Lee Hsien Loong

Proactive Vulnerability Discovery in Agentic LLMs Through Automated Indirect Prompt Injection Testing

PI-Hunter is an automated red-teaming framework designed to expose latent prompt injection vulnerabilities in LLM agents that interact with external tools and environments. The system constructs realistic test cases and iteratively evolves them through feedback-driven exploration to reveal how malicious instructions embedded in external sources can be retrieved and executed. Unlike existing defenses that focus on inference-time blocking, PI-Hunter proactively identifies attack surfaces and demonstrates effectiveness even against existing prompt injection defenses.

Spatiotemporal Feature Maps Enable Long-Horizon Mobile Manipulation Through Unified Environment-Robot Representations

SERF introduces a joint spatiotemporal feature map that represents both environment and robot state as neural points in a shared latent space, enabling vision-language-action policies to reason more effectively over long-horizon manipulation tasks. By maintaining updated representations of object positions (via rigid tracking) and robot configuration (via forward kinematics), the approach provides VLA models with multi-scale, multi-frame context that outperforms image-only baselines in trajectory efficiency, robustness to scene changes, and failure recovery on household manipulation benchmarks.

Geometry-Aware Vision Networks Improve Precision Guidance for Robot-Assisted Endoscopic Surgery

GeoCFNet addresses visual guidance challenges in robot-assisted endoscopic submucosal dissection by modeling dissection regions as geometry-aware confidence fields rather than discrete segmentation masks. The approach integrates pretrained vision foundation models (DINOv3) with specialized spatial regularization to maintain geometric coherence despite dynamic endoscopic artifacts like smoke and tissue deformation, achieving quantitatively stable confidence field estimation (RMSE 0.0480, SSIM 0.3397) suitable for surgical workflow integration.

Integrated Sensing-Communication Framework Mitigates UAV Beam Misalignment Through Active AoA Estimation

UAV communications over cellular networks suffer performance degradation from platform jittering-induced beam misalignment. The paper proposes a two-stage active sensing framework that alternates between deterministic sensing signals and information-bearing communication, with optimal time allocation derived in semi-closed form. A fundamental sensing-communication tradeoff is characterized, with the communication-oriented scheme outperforming when signal quality constraints exceed a specific threshold, and performance approaching jitter-free limits at high transmit power.

STCC: Unified Semantic Token Coding Framework for Foundation Model Communication

The Semantic Token Codebook Communication (STCC) framework is proposed for transmitting discrete semantic tokens from foundation models over noisy channels. Unlike previous Deep JSCC methods that operate on raw analog data, STCC's Semantic Token Codec (STC) uses an MLP-based encoder to learn token constellations optimized for channel topology. This approach ensures that channel noise induces topological or semantic errors rather than random data corruption, leading to superior performance in low-SNR environments without receiver modifications.

CORE-Bench: A New Benchmark for Agentic Code Retrieval Challenges Traditional Metrics

This paper introduces CORE-Bench, a novel benchmark addressing the limitations of existing evaluations in code retrieval for agentic coding. Unlike traditional methods focusing on snippet matching, CORE-Bench assesses a more complex, repository-aware retrieval process involving code understanding, issue localization, and contextual information gathering. Experimental results indicate a performance gap for current embedding models in this setting, though supervised fine-tuning shows potential for significant improvement.

Reconfigurable Antennas: Enabling Technologies and Future Directions for 6G Mobile Networks

Reconfigurable antennas (RAs) are critical for enabling 6G mobile networks by providing dynamic control over RF characteristics like radiation patterns and polarization. This survey categorizes RAs into fluid, movable, pinching, and reconfigurable holographic types, detailing their applications, channel modeling, performance analysis, resource allocation, and integration with other wireless technologies. The paper aims to guide future research by analyzing current challenges and future directions in RA development for advanced mobile communications.

Structural Properties of Constrained Graph Homomorphism Orders

This paper introduces and analyzes partial orders derived from constrained graph homomorphism types, extending standard properties like canonical cores and finiteness to these variants. It identifies specific cores for full and surjective homomorphisms, links them to point-determining graphs, and characterizes gaps in the full homomorphism order. The research also details properties of locally injective homomorphisms on connected graphs, including their core structure and universality.

Capability-Aligned Hierarchical Learning for Enhanced LLM Tool Use

This paper introduces Capability-Aligned Hierarchical Learning (CAHL), a novel approach for tool-augmented Large Language Models (LLMs). CAHL addresses the planner-executor misalignment issue prevalent in prior hierarchical methods by jointly optimizing high-level planning and low-level tool execution policies using RLVR. Experimental results on API-Bank, BFCL, and Bamboogle benchmarks validate the effectiveness of CAHL in improving LLM performance on tool-use tasks.

CapRL++: Unified Reinforcement Learning with Verifiable Rewards for Dense Image and Video Captioning

CapRL++ is a novel framework for training Large Vision-Language Models (LVLMs) for dense image and video captioning that moves beyond supervised fine-tuning (SFT). It employs reinforcement learning with verifiable rewards, redefining caption quality based on a caption's utility in enabling a separate language model to answer questions about visual content. This approach, featuring a decoupled two-stage pipeline, has demonstrated substantial downstream gains and enables compact models to achieve performance comparable to significantly larger models.

Complex Temperature-dependent Thermal Conductivity in a Sawtooth Chain Magnet Fe$_\mathrm{2}$SiSe$_\mathrm{4}$

An Explicit $O(r\log r)$ Threshold for Attaining the Semple--Steel Bound with $r$-State Characters

From Sampled Outcomes to Capability Distributions: Rethinking Supervision for LLM Routing

DaX: Learning General Pathology Representations Across Scales

Efficient $(α,β)$-core Computation and On-the-fly Query at Billion Scale with GPUs

Beyond Alignment: Value Diversity as a Collective Property in Multicultural Agent Systems

Spinning bulk-to-boundary correlators in the massless theories with Poincaré symmetry

Ensuring Interaction Safety in Multitask Exoskeleton Control: A Simulation-Trained Variable Impedance Framework

1/3 Fractional and Gapless Integer Quantum Anomalous Hall States in Rhombohedral Graphene

NICER detection of a new candidate cyclotron line in the bursting X-ray pulsar GRO J1744-28

Towards Accurate Model Selection in Deep Unsupervised Domain Adaptation

Physics-Informed Video Generation via Mixture-of-Experts Latent Alignment

UAT: Unified Audio-Text Diffusion for Audio Generation, Editing, and Captioning

Dual-Stream MLP is All You Need for CTR Prediction

EmoDistill: Offline Emotion Skill Distillation for Language Model Agents in Adversarial Negotiation

The spectral energy distribution of YSES 1 b and its circumplanetary disc

Generating Logically Consistent Synthetic Supply Chain Data with LLM-Driven Knowledge Graph Reasoning

Helicase: Uncertainty-Guided Supply Chain Knowledge Graph Construction with Autonomous Multi-Agent LLMs

Target-Aligned Bellman Backup for Cross-domain Offline Reinforcement Learning

The Quantum Homomorphism Orders are Universal

Beyond Imitation: Learning Safe End-to-End Autonomous Driving from Hard Negatives

Computing the Arc-Deletion Distance to Orchard Networks is NP-hard

SetCon: Towards Open-Ended Referring Segmentation via Set-Level Concept Prediction