Source
Dasgupta, Schulz, Goodman and Gershman (2018), LIT-1074 (Cognition 178, 67–81), read in NOTE-875: Experiments 1–3, the simulations of Figure 3, the General Discussion. Gershman and Goodman (2014), LIT-1056 (CogSci 2014, 517–522), read in NOTE-886: Experiments 1 and 2, Figure 7. Dasgupta and Gershman (2021), LIT-1077 (Trends in Cognitive Sciences 25, 240–251), a review, read in NOTE-865.
What was actually shown
Dependence on an earlier sub-answer (2014). On a five-node network of defecting officers, each subject’s answer to a sub-query, put through Bayes’ rule, predicted their answer to a complex query when the sub-query came first (r = 0.28, n = 73) and not when it came after (r = −0.11). The difference between the correlations was not tested. A structurally matched control query showed no correlation in either order. In speeded choices, the complex query was answered faster relative to the control when the sub-query came first (median log-RT facilitation about 0.47 against 0.23; z = 2.09), analysing only subjects who gave the more likely answer.
Carry-over of an induced bias (2018). Unpacking a first query with typical or atypical examples raised or lowered the judged probability. A second, always packed query carried the effect (Experiment 1, N = 78 analysed: regression 4.39 and −7.86), and errors on the two were correlated (r = 0.67 aggregated, 0.55 within participants). With no shared letter between the queries, nothing carried over, which rules out simple numerical anchoring (Experiment 2, N = 80). Load on the first query enlarged the carry-over, which the authors’ simulations read as reuse of a stored summary rather than of samples. With a new cue, carry-over appeared for subadditivity only, when the cue’s posterior was similar to the first (Experiment 3; similar against dissimilar, p = 0.07).
Efficiency, in 2018. No response times were recorded. With load on the second query, probe accuracy was higher on reusable trials (t(36) = 3.16). This is the paper’s only evidence that reuse saves anything.
The framing. “An inference engine may gain efficiency by incurring some amount of bias due to reuse of past computations—a strategy we will refer to as amortized inference”. The 2014 paper calls it caching or “memoizing” of “sub-computations that are invoked by multiple queries”. The 2021 review extends it: a learned recognition model is “a distributed form of memoization called amortization”, and the planning literature supplies decisions based on cached values that have not been updated after devaluation.
What could have come out otherwise. Carry-over could have appeared without overlap, which would make it anchoring; it did not. The order effect could have been absent; it was present, but rests on one significant and one non-significant correlation.
What this does not say
- That reuse is a pull toward earlier answers. What carries is the earlier computation’s error, and only where its result is usable. The no-overlap condition was built to rule out a numerical pull, and found none.
- That reuse is faster, in general. Speed is shown in one small study with selection on the outcome query; the 2018 paper measures no times, and the review assumes reuse is cheaper throughout.
- That retrieval is computation. The models give storage and retrieval no cost; the review mentions retrieval demands once. It says memory replaces computation, not that retrieval is a kind of it.
- That the reused results went stale. In the experiments the error is present when the first answer is computed, or comes from applying it to a different query. A result that was right and has since gone out of date appears only in the review’s account of devaluation studies the record has not read.
- When reuse pays. The 2014 paper lists trade-offs (complexity, decomposability, frequency) in prose; the 2018 paper asks “What objective function is being optimized by amortization?”; the review lists when to cache as open.
- That the brain uses either tested scheme. “We do not claim that either extreme is plausible”.
- Anything about stores outside the head. None of the three cites an offloading or distributed-cognition source.
Connections
- QUESTION-095: the “precompute and leave a result” strategy with the store in the head, though memoization here is lazy: results are cached when first computed, not ahead of a predicted query as the speed bugs are.
- CLAIM-303’s rider, that “persistence can preserve errors as efficiently as valid discoveries”, has an empirical counterpart here, for errors present from the start.
- CLAIM-295: keeping a summary in place of samples adds no information and narrows the queries it serves (NOTE-875).
- THEORY-355 and THEORY-346 treat stores in the world.
Note of 2026-10-30: what holds it back
The title states a general effect in people from five small online experiments by one group, so it keeps its condition (ADR-065 §6): the missing result is a preregistered replication outside the original paradigm that measures, in the same participants, both the carry-over of an induced error and the time or effort saved. The first finding also says more than the 2014 study tested: the sub-answer predicted the later answer when asked first (r = 0.28) and not when asked after (r = −0.11), but the difference between the two correlations was not tested (NOTE-886), so “only when asked first” is not shown.