Mind the Sim-to-Real Gap & Think Like a Scientist
Framework for optimal simulator-experiment allocation when deploying pre-trained simulators; decomposes value error into calibration drift and parametric residual.
Search the full wire by company, model, lab, or keyword. Every story we have ever aggregated.
Framework for optimal simulator-experiment allocation when deploying pre-trained simulators; decomposes value error into calibration drift and parametric residual.
Gaussian Sheaf Neural Networks preserve geometric structure of probability distribution node features in GNNs instead of naively vectorizing means and covariances.
Semiparametric debiasing theory for bilevel gradient estimation using efficient influence functions.
Input symmetries in stimulus space confound representational similarity analysis of neural network codes.
Advantage Collapse Rate metric diagnoses GRPO training failure in RL preference optimization for LLM reasoning.
HORST: modular optimizer composing L∞ and L1 geometries for stable, sparse transformer training.
Hybrid ML-physics model for forest height estimation from TanDEM-X interferometry, extending feature selection to resolve structural ambiguities in remote sensing data.
Hi, I'm interested in geometric deep learning (due to Michael M. Bronstein's book and Maurice Weiler's PhD thesis), and in order not to write projects to nowhere, I decided to keep a technical blog. I started with a short note about machine learning on spherical manifolds, but…
Reddit user claims Gemini 3.5 Flash has higher inference costs and lower performance than 3.1 Pro; unverified observation without detailed metrics.
OpenAI model disproved 80-year-old unit distance conjecture in discrete geometry, demonstrating AI capability in mathematical research.
Quantization benchmark study on Qwen 27B comparing TurboQuant, TCQ, and symmetric q8 via PPL/KLD metrics on RTX 3090 with 64k-128k context.
Theoretical analysis of representation geometry in JEPAs showing isotropic regularization is suboptimal for structured downstream tasks.
Probability-Conserving Flow Guidance reformulates diffusion guidance through continuity equations to preserve learned manifold geometry under strong conditioning.
Minimalist visual-inertial odometry uses four photodiodes with Gabor masks and IMU for differential-drive robot motion estimation.
Dynamic layer-wise optimizer geometry selection via Schatten-p norms unified under Linear Minimization Oracle theory.
Memory-augmented RL framework for CAD generation agents handling long operation sequences and geometric constraints with error correction.
Automated neighborhood generation for local search via constraint symmetry analysis in the IDP system.
Framework for validating interpretable surrogate models using normalized Loss of Interpretability metric grounded in power divergence.
Bayesian nonparametric LABS regression establishes posterior contraction rates in Besov spaces for adaptive spline modeling.
Proposes metric for evaluating XAI methods via continuous input perturbation, assessing sufficiency and necessity without ground-truth labels.
Proxy metrics from token-level statistics predict downstream LLM capabilities faster than cross-entropy loss.
PACE-FNO uses Lie-algebra symmetries to improve neural operator generalization on PDE solution maps.
Pointwise Riemannian Dimension framework characterizes deep network generalization via learned feature representation geometry.
**World models** learn compact latent representations for planning without pixel reconstruction. LeWorldModel (LeWM), from LeCun's group at NYU, achieves stable end-to-end JEPA training by enforcing an isotropic Gaussian prior over the full latent space. **The flaw:** real…
Audit of SAEBench finds two major SAE evaluation metrics (TPP, SCR) unreliable, questioning interpretability benchmark validity.
Framework integrates fixed external RGB cameras as geometric priors for active 3D scene graph generation in robotic systems.
SAGE framework improves VLM spatial reasoning via geometric logic consistency and duality operations in GRPO training.
First source attribution benchmark for generative 3D models across multi-view, geometric, and frequency-domain fingerprints.
Been writing backend for 11 years. last 8 months I've moved most of my work into claude code. I want to ask something and I'm not sure how to phrase it. when I spend a full day in claude code and ship 3 or 4 PRs, do I actually feel like I worked? or do I feel like I…
Symmetry-compatible optimizer design principle ensuring equivariance under parameter space symmetries for embeddings and MoE routers.