{"id":"dc2d919c-2feb-4c22-932d-efd2a2fd5955","arxiv_id":"2608.09572","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"The paper proposes projecting continual updates away from old-task spatial directions in hyperbolic multimodal models, claims this is theoretically required to prevent forgetting, but the necessity claim is unproven and the experiments are confounded by a step-pullback.","lead":"Hyperbolic Multimodal Continual Learning proposes a geometry-preserving update rule for continual learning in hyperbolic vision-language models and claims a theoretical characterization based on shared isometries. The experiments show large reductions in forgetting, but the reported gains are confounded by an extra step-pullback mechanism not applied to the baselines, and the central theory overclaims what the theorem proves.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Step-pullback regularizer may explain HMCL's forgetting gains: baselines are not reported to use it, so the ablation does not isolate the geometric projection from this strong weight-anchoring mechanism.","rationale":"The paper's headline claim has two pillars: the Theorem 1 characterization and the experimental demonstration that HMCL reduces forgetting. The proof of Theorem 3 contains an unjustified use of the origin embedding, but this is repairable with the rank argument under the stated non-degeneracy, so it is not the most damaging issue. The more load-bearing weakness is the experimental confound: the step-pullback is a post-hoc regularizer not derived from the geometry, and the baselines are not reported to use it. Since the ablation 'w/o block' retains the pullback, no experiment isolates the projection. If the pullback is responsible for the BWT gains, the paper's empirical support for the geometric mechanism collapses, leaving only a theoretical result conditional on the P1–P3 formalization. The reader's verdict is CONDITIONAL and explicitly requested an ablation without pullback; this concern reinforces that condition rather than changing the verdict, so no adjustment is needed.","tokens_in":32798,"tokens_out":9529,"duration_ms":87319,"concrete_test":"Re-run the Section 6 protocol with the step pullback applied identically to the Vanilla baseline (and preferably to EWC, GEM, and C-FLAT) using the same β values and schedule as HMCL. If the BWT gap between HMCL and Vanilla collapses or narrows substantially, the gains are attributable to the pullback. Also run HMCL with β=0 (pullback disabled) while retaining the projection; if BWT degrades to near-Vanilla levels, the pullback, not the projection, is the active ingredient.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 6.1 states that 'after every projected AdamW optimizer step ... apply the step pullback W←(1−β)W+βW_{t,0}' with β=0.10 for MERU backbones and β=0.20 for HyCoCLIP-B. The baseline paragraph says baselines share backbone, task stream, optimizer, and learning-rate schedule, but does not state that they receive this pullback; Table 1's caption says 'HMCL uses the reported step-pullback configuration.' The pullback is a strong regularizer that anchors the head to the task-start snapshot, and by itself can reduce backward transfer. The supplementary ablation in Figure 5 (HMCL w/o block) still contains the pullback, so it only removes the block constraint, not the pullback. Consequently, the experimental comparison does not isolate the null-space projection derived from Theorem 1; the reported BWT improvements could be driven by the pullback rather than by geometry-preserving updates. This directly affects the paper's central empirical claim that the geometric update mechanism is effective.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies continual learning for multimodal representations embedded in hyperbolic (Lorentz) space. It defines stability as exact preservation of three geometric invariants: intra-modal Lorentzian inner products (P1), inter-modal Lorentzian inner products (P2), and spatial norms related to hierarchy (P3). The main theoretical result, Theorem 1, states that under a non-degeneracy condition these preservation conditions hold if and only if the old-task embeddings are transformed by a single shared spatial rotation acting identically across modalities. From this characterization the authors derive a first-order constraint on parameter updates, leading to a closed-form null-space projection method called HMCL. Experiments on classification and retrieval benchmarks with three hyperbolic backbones report substantially reduced backward transfer and geometric drift relative to Euclidean continual-learning baselines.","tokens_in":33007,"tokens_out":6974,"duration_ms":66126,"significance":"If the theoretical characterization and the experimental isolation were both sound, this would be a valuable first geometric study of hyperbolic multimodal continual learning: the update rule is simple, introduces no additional learnable parameters, and has memory overhead constant in the number of tasks. The paper also goes beyond aggregate metrics by measuring radial, angular, cross-modal, and paired-distance drift, which is a useful diagnostic connection between geometry and forgetting. However, the central empirical claim is currently confounded by the step-pullback regularizer, and the proof of the main theorem contains a gap concerning the use of the hyperbolic origin. The theoretical framework is also conditional on the modeling assumption that preservation of P1–P3 is the correct formalization of 'preventing forgetting.' These issues are locally addressable but are load-bearing for the paper's main claims.","major_comments":[{"comment":"The proof of Theorem 3 ('Boost Elimination') applies condition (P3) to the specific embedding z=(1,0,...,0), i.e., the hyperbolic origin. The origin is not an old-task embedding and is not guaranteed by the non-degeneracy condition, which only ensures that the rows of the joint embedding matrix span R^{1,d}. Because (P3) is a per-point nonlinear constraint, showing that a boost fails to preserve the origin's spatial norm does not imply that the boost fails on a spanning set of old-task embeddings. Consequently the 'only if' direction of Theorem 1 is not established as stated. A correct proof needs a lemma showing directly that if L in SO+(1,d) preserves spatial norms on a spanning subset of H^d, then L is a pure spatial rotation; if such a lemma is false, the theorem needs an explicit additional assumption.","section":"Appendix E.1, Theorem 3"},{"comment":"The step pullback W←(1−β)W+βW_{t,0} is applied to HMCL but is not reported for the baselines, and the ablation 'HMCL w/o block' in Figure 5 still contains the pullback. This pullback is a strong weight-anchoring regularizer that can reduce backward transfer independently of the geometric projection. The baseline paragraph states that baselines share the backbone, task stream, optimizer, and learning-rate schedule, but not the pullback. Therefore the experiments do not isolate the null-space projection derived from Theorem 1, and the claim that 'geometry-aware constraints are key to stable multimodal continual learning' is not yet supported. The authors should add control baselines trained with the same pullback and an HMCL variant without the pullback.","section":"Section 6.1, Table 1, Figure 5"},{"comment":"The paper defines stability as the preservation conditions (P1)–(P3) and then concludes that preventing forgetting requires a shared hyperbolic isometry. This conclusion is relative to that definition: exact preservation of all pairwise Lorentzian inner products and spatial norms is sufficient but not shown to be necessary for maintaining task performance, which could in principle be preserved by replay, task-specific heads, or other mechanisms. The abstract's stronger phrasing ('preventing forgetting requires cross-modal invariance') should be qualified, or the paper should supplement Theorem 1 with an argument linking (P1)–(P3) to downstream task performance.","section":"Section 3.2 and Theorem 1"}],"minor_comments":[{"comment":"The 'non-degeneracy condition' is not defined in the main text; it appears only in the appendix proofs as a rank condition on the joint embedding matrix. Please state the condition explicitly in the theorem statement.","section":"Section 4.1, Theorem 1"},{"comment":"Corollary 1 says 'If the time-like coordinate is required to be first-order invariant, i.e. (P3),' but (P3) is an exact equality, not a first-order condition. Please rephrase to distinguish the exact preservation condition from its first-order implication.","section":"Section 4.2, Corollary 1"},{"comment":"The drift reductions are reported as aggregate percentages with a p-value but without confidence intervals in the main text. Please report per-dataset error bars or a table with confidence intervals so the reader can assess the variability across old tasks.","section":"Section 6.3, Figure 3, Table 7"},{"comment":"The caption of Figure 5 should state explicitly that the 'w/o block' variant retains the step pullback; currently the reader must infer this from Section 6.1, and the caption could easily be misread as an ablation of the full geometric mechanism.","section":"Appendix G.2, Figure 5"}],"recommendation":"major_revision","confidential_remarks":"This is a promising submission whose central claims are not yet fully supported. The most urgent issue is the step-pullback confound: without a control condition, the empirical section does not test the proposed geometric mechanism. The proof gap in Theorem 3 is likely fixable but must be addressed before the theoretical characterization can be accepted. I recommend major revision rather than rejection, because both issues appear addressable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":null,"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","53B30","15A63"],"pacs":[],"model":"deepseek-v4-flash","headline":"Preventing forgetting in hyperbolic multimodal learning is equivalent to preserving old-task embeddings under one shared spatial rotation, and the paper derives a closed-form null-space projection that implements this constraint.","keywords":["hyperbolic geometry","continual learning","multimodal learning","catastrophic forgetting","Lorentz model","contrastive learning","representation preservation","null-space projection"],"falsifier":"Train a hyperbolic multimodal model on a task stream using pure replay of old samples, with no geometric constraints, and measure both old-task accuracy and the extended Lorentz Gram matrices of old-task embeddings; if accuracy on previous tasks remains flat while those Gram matrices change, the paper's necessity claim would be refuted, whereas if accuracy drops whenever the Gram matrices change, the claim would be supported.","tokens_in":32579,"feed_emoji":"📐","tokens_out":6960,"duration_ms":57592,"temperature":0.7,"pith_summary":"This paper studies what happens to multimodal representations that live in hyperbolic space when a model is trained on a stream of tasks. It claims that preventing catastrophic forgetting in this setting requires old-task embeddings to stay invariant under one shared hyperbolic isometry: the same spatial rotation applied to every modality. The paper proves a necessary-and-sufficient characterization: under a non-degeneracy condition, preserving intra-modal, inter-modal, and hierarchical structure across continual stages is equivalent to the existence of a shared rotation $R=\\mathrm{diag}(1,\\tilde R)$ with $\\tilde R\\in SO(d)$ applied to all old-task embeddings. From this, it derives HMCL, which projects each parameter update onto the null space of old-task spatial representations, and reports large reductions in backward transfer on classification and retrieval benchmarks. If the proof holds, it gives a geometric first principle for designing continual learners for hyperbolic multimodal models rather than adapting Euclidean strategies.","feed_headline":"Old tasks survive only under a shared spatial rotation","feed_subtitle":"A new theorem ties old-task retention to one shared spatial rotation; the HMCL update is the closed-form projection it implies.","key_machinery":"The central object is the extended Lorentz Gram matrix $G^s_{t,\\mathrm{ext}}=Z^s_{t,\\mathrm{all}}G(Z^s_{t,\\mathrm{all}})^{\\top}$ formed from the joint embedding matrix of all modalities; preserving it encodes both intra-modal and inter-modal invariance. The paper's argument runs through three geometric tools: Witt's extension theorem to show that equal Gram matrices force a global isometry, a uniqueness result for the proper orthochronous Lorentz group $SO^+(1,d)$ acting identically across modalities, and a boost-elimination argument showing that the spatial-norm condition (P3) forces that isometry to be a pure spatial rotation. The algorithmic mechanism that carries the result is the null-space projector $I-P^{t-1}$ onto the orthogonal complement of the old-task spatial subspace, computed once per stage from a low-rank basis; it is what turns the theorem into a closed-form update rule.","core_discovery":"On its own terms, the paper establishes that the preservation conditions (P1), (P2), and (P3) — exact preservation of intra-modal and inter-modal Lorentzian inner products and of spatial norms for old-task embeddings — hold between stage $t-1$ and stage $t$ if and only if there is a single spatial rotation $R=\\mathrm{diag}(1,\\tilde R)$, $\\tilde R\\in SO(d)$, with $Z^{m,t}_{t-1}=Z^{m,t-1}_{t-1}R^{\\top}$ for every modality $m$. This means a Lorentz boost, which changes the spatial norm and therefore hierarchical specificity, is incompatible with preserving old-task knowledge, while a rotation shared across modalities preserves both cross-modal similarity and entailment-based hierarchy. The proof combines preservation of the extended Lorentz Gram matrix, uniqueness of the proper orthochronous Lorentz isometry under full-rank embeddings, and elimination of boosts via the spatial-norm condition. The paper then derives the first-order condition that admissible parameter updates are exactly those whose spatial part lies in the null space of the old-task spatial representations, with the time-like block fixed, yielding the closed-form update $\\Delta W^m_s = \\delta^m_s(I-P^{t-1})$.","pith_inferences":["The necessity claim is conditional on the paper's choice to define forgetting as exact preservation of all pairwise Lorentzian inner products and spatial norms; replay-based or task-specific-head methods could in principle maintain task accuracy while letting those invariants drift.","The same Gram-matrix-plus-null-space argument should transfer to other pseudo-Riemannian representation spaces; a natural test is to replace the Lorentz metric with another nondegenerate bilinear form and check whether the shared-isometry theorem still holds.","The first-order projection is designed for a linear Lorentz transformation layer; extending the projection to full-backbone hyperbolic fine-tuning would require an approximate tangent-space projection, and the paper's own appendix notes such higher-order residuals are not exactly controlled.","A cheap diagnostic suggested by the analysis: monitor time-like (radial) drift alone as an early warning for hierarchy collapse, since first-order hierarchy preservation forces spatial updates orthogonal to the current direction."],"forward_implications":["Any continual update that preserves old-task hyperbolic knowledge must act as one shared spatial rotation across modalities; a boost that changes spatial norms will break the hierarchy condition.","The admissible-update rule is a closed-form projection: unconstrained gradients are multiplied by $I-P^{t-1}$, so no extra parameters or per-task bases are needed and memory stays constant.","Geometry-agnostic continual methods such as EWC, GEM, and C-FLAT leave the Lorentz invariants unprotected, which explains their limited gains in the reported experiments.","In the reported benchmarks HMCL improves overall performance and reduces backward transfer substantially; for example classification BWT on MERU-L moves from $-6.46$ to $-0.74$.","Preserving the invariants also preserves the entailment cone partial order, so the image-to-root traversals remain semantically coherent after subsequent tasks."],"supporting_citations":[{"why":"Supplies the MERU hyperbolic vision-language backbones and the origin-distance formulation of semantic hierarchy used throughout the experiments.","marker":"(Desai et al., 2023)"},{"why":"Defines the Lorentz transformation layer $f(W;z)$ that HMCL updates and constrains.","marker":"(Chen et al., 2022)"},{"why":"Provides Witt's extension theorem used in the proof that equal extended Lorentz Gram matrices imply a global isometry.","marker":"(Lam, 2005)"},{"why":"Supplies the generation of $SO^+(1,d)$ by rotations and boosts, used to eliminate boosts via condition (P3).","marker":"(Moretti, 2002)"},{"why":"Establishes the Lorentz model of hyperbolic space used for embeddings and distance.","marker":"(Nickel & Kiela, 2018)"},{"why":"Contributes the SVD/PCA projection strategy and the GEM baseline that HMCL's projector adapts.","marker":"(Lopez-Paz & Ranzato, 2017)"},{"why":"Frames catastrophic forgetting and provides the EWC baseline compared in experiments.","marker":"(Kirkpatrick et al., 2017)"},{"why":"Identifies intra-modal rotation and inter-modal deviation in continual vision-language learning, the empirical target HMCL addresses.","marker":"(Ni et al., 2023)"}],"fun_headline_variants":["Only a shared spatial rotation keeps old tasks alive","No forgetting without one rotation shared by all modalities","Shared rotation: the only way to preserve old tasks in hyperbolic learning","Preserving old tasks requires one shared cross-modal rotation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that avoiding forgetting is exactly the same as preserving all pairwise Lorentzian inner products and spatial norms of old-task embeddings; if task accuracy could survive while these geometric quantities change, the shared-isometry requirement would be overkill.","fun_headline_variants_meta":{"raw":{"variants":["Only a shared spatial rotation keeps old tasks alive","No forgetting without one rotation shared by all modalities","Shared rotation: the only way to preserve old tasks in hyperbolic learning","Preserving old tasks requires one shared cross-modal rotation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001314,"raw_usage":{"total_tokens":5342,"prompt_tokens":923,"completion_tokens":4419,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":4355}},"tokens_in":539,"tokens_out":4419,"duration_ms":28283,"temperature":1.0,"reasoning_tokens":4355,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:51:51.026920+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a hyperbolic multimodal model on a task stream using pure replay of old samples, with no geometric constraints, and measure both old-task accuracy and the extended Lorentz Gram matrices of old-task embeddings; if accuracy on previous tasks remains flat while those Gram matrices change, the paper's necessity claim would be refuted, whereas if accuracy drops whenever the Gram matrices change, the claim would be supported.","supporting_citations":[{"cited_title":"2005 , publisher=","cited_arxiv_id":null,"evidence_quote":"Provides Witt's extension theorem used in the proof that equal extended Lorentz Gram matrices imply a global isometry."}],"review_version":1}