Xuyang Ge

Rui Lin, Zhenyu Jin, Guancheng Zhou et al.

While modern transformer neural networks achieve grandmaster-level performance in chess and other reasoning tasks, their internal computation process remains largely opaque. Focusing on Leela Chess Zero (LC0), we introduce a sparse decomposition framework to interpret its internal computation by decomposing its MLP and attention modules with sparse replacement layers, which capture the primary computation process of LC0. We conduct a detailed case study showing that these pathways expose rich, interpretable tactical considerations that are empirically verifiable. We further introduce three quantitative metrics and show that LC0 exhibits parallel reasoning behavior consistent with the inductive bias of its policy head architecture. To the best of our knowledge, this is the first work to decompose the internal computation of a transformer on both MLP and attention modules for interpretability. Combining sparse replacement layers and causal interventions in LC0 provides a comprehensive understanding of advanced tactical reasoning, offering critical insights into the underlying mechanisms of superhuman systems. Our code is available at https://github.com/JacklE0niden/Leela-SAEs.

Aili Chen, Xuyang Ge, Ziquan Fu et al.

As global tourism expands and artificial intelligence technology advances, intelligent travel planning services have emerged as a significant research focus. Within dynamic real-world travel scenarios with multi-dimensional constraints, services that support users in automatically creating practical and customized travel itineraries must address three key objectives: Rationality, Comprehensiveness, and Personalization. However, existing systems with rule-based combinations or LLM-based planning methods struggle to fully satisfy these criteria. To overcome the challenges, we introduce TravelAgent, a travel planning system powered by large language models (LLMs) designed to provide reasonable, comprehensive, and personalized travel itineraries grounded in dynamic scenarios. TravelAgent comprises four modules: Tool-usage, Recommendation, Planning, and Memory Module. We evaluate TravelAgent's performance with human and simulated users, demonstrating its overall effectiveness in three criteria and confirming the accuracy of personalized recommendations.

Zhengfu He, Wentao Shu, Xuyang Ge et al.

Sparse Autoencoders (SAEs) have emerged as a powerful unsupervised method for extracting sparse representations from language models, yet scalable training remains a significant challenge. We introduce a suite of 256 SAEs, trained on each layer and sublayer of the Llama-3.1-8B-Base model, with 32K and 128K features. Modifications to a state-of-the-art SAE variant, Top-K SAEs, are evaluated across multiple dimensions. In particular, we assess the generalizability of SAEs trained on base models to longer contexts and fine-tuned models. Additionally, we analyze the geometry of learned SAE latents, confirming that \emph{feature splitting} enables the discovery of new features. The Llama Scope SAE checkpoints are publicly available at~\url{https://huggingface.co/fnlp/Llama-Scope}, alongside our scalable training, interpretation, and visualization tools at \url{https://github.com/OpenMOSS/Language-Model-SAEs}. These contributions aim to advance the open-source Sparse Autoencoder ecosystem and support mechanistic interpretability research by reducing the need for redundant SAE training.

Guancheng Zhou, Yisi Luo, Zhengfu He et al.

Most current paradigms in visual mechanistic interpretability (MI) remain confined to interpreting internal units of the vision model via heuristic methods (e.g., top-$K$ activation retrieval or optimization with regularization). In this work, we establish a theoretical distributional view for visual MI, which models the influence of a feature activation on the natural image distribution, thereby formulating a Kullback-Leibler (KL)-minimal optimization problem to model the MI task. Under this framework, statistical biases are identified within previous MI paradigms, which reveal that they may either be perceptually uninterpretable to humans (i.e., deviate from the natural image distribution), or mechanistically unfaithful to the vision models (i.e., unable to activate model features). To resolve the biases under the distributional view, we propose a model with a KL-minimal soft-constraint principle for visual MI that theoretically balances interpretability and faithfulness. We realize this principle via energy-guided diffusion posterior sampling. Extensive experiments validate the theoretical soundness of the proposed distributional view and demonstrate the practical effectiveness of our paradigm on the DINOv3 vision model.

Zhengfu He, Xuyang Ge, Qiong Tang et al.

Sparse dictionary learning has been a rapidly growing technique in mechanistic interpretability to attack superposition and extract more human-understandable features from model activations. We ask a further question based on the extracted more monosemantic features: How do we recognize circuits connecting the enormous amount of dictionary features? We propose a circuit discovery framework alternative to activation patching. Our framework suffers less from out-of-distribution and proves to be more efficient in terms of asymptotic complexity. The basic unit in our framework is dictionary features decomposed from all modules writing to the residual stream, including embedding, attention output and MLP output. Starting from any logit, dictionary feature or attention score, we manage to trace down to lower-level dictionary features of all tokens and compute their contribution to these more interpretable and local model behaviors. We dig in a small transformer trained on a synthetic task named Othello and find a number of human-understandable fine-grained circuits inside of it.

Xuyang Ge, Fukang Zhu, Wentao Shu et al.

Circuit analysis of any certain model behavior is a central task in mechanistic interpretability. We introduce our circuit discovery pipeline with Sparse Autoencoders (SAEs) and a variant called Transcoders. With these two modules inserted into the model, the model's computation graph with respect to OV and MLP circuits becomes strictly linear. Our methods do not require linear approximation to compute the causal effect of each node. This fine-grained graph identifies both end-to-end and local circuits accounting for either logits or intermediate features. We can scalably apply this pipeline with a technique called Hierarchical Attribution. We analyze three kinds of circuits in GPT-2 Small: bracket, induction, and Indirect Object Identification circuits. Our results reveal new findings underlying existing discoveries.

Xintao Wang, Jiangjie Chen, Nianqi Li et al.

In the rapidly advancing research fields such as AI, managing and staying abreast of the latest scientific literature has become a significant challenge for researchers. Although previous efforts have leveraged AI to assist with literature searches, paper recommendations, and question-answering, a comprehensive support system that addresses the holistic needs of researchers has been lacking. This paper introduces SurveyAgent, a novel conversational system designed to provide personalized and efficient research survey assistance to researchers. SurveyAgent integrates three key modules: Knowledge Management for organizing papers, Recommendation for discovering relevant literature, and Query Answering for engaging with content on a deeper level. This system stands out by offering a unified platform that supports researchers through various stages of their literature review process, facilitated by a conversational interface that prioritizes user interaction and personalization. Our evaluation demonstrates SurveyAgent's effectiveness in streamlining research activities, showcasing its capability to facilitate how researchers interact with scientific literature.

Zhengfu He, Junxuan Wang, Rui Lin et al.

We propose Low-Rank Sparse Attention (Lorsa), a sparse replacement model of Transformer attention layers to disentangle original Multi Head Self Attention (MHSA) into individually comprehensible components. Lorsa is designed to address the challenge of attention superposition to understand attention-mediated interaction between features in different token positions. We show that Lorsa heads find cleaner and finer-grained versions of previously discovered MHSA behaviors like induction heads, successor heads and attention sink behavior (i.e., heavily attending to the first token). Lorsa and Sparse Autoencoder (SAE) are both sparse dictionary learning methods applied to different Transformer components, and lead to consistent findings in many ways. For instance, we discover a comprehensive family of arithmetic-specific Lorsa heads, each corresponding to an atomic operation in Llama-3.1-8B. Automated interpretability analysis indicates that Lorsa achieves parity with SAE in interpretability while Lorsa exhibits superior circuit discovery properties, especially for features computed collectively by multiple MHSA heads. We also conduct extensive experiments on architectural design ablation, Lorsa scaling law and error analysis.

Xuyang Ge, Wentao Shu, Jiaxing Wu et al.

Language models obtain extensive capabilities through pre-training. However, the pre-training process remains a black box. In this work, we track linear interpretable feature evolution across pre-training snapshots using a sparse dictionary learning method called crosscoders. We find that most features begin to form around a specific point, while more complex patterns emerge in later training stages. Feature attribution analyses reveal causal connections between feature evolution and downstream performance. Our feature-level observations are highly consistent with previous findings on Transformer's two-stage learning process, which we term a statistical learning phase and a feature learning phase. Our work opens up the possibility to track fine-grained representation progress during language model learning dynamics.

Junxuan Wang, Xuyang Ge, Wentao Shu et al.

Transformer architectures, and their attention mechanisms in particular, form the foundation of modern large language models. While transformer models are widely believed to operate in high-dimensional hidden spaces, we show that attention outputs are confined to a surprisingly low-dimensional subspace, where about 60\% of the directions account for 99\% of the variance--a phenomenon that is consistently observed across diverse model families and datasets, and is induced by the attention output projection matrix. Critically, we find this low-rank structure as a key factor of the prevalent dead feature problem in sparse dictionary learning, where it creates a mismatch between randomly initialized features and the intrinsic geometry of the activation space. Building on this insight, we propose a subspace-constrained training method for sparse autoencoders (SAEs), initializing feature directions into the active subspace of activations. Our approach reduces dead features from 87\% to below 1\% in Attention Output SAEs with 1M features, and can further extend to other sparse dictionary learning methods. Our findings provide both new insights into the geometry of attention and practical tools for improving sparse dictionary learning in large language models.

Siyu Yuan, Jiangjie Chen, Xuyang Ge et al.

The vital role of analogical reasoning in human cognition allows us to grasp novel concepts by linking them with familiar ones through shared relational structures. Despite the attention previous research has given to word analogies, this work suggests that Large Language Models (LLMs) often overlook the structures that underpin these analogies, raising questions about the efficacy of word analogies as a measure of analogical reasoning skills akin to human cognition. In response to this, our paper introduces a task of analogical structure abduction, grounded in cognitive psychology, designed to abduce structures that form an analogy between two systems. In support of this task, we establish a benchmark called SCAR, containing 400 scientific analogies from 13 distinct fields, tailored for evaluating analogical reasoning with structure abduction. The empirical evidence underlines the continued challenges faced by LLMs, including ChatGPT and GPT-4, in mastering this task, signifying the need for future exploration to enhance their abilities.

Siyu Yuan, Jiangjie Chen, Ziquan Fu et al.

In everyday life, humans often plan their actions by following step-by-step instructions in the form of goal-oriented scripts. Previous work has exploited language models (LMs) to plan for abstract goals of stereotypical activities (e.g., "make a cake"), but leaves more specific goals with multi-facet constraints understudied (e.g., "make a cake for diabetics"). In this paper, we define the task of constrained language planning for the first time. We propose an overgenerate-then-filter approach to improve large language models (LLMs) on this task, and use it to distill a novel constrained language planning dataset, CoScript, which consists of 55,000 scripts. Empirical results demonstrate that our method significantly improves the constrained language planning ability of LLMs, especially on constraint faithfulness. Furthermore, CoScript is demonstrated to be quite effective in endowing smaller LMs with constrained language planning ability.