Open reviews
July 11, 2026 — Open review of Yan and colleagues:
Yan, X., Li, J. A., Franch, M., Zhu, H., Cowan, R., Belanger, J., … & Sheth, S. A. (2026). Polysemanticity in human hippocampal neurons. bioRxiv. DOI
Yan and colleagues analyze a rare dataset of single-neuron recordings in human hippocampus while patients listened to a variety of different naturalistic spoken story stimuli. They use semantic embeddings to show that individual neurons are densely tuned to unrelated semantic features, with tuning that is more dispersed than would be expected by chance and highly inflected by narrative and lexical context. This kind of feature encoding is congruent with the mixed selectivity reported in prefrontal and parietal areas, as well as the polysemanticity / superposition observed in large language models. They relate this feature tuning to historical models like concept cells and population sparse. I like this paper a lot! The methods seem rigorous to me and are very (extremely!) well detailed. To me, the results are novel and interesting. The figures are dense but well designed. Most of my comments are relatively minor clarification questions or suggestions.
Major comments:
1. First, I would encourage the authors to do whatever they can to avoid presenting the simulated concept cell model as a straw man. As I was reading, I had the thought “nobody really believes that a neuron could respond in such a conveniently simple way, right?” I think plenty of people actually do probably have this mental model for hippocampal cells (at least implicitly), so I think it’s important for the authors to contrast their findings with the concept cell model. But the more cartoonish this control model seems, the weaker the argument for the authors’ own alternative model. I’m honestly not sure how else to fortify against this possible objection, however. As far as I can tell, the authors do a good job deriving the simulation from the literature, although this is pretty deep in the methods; maybe up-fronting a bit more of this “validation” of the concept cell simulation in the main text would help? Here’s two other ideas: (1) I’m not sure the current implementation, where the “concepts” are operationalized as unique nouns, fully captures the classic concept cell results. Does this “unique nouns” list include proper nouns like “Applebee’s” that were excluded from other analyses? Will this analysis capture “Jennifer Aniston”-style responses? For example, I could imagine formulating this analysis differently: identify unique proper noun-style entities in the narrative (e.g., an actual person or place in the story), then test if certain cells respond to references to that entity, even indirectly—for example a pronoun referring back to a specific character in the story. For example, I’m imagining a cell that responds to “Applebee’s” and also responds to “so there we were” when referring to an event happening at Applebee’s. I’m not an expert on concept cells, but this seems like it could be an intuitive treatment of them in a naturalistic story. (2) Could it strengthen the comparison to treat concept cells as a proportion of cells within a larger, more complex population? This would allow you to vary the the proportion of simulated concept cells; e.g., a 100% concept cell population looks somewhat cartoonish compared to the empirical data, but a 20% concept cell population looks more realistic (but I suspect still doesn’t fully mimic the overdispersed quality of the empirical data). Anyway, I’m not demanding the authors perform either of these specific analyses; I just want to suggest that they make sure the concept cell model they use for comparison is taken seriously.
2. Second, somewhat related point: There were certain places in the Results where the authors report that a certain property of the neural tuning is greater than “chance”; e.g., “more regular spacing than chance” (line 274), “more isotropic than chance” (line 280). On my first read, I had a difficult time imagining what “chance” actually means in these comparisons. Some kind of randomization? Some kind of Gaussian distribution? I think the authors do a pretty good job of explaining this later on, but the explanation is pretty deep in the Methods. Can you add a little parenthetical when reporting these results in the main text that indicates what you mean by chance?
3. In your 5-fold cross-validation procedure for ridge regression (line 706), please specify whether your held-out tests sets comprised words that were randomly interspersed with the training words, or whether the training and test folds were temporally contiguous chunks (cf. Hadidi, Feghhi et al., 2026). Temporally contiguous cross-validation folds are probably best practice for continuous narrative stories like this, although it will make for a more stringent generalization test; if you really want to use random splits in the context of a continuous, autocorrelated narrative, you should justify why. Can you also explain how this cross-validation procedure relates to the multiple story stimuli? Do you concatenate all the words together across stories and then split? Do you cross-validate separately within each story and then aggregate the results?
4. At line 76, you mention that superposition is a strategy for a network to “represent more independent features than it has available neurons.” I’m not arguing with this point at all, but even in the Elhage et al. (2022) paper, I find this kind of opaque. What are these “features” when it comes to natural language? How could we hope to quantify them? The success of LLMs is a compelling demonstration that even all of the linguistic patterns described by linguists are likely a vast underestimate; there’s all sorts of graded, multidimensional features of contextual meaning that we can’t so easily define (relative to, e.g., syntactic, patterns). My point is that the “features” of language comprise all sorts of morphological, syntactic, semantic, pragmatic patterns, and so on, as well as their interactions and permutations as they co-occur in real-world contexts. Actionably, it might help to add a sentence somewhere in this vicinity that gives that non-language-oriented reader an intuition as to what you mean by “features” and how vast the space features might be in natural language.
5. I had some difficulty understanding how the analyses in the “pattern separation and pattern completion” section actually map onto classical notions of pattern separation/completion. Any additional hand-holding for folks who are less familiar with those hippocampal theories, and how you hope to translate them into natural language comprehension, would be helpful.
6. There were a couple connections to the surrounding literature that could be worth developing in the Discussion. For example, I wonder if the dialogue between Bowers (2009) and Plaut & McClelland (2010) is worth mentioning. The authors might also consider adding a note to the Discussion about how we should interpret these results within the larger framework for understanding how the brain performs language processing. For example, the hippocampus is completely outside the putative language network (Fedorenko et al., 2024). Is the hippocampus even necessary for language/narrative comprehension (e.g., Zuo et al., 2020)? Do the authors think the hippocampus is unique in using this coding scheme, and that other language areas may have different tuning properties? How do these single-unit results in the hippocampus compare to the recent batch of single-unit results acquired in more conventional language areas during natural language processing tasks (e.g., Jamali et al., 2024; Khanna et al., 2024; Cai et al., 2026)? (My sense is that most of those findings are also suggestive of polysemanticity or mixed selectivity, even if that’s not how those authors want to frame their results…)
Minor comments:
Line 113: “Layer 36” implies that you’re using a larger-than-base GPT-2 model. I would add a parenthetical at line 112 that you’re using GPT-2-large so “layer 36” doesn’t come as a surprise; maybe also add that this is the final layer in a parenthetical.
Figure 2, panel C: “encode” > “encoded” (panel title)
Figure 3, panel I: “three hypothesis” > “three hypotheses”
Line 369: The term “word identity” seems like an odd label for a FastText embedding to me; “word identity” makes me think of a meaningless ID number (e.g. a one-hot encoding of each word), whereas the FastText embedding has a lot of (lexical-)semantic information in it… Maybe say “lexical embedding” or “static semantic embedding”?
Line 535: “Classic concept cell protocols”—you mean the experimental paradigms?
Line 673: In describing the FastText embeddings in the Methods, I would further emphasize that they are “static” or “global” embeddings that are not contextualized by preceding words (i.e., a given word gets the same embedding no matter the context in which it occurs).
Line 689: I was a bit surprised to find that your “context” here is only local sentence-level context, which may not include larger-scale discourse/narrative context. Including a richer context might even strengthen the findings related to context sensitivity, but will also complicate the analysis pipeline; I assume this choice was made for the sake of simplicity.
Line 724: “pattern” > “patterns”
Line 828: Is it really fair to rectify the neuronal tuning vectors estimated by ridge regression? Aren’t the negative weights just as important as the positive weights? Are there any less “damaging” adjustments to make the weight vectors play nicely with downstream analyses? (For example, you use a min-shifting approach in the earlier NMF analysis.)
Line 994: “Dissemination” > “Decimation”?
Line 1025: Not sure “diverging” is the right word for this kind of rectified colormap, given that it only has colors going in one direction…
Line 1033: Is the term “lobe” some kind of jargon from the literature? You use this term “semantic lobes” in these sections, but not in the main text… Maybe just define this term when you first introduce it.
Line 1180: “fitted” > “fit”
Line 1235: Why only 3 PCs for the M1 vs. M2 vs. M3 analysis? I wonder how much semantic or contextual structure is actually captured in only 3 PCs when 100 PCs only captures ~61% of variance.
When I was reading through the Results corresponding to Figure 1, a particular analysis idea came to mind. One historically compelling argument against a localist code was the decoding analysis by Haxby et al. (2001) where they identify a subset of voxels with the strongest activation for one or more classes of stimuli (similar to a typical localizer analysis), and then show that these stimuli can still be classified from the population activity with similar accuracy even after excluding those most-activated voxels. I could imagine a similar neuron-level analysis for this dataset. I think the authors’ results are already well supported and they don’t actually need to perform this analysis; just wanted to share the idea.
References
Bowers, J. S. (2009). On the biological plausibility of grandmother cells: implications for neural network theories in psychology and neuroscience. Psychological Review, 116(1), 220–251. DOI
Cai, J., Kfir, Y., Jamali, M., Huang, H., Kim, Y. J., Cash, S. S., & Williams, Z. M. (2026). Mapping the neuronal building blocks of human language with language models. Nature. DOI
Jamali, M., Grannan, B., Cai, J., Khanna, A. R., Muñoz, W., Caprara, I., … & Williams, Z. M. (2024). Semantic encoding during language comprehension at single-cell resolution. Nature, 631(8021), 610–616. DOI
Khanna, A. R., Muñoz, W., Kim, Y. J., Kfir, Y., Paulk, A. C., Jamali, M., … & Williams, Z. M. (2024). Single-neuronal elements of speech production in humans. Nature, 626(7999), 603–610. DOI
Hadidi, N., Feghhi, E., Song, B. H., Blank, I. A., & Kao, J. C. (2026). Spurious alignment between large language models and brains can emerge from non-robust methods and overlooked confounds. Nature Communications, 17, 5769. DOI
Haxby, J. V., Gobbini, M. I., Furey, M. L., Ishai, A., Schouten, J. L., & Pietrini, P. (2001). Distributed and overlapping representations of faces and objects in ventral temporal cortex. Science, 293(5539), 2425–2430. DOI
Plaut, D. C., & McClelland, J. L. (2010). Locating object knowledge in the brain: comment on Bowers’s (2009) attempt to revive the grandmother cell hypothesis. Psychological Review, 117(1), 284–288. DOI
Zuo, X., Honey, C. J., Barense, M. D., Crombie, D., Norman, K. A., Hasson, U., & Chen, J. (2020). Temporal integration of narrative information in a hippocampal amnesic patient. NeuroImage, 213, 116658. DOI