{"id":"9bafcc8a-05d6-4f86-b126-63a9eaac73c3","arxiv_id":"2507.02085","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"GeoAda uses SE(3)-equivariant adapter blocks with zero-initialized convolutions to fine-tune frozen geometric diffusion models for new control tasks with few parameters.","lead":"GeoAda adds lightweight, trainable adapter modules to frozen 3D diffusion models so they can be retrained for new tasks with extra controls, without touching the pretrained network. In tests on particles, molecules, and human motion, it matches or beats full retraining while keeping the model's original skills intact.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Frame-control coupling (Eq. 5) changes the input temporal dimension; the paper never shows the frozen backbone or its copied layers can process a different number of conditioned frames, so the 'no architectural modification' claim is unverified.","rationale":"I read the paper in good faith and its central claim is that GeoAda is a parameter-efficient, SE(3)-equivariant adapter framework that fine-tunes frozen geometric diffusion models without altering the backbone. The empirical results are consistently favorable, and the adapter design (zero-initialized equivariant convolutions, trainable copies of pretrained layers) is coherent. However, the frame-control experiment is the first and most explicitly tested instantiation of the claim, and Equation (5) changes the number of frames fed to the network. The paper does not document whether the backbone's operations are length-agnostic, and the proof of Proposition 4.1 only addresses equivariance, not input-space compatibility. This is exactly the kind of hidden assumption that can invalidate a 'no architectural modification' claim. The presence of numeric results in Table 1 suggests the authors' implementation may handle the length change, perhaps because EGTN is built over a variable-size graph, but the paper as written does not establish this. Therefore the concern does not force a rejection; it makes the central claim conditional on an architectural property that must be stated and verified. The reader's verdict of CONDITIONAL with medium risk is appropriate and I see no reason to adjust it.","tokens_in":22525,"tokens_out":11387,"duration_ms":135285,"concrete_test":"Inspect the model definition of the pretrained backbone (GeoTDM/EGTN) to determine whether the first operation on coordinates is a linear map with input dimension proportional to the number of frames T (e.g., an equivariant linear layer over the flattened 3T coordinate vector). If yes, load the released checkpoint and run the Section 5.1.1 conditional experiment with 15 conditioned frames while keeping the first layer's weights at pretrained values; a shape mismatch in the adapter's copied first layer would confirm that GeoAda requires architectural modification (or an unstated input projection), invalidating the 'without architectural modification' claim. If the backbone instead treats frames as a variable-length graph dimension (message passing over N×T nodes, no learned per-frame positions), the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim is that GeoAda adapts a frozen pretrained geometric diffusion model to new controls 'without modifying the original model architecture' (Abstract; §4.2). For frame control, the coupling operator in Eq. (5) concatenates the noised trajectory x_i^[T] with the control frames x_i^[T̃] along the temporal axis, producing an input with T+T̃ frames. Section 5.1.1 changes the number of conditioned frames from 10 (pretraining) to 15 (finetuning) in the conditional setting, and from 0 to 10 in the unconditional setting. Unless the backbone (EGTN) and its trainable copy ϵθ' are natively variable-length (e.g., treating each (particle, frame) as a graph node or using length-agnostic temporal attention), the first linear/embedding layer of the pretrained network has a weight matrix whose input dimension is tied to T, so neither the frozen model nor the copied pretrained layers can consume the concatenated input. The paper neither states that the backbone supports variable T nor describes how the copied layers are adapted (e.g., projection, re-initialization). The proof of Prop. 4.1 assumes f is SE(3)-equivariant but sidesteps dimensional compatibility; equivariance alone does not imply the input lies in the pretrained layer's domain. This is load-bearing because frame control is one of the three control types and the first experimental section; if the backbone is fixed-length, the claimed plug-and-play property breaks for this control.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GeoAda, a parameter-efficient adapter framework for fine-tuning frozen SE(3)-equivariant geometric diffusion models. Control signals of three types (global type, subgraph, frame) are injected through coupling operators, processed by trainable copies of selected pretrained layers, and projected back through decoupling operators followed by equivariant zero-initialized convolutions. The authors claim that this design preserves SE(3)-equivariance, avoids architectural modification of the pretrained backbone, mitigates overfitting and catastrophic forgetting, and achieves state-of-the-art downstream performance across particle dynamics (CHARGED PARTICLES), molecular dynamics (MD17), human motion prediction (CMU Mocap), and pocket-conditioned molecule generation (CrossDocked2020). A proof of equivariance is given in the appendix, and ablations study the number of adapter blocks, zero convolution, and trainable copies.","tokens_in":22812,"tokens_out":8236,"duration_ms":97744,"significance":"If the technical gaps identified below are resolved, GeoAda would be a useful contribution to geometric diffusion fine-tuning. The framework adapts the ControlNet recipe to SE(3)-equivariant diffusion models, and the zero-convolution device is a sensible way to protect the pretrained score at initialization. The empirical evaluation is broad, includes standard deviations, reports parameter counts, and contains ablations that support the importance of the zero-convolution and trainable-copy design choices. The paper also documents an interesting sudden-convergence phenomenon. However, the central 'plug-and-play, no architectural modification' claim is currently under-supported for two of the three control types, and the equivariance proof is more asserted than derived.","major_comments":[{"comment":"The frame-control coupling in Eq. (5) concatenates the noised trajectory x_i^[T] with the control frames x_i^[\\tilde T] along the temporal axis, producing an input with T+\\tilde T frames. Section 5.1.1 changes the number of conditioned frames from 10 at pretraining to 15 at finetuning (and from 0 to 10 in the unconditional setting). The paper never states whether the EGTN backbone and its copied layers are variable-length in the temporal dimension. If the first linear layer or temporal attention has weights tied to T, the frozen model cannot consume the concatenated input, and the claim of adaptation 'without modifying the original model architecture' fails for this control type. Please specify the backbone's length handling, or describe any projection or resizing applied to the copied layers; the equivariance proof in Appendix 7 does not address this domain-compatibility issue.","section":"§4.2, Eq. (5); §5.1.1"},{"comment":"The QM9-pretrained GeoAda model is trained with atom types (H, C, N, O, F), while CrossDocked2020 is processed with eight atom types (H, C, N, O, F, P, S, Cl). The paper does not explain how a frozen denoiser whose node-feature vocabulary was learned over five types can consume eight-dimensional one-hot vectors. If a new embedding or projection is introduced at finetuning, that is an architectural modification that must be described and justified; if the backbone is type-agnostic, that property should be stated explicitly. As written, the subgraph-control experiments do not establish the plug-and-play claim for the QM9 checkpoint.","section":"§5.3, Table 5"},{"comment":"The proof of Proposition 4.1 is a sequence of assertions rather than a derivation. It does not define an SE(3) action on each control type (notably the global-type vector \\tilde c, which is invariant under the group action), and it states without proof that the coupling operators 'respect the SE(3) structure.' A rigorous proof should verify f(h·G, h·C)=h·f(G,C) and g(h·G'')=h·g(G'') for each of the three controls, and should also specify how the equivariant zero-convolution in Eq. (6) behaves under translations in the centered representation used by the backbone.","section":"Appendix 7, Prop. 4.1"},{"comment":"The abstract and §4.4 claim that GeoAda 'consistently matches or outperforms full fine-tuning baselines' and avoids catastrophic forgetting. For subgraph control, however, the comparison in Tables 5–7 is against task-specific methods trained from scratch on CrossDocked2020 (liGAN, GraphBP, AR, Pocket2Mol, TargetDiff); no Full FT, PARTIAL-k, prompt, or adapter ablation is reported for this control type. The subgraph experiment therefore supports competitiveness against specialized generators but not the paper's central fine-tuning claim, and should be supplemented or the claim narrowed.","section":"§5.3, Tables 5–7"}],"minor_comments":[{"comment":"There are several typos and grammar issues, including 'Frame Fontrol', 'freezed', and 'sythetic accessibility'; the manuscript should be proofread.","section":"Abstract and Section 1"},{"comment":"Table 1 contains malformed entries such as '3924 3950' and isolated 'nan' cells; please reformat the table so that missing values and standard deviations are presented consistently.","section":"Table 1"},{"comment":"The ring-size row for size 7 reports '23.5' without a percent sign, unlike the other entries in the table.","section":"Table 7"},{"comment":"The notation for W_{t,s} and γ is confusing: W_{t,s} is said to be in R^N while γ is a vector over T, and the construction of the weights w(t,s) should be defined with explicit dimensions.","section":"Appendix 8.4.1"},{"comment":"The PARTIAL-k baseline is cited as [10, 14, 43], but reference [43] appears to be a colorization paper rather than a partial fine-tuning method; please check the citation.","section":"Section 5, Baselines"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely from the same group as the GeoTDM/EGTN backbone, and the experiments depend heavily on that backbone. This is not improper, but the authors should be pushed to disclose and verify the backbone's variable-length and atom-type vocabulary properties, since the central plug-and-play claim rests on them. The empirical work is otherwise substantial, and the missing baselines in the subgraph experiment are fixable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper you asked about is GeoAda. Short version: it works, the empirical case is solid, and it deserves a real referee. The idea is a ControlNet-style adapter for geometric diffusion models: freeze the pretrained denoiser, add trainable copies of selected layers with coupling/decoupling operators for the control, and gate with zero-initialized equivariant convolutions. That combination is new for geometric diffusion, and the paper shows it convincingly across four domains (particles, MD17, CMU Mocap, QM9/CrossDocked). The key empirical finding—GeoAda matches or beats full fine-tuning while preserving the pretraining task—holds across the board, which is the central claim. I believe it.\n\nThe equivariance proof (Prop. 4.1, Appendix 7) is more of an assertion than a derivation. The individual steps are correct: f and g are equivariant for the three control types, and the composition inherits equivariance. But the paper doesn't actually spell out why the coupling operators are equivariant; it just states it. A referee should ask for a clean, case-by-case proof.\n\nThe frame-control worry from the stress-test—that changing from 10 to 15 conditioning frames breaks the frozen backbone because the input dimension changes—does not hold up if the backbone is what I think it is. GeoTDM/EGTN treats a trajectory as a spatiotemporal graph whose nodes are (particle, frame) pairs. Concatenating frames adds nodes; it does not change the per-node feature dimension. The paper never says this explicitly, which is a clarity problem, not a correctness problem. The authors should add a sentence saying the backbone is length-agnostic with respect to the number of frames.\n\nOther soft spots: the tables are littered with formatting errors, \"nan\" entries, and at least one nonsensical value (\"3924 3950\"). No code is released. These are fixable. The limitations section is honest about the handcrafted coupling operators and the lack of large-scale validation.\n\nBottom line: the contribution is moderate in novelty but high in usefulness, and the evidence supports the claims. This should go to peer review. I would cite it if I work on geometric diffusion or equivariant adapters.","headline":"A solid, well-empirically-supported equivariant adapter method for geometric diffusion models that deserves peer review, with a sketchy proof and presentation issues that are fixable.","tokens_in":23361,"tokens_out":4539,"would_cite":true,"duration_ms":54299,"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":"GeoAda shows that a frozen geometric diffusion model can be steered to new conditioning tasks by training only small SE(3)-equivariant adapter modules, which preserve the model's symmetries and its original task performance.","keywords":["geometric diffusion models","SE(3)-equivariant adapters","parameter-efficient fine-tuning","geometric control","catastrophic forgetting","equivariant zero-convolution","molecule generation","trajectory prediction"],"falsifier":"Fine-tune GeoAda on CHARGED PARTICLES with the conditioning length raised from 10 to 50 frames instead of 10 to 15: if the frozen backbone errors on the longer input, ignores the added frames, or shows no gain over the 10-frame baseline, then the claim that frame control adapts without architectural modification is falsified.","tokens_in":22289,"feed_emoji":"🧩","tokens_out":11187,"duration_ms":109344,"temperature":0.7,"pith_summary":"GeoAda is a recipe for adapting a pretrained geometric diffusion model to downstream tasks with new conditioning signals while leaving the original network frozen. The paper argues that coupling operators can inject three kinds of geometric control — global type vectors, subgraphs such as protein pockets, and extra observed frames of a trajectory — into a trainable copy of selected denoiser layers, and that an equivariant zero-initialized convolution can merge the result back into the frozen score. Because only the lightweight adapter blocks are trained, the model keeps its geometric inductive bias, avoids overfitting on small downstream datasets, and does not forget the task it was pretrained on. The authors prove the composition stays $\\mathrm{SE}(3)$-equivariant and show across particle dynamics, molecular dynamics, human motion, and molecule generation that the adapted model matches or beats full fine-tuning on the new task while retaining original-task accuracy.","feed_headline":"Equivariant adapters retune frozen diffusion models for new tasks","feed_subtitle":"GeoAda's lightweight adapter blocks let frozen diffusion models learn new controls without forgetting old ones.","key_machinery":"The carrying object is the equivariant adapter block, a four-stage module: a coupling operator $f$ that injects the control into the noised input without breaking its geometric structure, a trainable copy $\\epsilon_{\\theta'}$ of a subset of the frozen denoiser layers (picking the first layer of every $K$ consecutive layers proved best), a decoupling operator $g$ that selects the predicted components belonging to the original input, and an equivariant zero-convolution $z_\\phi(\\{x_i\\},\\{h_i\\}) = (\\phi_x (x_i - \\bar{x}), \\phi_h h_i)$ whose learnable weights start at zero. The zero initialization makes the adapter output vanish exactly at initialization, so the pretrained score is untouched at the start of fine-tuning while the gradients through $\\phi_x$ and $\\phi_h$ are nonzero, letting the control path learn smoothly; the center-of-mass subtraction $\\bar{x}$ makes the output translation-invariant, and rotation equivariance is carried by the geometric feature transformations.","core_discovery":"The paper's central claim is that a pretrained geometric diffusion model can be redirected to new controlled-generation tasks without touching its weights: the frozen denoiser $\\epsilon_\\theta$ is supplemented by an equivariant adapter $s_{\\theta',\\phi}$ whose output is added to the original score at sampling time. The adapter encodes the control $C$ through a coupling operator $f$ — node-wise addition of a lifted global vector for global type control, graph union for subgraph control, and frame concatenation for frame control — processes it with a trainable copy $\\epsilon_{\\theta'}$ of selected pretrained layers, extracts the relevant nodes or frames with a decoupling operator $g$, and scales the result with an equivariant zero-initialized convolution $z_\\phi$. Proposition 4.1 states that $g \\circ \\epsilon_\\theta \\circ f$ is $\\mathrm{SE}(3)$-equivariant for every control type, so the adapted score inherits the invariant marginal of Eq. (1). On CHARGED PARTICLES, five MD17 molecules, five CMU Mocap actions, and QM9/GEOM-Drugs to CrossDocked2020 pocket-conditioned generation, GeoAda matches or outperforms full fine-tuning on the downstream metric while the pretraining-task metrics stay at their original values, where full fine-tuning, partial fine-tuning, and prompt baselines degrade or collapse.","pith_inferences":["The adapter recipe should port to other equivariant generative paradigms such as score-based SDEs or equivariant flow matching, since the proof only relies on the frozen network being $\\mathrm{SE}(3)$-equivariant; the paper does not claim this.","If implicit regularization is the mechanism, the gap between GeoAda and full fine-tuning should widen as downstream data shrink; a controlled data-amount sweep would test this directly.","The three handcrafted coupling operators are the least general component: controls with high-dimensional or structured semantics (for example, text embeddings) would likely need learned coupling, a direction the paper's limitation section leaves open.","Frame control is the fragile case: increasing the conditioning length from 10 to 15 frames changes the backbone's input dimension, and the paper never verifies that the frozen architecture truly handles variable-length inputs."],"forward_implications":["Adapting to a new control with GeoAda costs about half the parameters of full fine-tuning (roughly 2.7 MB tunable on the tested datasets), which makes transfer practical for small downstream datasets.","The pretraining task survives adaptation: GeoAda keeps pretrain metrics at their original values on CHARGED PARTICLES and MD17, while full fine-tuning, partial fine-tuning, and prompt baselines degrade or collapse.","Training is punctuated by a 'sudden convergence': loss and validation MSE drop abruptly around steps 4500–4700, with ADE and FDE falling by 68.3% and 73.4%, so early-stopping rules based on gradual improvement can be misleading.","The choice of which pretrained layers to copy matters: selecting the first layer of every $K$ consecutive layers outperforms choosing only early or only late layers under the same parameter budget, making layer selection itself a tunable design choice."],"supporting_citations":[{"why":"The ControlNet design of frozen-backbone adapter branches plus zero-initialized convolution that GeoAda extends to SE(3)-equivariant geometric denoisers.","marker":"[42]"},{"why":"Supplies the pretrained geometric trajectory diffusion model and equivariant backbone that GeoAda freezes and adapts.","marker":"[9]"},{"why":"Establishes SE(3)-equivariant denoising for 3D molecule generation, the equivariance guarantee that GeoAda's adapters must preserve.","marker":"[12]"},{"why":"Provides the denoising objective and linear noise schedule used to train the adapters via the fine-tuning loss of Eq. (7).","marker":"[11]"},{"why":"The universal graph prompt baseline that GeoAda compares against and outperforms on downstream tasks.","marker":"[5]"},{"why":"TargetDiff, the strongest task-specific baseline, and the CrossDocked2020 protocol used for pocket-conditioned molecule generation.","marker":"[8]"}],"fun_headline_variants":["Frozen diffusion models get new tasks via equivariant adapters","SE(3)-equivariant adapters fine-tune diffusion without forgetting","Adapters add new controls to frozen diffusion models","GeoAda adapts SE(3) diffusion with minimal new parameters"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The frame-control setup assumes the frozen pretrained denoiser can accept a trajectory whose number of conditioned frames changes from 10 at pretraining to 15 at fine-tuning, since the coupling operator just concatenates the extra frames and the paper never shows how the backbone's time embeddings or attention layers handle the new input dimension.","fun_headline_variants_meta":{"raw":{"variants":["Frozen diffusion models get new tasks via equivariant adapters","SE(3)-equivariant adapters fine-tune diffusion without forgetting","Adapters add new controls to frozen diffusion models","GeoAda adapts SE(3) diffusion with minimal new parameters"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000659,"raw_usage":{"total_tokens":3077,"prompt_tokens":1068,"completion_tokens":2009,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":684,"completion_tokens_details":{"reasoning_tokens":1938}},"tokens_in":684,"tokens_out":2009,"duration_ms":16144,"temperature":1.0,"reasoning_tokens":1938,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:38:49.913432+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fine-tune GeoAda on CHARGED PARTICLES with the conditioning length raised from 10 to 50 frames instead of 10 to 15: if the frozen backbone errors on the longer input, ignores the added frames, or shows no gain over the 10-frame baseline, then the claim that frame control adapts without architectural modification is falsified.","supporting_citations":[{"cited_title":"Adding conditional control to text-to-image diffusion models","cited_arxiv_id":null,"evidence_quote":"The ControlNet design of frozen-backbone adapter branches plus zero-initialized convolution that GeoAda extends to SE(3)-equivariant geometric denoisers."},{"cited_title":"Geometric Trajectory Diffusion Models","cited_arxiv_id":"2410.13027","evidence_quote":"Supplies the pretrained geometric trajectory diffusion model and equivariant backbone that GeoAda freezes and adapts."},{"cited_title":"Equivariant diffusion for molecule generation in 3d","cited_arxiv_id":null,"evidence_quote":"Establishes SE(3)-equivariant denoising for 3D molecule generation, the equivariance guarantee that GeoAda's adapters must preserve."},{"cited_title":"Denoising diffusion probabilistic models","cited_arxiv_id":null,"evidence_quote":"Provides the denoising objective and linear noise schedule used to train the adapters via the fine-tuning loss of Eq. (7)."},{"cited_title":"Universal prompt tun- ing for graph neural networks","cited_arxiv_id":null,"evidence_quote":"The universal graph prompt baseline that GeoAda compares against and outperforms on downstream tasks."}],"review_version":1}