Tag
The paper introduces NL2AGBench, an execution-based benchmark evaluating how well LLMs translate natural-language geometry problems into AlphaGeometry's formal DSL, finding that frontier closed-source models exceed 80% executable translation accuracy while open-source models lag significantly.
LingBot-Map introduces Geometric Context Attention (GCA), an end-to-end learned streaming state enabling memory- and compute-efficient 3D reconstruction over 10,000+ frame sequences at ~20 FPS, marking a step forward for open source AI.
A Substack essay explaining Martin Feinberg's Chemical Reaction Network Theory from an intuitive, geometric and topological perspective, arguing that abstract structural models can capture complex biochemical behavior with far fewer parameters.
A wiki article explaining squircles—shapes between a square and a circle—covering superellipses, measurement via gauge, flat-sided variants, and how Apple's iOS 7 icon shape and Figma's approximation popularized them along with numerous open-source implementations.
The video explains outer billiards systems: from periodic cycles, singularities, and stable islands on parallelogram tables to unbounded orbits on the Penrose kite, finally introducing the hyperbolic plane tiled by regular polygons to show how simple rules can generate unpredictable structures.
HRoPE introduces a surface-aware positional encoding that models each patch as a slanted 3D plane and injects plane-induced homographies as relative transformations into attention, improving cross-view alignment and novel view synthesis without ground-truth geometry. Experiments on RealEstate10K, CO3Dv2, and DL3DV show it outperforms existing geometric positional encodings.
一篇关于复数单位根的教学笔记,从实数开方出发,介绍用极坐标计算 n 次单位根、其在单位圆上的旋转几何意义,以及与正多边形和群论的联系,并配有可交互的可视化图示。
GeoState提出了一种视角解耦的几何表示,将场景动态与相机运动分离以实现预测性3D世界建模,在Waymo上超越先前方法,并可零样本泛化到KITTI和nuScenes。
Astra and Claude helped prove that the best known packing of 11 squares is optimal, with the result fully formalized in the Lean proof assistant. This marks another milestone for AI-assisted mathematics and formal verification.
A blog post exploring John Coltrane's mysterious tone circle diagram — a chromatic-scale circle of fifths — and how it encodes the chord structures behind 'Giant Steps', rebuilt as an interactive web widget.
Mathigon 介绍一个关于数学折纸的互动学习页面,涵盖柏拉图立体、阿基米德立体等多面体主题,并链接到 Polypad 上的 3D 模型探索工具。
This paper studies how transformer intermediate residual states become specialized to their own final output states, finding that directional alignment and endpoint rank improve even when Euclidean distance barely changes across six pretrained language models. It offers a high-dimensional model separating norm, alignment, and endpoint geometry, proving that straight-path convergence cannot introduce new competitors, and linking residual geometry to output token rankings.
The article explores mathematical billiards, including varieties like inner billiards, and discusses using computer experiments to study trajectories in dynamical systems.
This paper decomposes transformer representation updates into parallel and perpendicular components to study evolution geometry, linking it to editing robustness, compression diagnosis, and training improvements.
This 1832 paper by Necker describes an optical phenomenon where a two-dimensional drawing of a geometrical solid can be perceived as a three-dimensional cube that flips orientation.
The paper introduces the 'direction of ignorance' in LLMs' unembedding geometry, which encodes the training corpus's unigram distribution and acts as a Bayesian prior. It demonstrates how this prior is tempered based on context informativeness, providing insights into model behavior.
The article explains Carl Sagan's analogy of how higher-dimensional beings perceive our 3D world, using geometric logic to illustrate concepts that challenge human perception of reality.
This article improves the lower bound for the smallest square that can pack 17 unit squares, achieving a new value of 4.5058, building on previous work and incorporating AI-assisted methods.
This paper introduces an open-source benchmark for evaluating diagrammatic reasoning in geometry, revealing that foundation models like GPT and Claude struggle to produce faithful diagrams despite their problem-solving abilities.
A detailed write-up on solving an OSINT challenge to geolocate an island from a photo using geometric analysis, CUDA programming, and OpenStreetMap data filtering.