The Archive

DeepWeb-Bench introduces harder evaluation for frontier LLMs on deep research requiring massive cross-source evidence and reasoning.

AiraXiv proposes AI-era publishing platform enabling human and AI authors with continuous feedback-driven iteration.

WikiVQABench benchmark combines Wikipedia images with Wikidata for knowledge-grounded visual question answering evaluation.

Interactive tool visualizing LLM token generation speeds from 5 to 800 tokens/second for practical latency understanding.

FROG enables learnable graph structure for relational deep learning on RDBs without fixed schema constraints.

Agent JIT compilation compiles task descriptions into executable code for web agents, reducing latency vs. sequential fetch-execute loops.

RLVR training exhibits rank-1 weight trajectory structure; minimal training captures performance gains via linear parameter evolution.

DelTA interprets RLVR updates as linear discriminators over token gradients, explaining token-level probability changes in reasoning model training.

LLM-based grammar adaptation for metamodel evolution in domain-specific languages; evaluated on Xtext DSLs.

Mem-π framework generates context-specific agent guidance on-demand via dedicated model rather than static retrieval-based memory.

ML framework for GNSS positioning error mitigation in urban environments using activation function-based weighted least squares.

HITL-D combines diffusion policies with human control for shared autonomous manipulation, conditioning on scene point clouds.

Framework for optimal simulator-experiment allocation when deploying pre-trained simulators; decomposes value error into calibration drift and parametric residual.

Rubric embeddings mitigate label bias in high-stakes prediction (hiring, admissions) by replacing black-box embeddings with interpretable representations.

Empirical study of AI-generated Python refactoring PRs from AIDev dataset; assesses maintainability, code quality, and security impact.

Survey of approximation theory for neural networks covering universal approximation, quantitative rates, depth/width efficiency over four decades.

Coming to your homescreen soon: your own app. | Photo: Allison Johnson / The Verge "There's an app for that" was the promise of the App Store from the very beginning. The app that will get your phone to do the thing you want it to? It's just a few taps away. The tagline wasn't strictly true - I'm still waiting for that one perfect grocery list app. Still, apps shaped the modern smartphone into what it is today. We spend all day, every day inside of apps - scrolling, listening, and tapping until we find what we want. But your next favorite app might just be one that you made yourself. If you w...

Study of Vision-Language-Action model robustness under sensor degradation in autonomous driving; Alpamayo R1 tested across 18K trials with noise, lighting, fog perturbations.

Benchmark distinguishing temporal vs. spatial glitch detection in VLMs for game quality assurance; finds temporal glitches substantially harder than frame-level anomalies.

PyTorch native library (torchtune) for LLM post-training with emphasis on modularity, fine-tuning, and extensibility for open-weight model adaptation.

Google's AI search evolution is accelerating at I/O 2026.

Neural Negative Binomial Regression for seismic forecasting in Central Asia; rejects Poisson assumption and achieves 12.5% lower CRPS than baseline.

Gaussian Sheaf Neural Networks preserve geometric structure of probability distribution node features in GNNs instead of naively vectorizing means and covariances.

A day after Elon Musk lost his lawsuit that threatened OpenAI's structure, leadership and finances, OpenAI is reportedly back to prepping for its IPO.

roto 2.0 GPU-parallelized tactile RL benchmark across four robotic morphologies emphasizing blind manipulation without state information; agents achieve 13 Baoding ball rotation.

Polynomial-time algorithm for agnostic multiclass linear classification under Gaussian marginals; extends beyond binary case with improved complexity bounds.

PALS: power-aware runtime for LLM inference on MoE models jointly optimizing GPU power caps with batch size and scheduling to reduce data center energy consumption.

Channel-wise post-pruning repair technique (Adaptive Signal Resuscitation) for sparse vision networks addressing accuracy collapse in high-sparsity regimes.

Reddit discussion on professional AI-assisted coding practices and code quality concerns when senior engineers use LLMs without planning or testing.

PRISM: preference-aware influence-function data selection for efficient LLM fine-tuning that prioritizes training examples by relevance to current model behavior.