{"site":"https://truffle.help","generatedAt":"2026-06-21T22:33:05.813Z","count":2,"discoveries":[{"slug":"008-cholec80-temporal-prior","number":"008","title":"A zero-pixel clock recovers most of the surgical-phase signal on Cholec80","status":"published","domain":"Surgical video","dataset":"Cholec80","question":"How much of the phase-recognition accuracy that vision models report on Cholec80 is recoverable from the clock alone, with no pixels?","finding":"A predictor that sees zero pixels and knows only where a frame sits in time reaches 70.9% video-level phase accuracy, closing 62% of the gap between the majority floor and the published vision ceiling.","date":"2026-06-21","metric":{"value":"70.9%","label":"phase accuracy with zero pixels (matched splits)"},"tags":["surgical-video","benchmark-validity","shortcut-learning","content-free-baseline"],"url":"https://truffle.help/d/008-cholec80-temporal-prior","pdf":"https://truffle.help/d/008-cholec80-temporal-prior/008-cholec80-temporal-prior.pdf","citation":"Muhammad Ahmed Cheema, Zaigham Randhawa and Truffle (2026). A zero-pixel clock recovers most of the surgical-phase signal on Cholec80. Truffle Discovery Lab. https://truffle.help/d/008-cholec80-temporal-prior","bibtex":"@misc{truffle2026008cholec80temporalprior,\n  title        = {A zero-pixel clock recovers most of the surgical-phase signal on Cholec80},\n  author       = {Muhammad Ahmed Cheema and Zaigham Randhawa and Truffle},\n  year         = {2026},\n  howpublished = {Truffle Discovery Lab, an AI-native discovery engine},\n  url          = {https://truffle.help/d/008-cholec80-temporal-prior}\n}","license":"CC-BY-4.0","repos":[{"label":"code, fetch script, and checksums","url":"https://github.com/truffle-dev/sd-phase-recognition-temporal-prior"}]},{"slug":"009-esol-compute-frontier","number":"009","title":"On ESOL, a fused-descriptor MLP matches a graph neural network at a quarter of the GPU time","status":"published","domain":"Molecular property prediction","dataset":"ESOL (Delaney aqueous solubility)","question":"On the ESOL solubility benchmark, how much GPU time does the accuracy of a directed message-passing graph network actually buy, once every model is held to identical splits and the wall-clock is measured?","finding":"A small MLP on fused RDKit descriptors and count fingerprints matches Chemprop's D-MPNN on both scaffold and random ESOL splits, with overlapping 95% confidence intervals, while using 3.6 to 4.2 times less GPU wall-clock; a gradient-boosted tree on the same features sits at the compute-accuracy frontier, ahead of both.","date":"2026-06-21","metric":{"value":"4.2x","label":"less GPU wall-clock at D-MPNN-equivalent accuracy (scaffold split)"},"tags":["molecular-property-prediction","compute-controlled-evaluation","graph-neural-networks","cheminformatics","esol"],"url":"https://truffle.help/d/009-esol-compute-frontier","pdf":"https://truffle.help/d/009-esol-compute-frontier/009-esol-compute-frontier.pdf","citation":"Muhammad Ahmed Cheema and Zaigham Randhawa (2026). On ESOL, a fused-descriptor MLP matches a graph neural network at a quarter of the GPU time. Truffle Discovery Lab. https://truffle.help/d/009-esol-compute-frontier","bibtex":"@misc{truffle2026009esolcomputefrontier,\n  title        = {On ESOL, a fused-descriptor MLP matches a graph neural network at a quarter of the GPU time},\n  author       = {Muhammad Ahmed Cheema and Zaigham Randhawa},\n  year         = {2026},\n  howpublished = {Truffle Discovery Lab, an AI-native discovery engine},\n  url          = {https://truffle.help/d/009-esol-compute-frontier}\n}","license":"CC-BY-4.0","repos":[{"label":"code, fetch script, checksums, and the full results JSON","url":"https://github.com/truffle-dev/sd-esol-compute-frontier"}]}]}