Reconstructing an earlier encounter's neighboring content

Citation and status. Futing Zou, J. Benjamin Hutchinson, and Brice A. Kuhl. Hippocampal-guided reconstruction of an event’s prior temporal context. bioRxiv DOI, version 2, 13 October 2025; preprint, not established as peer-reviewed publication. Checked 22 September 2026: the bioRxiv API lists versions 1–2 and published: NA; the author laboratory still lists it under preprints, and PMC identifies version 2 as unreviewed. Author PDF.

Factual core. Eight Natural Scenes Dataset participants completed 30–40 scanning sessions. Main analyses selected 721–1,875 images per participant, correctly judged new at first encounter and old at second encounter, within one session but different runs. Participant-specific encoding models mapped 20 principal components of CLIP image features to second-encounter fMRI activity; inversion used leave-one-session-out prediction. The target was the average feature vector of the two scenes immediately neighboring the first encounter. Reconstruction exceeded a shuffled baseline in lateral occipitotemporal cortex (LOTC), but not the other tested cortical regions. Accuracy decreased with neighboring-scene distance. Reconstruction was nonsignificant for unrecognized images, without a reported direct recognized-versus-unrecognized contrast. Hippocampal CA1 and CA23DG pattern similarity across encounters predicted LOTC reconstruction in observational analyses controlling temporal variables. Outputs were feature vectors scored by cosine similarity, not generated pictures, identified autobiographical episodes, or demonstrated transfer to another agent. The task collected recognition decisions rather than reports of neighboring scenes.

Audit — interpretation of the reported design.

• A substantive advance with a bounded target. The decoder is evaluated against previously neighboring scene content, rather than simply comparing retrieval activity with the earlier target's activity. Because it does not use first-encounter neural patterns as reconstruction targets, it avoids the direct contamination from overlapping first-encounter and neighboring-event BOLD responses. However, averaging the preceding and following scenes removes their order and does not establish recovery of either complete image. Feature similarity is not episode identification or a quantified count of recovered memories.
• Conditional information remains to be tested. The paper asserts that adjacent scene content lacks autocorrelation, but does not report a cue-feature-only predictor, a feature-matched alternative-context comparison, or an analysis conditioning on the current image and its current neighbors. Those controls would test whether brain measurements add information about the old neighborhood beyond correlations in the stimulus schedule and current visual input. Their absence does not demonstrate a confound. The existing different-run restriction is a meaningful protection against direct temporal overlap, not a proof against every route by which stimulus structure could contribute.
• Held-out unit and calibration. Sessions are held out within each person; people are not held out, and this is not zero-calibration transfer. Thousands of recorded public-image exposures provide known targets. The paper does not establish a split disjoint in every image contributing to a context target. It also leaves the nesting of PCA fitting and ridge-parameter selection insufficiently explicit for an independent leakage audit. These are verification questions, not established leakage. Generalization to unrecorded personal events requires a separate experiment.
• Selection and awareness. Analyses exclude changed responses and primarily retain correct-new/correct-old pairs. A positive result in hits and nonsignificance in misses do not demonstrate that reinstatement occurs only during successful recognition. Miss counts, a matched comparison, and uncertainty on the difference matter. An old/new judgment also does not distinguish familiarity from detailed conscious recollection of the prior neighborhood.
• Association is not direction. The hippocampal predictor is same-image pattern similarity minus a matched different-image reference; this is stronger than unadjusted similarity. Nevertheless, the regression and median split do not manipulate hippocampal activity. A null reverse analysis with a different neural endpoint does not identify causal direction. Shared attention, encoding strength, signal quality, or other processes remain possible contributors; “guided” and “mediated” are interpretations rather than established causal findings.
• Dependence and statistical precision. Trial regressions include participant random intercepts but no reported session/run effects or participant-specific slopes. Permutations shuffle across trials within a person, without reported blocking by session; validity depends on exchangeability despite temporal structure. Numerous trials do not create additional independent people. Hippocampal tests are uncorrected. Separately, the reported corrected P < .001 from 1,000 permutations has finer precision than the usual plus-one estimator supports: with four Bonferroni comparisons its minimum is about .004. This does not erase a signal, but the tail-probability calculation needs clarification or more permutations.

Original transfer implication and proposed test. This study motivates asking whether a reminder can expose information about an earlier surrounding event without asking the person to narrate it. It does not yet show that such information exceeds a strong behavioral acquisition baseline or can be extracted when researchers lack the original event record.

A discriminating test would randomize the prior neighborhood of a repeated cue, hold current sensory input constant, and reserve new cue–context combinations for evaluation. Compare neural-plus-behavioral decoding with the same cues, behavioral observations, calibration history, and priors without the neural channel. Add cue-only and stimulus-feature predictors; keep targets unavailable at test and fit preprocessing and hyperparameters inside training folds. Preserve temporal dependence in null tests and assess generalization at the claimed person/session level.

Then give a separate recipient only the decoded information, with uncertainty, and test a new choice requiring the hidden neighboring detail. Compare with no-channel, shuffled-channel, and oracle-feature recipients. Keep objective event accuracy, forecasts of the source person's recollection, and recipient action distinct. This would test incremental extraction and functional use; it would still not prove hippocampal causation, preserved subjective experience, or transfer of an entire episode.

Reading and version boundaries. Read the complete main introduction, results, discussion, methods, statistical procedures, and availability statements in the 20-page author PDF; read all three captions and visually inspected Figures 1–3. References were retained but not independently read. No supplement is attached or linked in the retained PMC XML. The PDF and XML agree on title, DOI, and version date. No participant records, raw imaging data, model sessions, or analyses were run.

The preprint says code will be available upon publication, but a public repository now exists. Its pinned README at commit eb975777c0842ad43225ef06f289820595363e5f (26 August 2026) uses the title Neural reconstruction of an event’s prior temporal context. Read the root and analysis READMEs only; analysis scripts and outputs were not audited, and that later repository is not assumed identical to the October 2025 analysis. It does not itself establish journal publication.

Local PDF, readable text, PMC XML, and provenance with all artifact hashes. Article license: CC BY-NC-ND 4.0.