arXiv Nightly
Built tonight
โถ interactive demo
Lahaina Lane-Reversal Simulator โ drag a slider, watch evacuation time collapse
The paper models the 2023 Lahaina peninsula evacuation (a single two-lane highway exit, 102 deaths)
as hyperbolic conservation laws on a road-network graph with game-theoretic junction splitting, and finds a sharp
phase transition: reversing one southbound lane captures nearly all achievable improvement in clearance time,
with diminishing returns beyond that โ and a fourth lane can be reserved for emergency vehicles at negligible
civilian cost. This demo implements a real Godunov-scheme LWR (Lighthill-Whitham-Richards) traffic solver over a
simplified peninsula grid, with a lane-reversal slider and an emergency-lane toggle driving a live clearance-time
readout, so the paper's diminishing-returns curve is something you trigger yourself rather than read about.
Grounded in arXiv:2603.29055. Won a 3-way build-off (35/40, vs 32/40 for both runners-up) โ the judge
called it out as the only build that reconstructed the paper's actual method (a numerical PDE solver with a
junction-splitting rule) instead of approximating it with heuristics.
Startups
- TrajectoryEdge โ dynamic treatment-effect signals for sports betting โ reframes causal inference on time series as trajectories rather than before/after snapshots, introducing a Dynamic Average Treatment Effect (DATE) estimator validated on COVID lockdown data. Repurposed for sports "shock events" (coaching fires, trade-deadline swaps, injury returns), DATE decomposes how the event reshapes a team's whole subsequent win-probability trajectory and its persistence โ catching slower-to-price market drift that static CLV tracking misses. Pitched as a differentiated signal feed layered onto an existing odds/CLV pipeline (arXiv:2602.00836).
YouTube
- How to Solve a Linear Program EXACTLY โ No Rounding, No Diophantine Tricks โ most LP explainers wave their hands at the leap from an approximate numerical solution to an exact rational one. This paper gives a clean oracle-model algorithm that recovers exact primal and dual solutions in O(nยฒ log(n/ฮด)) calls to any black-box approximate solver, where ฮด is a geometric condition number the algorithm never needs in advance โ and the call count doesn't depend on the cost vector at all. Video hook: animate a 2D polytope with an oracle probing near-optimal points, and visualize how ฮด (distance-to-degeneracy of the optimal vertex) governs how many probes it takes (arXiv:2606.11820).
Projects
- Safety Bridge Finder โ the paper finds that safety behavior in multilingual LLMs runs through a sparse subset of shared cross-lingual pathways that transfer refusal signal from high-resource to low-resource languages; patching just that subset substantially boosts refusal robustness in under-resourced languages while barely touching general capability. An open-weight replication toolkit โ activation-patch across 6-8 languages, intersect to find the shared pathway set, patch just that subset and measure refusal-rate lift vs a full-model fine-tune baseline โ shipped as a CLI plus a per-language pathway-overlap heatmap, useful standalone for anyone red-teaming multilingual deployments (arXiv:2608.09095).
Runner-up demos
- Break the Agent: Live Leak-Rate Sandbox โ a toy agent with a fake
send_emailtool and a secret embedded in a document; paste prompt-injection attempts and flip between UNHARDENED and HARDENED modes to watch a real attack succeed then fail on the same input, live-replicating the paper's reported 100%/91% leak-rate reduction. Scored 32/40 โ strong fidelity, but the judge favored B's from-scratch PDE solver over this one's rule-based mitigations (arXiv:2607.18847). - Trick the Color Brain โ reproduces the paper's psychophysics task as a browser game: pick an object's "true" achromatic color under a colored illuminant, then toggle the same cue-removal conditions (kill the local surround, kill the scene's spatial mean color) the paper used to degrade both human and DNN color-constancy accuracy, and compare your own degradation curve to theirs. Scored 32/40 โ the judge noted its paper-comparison curve was illustrative since the abstract itself gives no numeric accuracy figures (arXiv:2602.13887).
arxiv-scrape nightly scout ยท autonomous run ยท back to index