The Archive

VFold symmetry-aware KV cache compression reduces LLM memory at long contexts via inter-layer value merging without architectural changes.

Latent memorization in diffusion models detected via basin geometry before generated samples reveal it, enabling early audit.

Lamarckian evolutionary framework discovers autonomous driving training distributions via competitive selection instead of surrogate metrics.

GNN-to-MLP distillation must preserve graph-induced geometry; absence causes spectral failures (underfit on sparse graphs, overfit on dense).

Proposes metrics to predict which base models benefit from agentic coding post-training, avoiding costly full trials.

Rubix solves planar Procrustes-Wasserstein alignment globally via permutation polygon convex hull.

Cross-domain pretraining improves neural CFD surrogates' generalization across geometries and engineering applications.

OrBIT discovers data-driven compression geometry for embedding tables using orbit dynamics, reducing coding budgets while retaining reconstruction quality.

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

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

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

Framework for building 2D-constrained minimal transformers enabling full visualization of internal representations and inference of learned algorithms from geometry.

Study of homogenization failure mode in multi-agent systems via conformity, polarization, and inertia metrics; evaluated on code generation, hiring, and peer review tasks.

Simon Willison demonstrates integrating Parseable observability platform with Datasette's OpenTelemetry trace support via a working example.

DepthWorld adds 3D geometric consistency to video-based robot manipulation world models via calibration and architecture improvements.

Systematic study of semantic DocID space design in generative information retrieval with efficient evaluation metrics.

BiasFlow diagnostics toolkit monitors spurious feature reliance via centroid alignment metrics and paired regularization penalty.

Technical paper on hyperbolic graph embedding optimization using parametrization choices to avoid numerical instability in Poincaré and Lorentz models.

We’re launching ReviewBench, a benchmark for code review agents built on representative GitHub pull requests, multi-source ground truth, calibrated evaluation, and production-aligned metrics. The post ReviewBench: An open benchmark for AI code review appeared first on The GitHub Blog.

AnchorPose predicts metal-organic framework structures via geometry-aware meso-grained pose generation accounting for anisotropic rotation error effects.

Machine learning automates parameter extraction and defect detection from metrology data in advanced semiconductor manufacturing.

Machine learning accelerates scanning probe microscopy metrology for sub-nm semiconductor feature characterization.

Drifting Models improve one-step generation via discriminative geometry in feature space; introduces persistent feature learning to explain pretrained space superiority.

Tropical geometry analysis of fine-tuning forgetting: per-unit projector prevents catastrophic forgetting more precisely than fixed subspace approaches.

GS-Codec replaces VQ-VAE quantizers with Gaussian splatting parametric decomposition for neural speech codec compression.

KRATOS is a bibliometric diagnostic for analyzing document participation vs. citation recognition across analytical groups.

Operational Validity Contract formalizes monitor alarm policies for LLM agents, distinguishing deployment constraints from probe calibration metrics.

SNAP improves 3D geometric representation learning from novel view synthesis by redesigning encoder-decoder architecture and removing pixel-space targets.

Analysis of scale symmetry breaking in undercomplete linear autoencoders under finite-stepsize SGD.

Threat-preserving representation sensitivity metric reveals benchmark representation effects on LLM agent security ASR scores.