Training Binary Neural Networks via Ising Model Equivalence
A new theoretical mapping from binary and ternary network training to high-order Ising models could enable scalable on-the-edge learning.
Research Digest
Preprints and papers, rewritten in plain English.
Diffusion Transformer Models Limit-Order Supply Probabilistically
A new arXiv paper introduces OrderFlowDiT, a transport-regularized diffusion transformer that models the conditional distribution of limit-order volume over the next five minutes.
Normalizing Flow Maps Paper Highlights Limits of Generative Models
An arXiv preprint discusses the state of generative modeling, but its abstract is truncated, leaving the proposed contribution unclear.
Jailbreak success rate misses hidden harmful outputs, arXiv study argues
A new arXiv paper argues that standard jailbreak evaluations based on attack success rate overlook harmful outputs that can be reached through decoding, proposing a reachability-based measure instead.
CADENCE: Adaptive Decoding That Predicts Multiple Tokens at Once
A new arXiv paper proposes a context-adaptive decoding method that learns how many tokens to generate at once rather than always stepping one at a time.
Two Kinds of Similarity in the Platonic Representation Hypothesis
A new arXiv paper asks whether larger models converge on shared representations, distinguishing native cosine geometry from linear recoverability.
Benchmarking LLM Agents for Causal Discovery
A new study asks whether LLM agents can go beyond running causal discovery algorithms to spotting and fixing their mistakes.
Polynomial Graph Filters Hit Spectral Limits on Long-Range Ising Benchmark
A new preprint examines how polynomial graph filters perform on a benchmark whose labels are constructed to require long-range dependence.
Study Tests Whether Calibration Updates Transfer Between Representations
New research asks how much of a scale-calibration update learned for one model representation can be reused by another representation of the same pretrained model.
Drift-Aware Streaming PCA Aims to Make Low-Rank LLM Training More Stable
A new optimizer called Ori tackles the drift problem in streaming principal component analysis for low-rank large language model training.
Shaping Latent Geometry to Improve Classifier-Free Guidance
A new arXiv paper proposes learning the latent geometry of diffusion models to make classifier-free guidance more effective.
Accuracy Can Hide What Transformer Programs Learn
A new arXiv preprint warns that token-level accuracy may conceal important details of what Transformer Programs actually learn.
Diffusion Models Need More Than EMA, arXiv Study Suggests
A new preprint argues that large diffusion models may benefit from learning-rate decay and Schedule-Free optimization, not just constant LR and EMA.
Alert-List Concordance: Better Gauge for ICU Imputation
A new arXiv paper argues that imputation strategies for ICU sepsis prediction should be judged by how consistently they trigger alerts, not just by AUROC or reconstruction error.
Switching Optimizers During LLM Training Can Carry a Performance Cost
A new arXiv preprint systematically examines the AdamW-to-Muon transition, a concrete case of the suspected but understudied cost of changing optimizers across LLM training stages.
RLVR Rollout Windows: Shorter Is Cheaper but Can Erase Reward Signal
A new arXiv paper warns that cutting rollout length in RLVR to save compute may destroy the correctness signal needed for learning.
Partial-Revelation Panel Forecasting: Predicting Vote Distributions from Subsets
A new framework predicts a dialogue panel’s next-turn vote distribution from only k of n previous annotator votes.
Feature Selection Layer Proposed for Neural Network Classification
A new arXiv paper introduces a feature selection layer for multi-class classification, aiming to improve neural network applicability in scientific domains.