Theory Of Mind QA Eval

theory-of-mind-qa-eval

Evaluating Theory of Mind in Question Answering — Nematzadeh et al. (2018) (arXiv:1808.09352, 2018)

What this evaluates

Evaluates a model's ability to track first-order and second-order false beliefs, distinguishing an agent's mental state from physical reality and memory. It probes whether systems can maintain consistent world-state representations when agents hold incorrect beliefs about object locations or events.

Datasets

• Sally-Anne & Icecream Van Tasks — total ?; splits: test (-1)

Metrics

• accuracy (primary) — range: [0, 1]
  • Proportion of correctly answered questions (belief, reality, memory, or second-order belief) out of the total number of questions. Computed as exact-match or semantic equivalence against the gold answer.

Input / output format

Input: A short narrative describing a scenario with agents, objects, and state/location changes, followed by a specific question (e.g., belief, reality, memory, or second-order belief question).

Output: A direct answer to the question, typically a location name (e.g., 'basket', 'box', 'church') or a yes/no response for control questions.

Scoring recipe

def compute_accuracy(predictions, golds):
    correct = sum(1 for p, g in zip(predictions, golds) if p.strip().lower() == g.strip().lower())
    return correct / len(golds)

Common pitfalls

• Models often answer based on the current physical reality instead of the agent's false belief.
• Performance degrades significantly when irrelevant or noisy sentences are introduced to the context.
• Control questions (reality/memory) are sometimes skipped, leading to chance-level performance on the primary belief question.

Evidence (verbatim from paper)

The participants are asked the following questions: - 'Where will Sally look for her marble?' (belief question) - 'Where is the marble really?' (reality question) - 'Where was the marble in the beginning?' (memory question)

Citation

@misc{nematzadeh2018evaluating,
  title={Evaluating Theory of Mind in Question Answering},
  author={Nematzadeh et al. (2018)},
  year={2018},
  note={arXiv:1808.09352}
}

• arXiv: 1808.09352