Yuzhuang Xu

Yuzhuang Xu, Shuo Wang, Peng Li et al. · ibm-research, tsinghua

Communication games, which we refer to as incomplete information games that heavily depend on natural language communication, hold significant research value in fields such as economics, social science, and artificial intelligence. In this work, we explore the problem of how to engage large language models (LLMs) in communication games, and in response, propose a tuning-free framework. Our approach keeps LLMs frozen, and relies on the retrieval and reflection on past communications and experiences for improvement. An empirical study on the representative and widely-studied communication game, ``Werewolf'', demonstrates that our framework can effectively play Werewolf game without tuning the parameters of the LLMs. More importantly, strategic behaviors begin to emerge in our experiments, suggesting that it will be a fruitful journey to engage LLMs in communication games and associated domains.

Yuzhuang Xu, Shuo Wang, Peng Li et al. · tsinghua

Although neural machine translation (NMT) models perform well in the general domain, it remains rather challenging to control their generation behavior to satisfy the requirement of different users. Given the expensive training cost and the data scarcity challenge of learning a new model from scratch for each user requirement, we propose a memory-augmented adapter to steer pretrained NMT models in a pluggable manner. Specifically, we construct a multi-granular memory based on the user-provided text samples and propose a new adapter architecture to combine the model representations and the retrieved results. We also propose a training strategy using memory dropout to reduce spurious dependencies between the NMT model and the memory. We validate our approach on both style- and domain-specific experiments and the results indicate that our method can outperform several representative pluggable baselines.

Baocai Shan, Yuzhuang Xu, Wanxiang Che · tsinghua

Mobile input method editors (IMEs) are the primary interface for text input, yet they remain constrained to manual typing and struggle to produce personalized text. While lightweight large language models (LLMs) make on-device auxiliary generation feasible, enabling deeply personalized, privacy-preserving, and real-time generative IMEs poses fundamental challenges.To this end, we present HUOZIIME, a personalized on-device IME powered by LLM. We endow HUOZIIME with initial human-like prediction ability by post-training a base LLM on synthesized personalization data. Notably, a hierarchical memory mechanism is designed to continually capture and leverage user-specific input history. Furthermore, we perform systemic optimizations tailored to on-device LLMbased IME deployment, ensuring efficient and responsive operation under mobile constraints.Experiments demonstrate efficient on-device execution and high-fidelity memory-driven personalization. Code and package are available at https://github.com/Shan-HIT/HuoziIME.

Yuzhuang Xu, Xu Han, Yuxuan Li et al.

Large language models (LLMs) achieve strong performance but incur high deployment costs, motivating extremely low-bit but lossy quantization. Existing quantization algorithms mainly focus on improving the numerical accuracy of forward computation to eliminate performance degradation. In this paper, we show that extremely quantized LLMs suffer from systematic smoothness degradation beyond numerical precision loss. Through a smoothness proxy, we observe that such degradation becomes increasingly severe as the quantization bit-width decreases. Furthermore, based on sequence neighborhood modeling, we find that quantized models exhibit a rapid reduction of effective token candidates within the prediction neighborhood, which directly leads to a sparser decoding tree and degraded generation quality. To validate it, we introduce a simple smoothness-preserving principle in both post-training quantization and quantization-aware training, and demonstrate that preserving smoothness brings additional gains beyond numerical accuracy. The core goal of this paper is to highlight smoothness preservation as an important design consideration for future extreme quantization methods. Code is available at https://github.com/xuyuzhuang11/FINE.

Xiaolong Wang, Yuanchi Zhang, Ziyue Wang et al. · tsinghua

Large language models (LLMs) have revolutionized the field of natural language processing, enabling remarkable progress in various tasks. Different from objective tasks such as commonsense reasoning and arithmetic question-answering, the performance of LLMs on subjective tasks is still limited, where the perspective on the specific problem plays crucial roles for better interpreting the context and giving proper response. For example, in certain scenarios, LLMs may perform better when answering from an expert role perspective, potentially eliciting their relevant domain knowledge. In contrast, in some scenarios, LLMs may provide more accurate responses when answering from a third-person standpoint, enabling a more comprehensive understanding of the problem and potentially mitigating inherent biases. In this paper, we propose Reasoning through Perspective Transition (RPT), a method based on in-context learning that enables LLMs to dynamically select among direct, role, and third-person perspectives for the best way to solve corresponding subjective problem. Through extensive experiments on totally 12 subjective tasks by using both closed-source and open-source LLMs including GPT-4, GPT-3.5, Llama-3, and Qwen-2, our method outperforms widely used single fixed perspective based methods such as chain-of-thought prompting and expert prompting, highlights the intricate ways that LLMs can adapt their perspectives to provide nuanced and contextually appropriate responses for different problems.

Haoyu Wang, Shuo Wang, Yukun Yan et al. · tsinghua

Open-source large language models (LLMs) have gained significant strength across diverse fields. Nevertheless, the majority of studies primarily concentrate on English, with only limited exploration into the realm of multilingual abilities. In this work, we therefore construct an open-source multilingual supervised fine-tuning dataset. Different from previous works that simply translate English instructions, we consider both the language-specific and language-agnostic abilities of LLMs. Firstly, we introduce a knowledge-grounded data augmentation approach to elicit more language-specific knowledge of LLMs, improving their ability to serve users from different countries. Moreover, we find modern LLMs possess strong cross-lingual transfer capabilities, thus repeatedly learning identical content in various languages is not necessary. Consequently, we can substantially prune the language-agnostic supervised fine-tuning (SFT) data without any performance degradation, making multilingual SFT more efficient. The resulting UltraLink dataset comprises approximately 1 million samples across five languages (i.e., En, Zh, Ru, Fr, Es), and the proposed data construction method can be easily extended to other languages. UltraLink-LM, which is trained on UltraLink, outperforms several representative baselines across many tasks.

Yuzhuang Xu, Xu Han, Yuxuan Li et al.

Although existing frameworks for large language model (LLM) inference on CPUs are mature, they fail to fully exploit the computation potential of many-core CPU platforms. Many-core CPUs are widely deployed in web servers and high-end networking devices, and are typically organized into multiple NUMA nodes that group cores and memory. Current frameworks largely overlook the substantial overhead of cross-NUMA memory access, limiting inference scalability and intelligence enabling on such platforms. To address this limitation, we build ArcLight, a lightweight LLM inference architecture designed from the ground up for many-core CPUs. ArcLight integrates efficient memory management and thread scheduling, and introduces finely controlled tensor parallelism to mitigate the cross-node memory access wall. Experimental results show that ArcLight significantly surpasses the performance ceiling of mainstream frameworks, achieving up to 46% higher inference throughput. Moreover, ArcLight maintains compatibility with arbitrary CPU devices. ArcLight is publicly available at https://github.com/OpenBMB/ArcLight.

Yijun Liu, Yixuan Wang, Yuzhuang Xu et al. · tsinghua

Large language models (LLMs) utilize key-value (KV) cache to store historical information during sequence processing. The size of KV cache grows linearly as the length of the sequence extends, which seriously affects memory usage and decoding efficiency. Current methods for KV cache eviction typically utilize the last window from the pre-filling phase as queries to compute the KV importance scores for eviction. Although this scheme is simple to implement, it tends to overly focus on local information, potentially leading to the neglect or omission of crucial global information. To mitigate this issue, we propose Judge Q, a novel training method which incorporates a soft token list. This method only tunes the model's embedding layer at a low training cost. By concatenating the soft token list at the end of the input sequence, we train these tokens' attention map to the original input sequence to align with that of the actual decoded tokens. In this way, the queries corresponding to the soft tokens can effectively capture global information and better evaluate the importance of the keys and values within the KV cache, thus maintaining decoding quality when KV cache is evicted. Under the same eviction budget, our method exhibits less performance degradation compared to existing eviction approaches. We validate our approach through experiments conducted on models such as Llama-3.1-8B-Instruct and Mistral-7B-Instruct-v0.3, using benchmarks including LongBench, RULER, and Needle-in-a-Haystack. Results indicate an improvement of approximately 1 point on the LongBench and over 3 points on RULER. This proposed methodology can be seamlessly integrated into existing open-source models with minimal training overhead, thereby enhancing performance in KV cache eviction scenarios.

Yuzhuang Xu, Xu Han, Zonghan Yang et al. · tsinghua

Model quantification uses low bit-width values to represent the weight matrices of existing models to be quantized, which is a promising approach to reduce both storage and computational overheads of deploying highly anticipated LLMs. However, current quantization methods suffer severe performance degradation when the bit-width is extremely reduced, and thus focus on utilizing 4-bit or 8-bit values to quantize models. This paper boldly quantizes the weight matrices of LLMs to 1-bit, paving the way for the extremely low bit-width deployment of LLMs. For this target, we introduce a 1-bit model compressing framework named OneBit, including a novel 1-bit parameter representation method to better quantize LLMs as well as an effective parameter initialization method based on matrix decomposition to improve the convergence speed of the quantization framework. Sufficient experimental results indicate that OneBit achieves good performance (at least 81% of the non-quantized performance on LLaMA models) with robust training processes when only using 1-bit weight matrices.

Kaiyu Huang, Fengran Mo, Xinyu Zhang et al. · tsinghua

The rapid development of Large Language Models (LLMs) demonstrates remarkable multilingual capabilities in natural language processing, attracting global attention in both academia and industry. To mitigate potential discrimination and enhance the overall usability and accessibility for diverse language user groups, it is important for the development of language-fair technology. Despite the breakthroughs of LLMs, the investigation into the multilingual scenario remains insufficient, where a comprehensive survey to summarize recent approaches, developments, limitations, and potential solutions is desirable. To this end, we provide a survey with multiple perspectives on the utilization of LLMs in the multilingual scenario. We first rethink the transitions between previous and current research on pre-trained language models. Then we introduce several perspectives on the multilingualism of LLMs, including training and inference methods, information retrieval, model security, multi-domain with language culture, and usage of datasets. We also discuss the major challenges that arise in these aspects, along with possible solutions. Besides, we highlight future research directions that aim at further enhancing LLMs with multilingualism. The survey aims to help the research community address multilingual problems and provide a comprehensive understanding of the core concepts, key techniques, and latest developments in multilingual natural language processing based on LLMs.

Yuzhuang Xu, Shiyu Ji, Qingfu Zhu et al. · tsinghua

Powerful large language models (LLMs) are increasingly expected to be deployed with lower computational costs, enabling their capabilities on resource-constrained devices. Post-training quantization (PTQ) has emerged as a star approach to achieve this ambition, with best methods compressing weights to less than 2 bit on average. In this paper, we propose Channel-Relaxed Vector Quantization (CRVQ), a novel technique that significantly improves the performance of PTQ baselines at the cost of only minimal additional bits. This state-of-the-art extreme compression method achieves its results through two key innovations: (1) carefully selecting and reordering a very small subset of critical weight channels, and (2) leveraging extended codebooks to relax the constraint of critical channels. With our method, we demonstrate a 38.9\% improvement over the current strongest sub-2-bit PTQ baseline, enabling nearer lossless 1-bit compression. Furthermore, our approach offers flexible customization of quantization bit-width and performance, providing a wider range of deployment options for diverse hardware platforms.

Yuzhuang Xu, Xu Han, Yuanchi Zhang et al. · tsinghua

Large Language Models (LLMs) with Mixture-of-Experts (MoE) architectures are distinguished by their strong performance scaling with increasing parameters across a wide range of tasks, yet they also suffer from substantial computational and storage overheads. Notably, the performance gains of MoE models do not scale proportionally with the growth in expert parameters. While prior works attempt to reduce parameters via expert-level pruning, merging, or decomposition, they still suffer from challenges in both performance and computational efficiency. In this paper, we address these challenges by introducing micro-expert as a finer-grained compression unit that spans across matrices. We first establish a more fundamental perspective, viewing MoE layers as mixtures of micro-experts, and present CAMERA, a lightweight and training-free framework for identifying micro-expert redundancy. Our analysis uncovers significant variance in micro-expert contributions during decoding. Based on this insight, we further propose CAMERA-P, a structured micro-expert pruning framework, and CAMERA-Q, a mixed-precision quantization idea designed for micro-experts. Extensive experiments on nine downstream tasks show that CAMERA-P consistently outperforms strong baselines under pruning ratios ranging from 20% to 60%. Furthermore, CAMERA-Q achieves superior results under aggressive 2-bit quantization, surpassing existing matrix- and channel-level ideas. Notably, our method enables complete micro-expert analysis of Qwen2-57B-A14B in less than 5 minutes on a single NVIDIA A100-40GB GPU.

Yixuan Wang, Shiyu Ji, Yijun Liu et al. · tsinghua

Large language models (LLMs) rely on key-value cache (KV cache) to accelerate decoding by reducing redundant computations. However, the KV cache memory usage grows substantially with longer text sequences, posing challenges for efficient deployment. Existing KV cache eviction methods prune tokens using prefilling-stage attention scores, causing inconsistency with actual inference queries, especially under tight memory budgets. In this paper, we propose Lookahead Q-Cache (LAQ), a novel eviction framework that generates low-cost pseudo lookahead queries to better approximate the true decoding-stage queries. By using these lookahead queries as the observation window for importance estimation, LAQ achieves more consistent and accurate KV cache eviction aligned with real inference scenarios. Experimental results on LongBench and Needle-in-a-Haystack benchmarks show that LAQ outperforms existing methods across various budget levels, achieving a 1 $\sim$ 4 point improvement on LongBench under limited cache budget. Moreover, LAQ is complementary to existing approaches and can be flexibly combined to yield further improvements.

Yixuan Wang, Yijun Liu, Shiyu ji et al. · tsinghua

Large language models (LLMs) suffer from high inference latency due to the auto-regressive decoding process. Speculative decoding accelerates inference by generating multiple draft tokens using a lightweight model and verifying them in parallel. However, existing verification methods rely heavily on distributional consistency while overlooking semantic correctness, thereby limiting the potential speedup of speculative decoding. While some methods employ additional models for relaxed verification of draft tokens, they often fail to generalize effectively to more diverse or open-domain settings. In this work, we propose Reflective Verification, a training-free and semantics-aware approach that achieves a better trade-off between correctness and efficiency. Specifically, we leverage the inherent reflective capacity of LLMs to semantically assess the correctness of draft tokens in parallel during verification. Using prompt-based probing, we obtain both the original and reflective distributions of draft tokens in a single forward pass. The fusion of these distributions enables semantic-level verification of draft tokens that incorporates both consistency and correctness. Experiments across multiple domain benchmarks and model scales demonstrate that our method significantly increases the acceptance length of draft tokens without compromising model performance. Furthermore, we find that the proposed Reflective Verification is orthogonal to existing statistical verification methods, and their combination yields additional 5$\sim$15\% improvements in decoding speed.

Bowen Ping, Shuo Wang, Hanqing Wang et al.

Fine-tuning is a crucial process for adapting large language models (LLMs) to diverse applications. In certain scenarios, such as multi-tenant serving, deploying multiple LLMs becomes necessary to meet complex demands. Recent studies suggest decomposing a fine-tuned LLM into a base model and corresponding delta weights, which are then compressed using low-rank or low-bit approaches to reduce costs. In this work, we observe that existing low-rank and low-bit compression methods can significantly harm the model performance for task-specific fine-tuned LLMs (e.g., WizardMath for math problems). Motivated by the long-tail distribution of singular values in the delta weights, we propose a delta quantization approach using mixed-precision. This method employs higher-bit representation for singular vectors corresponding to larger singular values. We evaluate our approach on various fine-tuned LLMs, including math LLMs, code LLMs, chat LLMs, and even VLMs. Experimental results demonstrate that our approach performs comparably to full fine-tuned LLMs, surpassing both low-rank and low-bit baselines by a considerable margin. Additionally, we show that our method is compatible with various backbone LLMs, such as Llama-2, Llama-3, and Mistral, highlighting its generalizability.