The Archive

First-order theory connecting optimization and sampling for SDE-based reverse diffusion models.

Microsoft's new Surface laptops forego Copilot+ PC branding.

Black-box post-processing method to align generative AI output distributions with user-specified attributes.

Belief Self-Distillation extracts and manipulates user models from LLMs via compact learnable representations.

LoRA composition method prevents interference by controlling adapter factorizations through read-only updates.

Analysis of training collapse in direct feedback alignment with tanh units via common-mode decomposition.

Benchmark and optimizer for code documentation quality in agent-based software issue resolution.

Trust Guided Decision Transformer uses next-state prediction error to detect context drift in offline RL.

OC-GS refines camera angles and geometry for 3D reconstruction from sparse, irregularly-spaced turntable views.

Self-training method using strategic diversity sampling (GROOT) to generate varied problem-solving approaches instead of IID filtering for correctness.

Neural ratio estimation framework for calibrating rough Heston volatility models with uncertainty quantification, applied to exotic option pricing.

Qualitative study of how elementary teachers adapt classroom practices when implementing ToyTalk conversational AI tool in ELA curriculum.

DeepEdu-v1: Vietnamese-native open-weights tutoring agent addressing data sovereignty and curriculum alignment for education in developing regions.

The findings highlight how AI-generated and vibe-coded apps can spill and expose users' data when not configured or secured properly.

Weight Pair Encoding (WeightPE) technique compresses int8 network weights via hierarchical grammar patterns using Re-Pair compression and straight-through estimators.

Analysis of multi-agent LLM scaling using Steiner's task taxonomy; plurality voting and averaging converge to modal answer and item bias in large teams.

Theoretical generalization bounds for OPTQ quantization algorithm and stochastic variant on calibration data with regularization analysis.

Learned latent Bayesian tracking for online learning in non-stationary environments, replacing manual low-dimensional subspaces with learned representations.

Frontier AI models are finishing Alan Turing's World War II codebreaking work.

MexHat: video dataset for hate speech detection in Mexican Spanish, addressing multimodal non-English resource scarcity for context-dependent content moderation.

Meta's Muse delivers first consumer-accessible agentic AI with persistent Linux VMs, raising questions about user understanding of autonomous agent capabilities.

EAServe: disaggregated serving system optimizing Encode-Prefill-Decode pipeline for multimodal LLMs with resource-aware request flow regulation.

Conformal inference methods for controlling false alerts in AI-generated text detection with adaptive document screening.

BeatGraph foundation model for infant ECG using heartbeat-level representations instead of fixed patches.

Flow matching framework unifies neural representational dissimilarity metrics for comparing neural codes across domains.

Touchstone framework determines which network automation tasks are suitable for local SLM inference using intrinsic checks.

NEXT addresses spectral bias and numerical instability in physics-informed neural networks for stiff PDEs.

HySTAR fixes structural target drift in multi-agent RL credit assignment using anchored hypergraphs.

Stratechery newsletter roundup covering Meta vs. Amazon competition and GM's CarPlay integration; no AI-specific technical content.

Sinter-PiNDiff integrates physics constraints and neural networks for retrainable sintering process modeling.