Detoxification in large language models (LLMs) remains a significant research challenge. Existing decoding detoxification methods are all based on external constraints, which require additional resource overhead and lose generation fluency. This work innovatively proposes Detoxification with Self-Constrained Decoding (DSCD), a novel method for LLMs detoxification without parameter fine-tuning. DSCD strengthens the inner token distribution of the safety layer while weakening that of hallucination and toxic layer during output generation. This effectively diminishes toxicity and enhances output safety. DSCD offers lightweight, high compatibility, and plug-and-play capabilities, readily integrating with existing detoxification methods for further performance improvement. Extensive experiments on representative open-source LLMs and public datasets validate DSCD’s effectiveness, demonstrating state-of-the-art (SOTA) performance in both detoxification and generation fluency, with superior efficiency compared to existing methods. These results highlight DSCD’s potential as a practical and scalable solution for safer LLM deployments.

Temporal knowledge graph question answering (TKGQA) addresses time-sensitive queries using knowledge bases. Although large language models (LLMs) and LLM-based agents such as ReAct have shown potential for TKGQA, they often lack sufficient temporal constraints in the retrieval process. To tackle this challenge, we propose TempAgent, a novel autonomous agent framework built on LLMs that enhances their ability to conduct temporal reasoning and comprehension. By integrating temporal constraints into information retrieval, TempAgent effectively discards irrelevant material and concentrates on extracting pertinent temporal and factual information. We evaluate our framework on the MultiTQ dataset, a real-world multi-granularity TKGQA benchmark, using a fully automated setup. Our experimental results reveal the remarkable effectiveness of our approach: TempAgent achieves a 41.3% improvement over the baseline model and a 32.2% gain compared to the Abstract Reasoning Induction (ARI) method. Moreover, our method attains an accuracy of 70.2% on the @hit1 metric, underscoring its substantial advantage in addressing time-aware TKGQA tasks.

“查询改写是一种通过优化查询从而提高检索结果质量的技术。传统的基于伪相关反馈的方法受限于伪相关文档的质量。本文提出了一种基于ChatGPT查询改写的文档检索方法。这种方法不依赖伪相关文档,可以避免伪相关文档质量不高的问题。首先,利用BM25模型进行检索,获得初次检索结果集;同时借助ChatGPT生成新查询;然后分别将原始查询和新查询作为输入,利用重排模型对初次检索结果集进行重排,得到各自的文档相关性得分;最后,将两个查询的文档相关性得分进行融合,得到最终的文档得分。在多个检索测试集上的实验结果表明,相比于基准模型,基于ChatGPT查询改写的文档检索方法在nDCG@10指标上平均提升了约4.5个百分点。”