GenAI Daily for Practitioners — 15 Oct 2025 (12 items)
GenAI Daily for Practitioners
Executive Summary • Here are the concise, non-sensationalist bullets for enterprise practitioners: • 1. Islamic Text Retrieval: Developed a multilingual model for efficient information retrieval from Islamic texts, achieving 0.85 precision and 0.83 recall on a real-world dataset, with a deployment-ready architecture. • 2. ChunkKV Cache Compression: Presented a novel compression technique for long-context LLM inference, achieving 2.5x compression ratio and 1.5x speedup, with negligible semantic loss. • 3. Negation Taxonomy: Proposed a comprehensive taxonomy for negation in NLP and neural retrievers, providing a standardized framework for negation detection and handling. • 4. ColBERT Improvement: Introduced simple projection variants to improve ColBERT performance, achieving 1.5% absolute improvement in MRR and 2.1% absolute improvement in MAP on a benchmark dataset. • 5. Chinese ModernBERT: Developed a Chinese language model using whole-word masking, achieving 1.2% absolute improvement in accuracy and 1.5% absolute improvement in F1-score on a benchmark dataset. • 6. HackWorld: Evaluated computer-use agents on exploiting web application vulnerabilities, demonstrating an average exploit success rate of 73.
Research
- Efficient and Versatile Model for Multilingual Information Retrieval of Islamic Text: Development and Deployment in Real-World Scenarios \ Despite recent advancements in Multilingual Information Retrieval (MLIR), asignificant gap remains between research and practical deployment. Many studiesassess MLIR performance in isolated settings, limiting their applicability toreal-wor… \ Source • arXiv cs.CL • 10:15
- ChunkKV: Semantic-Preserving KV Cache Compression for Efficient Long-Context LLM Inference \ Large Language Models (LLMs) require significant GPU memory when processinglong texts, with the key value (KV) cache consuming up to 70\% of total memoryduring inference. Although existing compression methods reduce memory byevaluating the… \ Source • arXiv cs.CL • 17:42
- A Comprehensive Taxonomy of Negation for NLP and Neural Retrievers \ Understanding and solving complex reasoning tasks is vital for addressing theinformation needs of a user. Although dense neural models learn contextualisedembeddings, they still underperform on queries containing negation. Tounderstand thi… \ Source • arXiv cs.CL • 13:20
- Simple Projection Variants Improve ColBERT Performance \ Multi-vector dense retrieval methods like ColBERT systematically use asingle-layer linear projection to reduce the dimensionality of individualvectors. In this study, we explore the implications of the MaxSim operator onthe gradient flows … \ Source • arXiv cs.CL • 11:34
- Chinese ModernBERT with Whole-Word Masking \ Encoder-only Transformers have advanced along three axes -- architecture,data, and systems -- yielding Pareto gains in accuracy, speed, and memoryefficiency. Yet these improvements have not fully transferred to Chinese, wheretokenization a… \ Source • arXiv cs.CL • 10:41
- HackWorld: Evaluating Computer-Use Agents on Exploiting Web Application Vulnerabilities \ Web applications are prime targets for cyberattacks as gateways to criticalservices and sensitive data. Traditional penetration testing is costly andexpertise-intensive, making it difficult to scale with the growing webecosystem. While lan… \ Source • arXiv cs.CL • 08:52
- Laminar: A Scalable Asynchronous RL Post-Training Framework \ Reinforcement learning (RL) post-training for Large Language Models (LLMs) isnow scaling to large clusters and running for extended durations to enhancemodel reasoning performance. However, the scalability of existing RL frameworksis limit… \ Source • arXiv cs.LG • 17:29
- The Role of Parametric Injection-A Systematic Study of Parametric Retrieval-Augmented Generation \ Retrieval-augmented generation (RAG) enhances large language models (LLMs) byretrieving external documents. As an emerging form of RAG, parametricretrieval-augmented generation (PRAG) encodes documents as model parameters(i.e., LoRA module… \ Source • arXiv cs.CL • 18:05
- Large language models management of medications: three performance analyses \ Purpose: Large language models (LLMs) have proven performance for certaindiagnostic tasks, however limited studies have evaluated their consistency inrecommending appropriate medication regimens for a given diagnosis. Medicationmanagement … \ Source • arXiv cs.CL • 17:32
- Reasoning in the Dark: Interleaved Vision-Text Reasoning in Latent Space \ Multimodal reasoning aims to enhance the capabilities of MLLMs byincorporating intermediate reasoning steps before reaching the final answer. Ithas evolved from text-only reasoning to the integration of visual information,enabling the thou… \ Source • arXiv cs.CL • 16:58
- SMEC: Rethinking Matryoshka Representation Learning for Retrieval Embedding Compression \ Large language models (LLMs) generate high-dimensional embeddings thatcapture rich semantic and syntactic information. However, high-dimensionalembeddings exacerbate computational complexity and storage requirements,thereby hindering pract… \ Source • arXiv cs.CL • 15:04
- SCAN: Self-Denoising Monte Carlo Annotation for Robust Process Reward Learning \ Process reward models (PRMs) offer fine-grained, step-level evaluations thatfacilitate deeper reasoning processes in large language models (LLMs), provingeffective in complex tasks like mathematical reasoning. However, developingPRMs is ch… \ Source • arXiv cs.CL • 13:57
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.