Yu Ge

Zhensheng Jin, Xinze Li, Yifan Ji et al.

Recent advances in Chain-of-Thought (CoT) prompting have substantially improved the reasoning capabilities of Large Language Models (LLMs). However, these methods often suffer from overthinking, leading to unnecessarily lengthy or redundant reasoning traces. Existing approaches attempt to mitigate this issue through curating multiple reasoning chains for training LLMs, but their effectiveness is often constrained by the quality of the generated data and prone to overfitting. To address the challenge, we propose Reasoning Compression ThroUgh Stepwise Trials (ReCUT), a novel method aimed at balancing the accuracy and length of reasoning trajectory. Specifically, ReCUT employs a stepwise exploration mechanism and a long-short switched sampling strategy, enabling LLMs to incrementally generate diverse reasoning paths. These paths are evaluated and used to construct preference pairs to train two specialized models (Gemini LLMs)-one optimized for reasoning accuracy, the other for shorter reasoning. A final integrated model is obtained by interpolating the parameters of these two models. Experimental results across multiple math reasoning datasets and backbone models demonstrate that ReCUT significantly reduces reasoning lengths by approximately 30-50%, while maintaining or improving reasoning accuracy compared to various baselines. All codes and data will be released via https://github.com/NEUIR/ReCUT.

Mingyan Wu, Zhenghao Liu, Yukun Yan et al.

Retrieval-Augmented Generation (RAG) enhances the performance of Large Language Models (LLMs) by incorporating external knowledge. However, LLMs still encounter challenges in effectively utilizing the knowledge from retrieved documents, often being misled by irrelevant or noisy information. To address this issue, we introduce RankCoT, a knowledge refinement method that incorporates reranking signals in generating CoT-based summarization for knowledge refinement based on given query and all retrieval documents. During training, RankCoT prompts the LLM to generate Chain-of-Thought (CoT) candidates based on the query and individual documents. It then fine-tunes the LLM to directly reproduce the best CoT from these candidate outputs based on all retrieved documents, which requires LLM to filter out irrelevant documents during generating CoT-style summarization. Additionally, RankCoT incorporates a self-reflection mechanism that further refines the CoT outputs, resulting in higher-quality training data. Our experiments demonstrate the effectiveness of RankCoT, showing its superior performance over other knowledge refinement models. Further analysis reveals that RankCoT can provide shorter but effective refinement results, enabling the generator to produce more accurate answers. All code and data are available at https://github.com/NEUIR/RankCoT.

Shuliang Liu, Xinze Li, Zhenghao Liu et al.

Retrieval-Augmented Generation (RAG) has proven its effectiveness in alleviating hallucinations for Large Language Models (LLMs). However, existing automated evaluation metrics cannot fairly evaluate the outputs generated by RAG models during training and evaluation. LLM-based judgment models provide the potential to produce high-quality judgments, but they are highly sensitive to evaluation prompts, leading to inconsistencies when judging the output of RAG models. This paper introduces the Judge-Consistency (ConsJudge) method, which aims to enhance LLMs to generate more accurate evaluations for RAG models. Specifically, ConsJudge prompts LLMs to generate different judgments based on various combinations of judgment dimensions, utilize the judge-consistency to evaluate these judgments and select the accepted and rejected judgments for DPO training. Our experiments show that ConsJudge can effectively provide more accurate judgments for optimizing RAG models across various RAG models and datasets. Further analysis reveals that judgments generated by ConsJudge have a high agreement with the superior LLM. All codes are available at https://github.com/OpenBMB/ConsJudge.

Weizhuo Li, Zhigang Wang, Yu Gu et al.

Recently the generative Large Language Model (LLM) has achieved remarkable success in numerous applications. Notably its inference generates output tokens one-by-one, leading to many redundant computations. The widely-used KV-Cache framework makes a compromise between time and space complexities. However, caching data generates the increasingly growing memory demand, that can quickly exhaust the limited memory capacity of the modern accelerator like GPUs, particularly in long-context inference tasks. Existing studies reduce memory consumption by evicting some of cached data that have less important impact on inference accuracy. But the benefit in practice is far from ideal due to the static cache allocation across different LLM network layers. This paper observes that the layer-specific cached data have very different impacts on accuracy. We quantify this difference, and give experimental and theoretical validation. We accordingly make a formal analysis and shows that customizing the cache size for each layer in a personalized manner can yield a significant memory reduction, while still providing comparable accuracy. We simulate the cache allocation as a combinatorial optimization problem and give a global optimal solution. In particular, we devise a mini- and sampling-based inference over a lightweight variant of the LLM model, so as to quickly capture the difference and then feed it into the personalized algorithms. Extensive experiments on real-world datasets demonstrate that our proposals can reduce KV cache memory consumption by 61.6% on average, improve computational efficiency by 2.1x and then increase the throughput by up to 5.5x.