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

Aligning early visual cortex representations in vision-language models to shield them from sycophantic manipulation.

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

A budget-aware token compression framework for ultra-high-resolution remote sensing imagery that achieves state-of-the-art performance by leveraging text-guided importance estimation and region-wise strategies.

An AI co-scientist framework for materials research that automates hypothesis validation through experimentation and debate.

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.

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

A GPU-friendly token merging and restoration technique that significantly accelerates Vision Transformers and image generation models with minimal performance loss.

Revolutionize academic conference peer reviews with AI-driven critique enhancement systems.

A fast, forward-only KL-divergence based sensitivity analysis for efficient mixed-precision quantization of SSM-Transformer models on edge devices.

ReSS bridges symbolic and neural reasoning for tabular data, using decision trees to scaffold LLMs for accurate and faithful natural-language explanations.

Diffusion language models integrated with CTC for improved speech recognition accuracy, with code and recipes available.

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

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

A framework for unlearning sensitive data from LLMs using memory graphs and synthesized supervision, without access to the original training data.

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

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

HINTBench: A benchmark for evaluating intrinsic risks in AI agents, revealing a significant capability gap in detecting and localizing latent failures.

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

An agent-driven system that automates the entire LLM training lifecycle, from research to execution, optimizing model performance.

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

MERRIN is a new benchmark for evaluating AI agents' ability to reason across multimodal, noisy web data, highlighting a critical gap in current search technologies.

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

A hierarchical robotic manipulation system that decouples planning from execution, preserving VLM reasoning while improving control for complex tasks.

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

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

UniRect-CoT is a training-free framework that enhances unified multimodal model generation by leveraging their inherent understanding to reflect and rectify intermediate results, inspired by human 'Thinking-While-Drawing'.

A reinforcement learning framework that aligns sentiment reasoning with human rationale, improving interpretability and performance in ABSA tasks.

A large-scale dataset and benchmark for distractor-free novel view synthesis, enabling robust radiance field development and a diffusion-based enhancement application.

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.

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

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

A unified generative framework for text-driven motion editing and structural retargeting, simplifying deployment and improving consistency.

A novel framework for unsupervised time-series anomaly detection that generates latent pseudo-anomalies to train a Transformer-based classifier.

A benchmark and evaluation framework for multilingual Text-to-SQL in Indian languages, revealing an 'Indic Gap' in current LLM performance.

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

A new method for optimizing LLM post-training by controlling data overlap between fine-tuning stages, significantly improving performance without extra compute.

A novel methodology quantifies uncertainty in Large Reasoning Models with statistical guarantees, using Shapley values for explainable subsets of training data.

A 3D segmentation framework for radiotherapy injuries using progressive prompting and a novel loss function, outperforming state-of-the-art on limited medical data.

A multimodal robotic manipulation system integrating LiDAR for robust 3D spatial perception, enhancing data collection and policy performance.

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

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

A physics-guided hybrid CNN-BiLSTM model that outperforms attention mechanisms for accurate solar irradiance forecasting.

A hybrid CNN-BiLSTM-Attention model with interpretable heatmaps for accurate and safe industrial Remaining Useful Life prediction.

Integrate LLM and GNN to enhance open-world question answering over knowledge graphs.

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

Introduces Predictive Representation Learning (PRL) as a new paradigm for self-supervised learning, demonstrating its potential through comparative analysis of existing methods.

Systematic reward ablation study for training Vision-Language Models on physical reasoning, revealing domain-specific behaviors and improving spatial relation accuracy.

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

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

Confidence-based voting (C-voting) enhances test-time scaling for recurrent models by selecting trajectories based on prediction confidence, outperforming existing methods on complex reasoning tasks.

A framework for efficient 3D asset editing that leverages a self-constructed dataset and lightweight modules to inject textual semantics while preserving local invariance.

MAny is a framework that merges task-specific knowledge for multimodal LLMs to prevent catastrophic forgetting in both perception and reasoning.

A survey proposing a new taxonomy for feed-forward 3D reconstruction models, focusing on design strategies rather than output representations.

Enhance coding agent capabilities by leveraging a shared memory pool for cross-domain learning.

Transform subjective user feedback into quantifiable metrics for LLMs through structured vibe-testing.

Proposes a Target Decoupling architecture for multi-timescale PPO that isolates short-term signals to prevent surrogate hacking and improve performance in delayed-reward tasks.

MambaTrack offers a real-time, adaptive RGB-Event object tracking system that overcomes limitations of static state transitions for robust cross-modal fusion.

A framework for lightweight GUI agents that uses multi-role orchestration to improve task scalability and performance on resource-constrained devices.

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.

Leveraging zero-shot learning for sentiment analysis in software engineering to overcome the challenge of scarce annotated datasets.

Develop fair AI agents by enabling emergent fairness properties through multi-agent collaboration and negotiation.

A theoretical framework and adaptive algorithm for minimax optimal majority-vote ensembles in Markov-dependent data, validated on diverse benchmarks.

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.

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

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.

Investigating the impact of batch composition and data scaling on CLIP-like vision-language models for zero-shot diagnosis of abdominal CT scans, finding that random sampling outperforms engineered balancing.

A framework to detect LLM hallucinations by analyzing internal model states during inference, reducing latency and computational overhead.

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

A theoretical framework for multi-stakeholder fairness in algorithmic decision-making that explicitly models utilities and welfare across different groups.

An evolutionary optimization approach addresses dynamic chance-constrained open-pit mine scheduling by maximizing profit and minimizing standard deviation.

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

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

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

A novel framework for privacy-preserving vertical federated learning using secure multiparty computation and differential privacy.

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

A novel method for localizing LLM behavior to specific internal components by patching weights between models with differing capabilities.

An analysis of failure modes in two-stage human-object interaction detection models to improve future research.

Adaptive conformal prediction for improving the factuality of large language model generations by enabling prompt-dependent calibration.

A latent diffusion model for simulating low-frequency atmospheric variability at monthly timesteps with modest computational requirements.

Investigating the impact of AI on team cognition in agile sprint planning to propose a hybrid AI-human framework.

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

A literature review on using Large Language Models to enhance Business Process Modeling, highlighting current trends, challenges, and future research directions.

Generate minimal and sharp explanation maps for medical imaging AI decisions to improve clinician trust and understanding.

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 theoretical framework for off-policy, entropy-regularised reinforcement learning using eligibility traces.

A dynamic-growing fuzzy neural controller with adaptive strategy for parallel robot position control.

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

Young people's perceptions and recommendations for conversational generative AI in youth mental health are explored through co-design workshops.

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