Beyond the Black Box: Theory and Mechanism of Large Language Models

The rapid emergence of Large Language Models (LLMs) has precipitated a profound paradigm shift in Artificial Intelligence, delivering monumental engineering successes that increasingly impact modern society. However, a critical paradox persists within the current field: despite the empirical efficacy, our theoretical understanding of LLMs remains disproportionately nascent, forcing these systems to be treated largely as ``black boxes''. To address this theoretical fragmentation, this survey proposes a unified lifecycle-based taxonomy that organizes the research landscape into six distinct stages: Data Preparation, Model Preparation, Training, Alignment, Inference, and Evaluation. Within this framework, we provide a systematic review of the foundational theories and internal mechanisms driving LLM performance. Specifically, we analyze core theoretical issues such as the mathematical justification for data mixtures, the representational limits of various architectures, and the optimization dynamics of alignment algorithms. Moving beyond current best practices, we identify critical frontier challenges, including the theoretical limits of synthetic data self-improvement, the mathematical bounds of safety guarantees, and the mechanistic origins of emergent intelligence. By connecting empirical observations with rigorous scientific inquiry, this work provides a structured roadmap for transitioning LLM development from engineering heuristics toward a principled scientific discipline.

Revisiting Transformers through the Lens of Low Entropy and Dynamic Sparsity

Compression has been a critical lens to understand the success of Transformers. In the past, we have typically taken the target distribution as a criterion to evaluate a model's compression performance. Nevertheless,it often remains challenging to precisely assess how well the model achieves compression and to compare the information content of the learned distribution with that of the target distribution during compression,as the target distribution is typically unknown and entropy computation often incurs exponential cost. In this work, we explore these issues under a controlled experimental setup. We find that Transformers exhibit a unique inductive bias in data compression: beyond approaching the target distribution, they tend to favor learning lower-entropy distributions, with this tendency becoming more pronounced as the model size increases. This preference prevents Transformers from perfectly aligning with the target distribution, instead further compressing its information content. Furthermore, we show that the FFN module plays a critical role in driving this bias. In addition, while models remove informational redundancy from data during compression, they also exhibit redundancy within their parameters, which enables compression and can be characterized through dynamic sparsity. However, the dynamic sparsity patterns in Transformers, particularly in attention and FFN modules, demand further exploration. As for this, we show that larger Transformers show stronger preferences for bypassing attention computations via residual connections and have lower proportion of active neurons. Interestingly, we also find that training instability in larger models strongly correlates with sudden increases in dead neurons. Our work contributes to a deeper understanding of Transformers from the lens of entropy and dynamic sparsity.

Unveiling the Mechanisms of Explicit CoT Training: How Chain-of-Thought Enhances Reasoning Generalization

Training large language models (LLMs) with high-quality Chain-of-Thought (CoT) annotations has become a widely adopted strategy due to its significant enhancement of reasoning capabilities. To fully comprehend this approach, two questions naturally arise: (Q1) What advantages does training with CoT offer compared to training without CoT? (Q2) If there are advantages, what are the underlying mechanisms of explicit CoT training? Analyzing the advantages and mechanisms of CoT training is challenging due to the many factors involved. To address this, we conduct a detailed analysis using clear and controllable data distributions and, for the first time, reveal that CoT training offers the following advantages: (1) Training with CoT markedly improves reasoning generalization, extending it from in-distribution (ID) to both ID and out-of-distribution (OOD) scenarios, while also speeding up convergence; (2) Even when training with CoT includes a certain range of erroneous reasoning steps, it still enables the model to learn reasoning patterns, leading to systematic generalization. We further explore the underlying mechanisms from a circuit perspective: (1) The data distribution (e.g., ratio $\lambda$ and pattern) plays a crucial role in influencing the model's systematic generalization; (2) CoT training (with two-hop facts) internalizes reasoning into a two-stage generalizing circuit, where the number of stages corresponds to the explicit reasoning steps during training. Our findings elucidate the mechanisms underlying explicit CoT training and offer critical insights into tuning strategies for LLMs to achieve robust generalization.

Can Mamba Always Enjoy the "Free Lunch"?

Transformers have been the cornerstone of current Large Language Models (LLMs); however, its linear growth in overhead during inference with respect to sequence length poses challenges for modeling long sequences. In this context, Mamba has gradually attracted attention due to its constant-level size during inference and existing empirical results have shown that it can perform comparably to Transformers in sequence modeling while offering significant savings. However, one may ask that, can Mamba always enjoy the ``free lunch"? In this paper, we focus on analyzing the expressive ability of Mamba from a theoretical standpoint. First, inspired by the connection between Mamba and linear attention, we investigate potential shortcomings of the Mamba when performing the COPY operation. Our results indicate that Mamba with constant size may encounter bottlenecks when handling COPY, while it can achieve perfect performance when the size scales linearly with sequence length. Based on this observation, we analyze Mamba's ability to tackle DP problems when equipped with Chain of Thought (CoT). Our findings suggest that to solve arbitrary DP problems, the total cost of Mamba is comparable to standard and efficient Transformers. However, similar to efficient Transformers, when facing DP problems with favorable properties such as locality, Mamba can provide savings in overhead. Our results contribute to a deeper understanding of Mamba.

In-context Learning with Transformer Is Really Equivalent to a Contrastive Learning Pattern

Pre-trained large language models based on Transformers have demonstrated amazing in-context learning (ICL) abilities. Given several demonstration examples, the models can implement new tasks without any parameter updates. However, it is still an open question to understand the mechanism of ICL. In this paper, we interpret the inference process of ICL as a gradient descent process in a contrastive learning pattern. Firstly, leveraging kernel methods, we establish the relationship between gradient descent and self-attention mechanism under generally used softmax attention setting instead of linear attention setting. Then, we analyze the corresponding gradient descent process of ICL from the perspective of contrastive learning without negative samples and discuss possible improvements of this contrastive learning pattern, based on which the self-attention layer can be further modified. Finally, we design experiments to support our opinions. To the best of our knowledge, our work is the first to provide the understanding of ICL from the perspective of contrastive learning and has the potential to facilitate future model design by referring to related works on contrastive learning.