL introduces a spatially grounded contextual image generation task that uses a unified visual input with rendered text instructions, eliminating the need for a separate text encoder. The method achieves improved image quality (FID from 14 to 11, PSNR from 16 to 20) and up to 52% reduction in inference TFLOPs and 44% faster runtime on a new benchmark of 477K images.

GenEvolve is a self-evolving framework for image generation agents that models each generation attempt as a tool-orchestrated trajectory. It compares multiple trajectories for the same request and extracts best-worst differences into structured visual experience, providing dense token-level supervision. Experiments show state-of-the-art performance on public benchmarks and the newly constructed GenEvolve-Bench.

PhysX-Omni is a unified framework for generating simulation-ready physical 3D assets across diverse categories including rigid, deformable, and articulated objects. It introduces a novel geometry representation tailored for Vision-Language Models that directly encodes high-resolution 3D structures without compression, significantly improving generation performance. The framework also presents the first general simulation-ready 3D dataset, PhysXVerse, and a comprehensive benchmark, PhysX-Bench, for evaluating both generative and understanding capabilities. Experiments demonstrate strong performance in generation and understanding, with potential applications in simulation-ready scene generation and robotic policy learning, advancing embodied AI and physics-based simulation.