Mechanistic interpretability aims to understand the inner workings of large neural networks by identifying circuits, or minimal subgraphs within the model that implement algorithms responsible for performing specific tasks. These circuits are typically discovered and analyzed using a narrowly defined prompt format. However, given the abilities of large language models (LLMs) to generalize across various prompt formats for the same task, it remains unclear how well these circuits generalize. For instance, it is unclear whether the models generalization results from reusing the same circuit components, the components behaving differently, or the use of entirely different components. In this paper, we investigate the generality of the indirect object identification (IOI) circuit in GPT-2 small, which is well-studied and believed to implement a simple, interpretable algorithm. We evaluate its performance on prompt variants that challenge the assumptions of this algorithm. Our findings reveal that the circuit generalizes surprisingly well, reusing all of its components and mechanisms while only adding additional input edges. Notably, the circuit generalizes even to prompt variants where the original algorithm should fail; we discover a mechanism that explains this which we term S2 Hacking. Our findings indicate that circuits within LLMs may be more flexible and general than previously recognized, underscoring the importance of studying circuit generalization to better understand the broader capabilities of these models.

Evaluating the creativity of large language models (LLMs) in story writing is difficult because LLM-generated stories could seemingly look creative but be very similar to some existing stories in their huge and proprietary training corpus. To overcome this challenge, we introduce a novel benchmark dataset with varying levels of prompt specificity: CS4 ($\mathbf{C}$omparing the $\mathbf{S}$kill of $\mathbf{C}$reating $\mathbf{S}$tories by $\mathbf{C}$ontrolling the $\mathbf{S}$ynthesized $\mathbf{C}$onstraint $\mathbf{S}$pecificity). By increasing the number of requirements/constraints in the prompt, we can increase the prompt specificity and hinder LLMs from retelling high-quality narratives in their training data. Consequently, CS4 empowers us to indirectly measure the LLMs' creativity without human annotations. Our experiments on LLaMA, Gemma, and Mistral not only highlight the creativity challenges LLMs face when dealing with highly specific prompts but also reveal that different LLMs perform very differently under different numbers of constraints and achieve different balances between the model's instruction-following ability and narrative coherence. Additionally, our experiments on OLMo suggest that Learning from Human Feedback (LHF) can help LLMs select better stories from their training data but has limited influence in boosting LLMs' ability to produce creative stories that are unseen in the training corpora. The benchmark is released at https://github.com/anirudhlakkaraju/cs4_benchmark.

Large Language Models (LLMs) have experienced a rapid rise in AI, changing a wide range of applications with their advanced capabilities. As these models become increasingly integral to decision-making, the need for thorough interpretability has never been more critical. Mechanistic Interpretability offers a pathway to this understanding by identifying and analyzing specific sub-networks or 'circuits' within these complex systems. A crucial aspect of this approach is Automated Circuit Discovery, which facilitates the study of large models like GPT4 or LLAMA in a feasible manner. In this context, our research evaluates a recent method, Brain-Inspired Modular Training (BIMT), designed to enhance the interpretability of neural networks. We demonstrate how BIMT significantly improves the efficiency and quality of Automated Circuit Discovery, overcoming the limitations of manual methods. Our comparative analysis further reveals that BIMT outperforms existing models in terms of circuit quality, discovery time, and sparsity. Additionally, we provide a comprehensive computational analysis of BIMT, including aspects such as training duration, memory allocation requirements, and inference speed. This study advances the larger objective of creating trustworthy and transparent AI systems in addition to demonstrating how well BIMT works to make neural networks easier to understand.