Circuit Component Reuse Across Tasks in Transformer Language Models

• URL: http://arxiv.org/abs/2310.08744v3
• Date: Mon, 6 May 2024 14:31:32 GMT
• Title: Circuit Component Reuse Across Tasks in Transformer Language Models
• Authors: Jack Merullo, Carsten Eickhoff, Ellie Pavlick,
• Abstract summary: We present evidence that insights can indeed generalize across tasks. We show that the process underlying both tasks is functionally very similar, and contains about a 78% overlap in in-circuit attention heads. Our results provide evidence that it may yet be possible to explain large language models' behavior in terms of a relatively small number of interpretable task-general algorithmic building blocks and computational components.
• Score: 32.2976613483151
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Recent work in mechanistic interpretability has shown that behaviors in language models can be successfully reverse-engineered through circuit analysis. A common criticism, however, is that each circuit is task-specific, and thus such analysis cannot contribute to understanding the models at a higher level. In this work, we present evidence that insights (both low-level findings about specific heads and higher-level findings about general algorithms) can indeed generalize across tasks. Specifically, we study the circuit discovered in Wang et al. (2022) for the Indirect Object Identification (IOI) task and 1.) show that it reproduces on a larger GPT2 model, and 2.) that it is mostly reused to solve a seemingly different task: Colored Objects (Ippolito & Callison-Burch, 2023). We provide evidence that the process underlying both tasks is functionally very similar, and contains about a 78% overlap in in-circuit attention heads. We further present a proof-of-concept intervention experiment, in which we adjust four attention heads in middle layers in order to 'repair' the Colored Objects circuit and make it behave like the IOI circuit. In doing so, we boost accuracy from 49.6% to 93.7% on the Colored Objects task and explain most sources of error. The intervention affects downstream attention heads in specific ways predicted by their interactions in the IOI circuit, indicating that this subcircuit behavior is invariant to the different task inputs. Overall, our results provide evidence that it may yet be possible to explain large language models' behavior in terms of a relatively small number of interpretable task-general algorithmic building blocks and computational components.

Related papers

• On the Similarity of Circuits across Languages: a Case Study on the Subject-verb Agreement Task [11.97725093458812]
  We study the circuits implemented by Gemma 2B for solving the subject-verb agreement task across two different languages. Both circuits are highly consistent, being mainly driven by a particular attention head writing a subject number' signal to the last residual stream. We demonstrate that this direction has a causal effect on the model predictions, effectively flipping the Spanish predicted verb number by intervening with the direction found in English.
  arXiv  Detail & Related papers  (2024-10-09T02:49:56Z)
• Circuit Compositions: Exploring Modular Structures in Transformer-Based Language Models [22.89563355840371]
  We identify and compare circuits responsible for ten modular string-edit operations within a language model. Our results indicate that functionally similar circuits exhibit both notable node overlap and cross-task faithfulness.
  arXiv  Detail & Related papers  (2024-10-02T11:36:45Z)
• Transformer Circuit Faithfulness Metrics are not Robust [0.04260910081285213]
  We measure circuit 'faithfulness' by ablating portions of the model's computation. We conclude that existing circuit faithfulness scores reflect both the methodological choices of researchers as well as the actual components of the circuit. The ultimate goal of mechanistic interpretability work is to understand neural networks, so we emphasize the need for more clarity in the precise claims being made about circuits.
  arXiv  Detail & Related papers  (2024-07-11T17:59:00Z)
• DeTra: A Unified Model for Object Detection and Trajectory Forecasting [68.85128937305697]
  Our approach formulates the union of the two tasks as a trajectory refinement problem. To tackle this unified task, we design a refinement transformer that infers the presence, pose, and multi-modal future behaviors of objects. In our experiments, we observe that ourmodel outperforms the state-of-the-art on Argoverse 2 Sensor and Open dataset.
  arXiv  Detail & Related papers  (2024-06-06T18:12:04Z)
• Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models [55.19497659895122]
  We introduce methods for discovering and applying sparse feature circuits. These are causally implicatedworks of human-interpretable features for explaining language model behaviors.
  arXiv  Detail & Related papers  (2024-03-28T17:56:07Z)
• Towards Interpretable Sequence Continuation: Analyzing Shared Circuits in Large Language Models [9.56229382432426]
  This research aims to reverse engineer transformer models into human-readable representations that implement algorithmic functions. By applying circuit interpretability analysis, we identify a key sub-circuit in both GPT-2 Small and Llama-2-7B. We show that this sub-circuit has effects on various math-related prompts, such as on intervaled circuits, Spanish number word and months continuation, and natural language word problems.
  arXiv  Detail & Related papers  (2023-11-07T16:58:51Z)
• Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small [68.879023473838]
  We present an explanation for how GPT-2 small performs a natural language task called indirect object identification (IOI) To our knowledge, this investigation is the largest end-to-end attempt at reverse-engineering a natural behavior "in the wild" in a language model.
  arXiv  Detail & Related papers  (2022-11-01T17:08:44Z)
• Distribution Matching for Heterogeneous Multi-Task Learning: a Large-scale Face Study [75.42182503265056]
  Multi-Task Learning has emerged as a methodology in which multiple tasks are jointly learned by a shared learning algorithm. We deal with heterogeneous MTL, simultaneously addressing detection, classification & regression problems. We build FaceBehaviorNet, the first framework for large-scale face analysis, by jointly learning all facial behavior tasks.
  arXiv  Detail & Related papers  (2021-05-08T22:26:52Z)
• Multi-task Supervised Learning via Cross-learning [102.64082402388192]
  We consider a problem known as multi-task learning, consisting of fitting a set of regression functions intended for solving different tasks. In our novel formulation, we couple the parameters of these functions, so that they learn in their task specific domains while staying close to each other. This facilitates cross-fertilization in which data collected across different domains help improving the learning performance at each other task.
  arXiv  Detail & Related papers  (2020-10-24T21:35:57Z)
• Probing Linguistic Features of Sentence-Level Representations in Neural Relation Extraction [80.38130122127882]
  We introduce 14 probing tasks targeting linguistic properties relevant to neural relation extraction (RE) We use them to study representations learned by more than 40 different encoder architecture and linguistic feature combinations trained on two datasets. We find that the bias induced by the architecture and the inclusion of linguistic features are clearly expressed in the probing task performance.
  arXiv  Detail & Related papers  (2020-04-17T09:17:40Z)

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.