v0.24.0rc2: Fix P/D with DP Supervisor (#46628)
This release candidate addresses a fix for Prefill/Decode (P/D) functionality in conjunction with the Data Parallelism (DP) Supervisor within the vLLM project.
AutoDex: An Automated Real-World System for Dexterous Grasping Data Collection
AutoDex is an automated system designed to close the loop of real-world dexterous grasping data collection by handling perception, execution, labeling, and reset without human intervention. It addresses the scalability issues of teleoperation and the lack of physical certification in simulation by generating candidate grasps and verifying them on real hardware.
Adaptive Hard-Soft Physics-Informed Neural Networks for Robust Boundary-Constrained PDE Solving
This study proposes a unified hard--soft physics--informed neural network (HSPINN) with adaptive loss weighting to address the slow convergence and inaccurate boundary enforcement of conventional PINNs. The framework enforces Dirichlet and periodic boundary conditions exactly through analytical lifting or masking, while treating PDE residuals and initial conditions as soft constraints balanced by an inverse-share softmax strategy.
Rethinking Molecular Graph Backdoors under Chemistry-aware Admission
The article introduces ChemGuard, an operational protocol that formalizes the overlooked admission stage of molecular learning pipelines by requiring sanitizable strings and consistent graph reconstruction. This framework reveals that many existing graph-based backdoors lose efficacy because their poisons are chemically invalid or representation-inconsistent.
Measuring & Mitigating Over-Alignment for LLMs in Multilingual Criminal Law Courts
This article addresses the challenge of over-alignment in large language models used within Swiss Federal Supreme Court criminal law contexts, where model guardrails frequently trigger refusals when processing sensitive case details. The authors introduce TF-RefusalBench, a multilingual benchmark derived from public rulings, to measure this phenomenon across French, German, Italian, and English.
Energy-Based Transformers as Predictors of Reading Difficulty
This study introduces energy-based transformers as a novel measure for predicting human reading difficulty, establishing a formal link between transformer models and associative memory literature like Hopfield networks.
Distribution-Aware Diffusion-LLM for Robust Ultra-Long-Term Time Series Forecasting
The authors propose Diffusion-LLM, a framework that integrates a conditional diffusion model into an LLM-based pipeline to address challenges in multimodal time series forecasting. This joint design enables the learning of future data distributions while improving semantic alignment within a shared latent space.
Fast medical RAG API to give your local LLMs access to facts
A developer has released a free, simple Retrieval-Augmented Generation (RAG) API powered by medical Wikipedia articles to provide local large language models with accurate factual information. The service aims for subsecond responses and currently runs on a single ARM VPS using approximately 2GB of RAM.
DGX Spark OS lifetime?
A user on Reddit asks whether Nvidia has disclosed the support lifecycle for the operating system running on DGX Spark hardware. The inquiry specifically concerns the duration of OS support and whether users will be forced to upgrade to new products in the near future, such as by 2028.
Automated Semantic Fault Localization in SysML v2 Using Knowledge-Graph Augmented LLMs
This paper presents a human-in-the-loop framework for automatically identifying and repairing semantic errors in SysML v2 models that compilers cannot detect. The approach combines fine-tuned Small Language Models with a domain knowledge graph to ground repair suggestions in valid engineering constraints.
Litmus: Zero-Label, Code-Driven Metric Specification for Evaluating AI Systems
Litmus is a zero-label system that designs evaluation and monitoring metrics for AI pipelines by eliciting evaluation intent from source code and targeted interrogation. Instead of assuming the evaluation target is known, it identifies what must be measured and why to construct a justified metric portfolio.
ReasoningLens: Hierarchical Visualization and Diagnostic Auditing for Large Reasoning Models
The emergence of Large Reasoning Models has introduced exceptionally long Chain-of-Thought traces, creating a transparency burden where critical logic is often buried under massive procedural text. To address this, the authors present ReasoningLens, an open-source framework designed for the hierarchical visualization and diagnostic auditing of complex reasoning chains.
HyperQuant: A Rate-Distortion-Optimal Quantization Pipeline for Large Language and Diffusion Models
HyperQuant is a unified post-training quantization pipeline designed for the weights and KV cache of large language and diffusion transformers, combining Hadamard transforms with optimal lattice quantization. The method outperforms recent schemes like HIGGS, TurboQuant, and OCTOPUS across various bit rates while maintaining near-lossless quality.
UnBias-Plus: Detect, Explain, and Rewrite Bias
UnBias-Plus is an open-source toolkit designed to address persistent bias in natural language by unifying detection, explanation, and neutral rewriting capabilities.
Detecting Malicious Agent Skills in the Wild using Attention
The authors present Locate-and-Judge, a two-stage detector designed to identify malicious skills in LLM agent marketplaces where traditional prompt-injection defenses fail.
Digital Humanism and Evolutionary Design
This paper examines the concepts of digital humanism and evolutionary design to identify their common structures, synergies, and challenges within the context of human-centered technological development.
GRINQH: Graded Input-based Quantization Hierarchy for Efficient LLM Generation
Researchers propose GRINQH, a weight-only post-training quantization framework that accelerates large language model decoding by unifying quantization and sparsification. The method leverages activation magnitudes to dynamically assign weight channels to different precision levels, addressing the memory-bound nature of the decoding stage.
STAITUS: Disentangling Appearance and Pose for Video Object Tracking
The article introduces STAITUS, a unified framework for unsupervised video object tracking that addresses the limitations of existing slot-based methods by explicitly disentangling appearance from geometric pose. This approach resolves conflicts between temporal consistency and object motion, preventing slots from locking onto static backgrounds.
Cross-Architectural Mixture-of-Experts with Adaptive Soft Routing for Plant Leaf Disease Classification
This study proposes an adaptive soft Mixture-of-Experts (MoE) framework that integrates EfficientNet-B0, DenseNet-121, and Swin-Tiny to address challenges in plant leaf disease classification under complex backgrounds and class imbalance.
What Does a Chemical Language Model Know About Molecules?
This study applies sparse autoencoders to MolFormer to mechanistically examine how molecular representations are built across layers, challenging the assumption that chemical language models only learn surface-level syntax.