The Archive

CUTS-GPR enables high-dimensional Gaussian process regression with near-linear scaling via additive kernels on incomplete grids.

PropSplat reconstructs RF propagation fields using 3D anisotropic Gaussians without maps or dense measurement campaigns.

Proposes DR-ME, a semiparametrically efficient test for detecting distributional treatment effects beyond means in causal inference.

You can stop Chrome from taking up 4GB of storage for local AI, but that shouldn't be your problem.

Test-time domain adaptation framework for PET image reconstruction generalizing from phantom to clinical data.

STARFlow2 unifies text-image generation via autoregressive normalizing flows, aligning causal masking with LLM architecture.

ADD-PINN applies domain decomposition to physics-informed neural networks for traffic flow estimation with sparse sensors.

Bash is one of the most flexible and powerful interfaces exposed to AI agents. In the right system, a model that emits grep, curl, tar, or a shell pipeline is... Bash is one of the most flexible and powerful interfaces exposed to AI agents. In the right system, a model that emits , , , or a shell pipeline is producing an executable action that can read files, mutate a workspace, open network connections, and chain tools together. For the NVIDIA AI Red Team, this makes command generation a useful research target. If smaller language models can be guided… Source

MPD²-Router implements learning-to-defer with multi-expert routing and mask awareness for glaucoma screening triage.

Reddit user claims Claude Opus 4.7 performs so well at coding tasks it displaces human software engineers.

Extends convexification theory to train spiking neural networks with exact gradients, removing surrogate gradient approximation.

Evaluates frontier LRMs on game learning behavior and brain activity prediction against human fMRI data during rule discovery.

Proposes action-credit RL for CLI agents using structured command attributes and selective observation over filesystems.

Audit finds mechanistic interpretability papers claim causality without stating identification assumptions; calls for transparency.

Probabilistic framework for abductive reasoning integrating LLMs with logic solvers, accounting for variable commonsense beliefs.

Stratechery weekly roundup covering Big Tech Q1 earnings, consumer tech interview, and sports commentary with minimal AI focus.

Susceptibilities technique extends neural network interpretability to deep RL agents, revealing parameter-space development undetectable from policy analysis alone.

Penalty-based first-order methods for bilevel minimax optimization problems with applications to emerging ML settings.

STEPS: test-time adaptation method for time-series forecasting under distribution shift using smooth error propagation on manifolds.

PSP-HDC applies graph-structured hyperdimensional computing for data-efficient, explainable process-structure-property prediction in materials science.

Bayesian sensitivity analysis for causal inference using evidence-based priors instead of worst-case assumptions.

Tool-calling decisions are linearly readable and steerable in LLMs; mean-difference activation patching switches tool selection at 77-100% accuracy.

Framework for standardizing AI evaluation scenarios via structured use-case elicitation and human-centered design to enable apples-to-apples comparisons.

Dooly: configuration-agnostic LLM inference profiler that exploits structural constraints to avoid redundant profiling across hardware/engine/model combinations.

Linear probing and activation patching reveal latent planning representations in Qwen3, Gemma-3, Llama-3; future constraints encoded at layer boundaries.

GLiGuard: 0.3B schema-conditioned encoder for LLM content moderation, 100-200× faster than 7B-27B autoregressive guardrail decoders.

Primer on susceptibilities and linear response theory for interpreting neural network posteriors via Bayesian influence functions.

Reddit discussion requesting comparative analysis of fragmented agent APIs and frameworks; community crowdsourcing of implementation experience.

Study of aleatoric uncertainty limits in ML-driven resource allocation for policy and humanitarian applications under irreducible individual vulnerability variance.