Papers where the original authors have engaged with our plain-language explanation.
For every paper on this page, at least one of the original authors has seen our plain-language explanation and engaged with it — either confirming it reads accurately or requesting corrections that we then applied. An endorsement does not mean the authors formally approve every sentence, but it does mean the explanation has passed the eyes of the people who wrote the paper.
949 papers reviewed by authors · 231–240 / 949
This paper proposes a robust SINDy framework called ILTS-SINDy that combines Iterative Least Trimmed Squares for outlier filtering with Sequentially Thresholded Least Squares for sparse regression, demonstrating superior performance in identifying nonlinear dynamics from datasets with up to 20% corrupted observations.
Building on Erdos's proof of the existence of these numbers for Sidon sets 70 years ago, this paper establishes upper and lower bounds on the asymptotic density of -Golomb rulers (sets where each positive difference occurs at most times), proving that their size is bounded above by a term proportional to and below by a term proportional to .
This paper introduces Trajectory-Invariant Latent Refinement (TILR), a training-free framework that identifies and manipulates stable, low-rank invariant directions within language model latent trajectories to significantly improve reasoning consistency and reduce sensitivity to paraphrasing and perturbations without sacrificing accuracy.
This paper provides a perspective on the theoretical landscape of cuprate high-Tc superconductivity after 40 years, focusing on the growing consensus around repulsive interaction mechanisms, outlining three main theoretical schools, and updating Anderson's "dogmas" in light of recent experimental findings.
This paper introduces a unified framework to independently evaluate and compare data-generating priors for tabular foundation models by training identical architectures on tasks derived from various priors, revealing that different priors significantly influence downstream performance and consistency beyond mere data-level similarity.
This paper identifies and quantifies "intervention bias" in zero-shot LLM educational advisors, which erroneously recommend unnecessary actions compared to an optimal oracle, and demonstrates that supervised policy learning using Decision Transformers or XGBoost effectively eliminates this bias while achieving high accuracy and low-latency decisions suitable for high-stakes deployment.
This paper systematically benchmarks five lightweight, CPU-feasible hallucination detection methods across question answering, dialogue, and summarization tasks, revealing that performance is highly task-dependent with ensemble methods excelling in QA, NLI detectors leading in dialogue, and all approaches failing significantly on summarization.
This paper introduces an evidence-informed framework for continual scientific discovery with LLMs that replaces static Bayesian surprise with non-stationary, belief-updated surprisal and employs filtering and diversity maximization to eliminate spurious rewards, thereby significantly improving hypothesis exploration across multiple domains.
SurrogateShield is a client-side proxy that enhances privacy-preserving LLM interactions by replacing detected PII with locally generated, type-consistent surrogate values before transmission and restoring the originals in responses, thereby eliminating real PII exposure while significantly improving semantic utility compared to traditional redaction methods.
This paper establishes that for any multiset of real numbers summing to zero, the Shannon entropy of the sum of randomly sampled elements is bounded below by the entropy of a Bernoulli random variable with mean , presenting a sharp result that serves as an entropic analogue to the Manickam-Miklós-Singhi conjecture.