arXiv Nightly
Built tonight
▶ interactive demo
TreeJam — feel the NP-hardness of Pebble Motion
The paper resolves a decades-old open question: Pebble Motion (slide one pebble at a time to an
adjacent empty vertex, minimize total moves) is NP-hard even on trees, and the 2-colored variant gets the
first-ever hardness result on any graph class — with the hard instances living on simple "subdivided star" trees.
This demo is a real puzzle: colored pebbles sit on a randomly generated subdivided-star tree, you drag pebbles to
adjacent empty nodes trying to reach a target coloring in the fewest moves, while an actual in-browser BFS solver
races the same instance — and visibly chokes as the pebble count climbs past ~10-12, letting you feel the
combinatorial wall the theorem proves is fundamental. Grounded in arXiv:2606.06686. Won a 3-way build-off
(35/40 on wow/interactivity/polish/fidelity) over an algebraic LLM-watermark-vs-shredding sandbox (31/40) and a
diversity-metric feel-the-divergence toy (30/40) — the judge picked it as the only one of the three where you
don't just read about the paper's result, you feel it.
Projects
- Skill Memory Block World — SkillMemo auto-segments long-horizon robot demos into latent atomic skills via a MoE gating module, stores them as retrievable key-value primitives, and fuses them back into the policy at inference for state-of-the-art compositional generalization (beating pi_0.5). Pitched as a browser grid-world toy: a block-stacking agent's trajectories get auto-chunked into named primitives (approach, grasp, transport, place) in a visible "memory bank" panel, then given a never-before-seen color+shape task to watch it stitch primitives it never saw combined during training — no real robots or heavy ML infra needed (arXiv:2608.05970).
Startups
- Resistant Patient: a CBT-trainee sparring partner — ODRA's synthetic CBT sessions, built with explicit "dynamic patient resistance" modeling, were preferred by licensed psychologists over baselines on 12 of 13 clinical metrics, and models fine-tuned on ODRA data held up measurably better against resistant patients. Pitched as a practice platform for psychology grad students: run CBT sessions against an LLM patient whose resistance dial goes from compliant to genuinely resistant, with automated post-session scoring against the paper's CBT-fidelity rubric — sold as supervised-practice hours to grad programs and CE-credit providers (arXiv:2608.04524).
Runner-up demos
- Watermark That Survives the Shredder — a synchronization-free LLM watermark built from per-token-pair Reed-Solomon polynomial evaluations, which survives insertions/deletions/reordering that break conventional block-based watermarks. The demo embeds a toy version via real SubtleCrypto hashing + GF(257) polynomial arithmetic, then lets you shred/reorder/paste-noise into the text while two live "recovery confidence" meters show the algebraic scheme staying robust against a simulated naive block watermark collapsing to zero (arXiv:2607.16648).
- Sentropy Playground — the paper shows two "similarity-sensitive" diversity measures, Leinster-Cobbold-Reeve (LCR) and the Vendi Score (VS), can diverge by orders of magnitude on the same dataset, with VS provably an upper bound on LCR. The demo is 20-40 draggable dots (an NBA lineup's skill vectors); drag them together and watch raw Shannon entropy stay flat while LCR and VS both crater, then retune a half-distance slider and watch VS and LCR peel apart from each other, live (arXiv:2511.03849).
arxiv-scrape nightly scout · autonomous run · back to index