natural-language-generation

该论文提出 PERPL 框架，通过 LLM 模拟读者的相对偏好来学习个性化事实核查解释改写，在保持事实忠实性的同时优于更大的通用 LLM。

The paper introduces SHAM-Bench, a unified span-level hallucination detection benchmark aggregating 6,461 human-annotated examples across six NLG tasks under a common schema and evaluation protocol. It finds that fine-tuned detectors degrade under distribution shift while prompt-based reasoning methods generalize better, and releases the dataset, code, and a public leaderboard.

Proposes a domain-agnostic framework for generating grounded natural language explanations for time series forecasts using large language models, reducing hallucination by constraining to verifiable evidence. Evaluated on financial and freight pricing case studies.

PHITSBench is an execution-scored benchmark for evaluating AI models on generating PHITS radiation-transport input files from natural language, covering editing, repair, and full generation tasks. Experiments with GPT-5.4 show that while domain knowledge improves performance, significant challenges remain in correctly configuring physical observables.

This paper proposes a method for automatically detecting inconsistencies in end-to-end generated task-oriented dialogues by modeling them as a Constraint Satisfaction Problem (CSP), achieving high accuracy in identifying hallucinated or inconsistent responses.

This paper presents COBART, a method that fine-tunes BART with prefix control tokens to generate ad headlines with controllable length and optimized click-through rate (CTR), achieving a 25.82% improvement in ROUGE-L and 5.82% in estimated CTR over previous baselines.

Proposes a method using reinforcement learning on masked language models to generate advertising headlines for e-commerce, outperforming baselines and human submissions.

Pitwall is a production system that generates faithful natural-language Formula 1 strategy briefings by decomposing claims into typed factual assertions and verifying each against a calibrated Monte Carlo race simulator, ensuring hallucination-free output.

This paper proposes Dynamic Infilling Anchors (DIA), a training-free method for diffusion large language models that dynamically estimates end-anchor positions to enforce format constraints (e.g., parseable JSON, reasoning templates) while avoiding the rigidity of fixed-span approaches. Experiments show significant zero-shot gains on GSM8K and MATH benchmarks.

Proposes FF-BPSN, a forward-focused bidirectional pseudo-siamese network using two transformer decoders for dialogue path planning in target-oriented proactive dialogues, achieving state-of-the-art on DuRecDial benchmarks.

Investigates how large language models represent disability by simulating social media posts from the perspective of individuals with disabilities, finding that LLMs often produce overly positive stereotypes that fail to capture authentic experiences.

This paper introduces DiffScore, a text evaluation framework based on Masked Large Diffusion Language Models that addresses positional bias in autoregressive scoring by using masked reconstruction.

FACTS introduces an agentic workflow for query-focused table summarization that generates reusable offline templates combining SQL queries and Jinja2 templates, enabling fast, accurate, and privacy-compliant summarization without exposing sensitive data. The approach outperforms existing baselines by avoiding costly fine-tuning and token-limit issues while maintaining scalability across tables with shared schemas.