The Score Hamiltonian: Mapping Diffusion Models to Adiabatic Transport
Research establishes an exact correspondence between score-based diffusion models and adiabatic transport in quantum mechanics.
arXiv cs.LG — Machine Learning·15 Jul
Search signals, briefings, company results, benchmarks and glossary terms.
Search signals, briefings, company results, benchmarks and glossary terms.
Use this view to inspect the underlying evidence corpus. For ranked developments, decision posture and interpretation, use Signals.
Raw feed or Signals?
Raw feed is chronological evidence. Signals ranks and interprets material change.
Research establishes an exact correspondence between score-based diffusion models and adiabatic transport in quantum mechanics.
TerraZero is a new procedural driving simulator and self-play training stack for robust autonomous driving agents, optimizing for scale, realism, and diversity.
Research identifies adversarial attack methods targeting online handwriting recognition models via salience-based temporal editing of pen trajectories.
Research explores efficient video representation learning, focusing on motion cues to reduce the prohibitive costs associated with scaling video models.
Researchers propose CROCS, a two-stage clustering framework for consumer segmentation using smart meter data for demand-side management.
Research shows that verifying robustness and satisfiability for Binarized Neural Networks (BNNs) is NP-complete, even under uniform image occlusion.
Research proposes constraint-aware aggregation rules for Federated Reinforcement Learning in microgrid energy coordination to ensure safe global behavior.
Research explores Kernel PCA for out-of-distribution (OoD) detection using non-linear feature subspaces to enhance deep neural network reliability.
ReDiTT, a retrieval-augmented conditional diffusion transformer, is proposed for predicting inter-event times and types in asynchronous time series.
New research proposes 'Angular Calibration' and refines Platt Scaling for provably optimal calibration in high-dimensional binary classifiers.
Research introduces a lightweight multi-scale anomaly detection method for time-series data, optimized for resource-constrained edge devices.
Propheticus is an ML framework for developing predictive models to enhance software reliability and security, addressing complexity in ML application.
Research proposes using autoregressive sequence models for uncertainty quantification and exploration in online decision-making by treating uncertainty as missing data.
Researchers introduced RAFP, a new method for robustly identifying the lineage of large language models even after finetuning, crucial for licensing compliance.
Research proposes a compliance-aware federated learning framework adapting differential privacy to varying institutional compliance and resources.
Researchers introduced Sample Efficient Generative Optimization (SEGO), a framework for molecular design optimizing chemical spaces with fewer evaluations.
Research explores the underlying computational basis of confidence signals in large language models to improve trustworthy deployment.
Research identifies 'Fisher Rank Inflation,' a spectral signature in last-layer gradients, indicating when deep networks memorize corrupted labels.
Research demonstrates Evolution Strategies (ES) can scale for LLM fine-tuning, challenging the dominance of Reinforcement Learning (RL) in this area.
Research finds higher embedding dimensions in transformers improve internal 'world model' fidelity and robustness for simple sorting tasks via RL.
Research indicates multi-agent debate (MAD) for LLM reasoning may underperform single-agent approaches, questioning its benefits.
Research explores auditable context-aware HFMD forecasting using structured LLM agents to provide explainable risk predictions for clinical settings.
Research explores efficient learning of branching networks for multitask algorithmic reasoning, aiming to perform multiple algorithmic tasks simultaneously.
Research investigates why Federated Averaging (FedAvg) performance degrades with non-IID data, focusing on representation preservation vs. utilization.
Research proposes a novel replicate-and-quantize strategy to address load imbalance in Sparse Mixture-of-Experts (SMoE) LLMs, improving efficiency.
TraceSynth uses diffusion models to generate synthetic kernel traces, addressing challenges of collecting real production data for ML-based system diagnostics.
Research investigates Low-Rank Adaptation (LoRA) as a parametric method for continuous knowledge updating in LLMs, addressing ICL/RAG limitations.
Research dissociates biological and artificial visual systems using rate-distortion theory, finding different compression geometries.
Research systematically investigates memorization behaviors in Rectified Flow generative models, focusing on generalization vs. data recall.
Researchers propose a novel risk scoring system that optimizes directly for net benefit and provides interpretability for decision-making.
New research proposes 'gated activation redirection' for inference-time machine unlearning in LLMs, aiming to remove data influence without retraining.
New research proposes a geometric approach to constrained online convex optimization, achieving optimal regret with improved cumulative constraint violation.
Research identifies how gradient noise structure and optimizer normalization jointly determine the effectiveness of learning-rate cooldowns in large model pretraining.
New research suggests adversarial non-robust features, not information dependency or rote memorization, cause training data exposure in image reconstruction attacks.
FlashDiff is a research paper introducing efficient regional execution and scheduling for diffusion models to reduce latency and increase throughput.
Research introduces Quantum Port-Hamiltonian Neural Networks (Q-pHNNs) to learn classical dynamics using parameterised quantum circuits.
New research proposes H-TDBU, a hierarchical hybrid framework for synthetic tabular data generation, addressing issues like data heterogeneity and rare events.
dMX is a new framework for differentiable mixed-precision quantization, allowing learnable floating-point bit-width assignment to LLM layers for efficiency.
Research proposes a unified theoretical framework to explain data mixing scaling laws, extending neural scaling laws to multi-domain data mixtures.
Research finds quantization errors are additive at 4-bit precision, challenging isolated sensitivity methods for mixed-precision model deployment.
Researchers introduced Prime Fourier Embeddings (PFE), a novel method encoding integers using prime-indexed (cos, sin) pairs to enhance modular arithmetic.
Research introduces a method for explaining Event-based Temporal Graph Neural Networks (ETGNNs) by tracing feature-induced information flow.
Research explores interaction-aware mixture-of-experts for structured health data, finding minimal performance gains but improved interpretability through view construction.
Airbnb developed "Proximity Features" for privacy-compliant cold-start personalization, leveraging geo-temporal proximity without user-level history.
Research paper proposes a multi-label classification method using classifier chains and SHAP for pathology test recommendations.
Research explores using Diffusion Models (DMs) for Out-of-Distribution (OoD) detection by assessing discrepancies in deep representations.
Research evaluates deep learning model robustness for PV power forecasting under realistic numerical weather prediction errors, focusing on spatiotemporal coupling.
Research details a reproducible audit of Sparse Autoencoders (SAEs), distinguishing decoder-geometry alignment from encoder-activation behavior.
Elenchos, a new generative framework, evaluates LLMs' abductive reasoning capacity by framing it as a structural inverse problem.
LatentFlow introduces a general framework for conditioning stochastic processes without learned neural approximations or training, addressing intractable conditional laws.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion