2
Follow the Latent Roadmap: Navigating Revocable Decoding for Diffusion LLMs with Anchor Tokens
This paper introduces a decoding strategy for diffusion language models designed to reduce error propagation and improve generation quality.
2
Attention-Discounted Adaptive Sampler for Masked Diffusion Language Models
Researchers have proposed a new sampling method for masked diffusion language models to improve token generation stability by addressing prediction coupling.
2
Egocentric Bias in Vision-Language Models
A new diagnostic benchmark called FlipSet reveals that current vision-language models struggle with visual perspective-taking tasks.
2
Declarative by Design, Assistable Only by Convention: Benchmarking Multi-Agent Frameworks for AI-Assistability
Researchers have proposed a new metric called AI-assistability to evaluate how effectively multi-agent frameworks support AI-driven software development.
2
Real-Time Model Checking for Closed-Loop Robot Reactive Planning
This study introduces a real-time model checking algorithm that enables robots to perform multi-step reactive planning to avoid obstacles more effectively.
2
ReLope: KL-Regularized LoRA Probes for Multimodal LLM Routing
This paper proposes a regularization technique for probe routing to improve the performance of multimodal large language model systems when balancing cost and accuracy.
2
Mobility-Aware Cache Framework for Scalable LLM-Based Human Mobility Simulation
Researchers have developed a caching framework to improve the computational efficiency of using large language models for large-scale human mobility simulations.
2
RippleBench: Capturing Ripple Effects Using Existing Knowledge Repositories
RippleBench-Maker is an automated pipeline developed to identify and measure unintended side effects that occur when editing or unlearning information in language models.
2
Real-time fall detection based on vision for low-power edge platforms
Researchers have developed a physics-informed vision system for real-time fall detection optimized for low-power edge computing devices.
2
Hallo4D: Multi-Modal Hallucination Mitigation for Consistent Spatio-Temporal Generation
Researchers have developed a new method to reduce spatial and temporal hallucinations in 4D generative models by improving geometric consistency.
2
Learning-based Probabilistic Load Forecasting with Post-hoc and In-model Uncertainty
This study evaluates methods for managing input uncertainty in machine learning models used for probabilistic load forecasting in smart buildings.
2
Explainable-by-Design Audio Deepfake Detection via Wiener-Hopf Linear Prediction
A new explainable audio deepfake detection framework utilizing Wiener-Hopf linear prediction has been developed to improve interpretability in multimedia forensics.
2
Lost in Visual Translation: A VLM-Assisted Perceptual-Semantic Coherence Framework for EEG-to-Image Reconstruction
A new framework utilizes vision-language models to improve the evaluation of EEG-to-image reconstruction by focusing on semantic coherence rather than just visual fidelity.
2
FAIR GraphRAG: A Retrieval-Augmented Generation Approach for Semantic Data Analysis
Researchers propose a graph-based RAG framework designed to improve semantic data analysis while adhering to FAIR data principles.
2
CANDI: Contextual Alignment for Niche Domains Question Answering
The authors introduce a new benchmark, CANDI-QA, to assess the contextual grounding and domain-specific knowledge of LLMs in fields like finance and medicine.
2
Do You Remember? Toward Memory-Centric Multimodal AI
Researchers propose a new architecture called DoYouRemember that enables multimodal large language models to utilize reconstructive memory rather than relying on one-shot processing.
2
The Computational Basis of Confidence in Large Language Models
This study investigates the computational mechanisms underlying confidence signals in large language models to improve their reliability.
2
Scale-Aware Attention for Scarce Neural Data: An RG-Flow Transformer on Sleep-EDF EEG
The authors introduce an RG-Flow Transformer that incorporates scale-aware inductive biases to improve neural data analysis in sleep EEG recordings.
2
Graph-Constrained Policy Learning for Extreme Clinical Code Prediction
This research introduces a graph-constrained policy learning approach to improve the accuracy of clinical code prediction from discharge summaries.
2
Self-Evolving In-Context Learning for Direct Pilot-to-Beamformer Design in MU-MISO Systems
A new in-context learning framework improves pilot-based beamforming in multi-user communication systems by adapting to various channel models without retraining.
2
Learning to Discretize: Diffusion-Based Adaptive Mesh with Spectral Guidance
Researchers propose a diffusion-based method for learning adaptive meshes to optimize spatial and spectral resolution in neural partial differential equation surrogates.
2
Removable Defects: The Economics and Limits of Deliberate Deficiency
Researchers propose a theoretical framework for intentionally maintaining model deficiencies as a design choice that can be bypassed when necessary.
2
An Empirical Analysis of Continual Learning for Heterogeneous Medical Visual Question Answering
This paper provides an empirical analysis of continual learning techniques applied to medical visual question answering systems to prevent catastrophic forgetting.
2
Representation and Reference Selection in Training-Free Synthetic Image Attribution
Researchers evaluate factors influencing the effectiveness of training-free methods for identifying the source generators of AI-created images.