A small ablation effect does not prove weak reward use
The paper targets a common move in cognitive modeling: fit a model to behavior, remove or replace reward information in its input, and read the resulting change in predictions as evidence about how much the behavior depended on reward. The inference assumes that the fitted model's dependence on information mirrors the dependence of the process that produced the choices. The study asks whether that assumption holds when choices are sequential.
Sequential choice makes the assumption fragile. Earlier choices are themselves consequences of earlier rewards and of persistent tendencies, so they can carry predictive information about the next choice even when rewards are unavailable to the predictor. A model can therefore predict well from choice history alone, and a fixed model's response to a reward perturbation can depend on its representation, its training distribution, and the exact perturbation procedure rather than on the generating weight alone.
Known generators turn fidelity into a measurable target
The study builds two synthetic bandit tasks whose generating policies are fully specified. In the four-armed restless bandit, a reward-learning component is mixed with a signed last-choice kernel; in the spatially correlated bandit, reward generalization over a 5×5 layout is mixed with a tendency to stay near the last visited location. Seven weights vary how much each mechanism contributes, shifting behavior from choice-history-led to reward-learning-led.
Because the generator is known, the authors can compute an oracle response: the change in next-choice probabilities produced by the true generating policy under the same input operation. Fitted models are then scored on how closely their probability changes match that oracle, using a response-fidelity measure that averages the total-variation distance between the model's and the oracle's probability-response vectors over participants, trials, and donor assignments. This is a different quantity from predictive accuracy, which is scored as mean choice negative log likelihood in nats with participants weighted equally.
Training without rewards is not the same test as replacing rewards
The design separates two operations that are easy to conflate. The first is training a predictor on histories with reward observations omitted and comparing its held-out negative log likelihood with a full-input fit; this measures the predictive benefit of reward access for that pipeline. The second leaves a trained predictor's parameters frozen and swaps in another participant's recorded reward sequence, a donor assignment, while keeping the recipient's actual choices fixed. Only the second operation is the ablation whose fidelity is being validated.
The replacement is a permutation of reward sequences between participants within the same weight condition, with no participant assigned to themselves and 20 fixed donor assignments shared across model families. It is not literal deletion of reward tokens, and because all choices remain those of the original data, it is a teacher-forced one-step-ahead evaluation rather than a closed-loop rollout in an altered environment.
Neural models predict well without ever seeing rewards
In the restless task, neural models trained without reward observations predicted held-out choices better than four simple training-fitted behavioral baselines. That result establishes that choice history retains substantial predictive information in this task. It does not establish that the generating process relied little on reward, because retained predictive information and the causal role of reward in the generator are different questions.
The paper is explicit that this comparison is about the fitted pipeline, not about the generating agent. Strong choice-only performance shows that information sufficient for prediction survives when feedback is withheld; it does not identify whether the model is recovering inferred reward structure or exploiting other regularities in the choice sequence.
Accurate predictors can respond far less than the generator
Under matched donor replacement at reward weight one in the restless task, the oracle's mean increase in negative log likelihood was 2.07 nats per choice, and a fitted same-family generator reached 1.87. The pooled RW + softmax baseline reached 1.29, the GRU 0.46, the fine-tuned LLaMA 0.19, and the Transformer 0.08. All of these models were evaluated on the same data with the same donor assignments.
The gap is the paper's central counterexample. A model can be an accurate predictor and still show a much weaker probability response to reward replacement than the process that generated its training data. The authors note that the RW baseline shares a value-update motif with the restless generator, which may partly explain its comparatively high response fidelity at this weight.
Prediction and fidelity rankings diverge across tasks
The ordering is not a simple story of neural models trading accuracy for faithfulness. At some reward weights, neural networks predicted better than a pooled reinforcement-learning model while showing less faithful changes in choice probabilities. In the spatial task, the pattern shifted: LLaMA had better predictive performance than RW (1.946 versus 2.182 negative log likelihood) and also lower response error (0.368 versus 0.487), with the LLaMA advantage in response error holding at every tested nonzero weight and paired participant-bootstrap intervals excluding zero.
The authors caution that the spatial task is a replication of the evaluation framework under a new mechanism combination, not zero-shot transfer, since each model was trained separately on the spatial data. The mixture weight also means something different in each task: the restless policy mixes utilities before an epsilon-greedy decision, while the spatial policy mixes two probability distributions directly. Neither absolute negative log likelihood nor weight should be read as a common mechanistic unit across the two tasks.
The oracle calibrates a response, not a mechanism
The oracle defines a reproducible response for the chosen assay, but it does not turn donor replacement into the causal effect of changing rewards in a live agent. Original choices stay fixed, including choices that would have changed under an altered environment, and donor histories disrupt the joint structure of actions and rewards. The reference therefore calibrates models against an explicitly defined conditional replay rather than a distribution-free measure of reward importance.
Agreement with the oracle is evidence about an observable response; disagreement cannot by itself locate the responsible latent computation. The paper also notes that model differences cannot be attributed to architecture alone, since pretraining, tokenization, precision, optimization, and access to individual calibration all vary across families, and that more training seeds, especially for the LLaMA fine-tuning, would strengthen robustness assessment.
Validate the ablation response before reading it as mechanism
The practical priority the paper derives is to make the explanatory target explicit and test it independently of fit. That means distinguishing retraining without an input from perturbing an already trained predictor, reporting intact and perturbed performance on common targets rather than an isolated loss difference, and, where a controlled simulator exists, comparing probability responses under identical histories and operations against both a fitted same-family reference and a pooled simple baseline.
The most decision-relevant point is that a small ablation effect licenses only a claim about the fitted predictor's insensitivity to a specified operation under a specified protocol. Before that observation is used to infer how observed choices were generated, the response itself needs its own validation, because in these sequential tasks predictive accuracy and response fidelity came apart and their model ordering depended on the task.
Sources used in this answer
Better Behavioral Prediction, More Faithful Model Ablations? Evidence from Sequential Choice
In two synthetic sequential bandit tasks with known generators, neural models could predict held-out choices well while responding far less than the true generator to matched reward replacement, and prediction and response-fidelity rankings differed between tasks.