[Paper Review] FIND: A Function Description Benchmark for Evaluating Interpretability Methods

Sarah Schwettmann, Tamar Rott Shaham+6 more|arXiv (Cornell University)|Sep 7, 2023

Explainable Artificial Intelligence (XAI)Computer Science3 citations

TL;DR

This paper introduces FIND, a benchmark for evaluating automated interpretability methods that generate natural language or code descriptions of black-box functions. Using procedurally generated functions with real-world complexities like noise and bias, it evaluates large language models (LLMs) and a novel Automated Interpretability Agent (AIA) that uses hypothesis-driven experimentation; results show LLMs capture global behavior but miss local details, highlighting the need for enhanced tools in automated interpretability.

ABSTRACT

Labeling neural network submodules with human-legible descriptions is useful for many downstream tasks: such descriptions can surface failures, guide interventions, and perhaps even explain important model behaviors. To date, most mechanistic descriptions of trained networks have involved small models, narrowly delimited phenomena, and large amounts of human labor. Labeling all human-interpretable sub-computations in models of increasing size and complexity will almost certainly require tools that can generate and validate descriptions automatically. Recently, techniques that use learned models in-the-loop for labeling have begun to gain traction, but methods for evaluating their efficacy are limited and ad-hoc. How should we validate and compare open-ended labeling tools? This paper introduces FIND (Function INterpretation and Description), a benchmark suite for evaluating the building blocks of automated interpretability methods. FIND contains functions that resemble components of trained neural networks, and accompanying descriptions of the kind we seek to generate. The functions span textual and numeric domains, and involve a range of real-world complexities. We evaluate methods that use pretrained language models (LMs) to produce descriptions of function behavior in natural language and code. Additionally, we introduce a new interactive method in which an Automated Interpretability Agent (AIA) generates function descriptions. We find that an AIA, built from an LM with black-box access to functions, can infer function structure, acting as a scientist by forming hypotheses, proposing experiments, and updating descriptions in light of new data. However, AIA descriptions tend to capture global function behavior and miss local details. These results suggest that FIND will be useful for evaluating more sophisticated interpretability methods before they are applied to real-world models.

Motivation & Objective

• To address the lack of standardized evaluation for automated interpretability methods that generate function descriptions.
• To create a benchmark for assessing how well models can infer and describe the behavior of black-box functions.
• To evaluate the effectiveness of large language models (LLMs) and a new interactive agent (AIA) in generating accurate, human-readable descriptions of functions.
• To identify limitations in current LLM-based approaches, particularly in capturing local function details despite global understanding.
• To guide future development of automated interpretability tools by characterizing performance on realistic, complex functions.

Proposed method

• FIND is a procedurally generated benchmark containing over 2,000 function interpretation problems across numeric, textual, and synthetic neural modules.
• Functions are designed with real-world complexities such as composition, approximation, bias, and noise to simulate challenges in real model interpretability.
• Baseline evaluation uses LLMs to generate both natural language and code-based descriptions of function behavior, compared against ground-truth implementations.
• An Automated Interpretability Agent (AIA) is introduced, which uses an LLM to iteratively hypothesize, select inputs, observe outputs, and refine descriptions—mimicking the scientific method.
• Evaluation uses metrics like string-matching for string functions and structural similarity for code, with success scores per function subcategory.
• The AIA method enables interactive probing, allowing the agent to generate data and update hypotheses dynamically, unlike static captioning baselines.

Experimental results

Research questions

• RQ1 Can LLM-based methods accurately describe the behavior of complex, black-box functions with real-world complexities such as noise and bias?
• RQ2 How effective is an interactive Automated Interpretability Agent (AIA) that uses hypothesis testing and data-driven refinement to infer function structure?
• RQ3 To what extent do LLMs capture global function behavior versus missing local or edge-case behaviors?
• RQ4 How do different LLMs (e.g., GPT-3.5 vs. GPT-4) compare in their ability to interpret diverse function types?
• RQ5 What are the limitations of off-the-shelf LLMs in interpreting highly specific or biased functions, and how can they be addressed?

Key findings

• GPT-4 and GPT-3.5 show comparable performance in interpreting entity-based functions, indicating that performance is not solely dependent on model size.
• LLM-based descriptions often capture the global behavior of functions but fail to detect local corruptions or edge cases, such as region-specific exceptions.
• The AIA method, which uses iterative hypothesis testing and data generation, outperforms static captioning baselines by enabling deeper exploration of function behavior.
• String functions are more reliably interpreted than complex or biased functions, with success scores varying significantly across subcategories.
• The benchmark reveals that off-the-shelf LLMs struggle with highly specific reference entities (e.g., 'The New York Times'), suggesting a need for better exemplar initialization strategies.
• Complex functions are generally harder to interpret than atomic functions, and success rates drop significantly with increasing composition and bias.