Zouying Cao

Zouying Cao, Yifei Yang, XiaoJing Li et al.

Large language models (LLMs) have gained broad applications across various domains but still struggle with hallucinations. Currently, hallucinations occur frequently in the generation of factual content and pose a great challenge to trustworthy LLMs. However, hallucination detection is hindered by the laborious and expensive manual annotation of hallucinatory content. Meanwhile, as different LLMs exhibit distinct types and rates of hallucination, the collection of hallucination datasets is inherently model-specific, which also increases the cost. To address this issue, this paper proposes a method called $\textbf{AutoHall}$ for $\underline{Auto}$matically constructing model-specific $\underline{Hall}$ucination datasets based on existing fact-checking datasets. The empirical results reveal variations in hallucination proportions and types among different models. Moreover, we introduce a zero-resource and black-box hallucination detection method based on self-contradiction to recognize the hallucination in our constructed dataset, achieving superior detection performance compared to baselines. Further analysis on our dataset provides insight into factors that may contribute to LLM hallucinations. Our codes and datasets are publicly available at https://github.com/zouyingcao/AutoHall.

Yifei Yang, Zouying Cao, Hai Zhao

Large language models (LLMs) based on transformer are witnessing a notable trend of size expansion, which brings considerable costs to both model training and inference. However, existing methods such as model quantization, knowledge distillation, and model pruning are constrained by various issues, including hardware support limitations, the need for extensive training, and alterations to the model internal structure. In this paper, we propose a concise layer-wise structured pruner called \textit{Layer Collapse (LaCo)}, in which rear model layers collapse into a prior layer, enabling a rapid reduction in model size while preserving the model structure. Comprehensive experiments show that our method maintains an average task performance of over 80\% at pruning ratios of 25-30\%, significantly outperforming existing state-of-the-art structured pruning methods. We also conduct post-training experiments to confirm that the \textit{LaCo} effectively inherits the parameters of the original model. Additionally, we perform ablation studies on various settings of \textit{LaCo}. Finally, we discuss our motivation from the perspective of layer-wise similarity and evaluate the performance of the pruned LLMs across various pruning ratios\footnote{\url{https://github.com/yangyifei729/LaCo}}.

Zhuoyun Du, Runze Wang, Huiyu Bai et al.

While natural language is the de facto communication medium for LLM-based agents, it presents a fundamental constraint. The process of downsampling rich, internal latent states into discrete tokens inherently limits the depth and nuance of information that can be transmitted, thereby hindering collaborative problem-solving. Inspired by human mind-reading, we propose Interlat (Inter-agent Latent Space Communication), a paradigm that leverages the last hidden states of an LLM as a representation of its mind for direct transmission (termed latent communication). An additional compression process further compresses latent communication via entirely latent space reasoning. Experiments demonstrate that Interlat outperforms both fine-tuned chain-of-thought (CoT) prompting and single-agent baselines, promoting more exploratory behavior and enabling genuine utilization of latent information. Further compression not only substantially accelerates inference but also maintains competitive performance through an efficient information-preserving mechanism. We position this work as a feasibility study of entirely latent space inter-agent communication, and our results highlight its potential, offering valuable insights for future research.

Yifei Yang, Zouying Cao, Qiguang Chen et al.

The development of large language models (LLMs) has significantly expanded model sizes, resulting in substantial GPU memory requirements during inference. The key and value storage of the attention map in the KV (key-value) cache accounts for more than 80\% of this memory consumption. Nowadays, most existing KV cache compression methods focus on intra-layer compression within a single Transformer layer but few works consider layer-wise compression. In this paper, we propose a plug-and-play method called \textit{KVSharer}, which shares the KV cache between layers to achieve layer-wise compression. Rather than intuitively sharing based on higher similarity, we discover a counterintuitive phenomenon: sharing dissimilar KV caches better preserves the model performance. Experiments show that \textit{KVSharer} can reduce KV cache computation by 30\%, thereby lowering memory consumption without significantly impacting model performance and it can also achieve at least 1.3 times generation acceleration. Additionally, we verify that \textit{KVSharer} is compatible with existing intra-layer KV cache compression methods, and combining both can further save memory.

Zouying Cao, Yifei Yang, Hai Zhao

Large Language Models (LLMs) suffer from huge number of parameters, which restricts their deployment on edge devices. Weight sharing is one promising solution that encourages weight reuse, effectively reducing memory usage with less performance drop. However, current weight sharing techniques primarily focus on small-scale models like BERT and employ coarse-grained sharing rules, e.g., layer-wise. This becomes limiting given the prevalence of LLMs and sharing an entire layer or block obviously diminishes the flexibility of weight sharing. In this paper, we present a perspective on head-wise shareable attention for large language models. We further propose two memory-efficient methods that share parameters across attention heads, with a specific focus on LLMs. Both of them use the same dynamic strategy to select the shared weight matrices. The first method directly reuses the pre-trained weights without retraining, denoted as $\textbf{DirectShare}$. The second method first post-trains with constraint on weight matrix similarity and then shares, denoted as $\textbf{PostShare}$. Experimental results reveal our head-wise shared models still maintain satisfactory capabilities, demonstrating the feasibility of fine-grained weight sharing applied to LLMs.

Zouying Cao, Runze Wang, Yifei Yang et al.

Large Language Model (LLM) agents have demonstrated impressive capabilities in handling complex interactive problems. Existing LLM agents mainly generate natural language plans to guide reasoning, which is verbose and inefficient. NL plans are also tailored to specific tasks and restrict agents' ability to generalize across similar tasks. To this end, we explore pseudocode-style plans (P-code Plan) to capture the structural logic of reasoning. We find that P-code Plan empowers LLM agents with stronger generalization ability and more efficiency. Inspired by this finding, we propose a pseudocode-style Planning Guided Preference Optimization method called PGPO for effective agent learning. With two planning-oriented rewards, PGPO further enhances LLM agents' ability to generate high-quality P-code Plans and subsequent reasoning. Experiments show that PGPO achieves superior performance on representative agent benchmarks and outperforms the current leading baselines. Analyses reveal the advantage of PGPO in reducing action errors and omissions during reasoning.

Yifei Yang, Zouying Cao, Xinbei Ma et al.

Training Large Language Models (LLMs) from scratch requires immense computational resources, making it prohibitively expensive. Model scaling-up offers a promising solution by leveraging the parameters of smaller models to create larger ones. However, existing depth scaling-up methods rely on empirical heuristic rules for layer duplication, which result in poorer initialization and slower convergence during continual pre-training. We propose \textbf{LESA}, a novel learnable method for depth scaling-up. By concatenating parameters from each layer and applying Singular Value Decomposition, we uncover latent patterns between layers, suggesting that inter-layer parameters can be learned. LESA uses a neural network to predict the parameters inserted between adjacent layers, enabling better initialization and faster training. Experiments show that LESA outperforms existing baselines, achieving superior performance with less than half the computational cost during continual pre-training. Extensive analyses demonstrate its effectiveness across different model sizes and tasks.