The Archive

Decision models outperform expert rules and LLMs for HVAC fault diagnosis with calibrated confidence and minimal retraining.

LoRA-adapted Whisper achieves robust Arabic ASR across dialects and code-switching with parameter-efficient fine-tuning.

Expected sample complexity framework for multi-armed bandits provides convergence guarantees in approximately correct in expectation setting.

KGATE provides unified training environment for knowledge graph embedding models supporting complete autoencoder architectures.

Multi-vector document indices used by vision-language retrievers enable page reconstruction via conditional image generation, raising privacy risks.

Hessian spectral clustering in deep networks arises from symmetry breaking in nearby reference configurations, not weight symmetries alone.

NeuralZip reuses statistical structures across model weights to improve lossless compression via shared Huffman codes and packed exponents.

RollVerify addresses GPU efficiency and accuracy trade-offs in long-tail RL rollouts for LLM reasoning by mitigating off-policy stale samples.

NeMo-Guardrails proxy constrains LLM-driven SOC incident response to prevent adversarial alerts from triggering unauthorized production commands.

Google launches Playground, an experimental gaming platform for creating and sharing custom games.

DoorDash, the leading food delivery app, processed 970 million orders in its second quarter this year and generated $4.5 billion in revenue. A 10-person startup called Bites is a blip in comparison: It has just around 300 restaurants signed up in the Bay Area, where it's operating as a pre-seed startup. But this summer, Bites caught DoorDash's attention. In August, a slew of restaurants in the Bay Area received a strange email from DoorDash, warning businesses that they may be listed without their consent on Bites. The form letter, copies of which were seen by The Verge, said that DoorDash ha...

OpenAI launches College Planner, flashcards, quizzes, and teen AI council within ChatGPT for Teens product.

CLARA learns joint Gaussian weather forecasts from station data alone using ~28K parameters, outperforming larger comparators on energy scores.

NL2Hull converts natural language to constrained ship hull designs via typed discrete decisions and NURBS geometry with constraint validation.

QCATS enables sparse Transformer inference in databases by exposing query context and relational metadata to neural execution planning.

Stackelberg game analysis shows when feedback from testing, red-teaming, and incident response sufficiently defends AI systems economically.

Scoping review of RLHF in human-robot collaboration maps 20 papers on safety, feedback quality, and bidirectional adaptation for Industry 4.0.

Outperformance inverse optimization learns objective weights for MILPs that induce solutions outperforming observed suboptimal actions.

Structural identifiability theorem for nonlinear dissipative ODE model of human learning via concept prerequisites, forgetting, and practice dynamics.

Bayesian neural networks learn which parameters should be stochastic via deep weight factorization, sparsifying stochasticity rather than capacity.

Layerwise probabilistic error analysis enables fast, robust mixed-precision post-training quantization robust to corrupted calibration data.

Painter-Thinker combines recursive latent reasoning with diffusion models to solve visual reasoning tasks like Sudoku and mazes on pixels.

LiveMACEBench evaluates frontier LLM agents' process-aware capabilities in live financial markets with persistent memory and multi-agent collaboration.

ML toolkit for seismic data interpretation and implicit geological modeling in subsurface characterization.

Multi-agent reinforcement learning scheduler for time-sensitive vehicular edge computing networks with heterogeneous latency requirements.

MUNITE latent-variable framework enables any-to-any multimodal generation from partial modal subsets using unified conditional flow model.

Global Average Precision (gSAP) metric for representation learning ranks all query-candidate pairs globally rather than per-query.

DeepTopoClustering derives hierarchical surface process taxonomy from 4D point clouds for topographic monitoring.

Caddie trains critic models to provide natural-language advice to LLM agents during task execution, optimized from outcome-level feedback.

Self-evolving agents trained via curriculum learning and RL with reference-anchored feedback to avoid consensus collapse and improve task generation quality.