A benchmark for general aviation agents that evaluates LLMs on safety-critical flight trajectory and attitude prediction, revealing a trade-off between precision and controllability.

A surprisingly simple and efficient AI weather forecaster that matches state-of-the-art performance with significantly reduced compute.

E3-TIR is a warm-up paradigm for LLM agents that enhances tool-use reasoning with less data and improved efficiency.

LMGenDrive is a unified framework for autonomous driving that combines multimodal understanding with generative world models for robust, end-to-end control.

Frequency-Enhanced Diffusion Models improve zero-shot skeleton action recognition by addressing spectral bias and recovering fine-grained motion details.

VLA-World is a Vision-Language-Action world model that unifies predictive imagination with reflective reasoning for improved autonomous driving foresight and safety.

DeepGuard enhances LLM code generation security by aggregating multi-layer representations to detect and mitigate vulnerabilities.

An open-source framework for seamless cross-provider LLM API translation using an intermediate representation.

A maturity model and framework for evolving AI-assisted coding into self-sustaining development systems, emphasizing feedback loops over AI models.

A benchmark and empirical study revealing how multi-agent system architectures amplify bias, even with neutral agents, and providing code for analysis.

RecaLLM addresses the 'lost-in-thought' phenomenon in LLMs by interleaving explicit in-context retrieval with reasoning, improving long-context performance without expensive training data.

HTNav is a hybrid navigation framework for urban aerial vision-and-language tasks, integrating imitation and reinforcement learning with a tiered structure for improved planning and control.

A benchmark and training method for AI agents to intelligently ask for help when faced with ambiguity, improving task completion and reducing errors.

HM-Bench is the first benchmark for evaluating multimodal large language models on hyperspectral image understanding, featuring a dual-modality framework and a large-scale dataset.

A reinforcement learning framework that decouples confidence calibration in large vision-language models into visual and reasoning components to reduce hallucinations and improve accuracy.

A foundation model and large dataset for universal defect generation in images, enabling reference-based and text-guided editing.

A framework for robust out-of-distribution object detection in LiDAR data for autonomous driving, significantly improving safety by identifying unexpected objects.

BERT-as-a-Judge offers a scalable and robust alternative to lexical methods for LLM evaluation, matching larger models at lower computational cost.

A massive synthetic dataset for training AI to understand and generate physically plausible videos, enabling advancements in world models.

VisionFoundry teaches VLMs visual perception skills using task-specific synthetic images generated by LLMs and text-to-image models.

SafeAdapt provides provable safety guarantees for updating reinforcement learning policies in non-stationary environments.

A novel graph neural network architecture that overcomes homophily assumptions by applying self-attention within local neighborhoods, with significant performance gains and code availability.

CLIP-Inspector detects backdoors in prompt-tuned CLIP models by inverting out-of-distribution triggers, enabling model vetting and repair.

A new benchmark for evaluating self-evolving agents that quantifies their ability to accumulate experience and optimize strategies across task boundaries.

VisPrompt enhances vision-language models' robustness to label noise by injecting visual semantics into prompt learning, improving performance on noisy datasets.

Camera Artist automates narrative video creation with advanced cinematic storytelling techniques.

A tutor-student multi-agent system that enhances LLM problem-solving performance in coding tasks through structured, role-differentiated interaction.

A framework that automatically optimizes agent skill bundles for software engineering tasks, improving success rates and reducing costs.

An efficient diffusion-based vision-language model for one-step chest X-ray report generation that significantly speeds up inference while maintaining clinical accuracy.

A personalized sequential recommendation model that integrates time, multi-interest, and explanation personalization for improved accuracy and explanation quality at lower computational cost.

A human-in-the-loop web application for systematically evaluating LLM text simplifications across diverse prompts and architectures.

An agentic framework for interactive ASR that uses LLM-as-a-Judge for semantic evaluation and multi-turn correction to improve recognition quality.

SPPO is a scalable algorithm for aligning LLMs in long-horizon reasoning tasks, offering improved sample efficiency and stability over standard PPO.

An agentic system and benchmark for predicting multilingual model performance in low-resource scenarios, demonstrating structured reasoning for evaluation.

Mosaic is a multimodal jailbreak framework that overcomes surrogate dependency to achieve state-of-the-art attack success rates against closed-source vision-language models.

A multimodal deep learning framework using Vision Transformers can predict chemotherapy response in ovarian cancer from CT scans, acting as a decision-support tool for multidisciplinary teams.

A framework to enhance visual focus and counteract visual forgetting in multimodal reasoning models.

ViSA-R2 infers analytical solutions for physical fields from visualizations, outperforming existing models and providing a benchmark for visual-to-symbolic reasoning.

NyayaMind is an open-source framework for transparent legal reasoning and judgment prediction in the Indian legal system, improving explanation quality and evidence alignment.

A reinforcement learning framework that optimizes LLM reasoning by incorporating stability metrics like autocorrelation and path efficiency.

Process Reward Agents provide step-wise rewards to frozen LLMs for improved reasoning accuracy in knowledge-intensive domains like medicine.

DRBENCHER is a novel benchmark generator for AI agents that require both web browsing and multi-step computation, revealing significant performance gaps in current frontier models.

A plug-and-play module that drastically reduces computational cost in monocular SLAM by predicting frame utility before expensive processing, achieving 5x throughput speedup.

A human-grounded dataset and LLM evaluation for personality-shaped emotional responses to text, addressing the 'personality illusion' in affective computing.

A framework for granularly controlling multimodal LLM activations to enhance safety without compromising general capabilities, using a curated concept dictionary and Sparse Autoencoders.

A text-to-speech system that achieves natural automated dubbing by synchronizing translated text phonetically and temporally with source speech.

A benchmark for evaluating customer service LLMs with dynamic dialogue graphs and adversarial testing, revealing an 'Execution Gap' in action derivation.

A novel architecture for complex table question answering that reconstructs tables into logical semantic trees and uses a dual-mode reasoning framework to achieve state-of-the-art performance.

A system for collecting, transcribing, and anonymizing Telegram data to build datasets for social engineering detection while complying with privacy regulations.

XFED is the first aggregation-agnostic, non-collusive model poisoning attack against Byzantine-robust federated classifiers, demonstrating significant security vulnerabilities.

SenBen provides a large-scale benchmark and a compact, efficient model for explainable content moderation by generating sensitive scene graphs.

A user-side method injects visual prompts into images to prevent multi-modal LLMs from analyzing them, protecting sensitive information.

An open-access dataset and benchmark for energy-aware LLM inference on heterogeneous GPUs, enabling significant energy savings.

AudioGuard provides comprehensive audio safety protection against diverse threats with low latency.

A training-free framework uses an evolving skill graph to improve vision-language geolocation and enable autonomous self-evolution.

eBandit is a kernel-resident reinforcement learning framework for adaptive video streaming that optimizes quality of experience by directly monitoring transport layer signals.

The Symbolic-Neural Consistency Audit (SNCA) framework measures the gap between LLMs' self-stated safety policies and their actual behavior, revealing systematic compliance gaps.

An adaptive dual residual U-Net with attention mechanisms for precise brain tumor segmentation.

Accelerating the enumeration of Minimal Unsatisfiable Subsets (MUSes) in constraint satisfaction problems using domain-agnostic Hypergraph Neural Networks.

PinpointQA provides a benchmark for improving AI's ability to understand small object locations in indoor videos.

CORA provides statistically grounded safeguards for mobile GUI automation to prevent harmful actions.

A lightweight router that improves two-hop question answering by intelligently selecting retrieval strategies based on question characteristics.

A Video Diffusion Model that jointly learns video frames and camera trajectories, enabling novel view synthesis and pose estimation.

A mathematical framework for simulating counterfactual student dropout policies using temporal engagement data.

A theoretical model for causal inference in graph representation learning that guarantees causal validity by operating on indivisible graph units, with an integrated module for existing pipelines.

AlignGuard is a new technique that detects critical correctness bugs in PyTorch's compiler, improving the reliability of LLM applications.

A plug-and-play method uses noise-aware in-context learning to reduce hallucinations in auditory large language models without fine-tuning.

A novel multilingual story moral generation task and dataset to evaluate the cultural alignment and diversity of LLMs.

An agent-based system for nuanced, inference-time consent verification in generative AI, balancing rights holder control with developer flexibility.

A framework for evidence-grounded reasoning that trains models to explicitly depend on provided evidence for decision-making, improving accuracy in domains like radiology.

Empower simulations and analytics by forecasting future event outcomes from static images.

A visual retrieval-augmented generation system improving on state-of-the-art, targeting visual data workflows.

SatIR is a scalable constraint-satisfaction method for clinical trial matching that uses LLMs to improve recall and interpretability over existing techniques.

TensorHub is a production-ready system for scalable and elastic weight transfer in LLM RL training, significantly reducing GPU stall time.

Spatial-Gym: A new environment for step-by-step evaluation of agent spatial reasoning capabilities.

A graph-based parser outperforms LLMs on supervised relation extraction for complex graphs, offering a lighter and more effective solution.

A computational model proposing that Von Economo neurons facilitate rapid social decision-making by implementing a biological speed-accuracy tradeoff.

A new Bayesian optimization framework that focuses on finding a single high-quality solution rather than approximating the entire Pareto front for many-objective problems.

Leveraging agentic language models and heuristic search to improve the generation of AI planning domains from natural language.

A theoretical framework for understanding communication between agents with different computational capacities, defining capacity-derived semantic spaces and identifying critical communication rates.

Integrates graph-based embeddings into event sequence models to improve user attribute prediction for fraud prevention and recommendations.

A novel architecture for spiking transformers that achieves ultra-high energy efficiency through multi-dimensional grouping.

A utility-aware framework for jailbreaking audio LLMs by selectively perturbing frequency bands, balancing attack success with transcription quality.

A survival-oriented benchmark for temporal student dropout risk modeling with harmonized representations.

Identifies limitations in GNN-based reinforcement learning for cloud scheduling under distribution shifts and proposes more robust representations.

A framework for language-based drug discovery agents that improves success rates by precisely diagnosing and correcting protocol violations at the set level.

A new benchmark evaluates large language models on conditional decision-making in complex, compositional action spaces.

An AI system uses computer vision to track players and predict their positions on a soccer field for tactical analysis.

A novel OSINT methodology to detect real-world AI scheming incidents by analyzing chatbot and command-line transcripts.

Developing linguistically-informed representations for English as a Second or Foreign Language (ESFL) to improve language acquisition research.

An uncertainty-routed Transformer for medical image classification that improves calibration and selective prediction.

RAMP is a hybrid DRL strategy for online learning of numeric action models in automated planning, outperforming PPO on standard IPC numeric domains.

A framework for robust and sample-efficient reinforcement learning in robotics by jointly generating and transferring experience from source to target tasks.

A deep learning framework for tracking and reconstructing the lineage of ligament breakup in liquid sheets from high-speed images.

A novel backdoor attack formulation targeting model-in-skill threats within agent ecosystems.

This work improves speech anonymization by enhancing the excitation stage of DDSP-QbE with explicit voicing detection and PolyBLEP correction for cleaner, more natural synthesized speech.

Investigating HiFloat4 format for language model pre-training on Ascend NPUs to improve computational and memory efficiency.

A reinforcement learning approach using graph neural networks for decentralized entanglement routing in dynamic satellite-assisted quantum networks.

PSIRNet is a deep learning method for rapid, free-breathing cardiac MRI that significantly reduces acquisition time while maintaining diagnostic image quality.

A pretraining framework for instructional videos that improves action-centric understanding by disentangling actions from objects and modeling temporal structure.

SafeMind is a differentiable control framework for quadrupeds that unifies safety guarantees with adaptive locomotion.

A dynamics-informed diffusion framework for stable long-horizon spatiotemporal prediction using PDE regularization and uncertainty-aware filtering.

Advantage-Guided Diffusion for Model-Based Reinforcement Learning improves sample efficiency and return by steering diffusion with advantage estimates.

AI models can significantly persuade people to take real-world actions like signing petitions and donating to charity, but attitudinal persuasion does not correlate with behavioral outcomes.

A new training strategy for LLMs that balances persona expressivity with task performance by mitigating a trade-off in reinforcement learning.

Physics-guided surrogate learning enables zero-shot control of turbulent wings, significantly reducing drag with drastically reduced training cost.

An advanced SE(3)-equivariant graph attention Transformer for efficient and expressive 3D atomistic modeling, achieving state-of-the-art results on multiple benchmarks.

A novel framework using tensor decomposition to quantify uncertainty in Large Language Model-based Multi-Agent Systems.

Evaluating two LLM-based agent scaffolding approaches ('structuring' vs. 'problematizing') for improving diagnostic reasoning in vocational training.

This paper investigates how preference data quality impacts language model reasoning, offering insights for more effective alignment techniques.

Benchmarking humor alignment in large language models by having them play Cards Against Humanity, revealing modest human preference alignment and significant agreement among models.

Formalizing how the environment can serve as an agent's memory in Reinforcement Learning, reducing the information needed to represent history through observable artifacts.

A tutorial arguing for a shift from relevance-centric to utility-centric retrieval in the LLM era, focusing on how retrieved information aids LLM generation.

This paper statistically analyzes the King Wen sequence and finds it does not improve neural network training, despite its anti-habituation properties.

This research identifies a distinct, unified mechanism for harmful content generation in LLMs, suggesting a path for more principled safety approaches.

This paper details the co-design process for an accessible, multi-modal 3D data visualization tool for blind and low-vision users.

This paper theoretically explores representing models of MSO formulas using decision diagrams, parameterized by graph treewidth and formula size, connecting to knowledge representation.

This research explores how humans adapt physical prediction and planning strategies under cognitive resource constraints, revealing a hierarchical, resource-rational architecture.

This research investigates how humans attribute responsibility when working with AI teammates, finding they assign more blame to humans in AI-human teams compared to human-human teams.

This paper introduces a framework to evaluate quantum-inspired document embeddings, finding they have structural limitations and are best used as auxiliary components.

Theoretical analysis of generalization and scaling laws for Mixture-of-Experts Transformers to optimize model size, data size, and compute tradeoffs.

Leveraging discarded AI models from a 'scrapyard' to enable frugal experimentation and reconfigure legacy systems for new applications.

LLMs exhibit both baseline and strategic algorithmic monoculture in coordination games, but lag humans in sustaining rewarded heterogeneity.

This paper details findings from a workshop on building better reinforcement learning environments for autonomous cyber defense.

Revisiting the capacity gap in Chain-of-Thought distillation, this paper proposes a more realistic evaluation protocol and offers practical guidance for selecting teacher-student pairs.