Yijiong Yu

Yijiong Yu, Huazheng Wang, Shuai Yuan et al.

Speculative Decoding (SD) accelerates low-concurrency LLM inference by employing a draft-then-verify paradigm. However, mainstream methods typically rely on multi-token prediction, which introduces escalating prediction difficulty and serial drafting latency. To address these, we propose Speculative Pipeline Decoding (SPD), a groundbreaking framework that unlocks the true potential of pipeline parallelism. By partitioning the target LLM into $n$ pipeline stages, SPD allows LLM to process $n$ tokens in parallel to accelerate decoding. To continuous fill the pipeline in single sequence decoding, a speculation module aggregates intermediate features across different pipeline depths to predict the next token, executing strictly in parallel with the target model's pipeline step, to realize bounded difficulty, higher acceptance rates, and zero latency bubbles. Our experiments demonstrate that SPD achieves a significantly higher theoretical speedup compared to mainstream baselines, offering a highly scalable solution for LLM decoding acceleration. Our code is available at https://github.com/yuyijiong/speculative_pipeline_decoding

Zhixiao Qi, Yijiong Yu, Meiqi Tu et al.

Currently, the construction of large language models in specific domains is done by fine-tuning on a base model. Some models also incorporate knowledge bases without the need for pre-training. This is because the base model already contains domain-specific knowledge during the pre-training process. We build a large language model for food testing. Unlike the above approach, a significant amount of data in this domain exists in Scanning format for domain standard documents. In addition, there is a large amount of untrained structured knowledge. Therefore, we introduce an incremental pre-training step to inject this knowledge into a large language model. In this paper, we propose a method for handling structured knowledge and scanned documents in incremental pre-training. To overcome the problem of machine hallucination, we constructe a knowledge graph to serve as an external knowledge base for supporting retrieval in the large language model. It is worth mentioning that this paper is a technical report of our pre-release version, and we will report our specific experimental data in future versions.

Yaolun Zhang, Yiran Wu, Yijiong Yu et al.

Large language model (LLM) agents are increasingly equipped with memory, which are stored experience and reusable guidance that can improve task-solving performance. Recent \emph{self-evolving} systems update memory based on interaction outcomes, but most existing evolution pipelines are developed for static train/test splits and only approximate online learning by folding static benchmarks, making them brittle under true distribution shift and continuous feedback. We introduce \textsc{Live-Evo}, an online self-evolving memory system that learns from a stream of incoming data over time. \textsc{Live-Evo} decouples \emph{what happened} from \emph{how to use it} via an Experience Bank and a Meta-Guideline Bank, compiling task-adaptive guidelines from retrieved experiences for each task. To manage memory online, \textsc{Live-Evo} maintains experience weights and updates them from feedback: experiences that consistently help are reinforced and retrieved more often, while misleading or stale experiences are down-weighted and gradually forgotten, analogous to reinforcement and decay in human memory. On the live \textit{Prophet Arena} benchmark over a 10-week horizon, \textsc{Live-Evo} improves Brier score by 20.8\% and increases market returns by 12.9\%, while also transferring to deep-research benchmarks with consistent gains over strong baselines. Our code is available at https://github.com/ag2ai/Live-Evo.

Yijiong Yu, Tao Wang, Kang Ran et al.

Due to the inevitable presence of quality problems, quality inspection of remote sensing images is indeed an indispensable step between the acquisition and the application of them. However, traditional manual inspection suffers from low efficiency. Hence, we propose a novel deep learning-based two-step intelligent system consisting of multiple advanced computer vision models, which first performs image classification by SwinV2 and then accordingly adopts the most appropriate method, such as semantic segmentation by Segformer, to localize the quality problems. Results demonstrate that the proposed method exhibits excellent performance and efficiency, surpassing traditional methods. Furthermore, we conduct an initial exploration of applying multimodal models to remote sensing image quality inspection.

Yijiong Yu, Shuai Yuan, Jie Zheng et al.

Soft context compression reduces the computational workload of processing long contexts in LLMs by encoding long context into a smaller number of latent tokens. However, existing frameworks apply uniform compression ratios, failing to account for the extreme variance in natural language information density. While adopting a density-aware dynamic compression ratio seems intuitive, empirical investigations reveal that models struggle intrinsically with operations parameterized by input dependent, continuous structural hyperparameters. To resolve this pitfall, we introduce Semi-Dynamic Context Compression framework. Our approach features a Discrete Ratio Selector, which predicts a compression target based on intrinsic information density and quantizes it to a predefined set of discrete compression ratios. It is efficiently jointly trained with the compressor on synthetic data, with the summary lengths as a proxy to create labels for compression ratio prediction. Extensive evaluations confirm that our density-aware framework, utilizing mean pooling as the backbone, consistently outperforms static baselines, establishing a robust Pareto frontier for context compression techniques. Our code, data and model weights are available at https://github.com/yuyijiong/semi-dynamic-context-compress

Yijiong Yu, Jiale Liu, Qingyun Wu et al.

The quadratic complexity of self-attention in Transformer-based Large Language Models (LLMs) renders long-context inference prohibitively expensive. While Sliding Window Attention (SWA), the simplest sparse attention pattern, offers a linear-complexity alternative, naively applying it to models pretrained with Full Attention (FA) causes catastrophic long-context performance collapse due to the training-inference mismatch. To address this, we propose Sliding Window Attention Adaptation (SWAA), a plug-and-play toolkit of recipes that adapt FA models to SWA without costly pretraining. SWAA systematically combines five strategies: (1) applying SWA only during prefilling; (2) preserving "sink" tokens; (3) interleaving FA/SWA layers; (4) chain-of-thought (CoT); and (5) fine-tuning. Our experiments demonstrate that while individual methods are insufficient, specific synergistic combinations can effectively recover original long-context capabilities. After further analyzing performance-efficiency trade-offs, we identify recommended SWAA configurations for diverse scenarios, which achieve 30% to 100% speedups for long-context LLM inference with acceptable quality loss. Our code is available at https://github.com/yuyijiong/sliding-window-attention-adaptation

Yijiong Yu, Huiqiang Jiang, Xufang Luo et al.

Large Language Models (LLMs) are increasingly applied in various real-world scenarios due to their excellent generalization capabilities and robust generative abilities. However, they exhibit position bias, also known as "lost in the middle", a phenomenon that is especially pronounced in long-context scenarios, which indicates the placement of the key information in different positions of a prompt can significantly affect accuracy. This paper first explores the micro-level manifestations of position bias, concluding that attention weights are a micro-level expression of position bias. It further identifies that, in addition to position embeddings, causal attention mask also contributes to position bias by creating position-specific hidden states. Based on these insights, we propose a method to mitigate position bias by scaling this positional hidden states. Experiments on the NaturalQuestions Multi-document QA, KV retrieval, LongBench and timeline reorder tasks, using various models including RoPE models, context windowextended models, and Alibi models, demonstrate the effectiveness and generalizability of our approach. Our method can improve performance by up to 15.2% by modifying just one dimension of hidden states. Our code is available at https://aka.ms/PositionalHidden.

Yijiong Yu

It has been well-known that Chain-of-Thought can remarkably enhance LLMs' performance on complex tasks. However, because it also introduces slower inference speeds and higher computational costs, many researches have attempted to use implicit CoT, which does not need LLMs to explicitly generate the intermediate steps. However, the invisible reasoning process leaves us a doubt that, can implicit CoT really be equal to explicit CoT? Therefore, in this study, we address this question through experiments. We probe the information of intermediate steps from the model's hidden states when it is either trained or prompted to perform implicit CoT. The results surprisingly indicate that when prompted, LLMs hardly think about intermediate steps, suggesting they may just rely on experience rather than strict step-by-step reasoning. But when trained, they indeed calculate intermediate steps. Moreover, in both situations, we find the effect of using implicit CoT is susceptible to the format of the problem, reaffirming the current deficiency of implicit CoT.

Yijiong Yu, Ziyun Dai, Zekun Wang et al.

Large language models (LLMs) have demonstrated remarkable capabilities, but their success heavily relies on the quality of pretraining corpora. For Chinese LLMs, the scarcity of high-quality Chinese datasets presents a significant challenge, often limiting their performance. To address this issue, we propose the OpenCSG Chinese Corpus, a series of high-quality datasets specifically designed for LLM pretraining, post-training, and fine-tuning. This corpus includes Fineweb-edu-chinese, Fineweb-edu-chinese-v2, Cosmopedia-chinese, and Smoltalk-chinese, each with distinct characteristics: Fineweb-edu datasets focus on filtered, high-quality content derived from diverse Chinese web sources; Cosmopedia-chinese provides synthetic, textbook-style data for knowledge-intensive training; and Smoltalk-chinese emphasizes stylistic and diverse chat-format data. The OpenCSG Chinese Corpus is characterized by its high-quality text, diverse coverage across domains, and scalable, reproducible data curation processes. Additionally, we conducted extensive experimental analyses, including evaluations on smaller parameter models, which demonstrated significant performance improvements in tasks such as C-Eval, showcasing the effectiveness of the corpus for training Chinese LLMs.

Yijiong Yu, Yongfeng Huang, Zhixiao Qi et al.

As Large Language Models (LLMs) continue to evolve, more are being designed to handle long-context inputs. Despite this advancement, most of them still face challenges in accurately handling long-context tasks, often showing the "lost in the middle" issue. We identify that insufficient retrieval capability is one of the important reasons for this issue. To tackle this challenge, we propose a novel approach to design training data for long-context tasks, aiming at augmenting LLMs' proficiency in extracting key information from long context. Specially, we incorporate an additional part named "paraphrasing the original text" when constructing the answer of training samples and then fine-tuning the model. Experimenting on LongBench and NaturalQuestions Multi-document-QA dataset with models of Llama and Qwen series, our method achieves an improvement of up to 8.48% and 4.48% in average scores, respectively, showing effectiveness in improving the model's performance on long-context tasks.

Yijiong Yu

Recent advances in reasoning models have demonstrated significant improvements in accuracy by employing detailed and comprehensive reasoning processes. However, generating these lengthy reasoning sequences is computationally expensive and time-consuming. To address this inefficiency, we leverage the inherent parallelizability of certain tasks to accelerate the reasoning process. Specifically, when multiple parallel reasoning steps exist, we decode multiple tokens per forward pass via a tree-like attention mask within a single sequence, avoiding additional memory usage. Experimental results show that our method achieves up to nearly 100\% speedup in decoding while basically maintaining the answer quality.

Wenjing Yin, Tianze Sun, Yijiong Yu et al.

Large language models (LLMs) play a crucial role in software engineering, excelling in tasks like code generation and maintenance. However, existing benchmarks are often narrow in scope, focusing on a specific task and lack a comprehensive evaluation framework that reflects real-world applications. To address these gaps, we introduce CoCo-Bench (Comprehensive Code Benchmark), designed to evaluate LLMs across four critical dimensions: code understanding, code generation, code modification, and code review. These dimensions capture essential developer needs, ensuring a more systematic and representative evaluation. CoCo-Bench includes multiple programming languages and varying task difficulties, with rigorous manual review to ensure data quality and accuracy. Empirical results show that CoCo-Bench aligns with existing benchmarks while uncovering significant variations in model performance, effectively highlighting strengths and weaknesses. By offering a holistic and objective evaluation, CoCo-Bench provides valuable insights to guide future research and technological advancements in code-oriented LLMs, establishing a reliable benchmark for the field.

Yijiong Yu

Recent advancements in the field of large language models, particularly through the Chain of Thought (CoT) approach, have demonstrated significant improvements in solving complex problems. However, existing models either tend to sacrifice detailed reasoning for brevity due to user preferences, or require extensive and expensive training data to learn complicated reasoning ability, limiting their potential in solving complex tasks. To bridge this gap, following the concept of scaling test-time, we propose a simple method by encouraging models to adopt a more patient reasoning style without the need of introducing new knowledge or skills. To employ a preference optimization approach, we generate detailed reasoning processes as positive examples and simple answers as negative examples, thereby training the model to favor thoroughness in its responses. Our results demonstrate a performance increase of up to 2.1% on GSM8k with training just on a lightweight dataset.

Yijiong Yu

Large Language Models (LLMs) have demonstrated remarkable capabilities in handling long texts and have almost perfect performance in traditional retrieval tasks. However, their performance significantly degrades when it comes to numerical calculations in the long-context. Numeric-involved long-context tasks typically cannot be addressed by current LLMs in normal settings due to their inherent limitations in simultaneously handling complex and massive information. Some CoT like prompting methods can improve accuracy but demands massive output tokens, which is costly and slow. To address this issue, we propose a workflow, which decompose a numeric-involved long-context task into 4 low-level subtasks: judging, extracting and processing with code and conclusion. The former 2 subtasks is relatively simple, which allows us to use smaller models for efficiently processing long context. When numerical calculations are required, we use code generated by LLMs to avoid the disadvantage of LLM not being good at calculations. The results in 2 numeric-involved long-context benchmarks demonstrate our workflow can not only improve accuracy, but also significantly reduce the cost of API calls.