{"id":"cf4fa6b2-e5bb-4a91-ba22-2118d000ba52","arxiv_id":"2501.16559","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"LoRA-X is a low-rank adapter constrained to a base model's singular subspace, enabling data-free transfer to closely related models via closed-form projection.","lead":"This paper introduces LoRA-X, an adapter that can be moved from one foundation model to another without retraining or access to the original training data. The method works by keeping the adapter inside the base model's mathematical subspace and projecting it onto the new model's subspace.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 3 and Algorithm 1 conflate full and truncated SVD; for the square attention weights used here, U_t U_t^T and V_t V_t^T are identity, making the claimed transfer projection either a no-op or an unstated rank-r truncation.","rationale":"The reader's weakest assumption was that source and target attention modules have sufficiently overlapping singular subspaces, and that manual module selection plus missing variance weaken the result. That is a fair empirical concern. My stress-test found a more basic formal issue: the projection operation, as written in Eq. 3 and as implemented in Algorithm 1, is the identity map for square attention matrices because full orthogonal SVD factors satisfy U U^T = I and V V^T = I. The paper's notation conflates full singular matrices U,V with the truncated \\tilde U,\\tilde V defined in Eq. 1. If the projection is truly the identity, then the claimed cross-model alignment is not happening, and the mechanism behind the method is different from what is described. If the authors intended a truncated projection, they need to say so explicitly, fix Eq. 3 and the pseudocode, and show that the truncated projection is what produces the reported numbers. This does not necessarily invalidate the empirical results, but it makes the central claim non-reproducible as written. Therefore I keep the reader's conditional verdict, but the condition is stronger: the paper must correct the SVD truncation inconsistency and demonstrate that the reported transfer results come from a non-trivial subspace projection.","tokens_in":19498,"tokens_out":8501,"duration_ms":81275,"concrete_test":"Check the shapes of all LoRA-X target weight matrices in the SD-v1.5→SD Eff-v1.0 and SDXL→SSD-1B experiments. For each square matrix A, compute U,S,Vh = torch.linalg.svd(A, full_matrices=False) and verify whether U@U.T and Vh.T@Vh are identity (they are for square A). Then re-run the BlueFire transfer with three variants: (a) the published Algorithm 1, (b) a truncated version using only the top-320 left/right singular vectors, and (c) no projection, i.e., adding the source adapter directly to the target. If (a) equals (c) and differs from (b), the reported projection is an artifact of unstated truncation. If (b) reproduces the table's HPSv2 of 0.3073, the paper must correct Eq. 3 and Algorithm 1 and explicitly state that transfer is a rank-r truncated-subspace projection, not the full-SVD identity.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim rests on projecting the source adapter onto the target's singular subspace. In Eq. 1 the authors define truncated matrices \\tilde U and \\tilde V of rank r, but Eq. 3 writes the transfer as ΔW_{t←s}=U_t U_t^T ΔW_s V_t V_t^T = U_t ΔΣ_{t←s} V_t^T. For the dimensions to match, U_t and V_t in Eq. 3 must be the r-column truncated versions. However, the surrounding text and Algorithm 1 use torch.linalg.svd(full_matrices=False), which for square matrices (the attention linear layers used throughout, e.g., 4096×4096 or 1024×1024) returns full orthogonal U and Vh. For such square matrices, U_t U_t^T = I and V_t V_t^T = I, so the projection is the identity map and the transferred adapter is simply the source adapter applied unchanged to the target. If instead truncation to the top r=320 singular vectors is intended, it must be stated and applied consistently; otherwise Table 5's contrast between \"subspace projection\" and \"copying ΔΣ_s\" cannot be reproduced from the paper. This ambiguity is load-bearing because every reported transfer and the core notion of \"operating within the subspace of the target\" depend on this operation. The reader's subspace-overlap concern is valid, but the paper has not yet specified a projection that is both well-defined and non-trivial on the models evaluated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LoRA-X, an adapter whose weight update is constrained to the top-r singular subspace of the base model weight matrix, so that only the corresponding singular-value changes are learned. For cross-model transfer, the source adapter is projected onto the target model's singular subspace via U_t U_t^T ΔW_s V_t V_t^T, and only modules with sufficiently high subspace similarity are transferred. The authors report experiments on Stable Diffusion v1.5 and SDXL families, including style-transfer datasets, an optimal-transport-based Adapter Transferability Cost (ATC), a comparison with X-Adapter, and a TinyLlama text-generation experiment, claiming that the training-free transferred adapter performs comparably to one trained directly on the target model.","tokens_in":19857,"tokens_out":6792,"duration_ms":61277,"significance":"If correct, the central claim is practically valuable: LoRA adapters trained on a deprecated base model could be reused on successor or distilled models without access to original or synthetic data and without gradient updates. The method is closed-form and uses only pre-trained weights, which is an appealing and falsifiable design. The paper also contributes a subspace-similarity metric and an ATC metric that could be useful for predicting transfer difficulty. However, the significance is currently limited by an ambiguity in the central projection operation: for the square attention matrices used in the experiments, the pseudocode's full SVD makes the projection the identity map, so the non-triviality of the claimed transfer is not established as written.","major_comments":[{"comment":"The transfer operation is not well-defined as written. In Eq. (3), the matrices U_t and V_t must be the r-column truncated singular matrices for the product U_t^T \\tilde U_s ΔΣ_s \\tilde V_s^T V_t to be r×r; with the full singular matrices used in the text and in Algorithm 1 (torch.linalg.svd(..., full_matrices=False) on square attention weights), U_t U_t^T = I and V_t V_t^T = I, making the projection the identity and the transferred adapter identical to the source adapter. The reported benefit of subspace projection in Table 5 therefore cannot be reproduced from the paper unless the rank-r truncation is stated and applied consistently in the pseudocode. Please clarify which SVD is used and modify Eq. (3) and Algorithm 1 accordingly.","section":"§4.2.1 and Algorithm 1"},{"comment":"The different-dimension alignment in §4.2.2 is dimensionally inconsistent. With \\tilde U_s ∈ R^{m×r} and U_t ∈ R^{m′×r}, the expression \\tilde U_s = U_t U_s^T (U_s U_s^T)^{-1} U_s does not have matching dimensions (U_t U_s^T is m′×m while the inverse is r×r), and it also redefines \\tilde U_s, which was already the truncated left singular matrix of the source. A correct least-squares alignment should be stated explicitly, or Section 4.2.2 should be restricted to the same-dimension case actually used in the experiments.","section":"§4.2.2"},{"comment":"The X-Adapter comparison in Table 4 is not controlled: the LoRA-X transfer uses SSD-1B as source while the X-Adapter transfer uses SD-v1.5, and the text attributes part of the DINOv2 gain to this source difference. Since source-family similarity is itself a major factor in the proposed method (Figure 4), the comparison does not isolate the transfer method. Please either use the same source for both methods or present the comparison as illustrative only and temper the corresponding claim.","section":"§5.4, Table 4"},{"comment":"The central claim that transferred LoRA-X performs comparably to trained LoRA-X rests on small HPSv2 differences, but the paper reports no variance or significance tests (only 30-seed averages), and several LPIPS-diversity drops are large (e.g., Table 1, SSD-1B: -8.4% and -13.2%; RealVisXL-v3.0: -6.1% and -9.6%). Without standard deviations, confidence intervals, or per-seed results, the 'comparable' conclusion is not yet supported. Please add error bars or significance analysis.","section":"§5.2, Tables 1 and 9"},{"comment":"Module selection is performed manually: the paper states that for low-similarity blocks the authors 'seek another transformer block' and that some up-block modules receive no transfer. This manual matching, rather than the stated similarity threshold, determines which layers are transferred and could introduce selection bias. To make the method reproducible and the ATC metric predictive, the selection rule should be automated and specified (e.g., a numeric threshold with the chosen value).","section":"Appendix B.1.1"}],"minor_comments":[{"comment":"The reference 'Klaudia Ba/suppress lazy' appears to contain corrupted text and should be corrected to the proper author name.","section":"References"},{"comment":"There is a typo: 'Oriami' should be 'Origami'.","section":"Appendix E.1"},{"comment":"In the sentence 'we project the source's LoRA-X onto the corresponding module in the target model using Equation equation 2', the reference should be to Eq. (3), since Eq. (2) only shows the source-side identity property.","section":"§5.2"},{"comment":"The statement that 'the matrix ΔΣ can be any arbitrary square matrix and does not need to be diagonal' conflicts with the earlier definition of ΔΣ as diagonal in Eq. (1); please clarify which form is used in the experiments.","section":"§4.1"},{"comment":"The phrase 'zeroing out the n−r smallest singular values from U and V' is imprecise: zeroing singular values does not change U and V; the intended operation is truncating the columns of U and V to the top r singular vectors.","section":"§4.1"}],"recommendation":"major_revision","confidential_remarks":"The SVD ambiguity in the central projection is the main correctness risk and must be resolved before the paper can be accepted. If the authors clarify the rank-r truncation and adjust Algorithm 1, the core idea may be salvageable, but the experimental validation will also need error bars and a fairer X-Adapter comparison. The manual module selection should be automated or at least given a precise threshold. The paper fits the journal's scope, but the current version is not yet reproducible from the text alone."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing to know: LoRA-X is a genuinely interesting idea with a load-bearing ambiguity in the math and code. The paper claims a training-free, closed-form transfer of LoRA adapters across base models by constraining the adapter to the source's singular subspace and projecting onto the target's. If the projection works as intended, it would remove a real friction point in the PEFT ecosystem. The subspace-constrained parameterization is not new (SVDiff and others fine-tune singular values), but the explicit transfer via subspace projection, plus the subspace-similarity-based module selection, is novel. The related work is handled honestly: they distinguish Trans-LoRA and X-Adapter, both of which need training or data, and they correctly identify that their own metric predicts cross-family transfer failure.\n\nWhat's good: the problem is real, the mechanism is conceptually clean, and the empirical results on same-family diffusion models are plausible. The ablation in Table 5 shows that the projection matters a lot compared to copying singular values, which supports the core mechanism. The subspace similarity metric and the Adapter Transferability Cost are sensible tools. They also include a small LLM transfer experiment, broadening the claim beyond image generation.\n\nThe soft spots: the stress-test note is on point. Equation 3 and Algorithm 1 are ambiguous about whether U_t and V_t are full or truncated to rank r. For the square attention matrices used here (e.g., 4096×4096), torch.linalg.svd(full_matrices=False) returns full orthogonal matrices, making U_t U_t^T and V_t V_t^T identity. In that case, the 'projection' is a no-op: the transferred adapter is just the source adapter applied to the target. That cannot be what produces the Table 5 improvement, so either the code is wrong or the intended truncation is unstated. This is not a minor presentational issue; it is the central mechanism. Without a clear definition of the projection, the results cannot be reproduced. The reader's additional concerns are valid: no error bars (averages over 30 seeds but no variance), manual module selection based on similarity plots, and the transferability cost method itself shows that cross-family transfer fails, which is fine but should be prominent in the abstract.\n\nThe math in Eq. 2 and Eq. 3 is correct if the truncated definitions are used consistently. The paper deserves a serious referee to fix the SVD truncation ambiguity, add error bars, and clarify the module selection. As written, I would not cite it in my own work until the projection is pinned down, and I would not bring it to reading group without a caveat. But the idea is worth engaging with; a revision could make it solid.\n\nRecommendation: send to peer review, but with a strong request to resolve the SVD ambiguity and provide reproducible details.","headline":"Subspace-constrained adapter transfer is a real idea, but the paper's core projection is ill-specified and the implementation appears to reduce to an identity for square matrices.","tokens_in":20338,"tokens_out":4583,"would_cite":false,"duration_ms":39504,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a LoRA adapter constrained to the singular subspace of its source base model can be moved to a related target base model by a closed-form projection, with no gradient updates and no access to original or synthetic…","keywords":["cross-model adapter transfer","training-free transfer","low-rank adaptation","singular value decomposition","subspace similarity","optimal transport","text-to-image generation","parameter-efficient fine-tuning"],"falsifier":"Find a source-target module pair with measured subspace similarity above the paper's working threshold and transfer LoRA-X into it without the manual filter; if the transferred adapter fails to match the trained baseline on image quality and diversity, the subspace-overlap criterion is not sufficient. Conversely, a high-cost cross-family pair that still transfers well would show the criterion is not necessary.","tokens_in":19313,"feed_emoji":"🔄","tokens_out":10377,"duration_ms":95661,"temperature":0.7,"pith_summary":"The paper is trying to establish that a fine-tuned adapter can outlive the base model it was trained on. It introduces LoRA-X, a low-rank adapter that is constrained to live in the singular subspace of its source model's weight matrices; when the base model is superseded, LoRA-X moves to the successor by a closed-form projection onto the successor's singular subspace. The transfer requires no original or synthetic data and no training, only the target model's weights and precomputed singular vectors. In text-to-image style transfer on Stable Diffusion v1.5 and SDXL families, and in a text-generation test, the transferred adapter performs about as well as an adapter trained from scratch on the target. The paper also proposes a subspace-similarity metric and an optimal-transport cost that predict which model and module pairs are cheap to transfer.","feed_headline":"One projection moves fine-tuned adapters to a new base model","feed_subtitle":"LoRA-X pins an adapter to a model's singular subspace; two matrix multiplications move it to a sibling model, no data.","key_machinery":"The load-bearing object is the truncated singular-value decomposition of each attention weight matrix. LoRA-X freezes the top-$r$ left and right singular vectors of the source weight, learns only the diagonal (or general) singular-value change $\\Delta\\Sigma_s$, and transfers by rotating that change into the target's singular frame through $\\Delta\\Sigma_{t\\leftarrow s} = U_t^\\top \\widetilde{U}_s \\Delta\\Sigma_s \\widetilde{V}_s^\\top V_t$. A subspace-similarity score and an optimal-transport cost over module pairs decide which target modules receive the transfer, and mismatched source-target dimensions are handled by a least-squares alignment of the singular frames. The whole operation is closed-form and cheap, so transfer is just two cached SVD-based matrix multiplications per module.","core_discovery":"The central claim is that adapter transferability is a geometric property of model weights rather than a data-recovery problem. If a fine-tuned delta is written as $\\Delta W_s = \\widetilde{U}_s \\Delta\\Sigma_s \\widetilde{V}_s^\\top$ using the source's truncated singular vectors, then placing $\\Delta\\Sigma_{t\\leftarrow s} = U_t^\\top \\widetilde{U}_s \\Delta\\Sigma_s \\widetilde{V}_s^\\top V_t$ into the target's singular frame produces a target delta $U_t \\Delta\\Sigma_{t\\leftarrow s} V_t^\\top$ that carries the same task modification, provided the source and target attention modules share enough subspace overlap. The paper measures that overlap with the subspace-similarity score $\\Phi_l(A,B) = \\|U_A^\\top U_B\\|_F^2 / n$ (and its right-singular analogue), aggregates the pairwise scores into an Adapter Transferability Cost via optimal transport, and applies the adapter only to target modules whose similarity passes a threshold. The experiments show transferred LoRA-X matching or coming close to the trained-from-scratch baseline on image quality and style-alignment metrics, while ablations show that omitting the subspace projection, or transferring ordinary LoRA deltas instead, degrades performance.","pith_inferences":["Editorial inference: if the transferability cost is a faithful predictor of downstream quality, adapter migration for an entire model family could be automated by computing the cost matrix once and transplanting every adapter with the same projection.","Editorial inference: because the projection operates on delta weights rather than on the SVD parameterization itself, it may generalize to any low-rank adapter whose learned delta lies near the target subspace; the paper's DoRA and FouRA results are consistent with this but do not prove it.","Editorial inference: a natural next test is automatic module selection from the transferability cost instead of the manual similarity plots used in the paper, which would scale the method to architectures with hundreds of attention blocks.","Editorial inference: the high cross-family cost suggests that orthogonal projection alone will not bridge unrelated architectures, so a data-free rotation or alignment that maps task directions into another family's subspace is the open next step."],"forward_implications":["A LoRA-X adapter trained once on a source model can be carried to a successor or distilled sibling model without retraining, original data, or synthetic data.","The subspace constraint is what makes transfer work: projecting ordinary LoRA deltas onto the target subspace loses 10 to 28 percent on quality metrics for the tested ranks, while LoRA-X transfers within a few percent of its trained baseline.","The same projection applies to other adapter designs such as DoRA and FouRA, so the mechanism is not limited to the diagonal-singular-value parameterization.","The optimal-transport transferability cost predicts family-level feasibility: same-family pairs have cost below 0.5 and cross-family pairs above 0.5, so the method is claimed to work within closely related model families.","Transfer is size-asymmetric and still works from a smaller source to a larger target, and in the tested text-generation case the transferred adapter improved on the trained baseline on BLEU and ROUGE."],"supporting_citations":[{"why":"Defines the LoRA formulation $\\Delta W = BA$ and the feature-amplification view of low-rank adaptation that LoRA-X re-parameterizes and extends.","marker":"Hu et al., 2022"},{"why":"Supplies the Stable Diffusion v1.5 base model used as source and family target in the main text-to-image experiments.","marker":"Rombach et al., 2022"},{"why":"Supplies the SDXL base model used as source for the SDXL-family transfer experiments to SSD-1B and RealVisXL.","marker":"Podell et al., 2024"},{"why":"Proposes SVDiff, the singular-value fine-tuning parameterization closest to LoRA-X; the paper differentiates by adding truncation, attention-only application, and transferability.","marker":"Han et al., 2023"},{"why":"X-Adapter is the training-based cross-model adapter transfer baseline that LoRA-X compares against; LoRA-X replaces its trained mapper with a closed-form projection.","marker":"Ran et al., 2023"},{"why":"Defines SSD-1B, a distilled SDXL-family target that demonstrates training-free transfer across differently sized same-family models.","marker":"Gupta et al., 2024"},{"why":"Provides the FouRA adapter and the style-transfer dataset setup used for training and evaluation, and the comparison of projection-based transfer across adapter types.","marker":"Borse et al., 2024"},{"why":"Cited as evidence that related base model versions have strongly correlated layers, the premise for expecting overlapping singular subspaces to carry adapters.","marker":"Samragh et al., 2023"}],"fun_headline_variants":["No data? No retrain: LoRA-X transfers adapters via subspace","LoRA-X: Transfer fine-tuned adapters across models, data-free","Move LoRA adapters between base models with just two matrices","Subspace trick lets LoRA jump models without training","LoRA-X: Cross-model adapter transfer without data or training"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the task-relevant direction of a fine-tuned adapter is contained in the directions that similar base models already share; if the target's attention modules do not overlap the source's singular subspace, projecting the adapter destroys the task.","fun_headline_variants_meta":{"raw":{"variants":["No data? No retrain: LoRA-X transfers adapters via subspace","LoRA-X: Transfer fine-tuned adapters across models, data-free","Move LoRA adapters between base models with just two matrices","Subspace trick lets LoRA jump models without training","LoRA-X: Cross-model adapter transfer without data or training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000618,"raw_usage":{"total_tokens":2940,"prompt_tokens":1091,"completion_tokens":1849,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":707,"completion_tokens_details":{"reasoning_tokens":1769}},"tokens_in":707,"tokens_out":1849,"duration_ms":11065,"temperature":1.0,"reasoning_tokens":1769,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T12:18:26.726775+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Find a source-target module pair with measured subspace similarity above the paper's working threshold and transfer LoRA-X into it without the manual filter; if the transferred adapter fails to match the trained baseline on image quality and diversity, the subspace-overlap criterion is not sufficient. Conversely, a high-cost cross-family pair that still transfers well would show the criterion is not necessary.","supporting_citations":[{"cited_title":"Svdiff: Compact parameter space for diffusion fine-tuning","cited_arxiv_id":null,"evidence_quote":"Proposes SVDiff, the singular-value fine-tuning parameterization closest to LoRA-X; the paper differentiates by adding truncation, attention-only application, and transferability."}],"review_version":1}