A training-free framework that enables frozen LLMs to adapt to distribution shifts by distilling supervision from their own inference experiences through an 'Explore-Reflect-Steer' loop.

A novel reinforcement learning approach that optimizes LLM reasoning by directly updating the pre-train space, leading to significant improvements in pruning incorrect reasoning and fostering reflective behaviors.

A novel method for on-policy knowledge distillation that significantly reduces memory usage and training time by intelligently selecting informative tokens, validated on multiple LLM architectures.

An automated platform for generating and visualizing threat intelligence for agentic systems, addressing critical security gaps.

An agentic multi-modal framework that autonomously generates semiconductor failure analysis reports from inspection images in under a minute by fusing vision, telemetry, and historical data.

HETA is a novel attribution framework for decoder-only LLMs that provides context-aware, causally faithful, and semantically grounded explanations of token contributions.

MIND is an LLM-driven co-scientist for material research, automating hypothesis validation through integrated in-silico experiments and a user interface.

Shielding vision-language models from manipulation by aligning their early visual cortex representations with human neural processing.

A GPU-friendly, training-free method for accelerating Vision Transformers and enhancing image synthesis by merging and restoring tokens using matrix operations.

WebXSkill enables web agents to autonomously perform complex browser tasks with improved success rates through executable skills.

A budget-aware vision-language model for ultra-high-resolution remote sensing that efficiently compresses visual tokens using query-guided importance estimation and region-wise strategies.

A systematic study and open dataset for synthesizing high-quality pretraining data for LLMs, reducing generation costs by up to 30x.

A Transformer-based framework that bridges multi-agent reinforcement learning to single-agent reinforcement learning for improved coordination and performance.

An efficient RGB-Event object tracking framework with adaptive state transitions and gated fusion for robust cross-modal integration.

A novel framework using LLMs to automate and optimize reward function design in reinforcement learning, reducing evaluation costs and improving performance.

A framework for efficient 3D asset editing that leverages self-constructed datasets and lightweight modules to enhance foundational generative models.

MAny merges task-specific knowledge in multimodal LLMs through cross-modal and low-rank parameter merging to prevent catastrophic forgetting without additional training.

An AI system that generates technically sound peer reviews, preferred by authors and reviewers over human reviews, for large-scale scientific conferences.

A fast, forward-only sensitivity analysis using KL divergence for mixed-precision SSM-Transformer models, enabling efficient LLM deployment on edge devices.

BenGER is an open-source web platform for end-to-end benchmarking of German legal LLM tasks, integrating task creation, annotation, execution, and evaluation.

CONCORD is a privacy-aware A2A framework enabling socially deployable proactive conversational agents through collaborative context recovery.

ReSS bridges symbolic and neural reasoning for tabular data by using decision trees to scaffold LLMs, generating faithful and consistent natural-language explanations for high-stakes domains.

Diffusion language models enhance speech recognition accuracy by integrating acoustic and language information for improved hypothesis rescoring and joint decoding.

A unified generative framework using flow matching to perform text-driven motion editing and structural retargeting with a single model.

A bilingual TTS system synthesizes speech for the Peruvian Constitution in Quechua and Spanish.

Confidence-based voting (C-voting) enhances test-time scaling for recurrent models by selecting latent trajectories based on prediction confidence, outperforming existing methods without requiring explicit energy functions.

A code-free, LLM-driven pipeline for accessible eye-tracking event detection that converts natural language instructions into end-to-end analysis, reducing technical overhead.

A dynamic benchmark and agent architecture for evaluating and improving tool-augmented LLMs in complex spatial analysis tasks.

Introducing LongCoT, a benchmark for evaluating long-horizon chain-of-thought reasoning in LLMs, revealing a significant gap in current model capabilities and providing a rigorous measure for future progress.

A framework for unlearning sensitive data from LLMs using memory-graph guided synthesis of supervision, minimizing user input and corpus reliance.

A hybrid LLM-GNN system for advanced question answering over incomplete knowledge graphs.

HINTBench: A benchmark and evaluation framework for intrinsic risk in AI agents, revealing significant capability gaps in existing models.

TREX automates and optimizes the lifecycle of LLM fine-tuning using agents and a tree-based exploration approach.

A recomputation-free KV caching framework for LLMs that uses trainable adapters to significantly reduce inference latency and computational overhead.

SafeHarness is a lifecycle-integrated security architecture for LLM agents that significantly reduces unsafe behavior and attack success rates.

TokenFormer unifies multi-field and sequential recommendation models, overcoming sequential collapse propagation with a novel attention scheme and representation method.

A benchmark for evaluating AI agents' ability to retrieve multimodal evidence and reason over noisy web environments.

A hierarchical robotic manipulation system that decouples planning from execution to preserve VLM reasoning while enabling independent component improvement.

Peer-Predictive Self-Training (PST) enables collaborative self-improvement of language models for reasoning without external supervision.

Quantifies numerical instability in LLMs, identifying a chaotic 'avalanche effect' and three distinct regimes of unpredictability based on floating-point precision.

LLM-generated annotations achieve comparable F1-Macro to human annotations for hostility detection at a fraction of the cost, with nuanced error profiles.

A realistic benchmark for evaluating GUI agents in e-commerce risk management, revealing a significant capability gap in current models and demonstrating agentic RL improvements.

A reinforcement learning framework that aligns sentiment analysis with human-like reasoning, improving interpretability and prediction accuracy.

A large-scale dataset and benchmark for distractor-free novel view synthesis, enabling robust radiance field development and improving image enhancement.

This research provides a mechanistic analysis of sim-and-real co-training for generative robot policies, identifying key effects and proposing a method to improve performance.

This research systematically investigates reward design for improving physical reasoning in vision-language models, demonstrating accuracy gains through targeted reward signals.

VLAJS jump-starts reinforcement learning for robotics by using vision-language-action models to bias exploration and improve learning efficiency, outperforming baselines by over 50%.

A framework that generates latent pseudo-anomalies for unsupervised time-series anomaly detection, outperforming state-of-the-art with LLM enrichment.

Analyzing how LLMs internally represent rhetorical questions using linear probing, revealing that these signals emerge early and are encoded by multiple, context-dependent directions.

A benchmark and framework for evaluating multilingual Text-to-SQL capabilities in Indian languages, addressing the 'Indic Gap' in LLM performance.

InfiniteScienceGym is a procedurally generated benchmark for evaluating LLMs' scientific reasoning capabilities, offering a controlled and unbounded environment without large static datasets.

An interpretable machine learning model that predicts IT incident risk for regulated environments, outperforming rule-based systems.

STOMP is a novel offline RL algorithm that extends direct preference optimization to the multi-objective setting, enabling principled alignment of models for multi-attribute tasks like protein engineering.

Optimizing data overlap in SFT-GRPO post-training for LLMs significantly improves autoformalization accuracy without additional compute.

A novel methodology quantifies uncertainty in Large Reasoning Models with statistical guarantees and provides interpretable explanations by identifying key training examples and reasoning steps.

A 3D SAM-based framework for multi-task segmentation of radiotherapy-induced normal tissue injuries in limited-data settings, outperforming state-of-the-art.

A multimodal extension for robotic manipulation data collection that integrates LiDAR for robust 3D spatial perception.

Syn-TurnTurk is a synthetic dataset for Turkish dialogue turn-taking prediction, enabling more natural human-machine interaction in Turkish chatbots.

A new method for medical image segmentation that uses uncertainty estimates to guide decisions, significantly reducing errors with minimal deferral.

A safety-constrained hierarchical reinforcement learning framework for power grid operation that ensures runtime safety and robust generalization.

An explainable fall detection system for elderly care using temporally stable SHAP to provide reliable insights for clinicians.

A physics-guided hybrid CNN-BiLSTM model for solar irradiance forecasting that outperforms complex attention models, enabling efficient renewable energy management.

A hybrid CNN-BiLSTM-Attention model for industrial Remaining Useful Life prediction that provides interpretable failure heatmaps, improving safety and maintenance decisions.

An automated system for assessing teachers' geometric content knowledge using LLMs and a fine-grained skills dictionary.

A survey proposing a new taxonomy for feed-forward 3D scene modeling, focusing on model design strategies agnostic to output format.

A diffusion model with adaptive constraint guidance for generating thermodynamically plausible inorganic crystal structures with targeted properties.

A novel algorithm-system co-design framework for load-balanced long context LLM training that improves accuracy and efficiency.

A multitask embedding approach using a pretrained ResNet-18 to automate blastocyst grading for IVF, improving consistency and reducing subjectivity.

A parallel monitoring architecture for LLM agents that detects and recovers from reasoning degradation, offering both LLM-based and zero-overhead probe-based companions.

A study analyzing failure modes in two-stage human-object interaction detection models to provide insights for future research.

Develop a Memory Transfer Learning tool to improve coding agent adaptability across diverse tasks.

LAMO framework enables lightweight LLMs to perform complex GUI automation through multi-role orchestration, balancing cost and scalability.

A novel defense mechanism for neural operators enhances robustness against adversarial attacks.

Introduces Predictive Representation Learning (PRL) as a new paradigm for self-supervised learning, demonstrating its potential through comparative analysis of BYOL, MAE, and I-JEPA.

A theoretical framework and method for identifying latent variables from degenerate Gaussian mixture models transformed by piecewise affine functions.

A training-free framework that enhances unified multimodal model generation by leveraging their inherent understanding for reflective rectification, inspired by human 'Thinking-While-Drawing'.

A bio-inspired, sensor-first architecture for physical AI that improves end-to-end accuracy and reduces remote inference calls.

Fine-tune large audio-language models for precise temporal event detection using audio-side time prompts and reinforcement learning.

A comprehensive exploration of Hybrid Quantum Neural Networks for Chronic Kidney Disease diagnosis, benchmarking 625 models to find optimal design choices.

Med-CAM generates minimal, sharp activation maps for interpretable medical AI decisions, improving clinician trust and understanding.

Leveraging zero-shot learning to perform sentiment analysis in software engineering, reducing the need for extensive annotated datasets.

Adaptive model selection framework for clinical text classification that intelligently defers to LLMs for improved accuracy and cost-efficiency.

An AI system that transforms standard kitchen photos into MCI-friendly designs, offering a low-cost, scalable solution for older adults and caregivers to visualize and implement DIY kitchen modifications.

Reframing multimodal anomaly detection as a cross-modal contextual inference problem to improve reliability in dynamic environments.

A new framework for algorithmic decision-making that explicitly models multi-stakeholder utilities to achieve optimal performance-fairness trade-offs.

A theoretical framework and adaptive algorithm for minimax optimal majority-vote ensembles in Markov-dependent data, improving time-series and RL applications.

Processing radar data in the frequency domain for robust multi-object tracking on fast-moving vehicles, demonstrating radar-only odometry.

An integrated MPC-RL framework for automated driving that balances safety and efficiency in multi-agent scenarios, outperforming standalone methods and showing improved generalization.

Investigating the impact of batch composition and data scaling on CLIP-like architectures for abdominal CT image-text alignment and zero-shot learning.

A framework to detect LLM hallucinations by analyzing internal model states, reducing latency and computational overhead for mission-critical applications.

A framework for robustly evaluating supervised machine learning models by addressing common pitfalls in metric selection and validation.

Developing a method to formalize user-centric evaluations of LLMs by translating subjective 'vibe-testing' into quantitative metrics.

Proposes a Target Decoupling architecture for multi-timescale PPO that overcomes surrogate hacking and myopic degeneration by isolating short-term signals for policy updates.

A workshop overview and challenge report for maritime computer vision, focusing on predictive accuracy and real-time feasibility with benchmark challenges and datasets.

Fairness in AI agents emerges from multi-agent collaboration, not single-model optimization, offering a new perspective on ethical AI.

A bi-objective evolutionary algorithm with a diversity-based change response mechanism optimizes open-pit mine scheduling under dynamic economic values and resource capacities.

An analysis of the creativity of AI agents, exploring functionalist and ontological perspectives and discussing the desirability and risks of artificial creativity.

Creo: A multi-stage text-to-image system that allows progressive, co-creative ideation with user control and decision locking.

Fragata is a semantic search system for HPC support tickets that uses hybrid RAG to improve knowledge reuse and overcome limitations of traditional search engines.

Systematically explore the impact of multilingual data on LLM performance across different scales and tasks, revealing benefits for low-resource languages and overall cross-lingual generalization.

Developing AlphaCNOT, a model-based reinforcement learning framework for minimizing CNOT gates in quantum circuits.

A Digital Twin architecture for hybrid autonomous-teleoperated medical response systems, enhancing situational awareness and decision-making.

An adaptive conformal prediction approach is proposed to improve the factuality of large language model generations by providing prompt-dependent uncertainty estimates.

A latent diffusion model for monthly climate emulation using a spherical Fourier neural operator-inspired architecture.

Investigating the impact of AI on cognitive offloading in Agile sprint planning to propose a hybrid AI-human framework.

A safe, lightweight, and user-friendly agentic framework for the autonomous execution of well-defined scientific tasks, enabling researchers to offload routine workloads.

A Bayesian mitigation strategy for reinforcement learning agents to prevent unintended high-reward strategies by incorporating a large negative penalty and a mentor override.

An active constraint acquisition approach for Earth Observation satellite scheduling that learns operational constraints interactively from an oracle.

A literature review on using LLMs to automate business process modeling, highlighting current trends and future research directions.

An experimental evaluation of window shapes and lengths for feature extraction in classifying heart sound signals using bidirectional LSTMs.

A privacy-preserving framework for vertically split federated learning using secure multiparty computation and differential privacy.

This research explores the interpretability of Vision Transformers through Cross-Layer Transcoders.

A novel parameter-space intervention method for localizing specific behaviors within Large Language Models.

This paper theoretically demonstrates that Ordinary Least Squares is a special case of the Transformer architecture, revealing a decoupled slow and fast memory mechanism.

A novel three-layer cognitive architecture for autonomous agents that decomposes intelligence across heterogeneous hardware to reduce latency and energy consumption.

A theoretical framework for multi-step off-policy soft Q-learning using eligibility traces for improved credit assignment in reinforcement learning.

Young people's perceptions and co-designed recommendations for conversational generative AI in youth mental health, focusing on humanizing care, transparency, and personalized integration.

A theoretical characterization of forgetting in continual learning by analyzing task distributions rather than random orderings.

A dynamic-growing fuzzy-neuro controller with an adaptive strategy for precise and stable control of parallel robots.