Tracking Equivalent Mechanistic Interpretations Across Neural Networks

Mechanistic interpretability (MI) is an emerging framework for interpreting neural networks. Given a task and model, MI aims to discover a succinct algorithmic process, an interpretation, that explains the model's decision process on that task. However, MI is difficult to scale and generalize. This stems in part from two key challenges: there is no precise notion of a valid interpretation; and, generating interpretations is often an ad hoc process. In this paper, we address these challenges by defining and studying the problem of interpretive equivalence: determining whether two different models share a common interpretation, without requiring an explicit description of what that interpretation is. At the core of our approach, we propose and formalize the principle that two interpretations of a model are equivalent if all of their possible implementations are also equivalent. We develop an algorithm to estimate interpretive equivalence and case study its use on Transformer-based models. To analyze our algorithm, we introduce necessary and sufficient conditions for interpretive equivalence based on models' representation similarity. We provide guarantees that simultaneously relate a model's algorithmic interpretations, circuits, and representations. Our framework lays a foundation for the development of more rigorous evaluation methods of MI and automated, generalizable interpretation discovery methods.

Circuit Stability Characterizes Language Model Generalization

Extensively evaluating the capabilities of (large) language models is difficult. Rapid development of state-of-the-art models induce benchmark saturation, while creating more challenging datasets is labor-intensive. Inspired by the recent developments in mechanistic interpretability, we introduce circuit stability as a new way to assess model performance. Circuit stability refers to a model's ability to apply a consistent reasoning process-its circuit-across various inputs. We mathematically formalize circuit stability and circuit equivalence. Then, through three case studies, we empirically show that circuit stability and the lack thereof can characterize and predict different aspects of generalization. Our proposed methods offer a step towards rigorously relating the generality of models to their interpretability.

Algorithmic Phase Transitions in Language Models: A Mechanistic Case Study of Arithmetic

Zero-shot capabilities of large language models make them powerful tools for solving a range of tasks without explicit training. It remains unclear, however, how these models achieve such performance, or why they can zero-shot some tasks but not others. In this paper, we shed some light on this phenomenon by defining and investigating algorithmic stability in language models -- changes in problem-solving strategy employed by the model as a result of changes in task specification. We focus on a task where algorithmic stability is needed for generalization: two-operand arithmetic. Surprisingly, we find that Gemma-2-2b employs substantially different computational models on closely related subtasks, i.e. four-digit versus eight-digit addition. Our findings suggest that algorithmic instability may be a contributing factor to language models' poor zero-shot performance across certain logical reasoning tasks, as they struggle to abstract different problem-solving strategies and smoothly transition between them.

Achieving Domain-Independent Certified Robustness via Knowledge Continuity

We present knowledge continuity, a novel definition inspired by Lipschitz continuity which aims to certify the robustness of neural networks across input domains (such as continuous and discrete domains in vision and language, respectively). Most existing approaches that seek to certify robustness, especially Lipschitz continuity, lie within the continuous domain with norm and distribution-dependent guarantees. In contrast, our proposed definition yields certification guarantees that depend only on the loss function and the intermediate learned metric spaces of the neural network. These bounds are independent of domain modality, norms, and distribution. We further demonstrate that the expressiveness of a model class is not at odds with its knowledge continuity. This implies that achieving robustness by maximizing knowledge continuity should not theoretically hinder inferential performance. Finally, to complement our theoretical results, we present several applications of knowledge continuity such as regularization, a certification algorithm, and show that knowledge continuity can be used to localize vulnerable components of a neural network.

On the Exploration of LM-Based Soft Modular Robot Design

Recent large language models (LLMs) have demonstrated promising capabilities in modeling real-world knowledge and enhancing knowledge-based generation tasks. In this paper, we further explore the potential of using LLMs to aid in the design of soft modular robots, taking into account both user instructions and physical laws, to reduce the reliance on extensive trial-and-error experiments typically needed to achieve robot designs that meet specific structural or task requirements. Specifically, we formulate the robot design process as a sequence generation task and find that LLMs are able to capture key requirements expressed in natural language and reflect them in the construction sequences of robots. To simplify, rather than conducting real-world experiments to assess design quality, we utilize a simulation tool to provide feedback to the generative model, allowing for iterative improvements without requiring extensive human annotations. Furthermore, we introduce five evaluation metrics to assess the quality of robot designs from multiple angles including task completion and adherence to instructions, supporting an automatic evaluation process. Our model performs well in evaluations for designing soft modular robots with uni- and bi-directional locomotion and stair-descending capabilities, highlighting the potential of using natural language and LLMs for robot design. However, we also observe certain limitations that suggest areas for further improvement.

A perception centred self-driving system without HD Maps

Building a fully autonomous self-driving system has been discussed for more than 20 years yet remains unsolved. Previous systems have limited ability to scale. Their localization subsystem needs labor-intensive map recording for running in a new area, and the accuracy decreases after the changes occur in the environment. In this paper, a new localization method is proposed to solve the scalability problems, with a new method for detecting and making sense of diverse traffic lines. Like the way human drives, a self-driving system should not rely on an exact position to travel in most scenarios. As a result, without HD Maps, GPS or IMU, the proposed localization subsystem relies only on detecting driving-related features around (like lane lines, stop lines, and merging lane lines). For spotting and reasoning all these features, a new line detector is proposed and tested against multiple datasets.