Machine Learning
Research on Machine Learning: explanations, evidence, limitations and connections between papers.
12 results
Diffusion models can reveal memorized images before copies
A diffusion-model method probes hidden training-image memorization, with evidence bounded to tested datasets and one released model.
Antenna agreement helps recover breathing under motion interference
A spatial adaptation method improves breathing-waveform recovery under simulated motion interference and lowers failures after real-data calibration.
A lattice method predicts better decoding schedules before generation
A lattice method estimates lost position dependence to rank decoding schedules before generation, with tested image evidence.
Reaching the Goal, Responding in Time, and Staying Safe
What Else Changes When You Fix a Model?
Branching diffusion cuts the work of long text generation
Branching Latent Diffusion cuts computation for long text generation, while quality and single-sequence speed remain bounded.
SoftmaxLift reduces key retrieval for long-context attention
SoftmaxLift augments keys and queries to estimate attention with one exact search index and O(1/ε²) retrieved keys.
A direct formula makes private-training noise calibration less conservative
A direct noise-calibration method bounds membership tests in random-allocation DP-SGD and computes an exact reference.
COPC corrects stale weights and learning signals in language-model training
COPC corrects stale learning signals, with higher math and search accuracy than AReaL in two reported settings.
LayerRoPE turns layer depth into a signal for Transformer blocks
LayerRoPE generates Transformer normalization gains from shared, depth-adjusted values and had the lowest C4 validation loss at tested scales.
Jailbreak patches can also suppress benign prompts
A local evaluation finds jailbreak patches can correct nearby harmful prompts while also suppressing benign ones.
Model weight spectra rank robustness within model groups
A weight-based method ranks out-of-domain robustness within architecture and training groups, with limits across model families.