Neural models predict choices well but vary in insights from removing information

Better Behavioral Prediction, More Faithful Model Ablations? Evidence from Sequential Choice

Computation and Language

Summary

Predicting how people make choices can help understand their thinking, but just knowing a model predicts well doesn’t mean it shows why choices happen. The authors tested different AI models on tasks where past actions influence future choices, comparing how these models react when some information is removed. They found that even accurate models sometimes react less like the real process behind the choices. This means it’s important to check how models change with less information before assuming those changes reveal how people decide.

What this means in practice

  • For cognitive modelers: Evaluate model changes under information removal to improve explanations of decision-making behaviors.
  • For ai system developers: Use ablation tests to verify which parts of predictive models truly reflect underlying decision processes in sequential tasks.

Authors

Hanbo Xie

Abstract

Using predictive models to explain cognition requires more than accurate behavioral predictions. Input ablations offer an appealing route: remove information from a model and interpret the resulting performance change as evidence of its importance for behavior. Yet this inference assumes that the model's dependence on information reflects the dependence of the process generating the behavior. We test it in two synthetic sequential bandit tasks with known generating policies, where past choices can remain informative when feedback is unavailable to a predictor. We compare GRUs and Transformers trained from scratch, a fine-tuned LLaMA model, and cognitive models across systematically varied reward contributions. Our analyses distinguish prediction after training without reward observations from the response of a fixed predictor to donor-reward replacement. Three findings emerge. First, in the restless task, neural models trained without rewards predict held-out choices better than four simple training-fitted behavioral baselines. Second, under matched donor replacement, accurate predictors can respond much less than the known generator. Third, at some reward weights, neural networks predict better than a pooled reinforcement-learning model but have less faithful changes in choice probabilities; the model ordering differs between the two tasks. These independent-test results separate information sufficient for prediction from response fidelity under a specified ablation in sequential choice. They motivate validating model-ablation responses independently of predictive performance before using them to infer how the observed behavior was generated.