Data-Driven Aggregation for Scalable Digital Twin Generation
A proposed data-driven approach aims to make digital twin and simulation model generation scalable for dynamic manufacturing systems.
Research Digest
Preprints and papers, rewritten in plain English.
Dynamical Kernel Attention Accounts for Survival of Past States
A new attention mechanism estimates how much of a past state persists to the present, rather than relying only on token similarity.
Small Language Models Predict CVSS and CWE from CVE Descriptions
A new arXiv study explores using fine-tuned small language models to automatically derive vulnerability severity and type from CVE text, potentially easing the burden on vulnerability databases.
Transformers as Adaptive Kernel Learners for In-Context Learning
A new arXiv paper proposes viewing in-context learning as adaptive kernel learning, though the available abstract is truncated before presenting results.
Certifying Policy Improvement: Sample Complexity from History-Dependent Logs
A new paper formalizes how many logged episodes are needed to certify that a candidate policy beats a deployed one when the logging policy is history-dependent.
Precision Router Lets Agents Call Full-Precision Model Only When Needed
A learned router decides when a low-precision agent should fall back to its full-precision model, aiming to cut compute while preserving task performance.
Improved Algorithms for Smoothed Online Learning
New algorithms achieve better competitive ratios and optimal dynamic regret with switching costs in smoothed online learning.
Exact Worst-Case nDCG Loss for Pairwise-Judged Reranking
New work pins down the precise worst-case nDCG loss when a reranking system replaces a ranking with a fixed proposal under pairwise relevance judgments.
Measuring Empowerment from Offline Data to Drive Exploration
A new arXiv paper proposes using offline data to measure empowerment, potentially improving intrinsically motivated exploration in agentic tasks.
Constructive Universal Approximation via Ridgelet Transform
A new arXiv paper offers a constructive route to universal approximation for neural networks that take functions as inputs, moving beyond standard existence proofs.
Normal-Anchored SPD Transport for Source-Free ECG Adaptation
A new geometric transport method adapts ECG beat classifiers to new patients without labeled target data, addressing the extreme class imbalance of real Holter and patch recordings.
Tensor Learning Framework Targets Complex Multimodal Health Data
A new arXiv preprint proposes a tensor-based method to model complex relationships in heterogeneous clinical data for decision support.
Muon Optimizer Shows Flat Minima and Edge-of-Stability Traits
New empirical evidence suggests the Muon optimizer aligns with the Flat Minima Hypothesis and may operate at the edge of stability.
Unfaithful Chain-of-Thought Still Monitorable Across 11 Models
A replication study suggests that even when an LLM's chain-of-thought is unfaithful, it can remain a useful monitoring tool if producing the unfaithful reasoning requires computation.
HiPPO Functional Memory Targets Irregular Time-Series Queries
A new arXiv paper proposes coordinate-queryable attention to avoid repeatedly resolving observation-dependent representations for irregular time-series queries.
Generative Retrieval for Reverse Target Fishing with a Proteome-Scale Benchmark
A new preprint proposes a generative retrieval method and a proteome-scale benchmark to address limitations of discriminative approaches in reverse target fishing.
TideKV: Chunked Prefill to Cut Peak Memory in Long-Context LLMs
A new inference method targets peak memory during prompt prefill for long-context models.
Structured Latent Dynamics Aim to Fix Long-Horizon World Models
A new arXiv paper proposes structuring latent dynamics to improve long-horizon prediction in embodied world models.