MIB: A Mechanistic Interpretability Benchmark

• URL: http://arxiv.org/abs/2504.13151v1
• Date: Thu, 17 Apr 2025 17:55:45 GMT
• Title: MIB: A Mechanistic Interpretability Benchmark
• Authors: Aaron Mueller, Atticus Geiger, Sarah Wiegreffe, Dana Arad, Iván Arcuschin, Adam Belfki, Yik Siu Chan, Jaden Fiotto-Kaufman, Tal Haklay, Michael Hanna, Jing Huang, Rohan Gupta, Yaniv Nikankin, Hadas Orgad, Nikhil Prakash, Anja Reusch, Aruna Sankaranarayanan, Shun Shao, Alessandro Stolfo, Martin Tutek, Amir Zur, David Bau, Yonatan Belinkov,
• Abstract summary: We propose MIB, a benchmark with two tracks spanning four tasks and five models.<n>Using MIB, we find that attribution and mask optimization methods perform best on circuit localization.<n>For causal variable localization, we find that the supervised DAS method performs best, while SAE features are not better than neurons.
• Score: 77.35046700898326
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: How can we know whether new mechanistic interpretability methods achieve real improvements? In pursuit of meaningful and lasting evaluation standards, we propose MIB, a benchmark with two tracks spanning four tasks and five models. MIB favors methods that precisely and concisely recover relevant causal pathways or specific causal variables in neural language models. The circuit localization track compares methods that locate the model components - and connections between them - most important for performing a task (e.g., attribution patching or information flow routes). The causal variable localization track compares methods that featurize a hidden vector, e.g., sparse autoencoders (SAEs) or distributed alignment search (DAS), and locate model features for a causal variable relevant to the task. Using MIB, we find that attribution and mask optimization methods perform best on circuit localization. For causal variable localization, we find that the supervised DAS method performs best, while SAE features are not better than neurons, i.e., standard dimensions of hidden vectors. These findings illustrate that MIB enables meaningful comparisons of methods, and increases our confidence that there has been real progress in the field.

Related papers

• AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders [73.37603699731329]
  We introduce AxBench, a large-scale benchmark for steering and concept detection. For steering, we find that prompting outperforms all existing methods, followed by finetuning. For concept detection, representation-based methods such as difference-in-means, perform the best.
  arXiv  Detail & Related papers  (2025-01-28T18:51:24Z)
• Min-K%++: Improved Baseline for Detecting Pre-Training Data from Large Language Models [15.50128790503447]
  We propose a novel and theoretically motivated methodology for pre-training data detection, named Min-K%++.<n>Specifically, we present a key insight that training samples tend to be local maxima of the modeled distribution along each input dimension through likelihood training.
  arXiv  Detail & Related papers  (2024-04-03T04:25:01Z)
• MiM-ISTD: Mamba-in-Mamba for Efficient Infrared Small Target Detection [72.46396769642787]
  We develop a nested structure, Mamba-in-Mamba (MiM-ISTD), for efficient infrared small target detection. MiM-ISTD is $8 times$ faster than the SOTA method and reduces GPU memory usage by 62.2$%$ when testing on $2048 times 2048$ images.
  arXiv  Detail & Related papers  (2024-03-04T15:57:29Z)
• Towards Reliable AI Model Deployments: Multiple Input Mixup for Out-of-Distribution Detection [4.985768723667418]
  We propose a novel and simple method to solve the Out-of-Distribution (OOD) detection problem. Our method can help improve the OOD detection performance with only single epoch fine-tuning. Our method does not require training the model from scratch and can be attached to the classifier simply.
  arXiv  Detail & Related papers  (2023-12-24T15:31:51Z)
• Winning Prize Comes from Losing Tickets: Improve Invariant Learning by Exploring Variant Parameters for Out-of-Distribution Generalization [76.27711056914168]
  Out-of-Distribution (OOD) Generalization aims to learn robust models that generalize well to various environments without fitting to distribution-specific features. Recent studies based on Lottery Ticket Hypothesis (LTH) address this problem by minimizing the learning target to find some of the parameters that are critical to the task. We propose Exploring Variant parameters for Invariant Learning (EVIL) which also leverages the distribution knowledge to find the parameters that are sensitive to distribution shift.
  arXiv  Detail & Related papers  (2023-10-25T06:10:57Z)
• Mutual-Information Based Few-Shot Classification [34.95314059362982]
  We introduce Transductive Infomation Maximization (TIM) for few-shot learning. Our method maximizes the mutual information between the query features and their label predictions for a given few-shot task. We propose a new alternating-direction solver, which speeds up transductive inference over gradient-based optimization.
  arXiv  Detail & Related papers  (2021-06-23T09:17:23Z)
• MuyGPs: Scalable Gaussian Process Hyperparameter Estimation Using Local Cross-Validation [1.2233362977312945]
  We present MuyGPs, a novel efficient GP hyper parameter estimation method. MuyGPs builds upon prior methods that take advantage of the nearest neighbors structure of the data. We show that our method outperforms all known competitors both in terms of time-to-solution and the root mean squared error of the predictions.
  arXiv  Detail & Related papers  (2021-04-29T18:10:21Z)
• MetricOpt: Learning to Optimize Black-Box Evaluation Metrics [21.608384691401238]
  We study the problem of optimizing arbitrary non-differentiable task evaluation metrics such as misclassification rate and recall. Our method, named MetricOpt, operates in a black-box setting where the computational details of the target metric are unknown. We achieve this by learning a differentiable value function, which maps compact task-specific model parameters to metric observations.
  arXiv  Detail & Related papers  (2021-04-21T16:50:01Z)
• Interpretable Multi-dataset Evaluation for Named Entity Recognition [110.64368106131062]
  We present a general methodology for interpretable evaluation for the named entity recognition (NER) task. The proposed evaluation method enables us to interpret the differences in models and datasets, as well as the interplay between them. By making our analysis tool available, we make it easy for future researchers to run similar analyses and drive progress in this area.
  arXiv  Detail & Related papers  (2020-11-13T10:53:27Z)
• Neural Methods for Point-wise Dependency Estimation [129.93860669802046]
  We focus on estimating point-wise dependency (PD), which quantitatively measures how likely two outcomes co-occur. We demonstrate the effectiveness of our approaches in 1) MI estimation, 2) self-supervised representation learning, and 3) cross-modal retrieval task.
  arXiv  Detail & Related papers  (2020-06-09T23:26:15Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.