The Archive

Direct-P algorithm optimizes FP4 attention on NVIDIA Blackwell via direct quantization path, achieving 2.13× BF16 throughput for causal inference.

NVFP4 W4A4 quantization on all 496 linear layers of Qwen3.8-27B hybrid LLM including Gated DeltaNet recurrent blocks maintains performance across evals.

AdaRoboVLG decouples foundation models from grasp policies via kinematic mapping and force-closure stability, enabling generalizable vision-language grasping across robotic hands.

DRACO distributes rubric-based rewards dynamically during training to enable fine-grained credit assignment across long-horizon agent tasks without ground-truth labels.

Causal optimal transport framework for conditioning degenerate diffusion models without smooth densities.

Metatheorem proves every finite syntactic system has theorems it cannot autonomously produce, with implications for AI security.

CORE: distill reranker compositional reasoning into MLLM embeddings via Rank-KL objective for scene understanding.

PatchBench audits AI agents' C/C++ vulnerability patching; finds 25% exhibit patch memorization or surface-level fixes.

TAP-Path compresses Virchow2 pathology encoder via task-adaptive block selection and token pruning for inference efficiency.

PreferenceEKF frames active RLHF reward learning as Bayesian filtering to reduce human preference query samples.

Group Relative Policy Optimization exhibits spurious advantage when guesses match correct answers in bounded-answer tasks.

LLMs over-edit code during repair; study of 400 BigCodeBench problems shows widespread over-editing even in frontier models.

Reframe low-resource dialect neural translation as multi-agent decision space to explore linguistically valid alternatives.

Dice Roll Method: standardized protocol for repeated-query auditing of LLM brand recommendations with variance decomposition.

Theoretical analysis establishes exact hierarchy of attention head complexity: k heads compute k-bit parity but cannot compute (k+1)-bit parity.

Editable Visual Design combines VLM creative reasoning with Coding Agent precision to generate layered, editable visual designs beyond flattened diffusion outputs.

The family-focused AI assistant wants access to the details of your everyday life, but says it won’t use that data to train AI models or share it with others.

IRWOZ 2.0 dataset: 390 LLM-enhanced dialogue annotations (Mistral, Claude-3.5) for industrial robot conversations across 4 domains with improved quality.

Deep learning + FEM approach predicts filament geometry in 3D concrete printing to improve structural stability assessment.

Instruction duplication: repeating procedural instructions at inference time improves deterministic compliance across 7 instruction-tuned models without retraining.

Representational alignment via prototype theory improves LLM safety robustness: aligning moral concept representations strengthens resistance to adversarial reframing across 23 models.

FLY-EVAL++: safety-focused evaluation protocol for LLMs in physics-governed domains; measures constraint violations and physical inconsistency beyond accuracy metrics.

AI agents are learning to do more by working together. A lead agent can break a complex task into smaller jobs and assign those jobs to specialized subagents.... AI agents are learning to do more by working together. A lead agent can break a complex task into smaller jobs and assign those jobs to specialized subagents. Additionally, users are starting to run multiple agent sessions at the same time. Multi-agent workflows for accomplishing complex tasks are also becoming more common. This breadth-first approach can improve the speed of task completion… Source

Nvidia is announcing its new Personal AI Router (PAIR), a free tool that syncs up your home computers for tackling local AI inference tasks with tools like Ollama and LM Studio. Let's get the obvious thing out of the way, despite what its name might imply: PAIR is not a hardware router. It's open-source software developed by Nvidia that discovers compatible PCs on a network, connects them, and prepares them for crunching numbers on agentic workflows. While the compatible devices are mostly Nvidia GeForce GPUs (PAIR works with RTX 20-series cards and newer, as well as RTX Pro GPUs and DGX Spar...

Google is rolling out AI-powered voice assistant modes for Gmail, Docs, and Keep that allow you to manage the apps by talking to them. The real time conversational capabilities are called Gmail Live, Docs Live, and Keep Live, and like the Gemini Live experience for Google's chatbot, aim to make it easier to note down or find specific information when you're on the move or unable to do so manually. Gmail Live is primarily designed to surface details from your inbox without you having to dig through lengthy email threads using keywords or subject lines. For example, Google says that you can ask...

Statistical method for location-invariant estimation of extreme quantile treatment effects in heavy-tailed distributions.

InSituMeasure benchmark evaluates MLLMs on continuous-valued measurement tasks in realistic industrial settings with gauge reading and instrument-specific context.

Zero-shot/few-shot LLM framework for chronic kidney disease screening using structured prompts on tabular clinical features; compares with traditional ML/DL baselines.

Hybrid differentiable models combining neural networks with physics for chemical transport, addressing initial condition uncertainty and bias in engineering simulations.

Diffusion-augmented LLMs enable parallel token generation from autoregressive distributions, decoupling NTP and diffusion weights for lossless speedup.