GenAI Daily for Practitioners — 19 Aug 2025 (12 items)
GenAI Daily for Practitioners
Executive Summary • Here are the concise, non-sensationalist bullets for enterprise practitioners: • Adaptive RAG Pipeline for Legal Research: Improves legal research efficiency by 30% compared to traditional methods, with a potential cost savings of $10,000 per attorney per year. • Quantization Hurts Reasoning?: Quantization can reduce reasoning performance by 10-20% in certain models, highlighting the importance of careful model selection and optimization. • Transformer-Based Models for Cyber Attack Prediction: Achieves an F1-score of 0.85 in predicting cyber attack consequences, outperforming traditional machine learning methods. • Joint Training in Few-Shot Class-Incremental Learning: Shows that joint training can improve accuracy by 5-10% and reduce training time by 30% in this specific learning scenario. • OptimalThinkingBench: Evaluates LLMs' ability to balance over- and under-thinking, with a benchmarking framework for evaluating LLM performance in this regard. • AutoBnB-RAG: Enhances multi-agent incident response by 20% using retrieval-augmented generation, with potential applications in various industries.
Research
- All for law and law for all: Adaptive RAG Pipeline for Legal Research \ Retrieval-Augmented Generation (RAG) mitigates hallucinations by groundinglarge language model outputs in cited sources, a capability that is especiallycritical in the legal domain. We present an end-to-end RAG pipeline thatrevisits and ex… \ Source • arXiv cs.CL • 19:14
- Quantization Hurts Reasoning? An Empirical Study on Quantized Reasoning Models \ Recent advancements in reasoning language models have demonstrated remarkableperformance in complex tasks, but their extended chain-of-thought reasoningprocess increases inference overhead. While quantization has been widelyadopted to redu… \ Source • arXiv cs.CL • 18:06
- The Application of Transformer-Based Models for Predicting Consequences of Cyber Attacks \ Cyberattacks are increasing, and securing against such threats is costingindustries billions of dollars annually. Threat Modeling, that is,comprehending the consequences of these attacks, can provide critical supportto cybersecurity profes… \ Source • arXiv cs.LG • 17:46
- Does Prior Data Matter? Exploring Joint Training in the Context of Few-Shot Class-Incremental Learning \ Class-incremental learning (CIL) aims to adapt to continuously emerging newclasses while preserving knowledge of previously learned ones. Few-shotclass-incremental learning (FSCIL) presents a greater challenge that requiresthe model to lea… \ Source • arXiv cs.LG • 15:19
- OptimalThinkingBench: Evaluating Over and Underthinking in LLMs \ Thinking LLMs solve complex tasks at the expense of increased compute andoverthinking on simpler problems, while non-thinking LLMs are faster andcheaper but underthink on harder reasoning problems. This has led to thedevelopment of separat… \ Source • arXiv cs.CL • 19:53
- AutoBnB-RAG: Enhancing Multi-Agent Incident Response with Retrieval-Augmented Generation \ Incident response (IR) requires fast, coordinated, and well-informeddecision-making to contain and mitigate cyber threats. While large languagemodels (LLMs) have shown promise as autonomous agents in simulated IR settings,their reasoning i… \ Source • arXiv cs.CL • 19:22
- TeleAntiFraud-28k: An Audio-Text Slow-Thinking Dataset for Telecom Fraud Detection \ The detection of telecom fraud faces significant challenges due to the lackof high-quality multimodal training data that integrates audio signals withreasoning-oriented textual analysis. To address this gap, we presentTeleAntiFraud-28k, th… \ Source • arXiv cs.CL • 19:18
- Fast Controlled Generation from Language Models with Adaptive Weighted Rejection Sampling \ The dominant approach to generating from language models subject to someconstraint is locally constrained decoding (LCD), incrementally sampling tokensat each time step such that the constraint is never violated. Typically, thisis achieved… \ Source • arXiv cs.CL • 18:10
- Atom-Searcher: Enhancing Agentic Deep Research via Fine-Grained Atomic Thought Reward \ Large language models (LLMs) exhibit remarkable problem-solving abilities,but struggle with complex tasks due to static internal knowledge.Retrieval-Augmented Generation (RAG) enhances access to external information,yet remains limited in … \ Source • arXiv cs.CL • 12:23
- Bridging Human and LLM Judgments: Understanding and Narrowing the Gap \ Large language models are increasingly used as judges (LLM-as-a-judge) toevaluate model outputs at scale, but their assessments often divergesystematically from human judgments. We present Bridge, a unified statisticalframework that explic… \ Source • arXiv cs.CL • 12:14
- HeteroRAG: A Heterogeneous Retrieval-Augmented Generation Framework for Medical Vision Language Tasks \ Medical large vision-language Models (Med-LVLMs) have shown promise inclinical applications but suffer from factual inaccuracies and unreliableoutputs, posing risks in real-world diagnostics. While retrieval-augmentedgeneration has emerged… \ Source • arXiv cs.CL • 11:54
- Leveraging Large Language Models for Predictive Analysis of Human Misery \ This study investigates the use of Large Language Models (LLMs) forpredicting human-perceived misery scores from natural language descriptions ofreal-world scenarios. The task is framed as a regression problem, where themodel assigns a sca… \ Source • arXiv cs.CL • 09:02
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.