All articles
39316 articles
Terry Pratchett's 'Horizontal Wealth': Spend Money on What You Love
Terry Pratchett's 'horizontal wealth': spend new income not on yacht-style vertical consumption but on what you loved before getting rich—books, a library, 'undirected research.'
A Single AI Model Simulates the Motion of Everything: From Flowing Sand to Protein Folding
The WorldParticle system uses a single AI model to learn the laws of matter motion, simulating behaviors from sand grain sliding to protein folding, with potential to aid soft robotics and drug discovery.
AI Automatically Discovers and Formally Verifies Quantum Programs in Lean, Substantially Outperforming Baselines on Multiple Metrics
A new study lets LLM agents automatically synthesize and formally verify quantum programs in Lean, achieving 290× gains in quantum arithmetic and 11×, 8×, and 2.5× in fusion, photochemistry, and drug metabolism simulations.
FLEX: Teaching Robots to Move Beyond Literal Instructions and Learn to Reason
FLEX is an AI-assisted learning platform designed to help robots go beyond literal instructions and gain reasoning and intent-understanding abilities, potentially advancing more intuitive, interactive home and assistive robots. This article explains its approach, significance, and open questions.
FlowAlign: A Flow Matching-Driven Approach for Target-Space Alignment
FlowAlign is a flow-guided target-space alignment method that uses conditional flow matching as a training signal while preserving global structure and instance-level correspondence, excelling in single-cell multi-omics and image-text alignment tasks.
Reference-Free Evaluation: Ensemble Learning Boosts Trustworthiness of LLM Code Summaries
An arXiv paper proposes an ensemble method to estimate confidence in LLM-generated COBOL code summaries without human references, using multi-model generation, LLM judge pairwise comparison, and calibrated aggregation.
Can Agentic Systems Bridge the Radiology Gap? Testing FDA-Approved Tools
A study proposes a radiology agent framework that simulates report generation with FDA-approved device tools, finding that clinical indications alone recover less than one-third of findings; tool availability and orchestration are complementary constraints, providing a benchmark for comprehensive AI-assisted radiology reporting.
CPU-ARC: Reuse LLM Reasoning with Caching, Cutting Costs by Up to 291x
CPU-ARC turns LLM reasoning traces into CPU-executable caches, matching frontier models on math and science tasks at 16-291x lower cost, fully interpretable and robust.
Is Majority-Vote Pseudo-Labeling Unreliable? PC-Dr.DPO Makes TTRL More Robust
This paper analyzes the reliability of majority-vote pseudo-labels in TTRL, decomposes error into population-majority bias and finite-rollout variation, and proposes PC-Dr.DPO with prompt-conditional robust reweighting and a cache mechanism to improve stability, outperforming baselines on multiple benchmarks.
GRAC: A New Graph-Based Retrieval-Augmented Framework for Surgical Scene Captioning
The GRAC framework models surgical scenes as graphs and combines retrieval-augmented generation to improve captions. On EndoVis-18, it achieves 5.88 CIDEr-D and 52.6% triplet accuracy, significantly outperforming SGT and Transformer baselines.
Grasp-TAGS: Critic Distillation Enables Task-Aware Grasp Selection
Grasp-TAGS is a task-aware grasp selection framework that learns a grasp utility network via critic distillation, boosting large-object rearrangement success to 89.84% in loco-manipulation—9.97 points above the best single-grasp strategy.
SSF-Net: A Spatial-Spectral Fusion Network for Accurate PDE Operator Learning
SSF-Net proposes a dual-branch spatial-spectral fusion block that uses multi-directional Mamba to capture local interactions and Fourier transforms plus block-diagonal MLPs to extract global features, achieving lower error with fewer parameters in PDE operator learning.
AI Data Center Load Forecasting: How Probabilistic Calibration and Temporal Structure Affect Grid Decisions
Measured data from 448 GPUs shows quantile LightGBM's nominal 90% interval covers only 63.4% after regime shifts; online calibration raises it to 87.3%, and rebuilding temporal structure cuts reserve costs by 7.4%-10.6%.
Composable Causality: A Benchmark Toolkit for Time-Series Causal Discovery and Intervention Effects
DTCD is a composable benchmarking toolkit for time-series causal discovery and intervention effects, combining topology, mechanisms, noise, missingness, and intervention policies via configuration to enable controlled experiments and paired counterfactual evaluation.
Ur-Nammu: Founder of the Sumerian Renaissance
Ur-Nammu, founder of the Third Dynasty of Ur, drove the Sumerian Renaissance. This article, inspired by a Hacker News post, analyzes his systematic rebuilding in law, architecture, economy, and administrative technology, and explores its implications for modern institutional design and standardization.
Surrogate Quality Score: A New Metric for Model Selection that Balances Accuracy and Complexity
Introducing the Surrogate Quality Score (SQS) composite performance metric from an arXiv paper, which integrates prediction accuracy and model complexity via a weighted product, allowing users to adjust the accuracy-complexity trade-off through a parameter and providing a systematic, quantifiable decision framework for engineering model selection.
OPO: A PPO Improvement That Replaces Fixed Clipping with Ordering
OPO revisits PPO's fixed clipping by sorting importance ratios and assigning adaptive per-sample clipping radii to control extreme ratios, with strong results on MuJoCo and Atari and lightweight implementation changes.
The Value of Weak Supervisors: What the Student Inherits, Not How Accurate the Supervisor Is
New research on weak supervision shows a supervisor's value depends not on its accuracy but on what the student model inherits from its errors. This article explains student-relative value theory, identifiability limits, and practical implications.
Energy-Based Reasoners: Teaching Recurrent Models When to Stop Thinking
Energy-Based Reasoners (EBRs) reformulate recurrent reasoning as gradient descent on an energy function, enabling adaptive stopping and 345x inference compute savings with 99.9% accuracy on Sudoku-Extreme.
MMAS: Dual-Branch Temporal Fusion for Continuous Music Emotion Estimation
New continuous music emotion recognition research proposes the MMAS model, using variance decomposition to reveal the evaluation blind spot of pooled correlation coefficients, and employing dual-branch temporal fusion with cross-modal attention to significantly improve within-song emotion dynamics tracking on DEAM and PMEmo, while introducing conformal prediction intervals to quantify uncertainty.