Exploring Training and Inference Scaling Laws in Generative Retrieval
This paper investigates the training and inference scaling laws in generative retrieval, focusing on how model size and data scale influence performance. It introduces novel evaluation measures and demonstrates that larger models yield substantial performance gains in retrieval tasks.
Statistical Proof of Execution (SPEX)
SPEX addresses the need for verifiable computing in automated decision-making, proposing a sampling-based protocol that significantly improves the efficiency of verifying ML/AI inferences with robustness against non-determinism.
Large Language Models Empowered Personalized Web Agents
This research formulates the task of developing personalized LLM-based web agents, constructing a benchmark, and proposing a framework to enhance action execution based on personalized user history.
AgentDropout: Dynamic Agent Elimination for Token-Efficient and High-Performance LLM-Based Multi-Agent Collaboration
AgentDropout enhances multi-agent collaboration by dynamically eliminating redundant agents in communication, achieving significant reductions in token consumption while improving task performance.
From Trust to Truth: Actionable Policies for the Use of AI in Fact-Checking in Germany and Ukraine
This policy paper explores the challenges and strategies for integrating AI into journalism and fact-checking, particularly focusing on disinformation for Germany and Ukraine, advocating for cohesive regulation.
Detection of Somali-written Fake News and Toxic Messages on the Social Media Using Transformer-based Language Models
The study addresses the challenges of fake news and toxic content in Somali social media by implementing a new transformer-based model for detection, contributing to low-resource language processing.
Statistical Proof of Execution (SPEX)
This paper introduces a novel, sampling-based protocol for verifiable computing, improving both speed and effectiveness in managing non-deterministic decision-making algorithms.
Superpixel Tokenization for Vision Transformers: Preserving Semantic Integrity in Visual Tokens
The authors propose a novel superpixel-based tokenization for Vision Transformers, improving robustness and accuracy by preserving semantic integrity across visual tasks.
Clarifying Misconceptions in COVID-19 Vaccine Sentiment and Stance Analysis and Their Implications for Vaccine Hesitancy Mitigation: A Systematic Review
This systematic review examines how sentiment analysis and stance detection methods influence understanding and mitigation of COVID-19 vaccine hesitancy, advocating for improved NLP methodologies.
Dynamic Task Vector Grouping for Efficient Multi-Task Prompt Tuning
The work introduces a novel method for dynamic task grouping in multi-task prompt tuning, optimizing knowledge transfer across various NLP tasks while minimizing negative transfer.