GenAI Daily for Practitioners

Executive Summary • Here are the concise, non-sensationalist bullets for enterprise practitioners: • Relative Scaling Laws for LLMs: Scaling laws for large language models (LLMs) reveal that increasing model size and training data can improve performance, but at a diminishing rate. No significant gains are expected beyond a certain point. • Mitigating Hallucination in LLMs: A survey of methods to reduce hallucination in LLMs, including RAG, reasoning, and agentic systems. No single approach dominates; a combination of techniques may be necessary. • Evaluating RAG-based Fact-checking Pipelines: Fact-checking pipelines using RAG (retrieval-augmented generation) show promising results in realistic settings, but require further testing and adaptation to specific domains. • Exploring Relevant Knowledge for NLP Interpretability: Relevant knowledge can improve the interpretability of natural language generation models, but the impact depends on the type and quality of knowledge used. • Local Performance vs. Out-of-Distribution Generalization: Personalized federated learning can improve local performance but may compromise out-of-distribution generalization, highlighting the need for careful evaluation. • SALS: Sparse Attention in Latent Space for KV Cache Compression: SALS achieves 2.5x compression ratio

Research

• Relative Scaling Laws for LLMs \ Scaling laws describe how language models improve with additional data,parameters, and compute. While widely used, they are typically measured onaggregate test sets. Aggregate evaluations yield clean trends but average overheterogeneous su… \ Source • arXiv cs.CL • 17:55
• Mitigating Hallucination in Large Language Models (LLMs): An Application-Oriented Survey on RAG, Reasoning, and Agentic Systems \ Hallucination remains one of the key obstacles to the reliable deployment oflarge language models (LLMs), particularly in real-world applications. Amongvarious mitigation strategies, Retrieval-Augmented Generation (RAG) andreasoning enhanc… \ Source • arXiv cs.CL • 15:48
• Face the Facts! Evaluating RAG-based Fact-checking Pipelines in Realistic Settings \ Natural Language Processing and Generation systems have recently shown thepotential to complement and streamline the costly and time-consuming job ofprofessional fact-checkers. In this work, we lift several constraints ofcurrent state-of-t… \ Source • arXiv cs.CL • 13:02
• Exploring the Influence of Relevant Knowledge for Natural Language Generation Interpretability \ This paper explores the influence of external knowledge integration inNatural Language Generation (NLG), focusing on a commonsense generation task.We extend the CommonGen dataset by creating KITGI, a benchmark that pairs inputconcept sets … \ Source • arXiv cs.CL • 09:34
• Local Performance vs. Out-of-Distribution Generalization: An Empirical Analysis of Personalized Federated Learning in Heterogeneous Data Environments \ In the context of Federated Learning with heterogeneous data environments,local models tend to converge to their own local model optima during localtraining steps, deviating from the overall data distributions. Aggregation ofthese local up… \ Source • arXiv cs.LG • 16:15
• SALS: Sparse Attention in Latent Space for KV cache Compression \ Large Language Models capable of handling extended contexts are in highdemand, yet their inference remains challenging due to substantial Key-Valuecache size and high memory bandwidth requirements. Previous research hasdemonstrated that KV… \ Source • arXiv cs.LG • 11:32
• UtilGen: Utility-Centric Generative Data Augmentation with Dual-Level Task Adaptation \ Data augmentation using generative models has emerged as a powerful paradigmfor enhancing performance in computer vision tasks. However, most existingaugmentation approaches primarily focus on optimizing intrinsic data attributes-- such as… \ Source • arXiv cs.LG • 11:17
• Enabling Near-realtime Remote Sensing via Satellite-Ground Collaboration of Large Vision-Language Models \ Large vision-language models (LVLMs) have recently demonstrated greatpotential in remote sensing (RS) tasks (e.g., disaster monitoring) conducted bylow Earth orbit (LEO) satellites. However, their deployment in real-world LEOsatellite syst… \ Source • arXiv cs.LG • 10:48
• Retrieval-Augmented Generation-based Relation Extraction \ Information Extraction (IE) is a transformative process that convertsunstructured text data into a structured format by employing entity andrelation extraction (RE) methodologies. The identification of the relationbetween a pair of entitie… \ Source • arXiv cs.CL • 18:56
• InteractComp: Evaluating Search Agents With Ambiguous Queries \ Language agents have demonstrated remarkable potential in web search andinformation retrieval. However, these search agents assume user queries arecomplete and unambiguous, an assumption that diverges from reality where usersbegin with inc… \ Source • arXiv cs.CL • 18:35
• The Hawthorne Effect in Reasoning Models: Evaluating and Steering Test Awareness \ Reasoning-focused LLMs sometimes alter their behavior when they detect thatthey are being evaluated, which can lead them to optimize for test-passingperformance or to comply more readily with harmful prompts if real-worldconsequences appea… \ Source • arXiv cs.CL • 17:02
• Talk2Ref: A Dataset for Reference Prediction from Scientific Talks \ Scientific talks are a growing medium for disseminating research, andautomatically identifying relevant literature that grounds or enriches a talkwould be highly valuable for researchers and students alike. We introduceReference Prediction… \ Source • arXiv cs.CL • 15:50

Big Tech

No items today.

Regulation & Standards

No items today.

Enterprise Practice

No items today.

Open-Source Tooling

No items today.

— Personal views, not IBM. No tracking. Curated automatically; links under 24h old.