ml
Research on ml: explanations, evidence, limitations and connections between papers.
49 results
Cyclone forecasts can locate storms while missing strength
A retrospective cyclone forecast comparison finds stronger location and spatial placement scores alongside persistent peak-wind bias.
Choose low precision models by the constraint the device imposes
A low precision model can score well yet still miss the arithmetic rules of its target device. A proposed cheaper operation may also lack a measured end-to-end hardware benefit. Choose the design that addresses the actual constraint, then read its quality result within the reported test. These studies examine different routes through computation, stored weights, and fixed accelerator rules; their results do not form a common speed ranking.
BitSparse stores sparse feed-forward activations with less memory
BitSparse compresses exact-zero feed-forward activations, reducing measured memory in two language models.
A survival model predicts event curves from censored data
SurvICL predicts survival curves from censored examples and led aggregate comparisons on 63 held-out datasets, with weaker results on larger datasets.
A reporting framework shows how return timing shifts financial NLP labels
An earnings-call audit finds return labels can shift when outcome clocks move to reflect estimated input completion.
A reusable cache keeps tabular predictions accurate after compression
QCOC compresses reusable tabular model memory and has higher mean accuracy than direct compression in tested datasets.
A finer weight grid brings multiplier-free LLM inference closer
A finer four-bit weight grid brings multiplier-free language-model inference close to uniform integer accuracy in most tested settings.
A learned route archive speeds exact vehicle routing
A learned route archive and GPU pricing cut root time in exact vehicle routing while preserving certified bounds.
A learned noise policy steers fast generators without changing them
A reusable noise policy steers tested image, language and protein generators while keeping each generator frozen.
Hyperbolic distance helps agents navigate maze routes
Hyperbolic contrastive learning better reflects maze routes and improves goal reaching across many tested tasks.
A structure score reduces DFT checks under an error budget
A structure-based selector reduced DFT checks across eight molecular model combinations under a shared expected-error budget.
A single calibration protects early-exit prediction sets
A single conformal calibration protects true-label coverage after adaptive early exit, under exchangeability and a frozen model path.
A training flow makes CNNs fit a fixed edge chip’s number rules
ShiftQAT builds MAX7800x scaling rules into CNN training, with comparable reported 8-bit accuracy drops in ImageNet tests.
A method separates data selection from support shifts
A construction-based method separates data selection from support shifts; retrieval evidence favors within-region selection in one data family.
A calibrated reward helps models follow rules and answer well
CaLIF raises added-rule compliance while keeping answer quality near base across five tested model configurations.
Reconstructing replies steadies jailbreak success reports
A reply reconstruction method steadies jailbreak evaluations, while human-labeled tests expose misses on incomplete answers.
Language models preserve useful past predictions with FSPO
Across five task accuracy averages, FSPO exceeded on-policy SDPO after five hours on two tested language models.
Direct updates usually beat tiled reuse for changing sparse inputs
Across tested moving inputs, direct column updates beat a tiled executor in most sessions; matrix structure shaped the exceptions.
LayerTrim compresses reasoning models while preserving their depth
LayerTrim trims selected model features and led tested pruning methods in accuracy across three reasoning models.
Near-one momentum improved validation accuracy in two tests
Bandit theory and tests on two language models find higher validation accuracy with longer-lasting gradient averages.
Choose evidence that matches the decision conditions
A model can look best because its test set scarcely represents conditions it will face. Across these studies, evidence supports a choice when it examines the relevant conditions at the level where that choice acts. Earlier clues can screen options within their tested scope, but the studies’ different tasks and outcomes cannot be combined into one performance ranking.
Why larger rollouts learn slower but fail on a similar schedule
A two-clock study finds larger rollouts need more responses to learn, while failure follows a separate pattern.
FlashMask-3 speeds structured attention across GPUs
FlashMask-3 coordinates structured attention masks across Hopper GPUs and raised throughput in one large-model training setup.
Interaction traces help judges assess software as people use it
Interaction traces improved tested website preference judgments, broken game checks, and repairs, with evidence limited to selected tasks.