All Articles
16750 articles total
In-context learning of closed form solution to simple linear regression task using transformer with linear self-attention
This paper demonstrates that transformers with linear self-attention can in-context learn the closed-form least squares estimate for simple linear regression using layer normalization.
RTL-Sequencer: Towards Scalable RTL Timing Prediction with the Sequence-based Paradigm
RTL-Sequencer is a new sequence-based paradigm for scalable RTL timing prediction that linearizes logic cones to improve early-stage timing optimization.
CAMMAR: Culture-Aware Matryoshka for Metaphorical Arabic Representations
CAMMAR is a representation learning framework designed to separate lexical, cultural, and metaphorical meanings in Arabic language models to prevent 'semantic smearing'.
Test-Time Noise Guided Adaptation for Realistic Autoregressive Video Generation
TANGO is introduced to prevent error accumulation in autoregressive video generation by using the diffusion model as a critic of its own noise distribution.
Agentic Synthesis against Counterexample-Supplemented Sketches
The authors propose a repository-native method for coding agents to learn policies through counterexample-supplemented sketches, reducing rework in software synthesis.
Conditional Reliability of Toxicity Signals for Multilingual and Code-Mixed Abuse Detection
ToxGate is a trust-fusion head proposed to improve multilingual and code-mixed abuse detection by treating external toxicity signals as conditional evidence.
EgoExoMoCap: Distributed Ego-Exo Human Motion Capture
EgoExoMoCap is a distributed framework for human motion capture using smart glasses, combining egocentric and exocentric signals for 3D motion reconstruction.
1-Bit LLM in the Browser
A discussion on implementing 1-Bit Large Language Models (LLMs) directly in the browser, focusing on extreme quantization for efficiency.
On the Structure of Address in Multi-Party Dialogue: From Discrete Labels to Continuous Levels
Research on treating addressee detection in multi-party dialogues as a continuous phenomenon rather than discrete labels to improve turn-taking prediction.
Toward a mechanistic understanding of inference in visual cortex and diffusion models
A model bridging neuroscience and ML by describing perceptual inference in the visual cortex as a minimal diffusion model based on sparse coding.
Efficient Difficulty-Aware Dynamic Routing for Diffusion-Based Real-World Image Super-Resolution
Introduction of a Difficulty-aware Dynamic Routing (DDR) strategy for image super-resolution to optimize efficiency by matching network capacity to image difficulty.
Map as a Prompt: Learning Multi-Modal Spatial-Signal Foundation Models for Cross-scenario Wireless Localization
SigMap, a multimodal foundation model for wireless localization using a 'map-as-prompt' framework to improve cross-scenario adaptation in 5G/6G.
Debiasing Text-to-Image Evaluation via Implicit Cultural Alignment Reward Modeling
A lightweight 4.2B parameter MLLM reward model designed to debias text-to-image evaluation by aligning with implicit cultural norms.
AuEmoChat: Authentic Emotion Understanding and Rendering for Conversational Speech Synthesis
AuEmoChat, a CSS framework using discrete emotion tokens and flow matching to create more authentic emotional conversational speech synthesis.
GeoChrono: Benchmarking and Rethinking Long-Term Temporal Understanding in Remote Sensing
GeoChrono and ChronoBench, an MLLM and benchmark for long-term temporal understanding in remote sensing of Earth's surface evolution.
Scaling Time Series Classification via XAI-Driven Data Reduction
drXAI, a methodology that uses XAI attribution methods for data reduction in time series classification to enable scaling of resource-intensive models.
AquaAugmentor: A Novel Feature Augmentation Algorithm for Water Potability Prediction
AquaAugmentor, a feature augmentation algorithm designed to improve water potability prediction in low-dimensional datasets.
Ask Twice, Look Twice: Prompt Echoing Resolves the Question-First Paradox in Vision-Language Models
Researchers propose 'question echoing' to resolve a paradox in VLMs where placing the question after the image improves performance, despite intuition suggesting the opposite.
Information-Directed Sampling for Causal Bandits
A new Bayesian framework for causal bandits is introduced to improve reward estimation in systems with non-manipulable variables.
MemoGuard: An Adaptive Runtime for Guarding Against Memory Traps in Communication-Limited Robot Navigation
MemoGuard is a lightweight adaptive runtime that prevents memory traps in communication-limited robot navigation by validating episodic memories before reuse.