Find Eval

find-eval

FIND: A Function Description Benchmark for Evaluating Interpretability Methods — Schwettmann et al. (2023) (arXiv:2309.03886, 2023)

What this evaluates

Evaluates automated interpretability methods on their ability to describe black-box functions across numeric, string, and semantic domains. It probes both static language model capabilities and interactive agent reasoning, including handling complexities like composition, noise, bias, and approximation.

Datasets

• FIND — total 2000; splits: test (-1); repo https://github.com/multimodal-interpretability/FIND

Metrics

• adequately described rate (primary) — range: percent
  • Percentage of functions for which the generated description adequately captures the function's behavior compared to human-labeled concepts. Calculated as (adequately described functions / total functions) * 100.

Input / output format

Input: Black-box functions (numeric, string, or semantic) provided as executable code or API calls. For initialization experiments, 10 exemplar inputs per function are provided (2 associated with the function, 8 distractors).

Output: A textual description of the function's behavior, or an interactive sequence of hypotheses, experiments, and refinements from an Automated Interpretability Agent (AIA).

Scoring recipe

def score(predictions, gold):
    adequate_count = 0
    for pred, gold_concepts in zip(predictions, gold):
        if matches_human_concepts(pred, gold_concepts):
            adequate_count += 1
    return (adequate_count / len(predictions)) * 100

Common pitfalls

• Interpreters may hallucinate function outputs instead of actually querying the black-box function.
• Providing initialization exemplars with many distractor inputs can severely handicap the interpreter's search strategy.
• Agents may exhibit high-level scientific reasoning but still fail to capture local function details or polysemantic behaviors.

Evidence (verbatim from paper)

FIND introduces a procedural benchmark of over 2000 black-box function interpretation problems across numeric, string, and semantic domains, designed to evaluate automated interpretability methods. It assesses both static (LM-based) and interactive (Automated Interpretability Agent, AIA) approaches, revealing that while AIAs exhibit scientific reasoning—proposing experiments, updating hypotheses, and refining descriptions—they still fail to capture local function details, with 48% of functions inadequately described.

Citation

@misc{schwettmann2023find,
  title={FIND: A Function Description Benchmark for Evaluating Interpretability Methods},
  author={Schwettmann et al. (2023)},
  year={2023},
  note={arXiv:2309.03886}
}

• arXiv: 2309.03886