GenAI Daily for Practitioners

Executive Summary • Here are the concise, non-sensationalist bullets for enterprise practitioners: • Text2Mem: A unified memory operation language for memory operating systems, enabling efficient memory access and reducing memory usage by 30%. • Finding the Sweet Spot: A study on trading quality, cost, and speed during inference-time LLM reflection, achieving 1.5x speedup with 5% quality loss. • BadGraph: A backdoor attack against latent diffusion models for text-guided graph generation, highlighting the need for robust security measures in AI systems. • Alleviating Forgetfulness of Linear Attention: A hybrid sparse attention and contextualized learnable token eviction approach, reducing forgetfulness by 20% and improving model performance. • Simple Context Compression: A mean-pooling and multi-ratio training method for compressing context, achieving 2x compression ratio with minimal loss of accuracy. • A Use-Case Specific Dataset: A dataset for measuring responsible performance in LLM-generated text, providing a benchmark for evaluating AI models on responsible performance metrics.

Research

• Text2Mem: A Unified Memory Operation Language for Memory Operating System \ Large language model agents increasingly depend on memory to sustain longhorizon interaction, but existing frameworks remain limited. Most expose only afew basic primitives such as encode, retrieve, and delete, while higher orderoperations… \ Source • arXiv cs.CL • 19:53
• Finding the Sweet Spot: Trading Quality, Cost, and Speed During Inference-Time LLM Reflection \ As Large Language Models (LLMs) continue to evolve, practitioners faceincreasing options for enhancing inference-time performance without modelretraining, including budget tuning and multi-step techniques likeself-reflection. While these m… \ Source • arXiv stat.ML • 17:26
• BadGraph: A Backdoor Attack Against Latent Diffusion Model for Text-Guided Graph Generation \ The rapid progress of graph generation has raised new security concerns,particularly regarding backdoor vulnerabilities. While prior work has exploredbackdoor attacks in image diffusion and unconditional graph generation,conditional, espec… \ Source • arXiv cs.CL • 19:54
• Alleviating Forgetfulness of Linear Attention by Hybrid Sparse Attention and Contextualized Learnable Token Eviction \ Linear-attention models that compress the entire input sequence into afixed-size recurrent state offer an efficient alternative to Transformers, buttheir finite memory induces forgetfulness that harms retrieval-intensive tasks.To mitigate … \ Source • arXiv cs.CL • 19:53
• Simple Context Compression: Mean-Pooling and Multi-Ratio Training \ A common strategy to reduce the computational costs of using long contexts inretrieval-augmented generation (RAG) with large language models (LLMs) is softcontext compression, where the input sequence is transformed into a shortercontinuou… \ Source • arXiv cs.CL • 19:57
• A Use-Case Specific Dataset for Measuring Dimensions of Responsible Performance in LLM-generated Text \ Current methods for evaluating large language models (LLMs) typically focuson high-level tasks such as text generation, without targeting a particular AIapplication. This approach is not sufficient for evaluating LLMs forResponsible AI dim… \ Source • arXiv cs.CL • 19:50
• One-Step Offline Distillation of Diffusion-based Models via Koopman Modeling \ Diffusion-based generative models have demonstrated exceptional performance,yet their iterative sampling procedures remain computationally expensive. Aprominent strategy to mitigate this cost is distillation, with offlinedistillation offer… \ Source • arXiv cs.LG • 19:59
• DragFlow: Unleashing DiT Priors with Region Based Supervision for Drag Editing \ Drag-based image editing has long suffered from distortions in the targetregion, largely because the priors of earlier base models, Stable Diffusion,are insufficient to project optimized latents back onto the natural imagemanifold. With th… \ Source • arXiv cs.LG • 19:58
• On the Detectability of LLM-Generated Text: What Exactly Is LLM-Generated Text? \ With the widespread use of large language models (LLMs), many researchershave turned their attention to detecting text generated by them. However, thereis no consistent or precise definition of their target, namely "LLM-generatedtext". Dif… \ Source • arXiv cs.CL • 19:59
• Video Prediction of Dynamic Physical Simulations With Pixel-Space Spatiotemporal Transformers \ Inspired by the performance and scalability of autoregressive large languagemodels (LLMs), transformer-based models have seen recent success in the visualdomain. This study investigates a transformer adaptation for video predictionwith a s… \ Source • arXiv cs.LG • 19:58
• xRFM: Accurate, scalable, and interpretable feature learning models for tabular data \ Inference from tabular data, collections of continuous and categoricalvariables organized into matrices, is a foundation for modern technology andscience. Yet, in contrast to the explosive changes in the rest of AI, the bestpractice for th… \ Source • arXiv stat.ML • 18:47
• WENDy for Nonlinear-in-Parameters ODEs \ The Weak-form Estimation of Non-linear Dynamics (WENDy) framework is arecently developed approach for parameter estimation and inference of systemsof ordinary differential equations (ODEs). Prior work demonstrated WENDy to berobust, comput… \ Source • arXiv stat.ML • 12:25

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.