{"id":"256d13ec-9b69-48ff-8f67-0859e04209f2","arxiv_id":"2608.06659","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"CellWorld predicts latent cell representations from spatial context plus a partial expression hint, and this latent prediction objective beats gene-reconstruction models on spatial transcriptomics benchmarks.","lead":"A new AI pretraining method, CellWorld, learns to guess the internal state of hidden cells from their surrounding tissue and a tiny hint of their own genes, instead of reconstructing raw gene counts. Across four test tissue datasets, it outperformed existing spatial-transcriptomics models on every studied benchmark, even when trained on a small slice of the data.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 10% expression hint in Eq. 5 may let the predictor solve Eq. 7 by hint-to-latent autoencoding; the paper's own Figure 4 and Table 3 show the no-hint Large model collapses, so the central scalability claim rests on an untested balance.","rationale":"Agreement_with_reader: agree. The reader's weakest_assumption identifies exactly the load-bearing point: the 10% partial-expression hint must simultaneously stabilize training and preserve the need for spatial context. This is not a manufactured concern; the paper itself flags the failure mode. In the Introduction they state that visible spatial context alone may not uniquely determine a masked cell's identity ('cell-level target ambiguity'), motivating the hint. More importantly, the controlled analysis section explicitly reports that no-hint Large collapses after about 3,000 steps (Figure 4) and that at Base the no-hint model is best on both spatial tasks (Table 3). These admissions show the method sits on a narrow balance rather than on a guarantee. The manuscript also reports in Appendix E that the no-hint predictor degenerates to a patch-conditioned mean, which is why the hint is introduced; however, this diagnostic quantifies what the hint adds, not what the context contributes when the hint is present. The missing piece is a direct measure of the predictor's sensitivity to context under the actual training objective. This is why I set verdict_should_be to UNCHANGED: the reader's conditional verdict is appropriate, and the requested diagnostic is the condition that would settle the concern. I considered two other concerns and found them secondary. First, the comparison omits SToFM and HEIST, the closest published methods; this is a real publication gap, but it can be repaired by adding those baselines and does not threaten the internal logic of the method. Second, the lack of error bars makes small margins hard to interpret; this affects confidence but not the central mechanism. The hint concern, by contrast, goes to whether the pretraining objective does what the paper says: if the predictor solves Eq. 7 by autoencoding the hint, the reported improvements may be attributable to a different mechanism than the one claimed, and the 'scalable route to foundation models' would need to be reframed. Credit should be given where due: the paper includes controlled ablations, permutation tests, and honest reporting of the collapse, which is exactly the kind of evidence that makes the concern testable rather than speculative.","tokens_in":11224,"tokens_out":11137,"duration_ms":114066,"concrete_test":"On the trained CellWorld-Large (kappa=0.10), evaluate the pretraining MSE (Eq. 7) on a fixed set of held-out patches under three input conditions: (a) true visible context + true hints; (b) true hints but visible context representations taken from a different patch (or set to zero); (c) true visible context but hints replaced by non-informative per-target tokens (e.g., the shared mask token m). If loss(b) is close to loss(a), the predictor does not use spatial context and the objective is effectively hint-to-latent autoencoding; if loss(c) is close to loss(a), the expression content of the hint is irrelevant to stabilization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"CellWorld's central claim is that latent-space prediction from visible context plus a 10% partial-expression hint is a scalable, transferable ST pretraining objective. The load-bearing assumption is that the hint disambiguates target cell identity without letting the predictor ignore spatial context. This assumption is not currently demonstrated. Because the hint (Eq. 5) is a random 10% sample of the masked cell's own expressed genes, and the target (Eq. 4) is the EMA encoder's output on the full patch including that cell, the MSE in Eq. 7 can in principle be minimized by mapping the hint almost directly to the target latent, with context playing a minor role. The paper's own controls show the risk is real: Table 3 reports that at Base scale the no-hint model is best on both spatial tasks (Region 0.571 vs 0.539, Niche 0.551 vs 0.531), indicating the hint trades away spatial reliance; Figure 4 shows the no-hint Large model collapses at about 3,000 steps, so the 0.10 hint is the only thing preventing failure at scale. The predictor diagnostics in Appendix E (no-hint predictor outputs near-identical values for all targets) confirm the hint is needed for target-specific variation, but they do not establish how much the predictor relies on spatial context when the hint is present. Without such a measure, the objective may be a form of denoising autoencoding rather than the spatial latent predictive learning that motivates the method, undermining the conceptual claim and the interpretation of the SOTA results.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CellWorld, a spatial transcriptomics foundation model pretrained by predicting latent representations of masked cells from visible spatial context plus a 10% partial-expression hint. The architecture uses a context encoder, an EMA-updated target encoder, and a hint-conditioned spatial predictor with 2D-ALiBi attention. Pretraining is performed on 46 million human cells spanning three platforms and 11 organs, at four model scales from 5.74M to 94.56M trainable parameters. Controlled experiments compare masking strategies, hint ratios, prediction objectives, and spatial permutation; scaling experiments vary model size and data fraction/coverage; comparisons are made against PCA, CellPLM, Nicheformer, and scGPT-spatial on four held-out datasets with FOV-level splits. The paper claims state-of-the-art results on all 11 linear-probe benchmarks and all seven fine-tuned spatial benchmarks, with a frozen 5% corpus pretraining variant outperforming fully fine-tuned baselines on the spatial benchmarks.","tokens_in":11588,"tokens_out":6100,"duration_ms":62262,"significance":"If the results hold, CellWorld provides a credible demonstration that latent-space predictive pretraining can scale for spatial transcriptomics, and the controlled scaling analysis is a useful contribution to how data diversity and optimization interact. The paper has real strengths: FOV-level splits reduce spatial leakage, three-seed averages are used, multiple ablations isolate design choices, and the scaling experiments explicitly probe confounds such as repeated exposure and source diversity. Code is promised. The main limitations are the unresolved role of the partial-expression hint, the absence of the closest reconstruction-based baselines from the comparison, and the lack of error bars on headline claims; these temper the strength of the state-of-the-art conclusion.","major_comments":[{"comment":"The hint in Eq. (5) is a random 10% sample of the masked cell's own expressed genes, and the target in Eq. (4) is the EMA encoder's output on the full patch containing that cell; therefore the MSE in Eq. (7) can in principle be minimized by a near-direct hint-to-latent mapping, with spatial context playing only a minor role. Table 3 shows that removing the hint improves both spatial tasks at Base scale (Region 0.571 vs 0.539, Niche 0.551 vs 0.531), and Figure 4 shows that no-hint Large training collapses after about 3,000 steps, so the 0.10 hint is load-bearing. The spatial-permutation control in Table 4 demonstrates some use of the expression-coordinate correspondence, but it does not quantify how much the predictor relies on visible context when the hint is present. Please add a context-free control (hint only, no visible context) and/or a direct measure of context contribution, such as attention or gradient analysis; without this, the conceptual claim that CellWorld learns from visible spatial context rather than performing hint-to-latent autoencoding is not established.","section":"Method, Eqs. (4)-(7); Controlled Design Analysis, Table 3"},{"comment":"The Related Work identifies SToFM and HEIST as closely related existing methods, and the Introduction states that even CellWorld-Small outperforms every existing method on all benchmarks, yet neither SToFM nor HEIST (nor BrainBeacon) appears in Table 6 or Figure 1. The actual comparison is limited to PCA, CellPLM, Nicheformer, and scGPT-spatial. Because the central claim is state-of-the-art status among ST foundation models, the omission of the closest reconstruction-based baselines is load-bearing. Please include these methods in the comparison, or precisely qualify the claim to the evaluated baselines.","section":"Related Work; Comparison with Existing Methods, Table 6"},{"comment":"All reported scores are point estimates averaged over three downstream seeds, with no standard deviations, confidence intervals, or significance tests shown in any table or figure. The claims that CellWorld outperforms every baseline on all benchmarks require evidence that the margins are not within seed variability; for example, some linear-probe margins are small and the fine-tuning comparison on Cell Brain shows NicheFormer at 0.958 versus CellWorld-Huge at 0.953. Please report per-seed values or error bars for at least the headline comparisons in Tables 3, 5, and 6, and indicate which differences are significant.","section":"Experimental Setup; Tables 3, 5, 6"}],"minor_comments":[{"comment":"The phrase 'all seven fine-tuned spatial benchmarks' is used repeatedly but never explicitly defined; please state that these are the Region and Niche tasks across the four held-out datasets, excluding the Cell annotation task, or list them explicitly.","section":"Experiments, benchmark definition"},{"comment":"The 'Steps (k)' column uses values '10', '2.5', and '0.5'; writing '10k', '2.5k', and '0.5k' would avoid ambiguity with the percentage column in the same table.","section":"Table 5"},{"comment":"The caption says 'Comparison with existing methods' while the Introduction claims 'every existing method'; please align the wording with the actual set of evaluated baselines, or add the missing methods.","section":"Figure 1 caption"},{"comment":"HEIST is described as incorporating spatial and contrastive objectives, but no downstream comparison or discussion of its performance on the chosen benchmarks is provided; a brief note on why it was not included would improve clarity.","section":"Related Work"}],"recommendation":"major_revision","confidential_remarks":"The core idea is interesting and the controlled experimental design is a genuine strength. The main risks are the hint-leakage ambiguity, which affects the interpretation of the central pretraining objective, and the omission of the closest baselines from the comparison; both are addressable with additional experiments and analyses. I would not reject the paper on the current evidence, but the revision needs to resolve these points before the state-of-the-art claims can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take on CellWorld: it does something genuinely new by applying JEPA-style latent prediction to spatial transcriptomics, with each cell as a token and a masked-cell prediction target. The experimental design is careful—FOV-level splits, seeded averages, ablations of masking, hint ratio, prediction objective, spatial permutation, and a scaling analysis that separates cell count from optimization steps and source diversity. That scaling analysis is the best part: it shows spatial transfer depends more on source diversity and optimization than raw cell count, which is a useful result for the field.\n\nThe main claims hold up directionally. Even the small model beats PCA and three ST foundation models on most benchmarks, and the frozen 5%-data Large model beating fully fine-tuned baselines on spatial tasks is striking. The paper is also honest about the no-hint collapse at Large scale and predictor degeneration at Base.\n\nSoft spots. First, no error bars anywhere in the headline comparisons. Three seeds averaged but no variance. With sometimes small margins, the \"outperforms every baseline on all 11 benchmarks\" claim needs standard deviations. Second, the closest baselines from the related work—SToFM and HEIST—are not included. SToFM reconstructs expression-derived cell embeddings, which is halfway to latent prediction; excluding it weakens the \"latent beats reconstruction\" claim. Third, the 10% hint is load-bearing: without it, the Large model collapses, and the paper doesn't directly measure how much the predictor relies on spatial context when the hint is present. The spatial permutation test on the context encoder is reassuring, but it doesn't fully close whether the objective is closer to denoising autoencoding than spatial latent prediction. That's a conceptual ambiguity, not a refutation. Finally, the shared hyperparameter configuration for CellWorld needs clarification—was it chosen on the held-out validation sets? That would be a subtle form of selection.\n\nNet: this paper deserves a serious referee. I'd send it out, but with a request for error bars, the missing baselines, and a clearer statement on hyperparameter selection. The core idea is sound, and the scaling analysis is a genuine contribution. I'd cite it in my own work once cleaned up.","headline":"The paper makes a real contribution—latent cell prediction for spatial transcriptomics with controlled scaling experiments—but the SOTA claims need error bars and the closest reconstruction-based baselines are missing.","tokens_in":12081,"tokens_out":3709,"would_cite":true,"duration_ms":35284,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Latent-space prediction of masked cells, not gene reconstruction, makes spatial transcriptomics foundation models that beat larger baselines at a fraction of the parameters.","keywords":["spatial transcriptomics","foundation model","latent prediction","joint-embedding predictive architecture","self-supervised learning","cell annotation","representation learning","spatial context"],"falsifier":"Train CellWorld-Large with hint ratio 0.10 but with the spatial coordinates of all cells randomly permuted during pretraining; if downstream spatial benchmarks (region and niche) remain at the same level as the unpermuted model, the claim that spatial context drives the gains would be falsified. Alternatively, a matched model pretrained with the hint but without any visible context (using only the hint to predict the target latent) that matches CellWorld's spatial benchmark scores would show the spatial context is unnecessary.","tokens_in":11036,"feed_emoji":"🧬","tokens_out":6511,"duration_ms":56624,"temperature":0.7,"pith_summary":"This paper argues that the best way to pretrain a foundation model for spatial transcriptomics is not to reconstruct masked gene measurements but to predict latent representations of masked cells from their visible spatial neighbors, aided by a small partial-expression hint. The authors build CellWorld, a joint-embedding predictive architecture with a context encoder, an EMA-updated target encoder, and a hint-conditioned spatial predictor, and pretrain it on 46 million human cells. They report that even the smallest CellWorld, with 5.74M trainable parameters, outperforms existing spatial transcriptomics foundation models on all 11 linear-probe benchmarks and all seven fine-tuned spatial benchmarks, and that a frozen Large model pretrained on only 5% of the corpus beats every fully fine-tuned baseline on all seven spatial benchmarks. The paper's broader claim is that latent prediction avoids reproducing assay-specific technical variation and that, under a fixed cell budget, broad biological source diversity and enough optimization matter more than raw cell count for spatial transfer.","feed_headline":"A 5.7M-parameter model wins every spatial benchmark","feed_subtitle":"By predicting masked cells' latent states instead of raw genes, CellWorld beats all baselines -- even with only 5% of the data.","key_machinery":"The load-bearing mechanism is the hint-conditioned spatial predictor with a 10% partial-expression hint. Each masked cell's query is built from a random 10% subset of its nonzero gene expression (plus organ and platform embeddings), projected into predictor space, and combined with a learnable mask token; the predictor then recovers the EMA target encoder's latent for that cell from the visible context. The hint is used only by the predictor, not by either encoder, and it resolves what the paper calls cell-level target ambiguity -- the fact that a cell's identity and state are not uniquely determined by its spatial neighborhood. The objective is plain mean-squared error in the latent space, and the paper shows that without the hint the Large model's predictor degenerates and training collapses around 3,000 steps, while with the hint the model stays stable and outperforms its no-hint counterpart.","core_discovery":"CellWorld's central discovery is that shifting the pretraining objective from gene-level reconstruction to latent cell prediction turns spatial transcriptomics into a tractable foundation-model setting. Given a local patch of cells, CellWorld masks 60% of cells, encodes the remaining visible cells with a spatial Transformer using 2D-ALiBi relative attention biases, and trains a spatial predictor to recover, from the visible context and a 10% random partial-expression hint, the latent representations that an EMA target encoder produces for the masked cells when it sees the complete patch. The target encoder and predictor are discarded after pretraining; only the cell tokenizer and context encoder transfer to downstream tasks such as cell annotation, region prediction, and niche-composition prediction. The paper shows across four held-out datasets that this objective yields state-of-the-art linear-probe and fine-tuning results at every model scale, and that a frozen encoder pretrained on a broadly sampled 5% of the corpus outperforms all fully fine-tuned baselines on spatial tasks. The authors interpret these results as evidence that predicting abstract cell states, not raw expression values, avoids encouraging the model to reproduce assay-specific technical variation and thus produces more transferable spatial representations.","pith_inferences":["A natural next test is whether the same latent-prediction objective transfers to temporal or perturbational spatial data, where the target is not a masked present cell but a future cell state; CellWorld's authors hint at this direction in their conclusion, but the experiments to check it have not been run.","Because the hint controls the trade-off between target specificity and reliance on spatial context, one could test whether the optimal hint ratio changes with tissue type or platform, or whether an adaptive hint that shrinks as training progresses yields better spatial representations.","The finding that a 5% broadly sampled subset matches the full corpus suggests that pretraining data might be selected by diversity rather than volume; a direct comparison would be to hold out entire organs and measure how quickly spatial-transfer performance drops as source diversity is removed.","If the predictor's degeneration at scale is the main obstacle, then techniques that stabilize joint-embedding training without hints, such as centering or variance regularization, could remove the need for the hint entirely and may further improve spatial transfer."],"forward_implications":["At every tested scale, CellWorld outperforms existing ST foundation models on all 11 linear-probe benchmarks and all seven fine-tuned spatial benchmarks, with performance generally improving as model size grows from 5.74M to 94.56M trainable parameters.","A frozen CellWorld-Large trained on 5% of the corpus (64 slides, 11 organs) beats every fully fine-tuned baseline across all seven spatial benchmarks, indicating that enough optimization and broad source diversity can substitute for more cells.","Under the fixed 10,000-step schedule, the 25% corpus subset matches the full corpus, while proportionally scaling steps with data size preserves cell annotation but degrades spatial transfer, implying spatial transfer needs repeated exposure to the same local contexts.","Concentrating the same cell budget in fewer slides and organs lowers spatial-task performance, so biological source diversity matters for spatial transfer more than raw cell count."],"supporting_citations":[{"why":"Supplies the joint-embedding predictive architecture (I-JEPA) that CellWorld adapts from images to spatial transcriptomics.","marker":"Assran et al. 2023"},{"why":"Establishes the principle of predicting latent representations rather than raw observations, the conceptual basis of CellWorld's objective.","marker":"Baevski et al. 2022"},{"why":"Provides the masked-autoencoder reconstruction baseline used as the matched control in the prediction-objective ablation.","marker":"He et al. 2022"},{"why":"Supplies the ALiBi attention-bias mechanism that CellWorld extends to two-dimensional relative spatial attention.","marker":"Press, Smith, and Lewis 2021"},{"why":"One of the founding spatial transcriptomics technologies that motivates modeling cells within their tissue context.","marker":"Ståhl et al. 2016"},{"why":"Nicheformer is a key baseline that CellWorld must outperform; it represents gene-identity reconstruction pretraining.","marker":"Tejada-Lapuerta et al. 2025"},{"why":"scGPT-spatial is a key baseline representing masked-expression reconstruction in spatial foundation models.","marker":"Wang et al. 2025"},{"why":"CellPLM is another expression-reconstruction spatial foundation model used as a baseline in the comparison.","marker":"Wen et al. 2024"}],"fun_headline_variants":["Latent cell prediction beats gene reconstruction in spatial models","CellWorld: small model, all benchmarks won","Frozen model with 5% data beats fine-tuned baselines","Predict cell states, not genes: CellWorld wins"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The 10% partial-expression hint has to be strong enough to prevent predictor collapse at scale but weak enough that spatial context still determines the prediction; if this balance is fragile, the method's success rests on a hyperparameter rather than on a robust principle.","fun_headline_variants_meta":{"raw":{"variants":["Latent cell prediction beats gene reconstruction in spatial models","CellWorld: small model, all benchmarks won","Frozen model with 5% data beats fine-tuned baselines","Predict cell states, not genes: CellWorld wins"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000857,"raw_usage":{"total_tokens":3764,"prompt_tokens":1033,"completion_tokens":2731,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":649,"completion_tokens_details":{"reasoning_tokens":2665}},"tokens_in":649,"tokens_out":2731,"duration_ms":17508,"temperature":1.0,"reasoning_tokens":2665,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:01:55.452001+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train CellWorld-Large with hint ratio 0.10 but with the spatial coordinates of all cells randomly permuted during pretraining; if downstream spatial benchmarks (region and niche) remain at the same level as the unpermuted model, the claim that spatial context drives the gains would be falsified. Alternatively, a matched model pretrained with the hint but without any visible context (using only the hint to predict the target latent) that matches CellWorld's spatial benchmark scores would show the spatial context is unnecessary.","supporting_citations":[],"review_version":1}