{"id":"5bd3e9ad-edb8-41c0-9473-ad9cbaef571b","arxiv_id":"2611.11081","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"HBAR demonstrates that hierarchical tokenization with structured neurobiological priors achieves higher fidelity compression of fMRI data than rate-matched flat tokenizers, especially at coarser spatial scales.","lead":"This paper introduces HBAR, a hierarchical tokenization framework that compresses resting-state fMRI data up to 415x while preserving multiscale brain signals better than flat compression methods. The framework organizes voxel-level data into a three-level hierarchy (voxel groups, parcels, and functional networks) and shows that using known anatomical atlases improves reconstruction, especially at intermediate parcel scales.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"The coarse-scale evaluation metric (parcel/network correlation) is aligned with HBAR's training objectives but absent from flat baseline losses, so the reported advantage may reflect multi-scale supervision rather than hierarchical tokenization itself.","rationale":"The reader identified atlas validity as the weakest assumption, but the more immediate threat to the central claim is the failure to control for the multi-scale training objective. The paper transparently acknowledges the loss difference but does not treat it as a confound. The shuffled grouping ablation at 52× is a useful control for grouping specificity, yet it does not isolate the hierarchy from the loss; the flat baselines are missing the losses entirely. A simple control experiment (concrete test) would clarify the source of HBAR's advantage. If the augmented flat baseline performs similarly, the paper's conclusion would shift to 'multi-scale losses help, and hierarchy may provide additional modest benefit.' The paper otherwise is thorough and well-executed, so the manuscript should be accepted conditionally on this control being performed or on a clear qualification of the claim.","tokens_in":23004,"tokens_out":4239,"duration_ms":42096,"concrete_test":"Train flat BSQ-415 with additional auxiliary losses on parcel-level and network-level MSE (same L_P and L_N as HBAR-415, using the DiFuMo-1024 and Yeo-17 targets) without changing the flat architecture. Compare parcel and network correlations on the test set. If the augmented flat baseline matches or exceeds HBAR-415, then the advantage stems from the multi-scale losses rather than from hierarchical tokenization. If HBAR-415 remains superior, the hierarchy itself contributes beyond the supervision.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that HBAR achieves higher parcel and network correlations than rate-matched flat tokenizers at 415× compression conflates hierarchical architecture with multi-scale supervision. HBAR is trained with explicit reconstruction losses at parcel and network levels (L_P, L_N in Eq. 2), while flat baselines have no such coarse-scale objectives. Consequently, when evaluated on the same atlas-based pooling used in training, HBAR is effectively tested on what it was directly optimized for, whereas flat tokenizers are evaluated on metrics they were never trained to improve. The paper acknowledges this loss difference (Appendix C.3) but does not control for it. The shuffled grouping ablation at 52× partially addresses this issue by keeping the same loss structure, showing a smaller parcel correlation drop from 0.920 to 0.886, but no equivalent control exists at 415× where the largest gains are reported. Without a flat baseline that also receives auxiliary coarse-scale losses (e.g., parcel and network MSE from atlas pooling), we cannot attribute HBAR's coarse-scale advantage to the hierarchical tokenization rather than to the presence of multi-scale supervision. The reader's weakest assumption (atlas validity) is related but distinct; the core experimental confound is the asymmetry in training objectives, not merely the correctness of the atlas.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper proposes HBAR (Hierarchical Brain Activity Representation), a hierarchical tokenization framework for voxel-level resting-state fMRI. It organizes voxel signals into a three-level hierarchy (sub-parcel groups, parcels from DiFuMo-1024, and Yeo-17 functional networks) and uses binary spherical quantization at each level to achieve up to 415× compression. On the HCP-YA dataset, HBAR is evaluated against rate-matched flat tokenizers (BSQ, RVQ) and a multiscale VQ-VAE baseline. At 415×, HBAR shows higher parcel and network correlations; at 52×, fixed atlas grouping improves parcel correlation over shuffled grouping. Additional experiments compare fixed, learned, and softly aligned hierarchies, and assess demographic prediction via linear probing and autoregressive readouts.","tokens_in":23421,"tokens_out":3747,"duration_ms":37336,"significance":"If the result holds, the paper provides a principled framework for constructing compact, multi-scale representations of high-dimensional neuroimaging data that respect known functional organization. The work is methodologically solid: it includes three-seed experiments, bootstrap confidence intervals, paired tests, and extensive ablation studies. The authors are transparent about several limitations (e.g., the 52× vs 415× comparison is not a pure rate ablation, compression ratios assume ideal bit packing, age prediction is limited by the narrow HCP-YA range). The code is released. The framework is flexible (fixed, learned, soft alignment) and could generalize to other structured scientific signals.","major_comments":[{"comment":"The central claim that HBAR’s hierarchical tokenization achieves higher parcel and network correlations than flat tokenizers at 415× compression is confounded by asymmetric training objectives. HBAR is trained with explicit parcel and network reconstruction losses (L_P, L_N in Eq. 2), while the flat baselines have no such coarse-scale supervision. Because the evaluation metric is the same atlas-based parcel/network correlation, the advantage may reflect the multi-scale objective, not the hierarchical aggregation per se. The shuffled grouping ablation at 52× (same loss structure, different grouping) partially addresses this, but no equivalent control exists at 415×. The authors should either provide an ablation where a flat baseline receives auxiliary coarse losses (e.g., parcel and network MSE from atlas pooling) or explicitly re-frame the claim to state that the advantage comes from the combination of hierarchy and multi-scale supervision, and discuss the confound clearly in the main text.","section":"§4.2 and Appendix C.3","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The 52× and 415× models differ not only in temporal token count but also in decoder capacity, code dimensions, the presence of a direct middle-residual loss, and the temporal pooling architecture. As the authors acknowledge in Appendix B, these are “two complete trained systems rather than the isolated effect of temporal token count.” The paper should therefore avoid language that implies a pure rate ablation (e.g., “enabling up to 415× compression”) without reiterating this caveat more prominently in the abstract and introduction, so that readers do not overinterpret the 415× results relative to the 52× results.","section":"§Appendix B.2 and §4.3","authors_acknowledge":"These changes in time representation, head capacity, and code allocation mean the 52×–415× comparison measures two complete trained systems rather than the isolated effect of temporal token count.","authors_acknowledge_p":0.92},{"comment":"The loss-component ablation shows that removing the temporal-difference terms (L_Δ) increases reconstruction MSE by 20% and nearly triples the atlas latent MSE. This suggests that L_Δ shapes the latent structure significantly. However, the flat baselines also include L_Δ (voxel and chunk temporal-difference losses). This is consistent with the first major comment: the coarse-scale supervision matters. The authors should clarify in the main text that HBAR’s coarse-scale advantage over flat baselines is not solely due to the hierarchical grouping; the auxiliary losses likely contribute.","section":"§Appendix C.5, Table 4","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"Line “preserving fine-grained spatial information beyond conventional parcellation”: This could be misinterpreted as claiming that HBAR reconstructs signals at a finer granularity than the atlas itself. The paper later makes clear that sub-parcel voxel groups are within fixed parcels, so the representation is still bounded by the atlas. Please rephrase to avoid implying that HBAR discovers new sub-parcel structure beyond the atlas boundaries.","section":"§Abstract and §1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The description of “voxel groups of size K=16”: Are these groups formed by deterministic ordering of voxel indices within each parcel? The paper states “voxel coordinates are normalized and ordered deterministically” — please specify the sorting scheme explicitly (e.g., lexicographic order along x, y, z) and note that groups are not spatially contiguous in the 3D volume.","section":"§3.3, “Voxel-level”","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The presentation of Table 1 would benefit from grouping the 415× methods separately from the 52× methods more clearly, and from adding a note that the MS-VQVAE baseline is a two-level latent model without atlas priors. The current footnote about “Bold and underlining” is helpful but could be complemented by a small gap between the rate blocks.","section":"§Table 1 and §5.1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The paper notes that compression ratios assume ideal fixed-length bit packing without entropy coding. This is a standard assumption, but it should also appear in the main text (e.g., in Section 4 or Section 5) so that readers who skip appendices are aware that real-world storage costs may differ.","section":"§Appendix D, compression accounting","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The evaluation uses Pearson correlation over flattened time×units. This metric can be sensitive to global scaling, though the oracle check shows near-perfect scale agreement. The paper could mention that this metric is used only for comparison and may not reflect signal-to-noise ratio or temporal fidelity equally across methods.","section":"§Appendix E, reconstruction metrics","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Table 7 shows that at 415× all models have error-to-input power ratios near 1 for frequencies above 0.10 Hz, indicating essentially no recovery of fast temporal dynamics. The authors note this limitation in the appendix text. Given how central the “415× compression with preserved multiscale signals” claim is, the spectral limitation should be mentioned briefly in the main paper’s results section (Section 5.1).","section":"§Appendix G, spectral analysis","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The limitation paragraph does not discuss the confound of asymmetric training objectives. Adding a sentence acknowledging that the comparison with flat tokenizers conflates hierarchical aggregation with multi-scale supervision would improve the paper’s fairness.","section":"§6, Limitations","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Thanks for pointing me to this one. The paper is a well-engineered study of hierarchical tokenization for fMRI, and the authors are admirably transparent about their limitations. The main claim is that fixed atlas-based hierarchical tokenization outperforms flat baselines at coarse scales under high compression, but the stress-test note hits the central experimental confound: HBAR is trained with parcel- and network-level reconstruction losses that the flat baselines simply do not receive. So the better parcel/network correlations at 415x may just reflect multi-scale supervision rather than the hierarchical architecture itself. The paper acknowledges the loss difference in Appendix C.3 but offers no control for it at the high compression rate where the biggest gains appear.\n\nThat said, this is a thorough piece of work. The systematic comparison of fixed, learned, and soft hierarchies is genuinely useful. The code is released. The experiments use three seeds, bootstrap confidence intervals, and paired tests. The authors explicitly document where comparisons are not pure rate ablations (52x vs 415x, Appendix B) and are honest about cohort and atlas limitations. The 52x shuffled grouping ablation partially addresses the loss confound by keeping the same loss structure, but the gap there is smaller (0.920 vs 0.886), and the key 415x results lack that control entirely.\n\nThe confound is load-bearing. Until a flat baseline is trained with the same coarse-scale losses (parcel and network MSE from atlas pooling), we cannot attribute HBAR's advantage at 415x to the hierarchical tokenization itself. The atlas used for grouping also sets the evaluation metrics, which is a mild circularity the paper acknowledges.\n\nThis paper is for neuroimagers working on representation learning, compression, or foundation models. It is not a breakthrough but a competent, useful study that deserves peer review – but a good reviewer will push hard on the confound. I would like to see the question addressed before citing it.\n\nYes, send it to review. The authors are serious thinkers and the work is reproducible; the main issue is fixable.","headline":"Solid engineering study of hierarchical fMRI tokenization, but evaluation confounded by asymmetric losses – the main claim needs a controlled baseline.","tokens_in":23919,"tokens_out":1658,"duration_ms":15469},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"HBAR: hierarchical tokenization preserves multiscale brain signals at 415x compression.","keywords":["fMRI","hierarchical tokenization","representation learning","compression","binary spherical quantization","resting-state","DiFuMo","Yeo networks"],"falsifier":"A study on the same dataset that replaces the atlas-based hierarchy with random grouping of voxels while matching compression rate and finds no significant difference in parcel reconstruction would falsify the claim that anatomical priors improve representation fidelity.","tokens_in":24521,"feed_emoji":"🧠","tokens_out":2756,"duration_ms":23471,"temperature":0.7,"pith_summary":"HBAR is a hierarchical tokenization framework for fMRI that compresses voxel-level brain activity up to 415 times while preserving multiscale structure. It organizes voxels into sub-parcel groups, parcels (DiFuMo-1024), and functional networks (Yeo-17), then learns discrete tokens at each level. At high compression, HBAR reconstructs parcel and network signals more accurately than rate-matched flat tokenizers. Using fixed anatomical atlases as priors improves reconstruction fidelity over fully learned hierarchies. The framework demonstrates that domain-specific organization can enhance representation learning in neuroimaging.","feed_headline":"fMRI hierarchy beats flat tokenizers at 415x compression","feed_subtitle":"HBAR organizes voxels into parcels and networks, retaining structure that flat methods lose.","key_machinery":"Hierarchical tokenization via three-level aggregation (voxel chunks → DiFuMo-1024 parcels → Yeo-17 networks) with binary spherical quantization (BSQ) applied independently at each level. The hierarchy can be fixed (atlas-based), fully learned (cross-attention pooling), or softly aligned. BSQ discretizes features into binary codes, enabling high compression without codebook collapse.","core_discovery":"HBAR demonstrates that hierarchical tokenization guided by neuroanatomical priors can compress resting-state fMRI to 415× of its original size while preserving multiscale brain signals better than flat tokenizers. At 415× compression, HBAR achieves parcel correlation of 0.746 and network correlation of 0.895, compared to 0.691 and 0.803 for flat RVQ. At 52×, preserving the anatomical chunk-to-parcel grouping raises parcel correlation from 0.886 (shuffled) to 0.920. These results show that explicit hierarchical organization through functional atlases provides a strong inductive bias for compact representation learning in high-dimensional neural data.","pith_inferences":["HBAR's advantage at coarse scales implies that downstream tasks sensitive to network-level connectivity (e.g., brain network analysis) may benefit more from hierarchical tokens than voxel-level tasks.","The modest age prediction results, partly due to narrow age range, suggest that the framework should be evaluated on lifespan datasets where age-related changes are more pronounced.","The 52× vs 415× comparison is confounded by architectural differences (not just compression rate), so a cleaner ablation varying only temporal tokens would clarify the effect of compression."],"forward_implications":["At 415× compression, HBAR enables storage and analysis of large-scale fMRI datasets that were previously impractical due to data volume.","The hierarchical tokens can be used as inputs for autoregressive models, potentially capturing temporal dynamics of brain activity at multiple scales.","The framework's design principle—embedding domain-specific atlases into tokenization—can be applied to other structured scientific signals like genomics or spatial transcriptomics.","The finding that fixed anatomical priors outperform learned hierarchies suggests that reconstruction objectives alone are insufficient to recover canonical brain organization, guiding future model design."],"supporting_citations":[{"why":"Provides the DiFuMo-1024 atlas used to define the parcel-level hierarchy and network assignments.","marker":"[7]"},{"why":"Defines the Yeo-17 functional networks used as the top-level hierarchy and for evaluation.","marker":"[28]"},{"why":"Supplies the HCP-YA resting-state fMRI dataset used for all experiments.","marker":"[30]"},{"why":"Introduces Binary Spherical Quantization (BSQ), the quantizer used at all hierarchical levels.","marker":"[41]"},{"why":"Foundational VQ-VAE work motivating discrete tokenization for representation learning.","marker":"[22]"}],"fun_headline_variants":["Hierarchical fMRI tokens beat flat ones at 415x compression","Anatomy-guided tokenization compresses fMRI 415x with better fidelity","HBAR: hierarchical tokens retain brain structure at 415x compression","Hierarchical tokenization boosts fMRI compression quality 415x"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The advantage of hierarchical tokenization depends on the atlas-based hierarchy (DiFuMo-1024 and Yeo-17) providing a meaningful and correct organization of brain function—if these atlases do not capture true functional structure, the benefit over flat tokenizers may diminish.","fun_headline_variants_meta":{"raw":{"variants":["Hierarchical fMRI tokens beat flat ones at 415x compression","Anatomy-guided tokenization compresses fMRI 415x with better fidelity","HBAR: hierarchical tokens retain brain structure at 415x compression","Hierarchical tokenization boosts fMRI compression quality 415x"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":6.3e-05,"raw_usage":{"cost":6.2708464e-05,"is_byok":false,"cost_details":{"upstream_inference_cost":6.2708464e-05,"upstream_inference_prompt_cost":3.1193344e-05,"upstream_inference_completions_cost":3.151512e-05},"total_tokens":1197,"prompt_tokens":864,"completion_tokens":333,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":256,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":259}},"tokens_in":864,"tokens_out":333,"cost_source":"openrouter_usage","duration_ms":3391,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":259,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-11-10T19:49:57.459962+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"A study on the same dataset that replaces the atlas-based hierarchy with random grouping of voxels while matching compression rate and finds no significant difference in parcel reconstruction would falsify the claim that anatomical priors improve representation fidelity.","supporting_citations":[{"cited_title":"The organization of the human cerebral cortex estimated by intrinsic functional connectivity","cited_arxiv_id":null,"evidence_quote":"Defines the Yeo-17 functional networks used as the top-level hierarchy and for evaluation."},{"cited_title":"The WU-Minn Human Connectome Project: An overview","cited_arxiv_id":null,"evidence_quote":"Supplies the HCP-YA resting-state fMRI dataset used for all experiments."},{"cited_title":"Image and Video Tokenization with Binary Spherical Quantization","cited_arxiv_id":null,"evidence_quote":"Introduces Binary Spherical Quantization (BSQ), the quantizer used at all hierarchical levels."},{"cited_title":"Neural Discrete Representation Learning","cited_arxiv_id":null,"evidence_quote":"Foundational VQ-VAE work motivating discrete tokenization for representation learning."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}