Chunyuan Deng

Feng Luo, Rui Yang, Hao Sun et al. · tencent-ai, tsinghua

Understanding human preferences is crucial for improving foundation models and building personalized AI systems. However, preferences are inherently diverse and complex, making it difficult for traditional reward models to capture their full range. While fine-grained preference data can help, collecting it is expensive and hard to scale. In this paper, we introduce Decomposed Reward Models (DRMs), a novel approach that extracts diverse human preferences from binary comparisons without requiring fine-grained annotations. Our key insight is to represent human preferences as vectors and analyze them using Principal Component Analysis (PCA). By constructing a dataset of embedding differences between preferred and rejected responses, DRMs identify orthogonal basis vectors that capture distinct aspects of preference. These decomposed rewards can be flexibly combined to align with different user needs, offering an interpretable and scalable alternative to traditional reward models. We demonstrate that DRMs effectively extract meaningful preference dimensions (e.g., helpfulness, safety, humor) and adapt to new users without additional training. Our results highlight DRMs as a powerful framework for personalized and interpretable LLM alignment. Our code is available at https://github.com/amandaluof/DRMs.

Chunyuan Deng, Xiangru Tang, Yilun Zhao et al. · gatech

Recently, large language models (LLMs) have evolved into interactive agents, proficient in planning, tool use, and task execution across a wide variety of tasks. However, without specific agent tuning, open-source models like LLaMA currently struggle to match the efficiency of GPT- 4, particularly given the scarcity of agent-tuning datasets for fine-tuning. In response, we introduce \textsc{Mimir}: a streamlined platform offering a customizable pipeline that enables users to leverage both private knowledge and publicly available, legally compliant datasets at scale for \textbf{personalized agent tuning}. Additionally, \textsc{Mimir} supports the generation of general instruction-tuning datasets from the same input. This dual capability ensures that language agents developed through the platform possess both specific agent abilities and general competencies. \textsc{Mimir} integrates these features into a cohesive end-to-end platform, facilitating everything from the uploading of personalized files to one-click agent fine-tuning.

Chunyuan Deng, Ruidi Chang, Hanjie Chen

Recent advancements in language models (LMs) have marked a shift toward the growing importance of post-training. Yet, post-training approaches such as supervised fine-tuning (SFT) do not guarantee the effective use of knowledge acquired during pretraining. We therefore introduce InfoSteer, a lightweight method that encourages parametric information utilization in LMs during post-training. Specifically, InfoSteer treats the feed-forward network (FFN) layer as associate key-value memory and promotes the use of stored memory vectors via forward-pass interventions or regularization during backpropagation. This simple guidance during post-training phase yields consistent performance improvements across diverse model families -- including Qwen, Gemma and Llama -- spanning 15 downstream tasks in both in-distribution (ID) and out-of-distribution (OOD) evaluations. Beyond performance gains, we also find that steered LMs can adaptively allocate information by placing more emphasis on generating semantically meaningful tokens, while using fewer resources on simple transition ones (e.g., `\texttt{,}' or `\texttt{and}'). Our work underscores that vanilla post-training does not fully exploit the potential gained during pre-training, and that steering LMs in latent representation space offers a promising approach to enhance both performance and interpretability. The code is available at: https://github.com/chili-lab/InfoSteer.

Chunyuan Deng, Zhiqi Li, Roy Xie et al.

Large Language Models (LLMs) are thought to struggle with arithmetic learning due to the inherent differences between language modeling and numerical computation, but concrete evidence has been lacking. This work responds to this claim through a two-side experiment. We first investigate whether LLMs leverage partial products during arithmetic learning. We find that although LLMs can identify some partial products after learning, they fail to leverage them for arithmetic tasks, conversely. We then explore how LLMs approach arithmetic symbolically by breaking tasks into subgroups, hypothesizing that difficulties arise from subgroup complexity and selection. Our results show that when subgroup complexity is fixed, LLMs treat a collection of different arithmetic operations similarly. By analyzing position-level accuracy across different training sizes, we further observe that it follows a U-shaped pattern: LLMs quickly learn the easiest patterns at the first and last positions, while progressively learning the more difficult patterns in the middle positions. This suggests that LLMs select subgroup following an easy-to-hard paradigm during learning. Our work confirms that LLMs are pure symbolic learners in arithmetic tasks and underscores the importance of understanding them deeply through subgroup-level quantification.

Ruidi Chang, Chunyuan Deng, Hanjie Chen

Superposition refers to encoding representations of multiple features within a single neuron, which is common in deep neural networks. This property allows neurons to combine and represent multiple features, enabling the model to capture intricate information and handle complex tasks. Despite promising performance, the model's interpretability has been diminished. This paper presents a novel approach to enhance model interpretability by regularizing feature superposition. We introduce SAFR, which simply applies regularizations to the loss function to promote monosemantic representations for important tokens while encouraging polysemanticity for correlated token pairs, where important tokens and correlated token pairs are identified via VMASK and attention weights respectively. We evaluate SAFR with a transformer model on two classification tasks. Experiments demonstrate the effectiveness of SAFR in improving model interpretability without compromising prediction performance. Besides, SAFR provides explanations by visualizing the neuron allocation within the intermediate layers.

Ruidi Chang, Chunyuan Deng, Hanjie Chen

Transformer-based language models have achieved state-of-the-art performance in natural language generation (NLG) tasks, yet their internal mechanisms for synthesizing task-relevant information remain insufficiently understood. While prior studies suggest that intermediate layers often yield more generalizable representations than final layers, how this generalization ability emerges and propagates across layers during training remains unclear. To address this gap, we propose InfoRidge, an information-theoretic framework, to characterize how predictive information-the mutual information between hidden representations and target outputs-varies across depth. Estimating this quantity enables us to trace the flow of task-relevant information throughout the model during training. Our experiments across various models and datasets reveal a consistent non-monotonic trend: predictive information peaks in upper-middle layers-forming a generalization ridge-before declining in final layers, reflecting a transition between generalization and memorization. To further investigate this phenomenon, we introduce residual scaling coefficients-trainable scalar parameters applied to each residual block-which serve as functional probes for assessing the relative importance of individual transformer layers. These coefficients reveal that, under distribution shift, models downweight final layers and increasingly rely on ridge layers, highlighting their role in generalization. Together, these findings offer new insights into the internal mechanisms of transformers and underscore the critical role of intermediate layers in supporting generalization.

Chunyuan Deng, Yilun Zhao, Yuzhao Heng et al.

Data contamination has garnered increased attention in the era of large language models (LLMs) due to the reliance on extensive internet-derived training corpora. The issue of training corpus overlap with evaluation benchmarks--referred to as contamination--has been the focus of significant recent research. This body of work aims to identify contamination, understand its impacts, and explore mitigation strategies from diverse perspectives. However, comprehensive studies that provide a clear pathway from foundational concepts to advanced insights are lacking in this nascent field. Therefore, we present a comprehensive survey in the field of data contamination, laying out the key issues, methodologies, and findings to date, and highlighting areas in need of further research and development. In particular, we begin by examining the effects of data contamination across various stages and forms. We then provide a detailed analysis of current contamination detection methods, categorizing them to highlight their focus, assumptions, strengths, and limitations. We also discuss mitigation strategies, offering a clear guide for future research. This survey serves as a succinct overview of the most recent advancements in data contamination research, providing a straightforward guide for the benefit of future research endeavors.