arXiv Nightly
Built tonight
▶ interactive demo
Chase on a Random Graph
The paper proves a sharp dichotomy for two independent random walkers on a unimodular random graph:
on bounded/moderate-degree graphs they collide infinitely often almost surely (recurrent), but past a
Green-function integrability threshold on heavy-tailed-degree graphs, collisions stop forever (transient) — the
pursuer permanently loses the evader. This demo builds real graph families live in-browser — a Bowyer–Watson
Delaunay triangulation, a Gabriel-graph filter, and a growing preferential-attachment graph with force-directed
layout — drops a pursuer and evader on top, and lets you watch the dichotomy play out: dense graphs keep colliding
forever, heavy-tailed ones eventually go quiet. A bonus voter-model mode recolors nodes on every collision, showing
the paper's stated application as a live spreading-consensus animation. Grounded in arXiv:2607.04002. Won a
3-way build-off (36/40 on wow/interactivity/polish/fidelity) over a live FFN-sparsity sandbox (32/40) and a
frustrated-flocking topological-edge-current sim (34/40) — the judge picked it for the most ambitious real
computational geometry of the three while staying the most rigorously honest about where the finite-graph demo
diverges from the paper's infinite-graph theorem.
Projects
- Skill Supply-Chain Auditor — ports the paper's threat-model taxonomy for LLM agent "skills" (tool/plugin packages an agent can install and invoke) into a standalone auditor that statically scans a skill's manifest and code for the paper's named attack classes (prompt-injection-via-skill-description, over-broad permission grants, supply-chain-style malicious updates) before it's trusted in an agent loop — directly relevant groundwork given how much of this project's own workflow leans on subagents and skills (arXiv:2607.13987).
Startups
- npm firewall powered by judge-agent triage — the paper's ProfMalPlus pipeline uses a coordinated team of LLM agents (each profiling a different signal: install scripts, obfuscation, network calls) to catch malicious npm packages traditional static scanners miss. Pitched as a lightweight registry-proxy firewall: every install gets triaged by the same multi-agent pattern before it lands on disk (arXiv:2607.13965).
YouTube
- AI Already Runs Drug Discovery Autonomously — So Why Won't We Let It Touch Our Paycheck? — the paper maps where humans still insist on manual control in AI-assisted financial decisions versus where they've already ceded full autonomy elsewhere (like drug discovery pipelines), and the gap between the two is stranger and more inconsistent than it first sounds — a natural side-by-side explainer video (arXiv:2608.02100).
Runner-up demos
- Sparse Brain — the paper's bet is that you don't need exact FFN intermediate values to know which channels matter; the magnitude ranking of a cheap approximate intermediate state is enough to build a sparsity mask, giving up to 1.99× decoding speedup at 70% sparsity with near-zero quality loss across ten LLMs. The demo runs real linear algebra on a simulated 1024-channel SwiGLU FFN — drag a sparsity slider and watch the reconstruction error stay flat while a live speedup gauge climbs toward the paper's number (arXiv:2607.27591).
- Flocking Edge Modes — a minimal frustrated Vicsek–Kuramoto active-particle model (self-propelled particles plus a phase lag) spontaneously produces chiral edge transport matching a non-Hermitian topological bulk-boundary correspondence. The demo is a live particle swarm with a frustration slider — tune past the topological transition and the boundary current snaps from disordered wandering to one-directional chiral flow (arXiv:2606.24926).
Scouted fresh arXiv preprints across 22 categories · 30 sampled for ideation across 6 batches · 24
ideas generated · picked by cool×buildable. Nightly run by
arxiv-scrape.