An arXiv paper introduces Drift Inspector, an open-source system that tracks how research fields change by having an LLM extract Atomic Contribution Claims (ACCs) from abstracts, then clustering them across years into an interactive map. Applied to six years of EMNLP, it shows NLP shifting toward LLM-era capabilities like reasoning and multimodality. The pipeline has also processed the full ACL Anthology, making it a reusable instrument for bibliometric analysis.
A scientometric study compares frequency-based and embedding-based approaches for detecting diachronic semantic change in Astrophysics and NLP corpora (2010–2024). Frequency methods slightly outperform semantic metrics against human judgments, but embedding metrics uniquely surface conceptually important terms, suggesting the two signals are complementary.
A scientometric study compares frequency-based and embedding-based approaches for detecting diachronic semantic change in Astrophysics and NLP corpora (2010–2024). Frequency methods slightly outperform semantic metrics against human judgments, but embedding metrics uniquely surface conceptually important terms, suggesting the two signals are complementary.
A new arXiv paper introduces a framework that combines large language models with dynamic knowledge graphs to overcome the limitations of traditional bibliometric indicators like citation counts and h-indexes. The approach aims to capture non-linear knowledge dynamics while mitigating LLM hallucination and opacity through structured grounding.
A new arXiv paper introduces a framework that combines large language models with dynamic knowledge graphs to overcome the limitations of traditional bibliometric indicators like citation counts and h-indexes. The approach aims to capture non-linear knowledge dynamics while mitigating LLM hallucination and opacity through structured grounding.
An arXiv paper introduces C-Unseen, a self-interpretable framework for detecting weak signals in Dynamic Temporal Knowledge Graphs. It defines weak signals as rare, semantically coherent subgraphs that proliferate across consecutive TKG snapshots, using LLM reasoning instead of keyword frequency, topic modeling, or untyped graph topology. The approach targets early, low-visibility indicators that precede significant changes.
An arXiv paper introduces C-Unseen, a self-interpretable framework for detecting weak signals in Dynamic Temporal Knowledge Graphs. It defines weak signals as rare, semantically coherent subgraphs that proliferate across consecutive TKG snapshots, using LLM reasoning instead of keyword frequency, topic modeling, or untyped graph topology. The approach targets early, low-visibility indicators that precede significant changes.
A new arXiv paper introduces Research Attention Prediction (RAP), a rolling benchmark evaluating whether LLM agents can track shifts in research attention. The authors argue reviews and research ideas lack uniquely verifiable outcomes, making existing evaluation difficult. RAP matters to practitioners building LLM research agents who need measurable, time-grounded evaluation tasks.
A new arXiv paper introduces Research Attention Prediction (RAP), a rolling benchmark evaluating whether LLM agents can track shifts in research attention. The authors argue reviews and research ideas lack uniquely verifiable outcomes, making existing evaluation difficult. RAP matters to practitioners building LLM research agents who need measurable, time-grounded evaluation tasks.
An arxiv paper introduces a method to automatically generate evolution trees of AI methods from citation graphs, addressing limits of manual surveys. It argues existing taxonomy-induction methods are leaf-bound and time-agnostic, misplacing transitional papers and inverting ancestor-descendant topology. The work matters to researchers tracking fast-moving subfields where surveys scale poorly.
An arxiv paper introduces a method to automatically generate evolution trees of AI methods from citation graphs, addressing limits of manual surveys. It argues existing taxonomy-induction methods are leaf-bound and time-agnostic, misplacing transitional papers and inverting ancestor-descendant topology. The work matters to researchers tracking fast-moving subfields where surveys scale poorly.