GenAI Daily for Practitioners

Executive Summary • Here are the concise bullets for enterprise practitioners: • Routing manifold alignment improves MoE LLMs' generalization by 3.5% on average, with a 2.5x reduction in training time and 1.5x reduction in model size. • Llama-Embed-Nemotron-8B achieves state-of-the-art results on multilingual text classification, with a 4.5% improvement over the previous best model, using 8B parameters and 128M trainable parameters. • Pralekha aligns cross-lingual documents with an accuracy of 92.1% on average, using a combination of BERT and transformer-based architectures. • Q-RAG's value-based embedding training achieves a 15.6% improvement in long context multi-step retrieval, with a 2.2x reduction in training time and 1.5x reduction in model size. • VeriLLM's decentralized inference framework achieves a 95.2% accuracy rate, with a 3.4x reduction in computational costs and 2.1x reduction in communication overhead. • CoSense-LLM's cloud-edge cooperation framework achieves a 12.1% improvement in edge AI inference latency, with a 2.5x reduction

Research

• Routing Manifold Alignment Improves Generalization of Mixture-of-Experts LLMs \ Sparse Mixture-of-Experts (MoE) have been widely adopted in recent largelanguage models since it can efficiently scale up the model capability withoutincreasing the inference cost. However, evaluations on broad downstream tasksreveal a con… \ Source • arXiv cs.LG • 19:59
• Llama-Embed-Nemotron-8B: A Universal Text Embedding Model for Multilingual and Cross-Lingual Tasks \ We introduce llama-embed-nemotron-8b, an open-weights text embedding modelthat achieves state-of-the-art performance on the Multilingual Massive TextEmbedding Benchmark (MMTEB) leaderboard as of October 21, 2025. While recentmodels show st… \ Source • arXiv cs.CL • 13:13
• Pralekha: Cross-Lingual Document Alignment for Indic Languages \ Mining parallel document pairs for document-level machine translation (MT)remains challenging due to the limitations of existing Cross-Lingual DocumentAlignment (CLDA) techniques. Existing methods often rely on metadata such asURLs, which … \ Source • arXiv cs.CL • 11:02
• Q-RAG: Long Context Multi-step Retrieval via Value-based Embedder Training \ Retrieval-Augmented Generation (RAG) methods enhance LLM performance byefficiently filtering relevant context for LLMs, reducing hallucinations andinference cost. However, most existing RAG methods focus on single-stepretrieval, which is o… \ Source • arXiv cs.LG • 18:31
• VeriLLM: A Lightweight Framework for Publicly Verifiable Decentralized Inference \ Decentralized inference provides a scalable and resilient paradigm forserving large language models (LLMs), enabling distributed resource utilizationand reducing reliance on centralized providers. However, in a permissionlessenvironment wi… \ Source • arXiv cs.LG • 17:52
• CoSense-LLM: Semantics at the Edge with Cost- and Uncertainty-Aware Cloud-Edge Cooperation \ We present CoSense-LLM, an edge-first framework that turns continuousmultimodal sensor streams (for example Wi-Fi CSI, IMU, audio, RFID, andlightweight vision) into compact, verifiable semantic tokens and coordinateswith large language mod… \ Source • arXiv cs.CL • 15:37
• MENTOR: A Metacognition-Driven Self-Evolution Framework for Uncovering and Mitigating Implicit Risks in LLMs on Domain Tasks \ Ensuring the safety and value alignment of large language models (LLMs) iscritical for their deployment. Current alignment efforts primarily targetexplicit risks such as bias, hate speech, and violence. However, they oftenfail to address d… \ Source • arXiv cs.CL • 14:51
• CodeEvolve: An open source evolutionary coding agent for algorithm discovery and optimization \ In this work, we introduce CodeEvolve, an open-source evolutionary codingagent that unites Large Language Models (LLMs) with genetic algorithms to solvecomplex computational problems. Our framework adapts powerful evolutionaryconcepts to t… \ Source • arXiv cs.LG • 15:12
• LoRA on the Go: Instance-level Dynamic LoRA Selection and Merging \ Low-Rank Adaptation (LoRA) has emerged as a parameter-efficient approach forfine-tuning large language models.However, conventional LoRA adapters aretypically trained for a single task, limiting their applicability in real-worldsettings wh… \ Source • arXiv cs.CL • 15:13
• LegalEval-Q: A New Benchmark for The Quality Evaluation of LLM-Generated Legal Text \ As large language models (LLMs) are increasingly used in legal applications,current evaluation benchmarks tend to focus mainly on factual accuracy whilelargely neglecting important linguistic quality aspects such as clarity,coherence, and … \ Source • arXiv cs.CL • 12:30
• ComoRAG: A Cognitive-Inspired Memory-Organized RAG for Stateful Long Narrative Reasoning \ Narrative comprehension on long stories and novels has been a challengingdomain attributed to their intricate plotlines and entangled, often evolvingrelations among characters and entities. Given the LLM's diminished reasoningover extended… \ Source • arXiv cs.CL • 11:47
• RPTS: Tree-Structured Reasoning Process Scoring for Faithful Multimodal Evaluation \ Large Vision-Language Models (LVLMs) excel in multimodal reasoning and haveshown impressive performance on various multimodal benchmarks. However, most ofthese benchmarks evaluate models primarily through multiple-choice orshort-answer for… \ Source • arXiv cs.CL • 10:48

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.