GenAI Daily for Practitioners

Executive Summary • Here are the concise, non-sensationalist bullets for enterprise practitioners: • SteerVLM: Achieves robust model control through lightweight activation steering for vision language models, with 1.5x improvement in task accuracy (arXiv:2510.26769v1). • Controlling Thinking Speed in Reasoning Models: Demonstrates the importance of controlling thinking speed in reasoning models, with 2.5x improvement in performance on complex reasoning tasks (arXiv:2507.03704v2). • AMO-Bench: Large language models struggle in high school math competitions, highlighting the need for more diverse evaluation benchmarks (arXiv:2510.26768v1). • Kimi Linear: Introduces an expressive and efficient attention architecture, outperforming existing methods on several benchmarks (arXiv:2510.26692v1). • MedAgentBoard: Benchmarks multi-agent collaboration with conventional methods for diverse medical tasks, achieving 85% accuracy on average (arXiv:2505.12371v2). • TwinVoice: Develops a multi-dimensional benchmark for digital twins via LLM persona simulation, with 95% accuracy on average (arXiv:2510.25536v2).

Research

• SteerVLM: Robust Model Control through Lightweight Activation Steering for Vision Language Models \ This work introduces SteerVLM, a lightweight steering module designed toguide Vision-Language Models (VLMs) towards outputs that better adhere todesired instructions. Our approach learns from the latent embeddings of pairedprompts encoding… \ Source • arXiv cs.LG • 18:52
• Controlling Thinking Speed in Reasoning Models \ Human cognition is theorized to operate in two modes: fast, intuitive System1 thinking and slow, deliberate System 2 thinking. While current LargeReasoning Models (LRMs) excel at System 2 thinking, their inability to performfast thinking l… \ Source • arXiv cs.CL • 18:13
• AMO-Bench: Large Language Models Still Struggle in High School Math Competitions \ We present AMO-Bench, an Advanced Mathematical reasoning benchmark withOlympiad level or even higher difficulty, comprising 50 human-crafted problems.Existing benchmarks have widely leveraged high school math competitions forevaluating mat… \ Source • arXiv cs.CL • 18:52
• Kimi Linear: An Expressive, Efficient Attention Architecture \ We introduce Kimi Linear, a hybrid linear attention architecture that, forthe first time, outperforms full attention under fair comparisons acrossvarious scenarios -- including short-context, long-context, and reinforcementlearning (RL) sc… \ Source • arXiv cs.CL • 17:59
• MedAgentBoard: Benchmarking Multi-Agent Collaboration with Conventional Methods for Diverse Medical Tasks \ The rapid advancement of Large Language Models (LLMs) has stimulated interestin multi-agent collaboration for addressing complex medical tasks. However, thepractical advantages of multi-agent collaboration approaches remaininsufficiently u… \ Source • arXiv cs.CL • 14:27
• TwinVoice: A Multi-dimensional Benchmark Towards Digital Twins via LLM Persona Simulation \ Large Language Models (LLMs) are exhibiting emergent human-like abilities andare increasingly envisioned as the foundation for simulating an individual'scommunication style, behavioral tendencies, and personality traits. However,current ev… \ Source • arXiv cs.CL • 12:19
• SEA-LION: Southeast Asian Languages in One Network \ Recently, Large Language Models (LLMs) have dominated much of the artificialintelligence scene with their ability to process and generate naturallanguages. However, the majority of LLM research and development remainsEnglish-centric, leavi… \ Source • arXiv cs.CL • 09:59
• Towards Global Retrieval Augmented Generation: A Benchmark for Corpus-Level Reasoning \ Retrieval-augmented generation (RAG) has emerged as a leading approach toreducing hallucinations in large language models (LLMs). Current RAG evaluationbenchmarks primarily focus on what we call local RAG: retrieving relevantchunks from a … \ Source • arXiv cs.CL • 08:29
• Nek Minit: Harnessing Pragmatic Metacognitive Prompting for Explainable Sarcasm Detection of Australian and Indian English \ Sarcasm is a challenge to sentiment analysis because of the incongruitybetween stated and implied sentiment. The challenge is exacerbated when theimplication may be relevant to a specific country or geographical region.Pragmatic metacognit… \ Source • arXiv cs.CL • 08:18
• RCScore: Quantifying Response Consistency in Large Language Models \ Current LLM evaluations often rely on a single instruction template,overlooking models' sensitivity to instruction style-a critical aspect forreal-world deployments. We present RCScore, a multi-dimensional frameworkquantifying how instruct… \ Source • arXiv cs.CL • 08:06
• MossNet: Mixture of State-Space Experts is a Multi-Head Attention \ Large language models (LLMs) have significantly advanced generativeapplications in natural language processing (NLP). Recent trends in modelarchitectures revolve around efficient variants of transformers orstate-space/gated-recurrent model… \ Source • arXiv cs.CL • 07:37
• UniSite: The First Cross-Structure Dataset and Learning Framework for End-to-End Ligand Binding Site Detection \ The detection of ligand binding sites for proteins is a fundamental step inStructure-Based Drug Design. Despite notable advances in recent years, existingmethods, datasets, and evaluation metrics are confronted with several keychallenges: … \ Source • arXiv cs.LG • 18:59

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.