{"id":"0bd192ab-ece0-4a6a-ae3f-7b4eb6538b6a","arxiv_id":"2508.05232","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":4.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":2,"one_line_summary":"Cross-LoRA transfers LoRA adapters across heterogeneous LLMs by aligning source and target base-model subspaces with rank-truncated SVD and a Frobenius-optimal linear map, data-free and training-free.","lead":"This paper describes Cross-LoRA, a method for moving a LoRA adapter from one language model to a different one without training data or extra training. If it works, one tuned adapter could be reused across many model families, saving fine-tuning compute.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim is unverifiable from the supplied manuscript: the body is an unrelated math.AP paper, and the abstract alone gives no per-benchmark results, baselines, or implementation details.","rationale":"The reader's verdict is UNVERDICTED at low confidence because the supplied full text does not match the target paper. My stress-test agrees that the central claim cannot be evaluated from the available evidence. The reader identifies the semantic comparability of singular subspaces as the weakest assumption; this is a plausible technical soft spot, but the more immediate issue is that the actual method description and experimental support are absent from the supplied manuscript. Thus I do not raise a new technical objection; I confirm that the evidence is insufficient. If the actual paper does define the method and report the experiments, the verdict could move toward ACCEPT or CONDITIONAL, but on the current record UNVERDICTED is the only defensible outcome. No change to the reader's verdict is needed.","tokens_in":15157,"tokens_out":3831,"duration_ms":42365,"concrete_test":"Retrieve the actual full text of arXiv:2508.05232 and verify that it (a) defines LoRA-Align via rank-truncated SVD and Frobenius-optimal linear transformation, (b) defines LoRA-Shift, and (c) includes per-dataset results with baselines for ARCs, OBOA, and HellaSwag. Then reimplement the method on one transfer pair (e.g., a LoRA trained on HellaSwag using Llama-2-7B transferred to Mistral-7B) and compare against a directly trained target LoRA and a random-projection baseline. If the actual paper is unavailable, the corresponding sections are absent, or the method does not clearly beat the random baseline, the central claim remains unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that Cross-LoRA transfers LoRA adapters across heterogeneous LLMs with gains up to 5.26% and performance comparable to directly trained LoRA—is not backed by the in-scope manuscript. The supplied full text is arXiv:2508.05220v2, a mathematics paper on parabolic abstract evolution equations; it contains no definitions, equations, experiments, or code for LoRA-Align or LoRA-Shift. The abstract is the only available evidence, and it is evidentially thin: 'up to 5.26%' is a selected best-case figure; 'comparable to directly trained LoRA' is stated without per-benchmark numbers, variances, baselines, or model pairs; and no code or artifact is referenced. There is also an unstated technical assumption that the top-r singular subspaces of the source and target base-model weight matrices are semantically comparable, so the Frobenius-optimal linear transformation from rank-truncated SVD preserves task-relevant LoRA directions. Without the actual manuscript, this assumption cannot be examined for correctness, and there is no way to tell whether the reported numbers are reproducible. The most load-bearing concern is therefore evidentiary insufficiency: the paper's argument cannot be checked from the supplied text.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript as submitted presents an abstract for 'Cross-LoRA', a data-free, training-free framework for transferring LoRA adapters between heterogeneous base LLMs. The proposed mechanism has two components: LoRA-Align (rank-truncated SVD of source and target base weights plus a Frobenius-optimal linear map) and LoRA-Shift (projecting source LoRA deltas into the target subspace). The abstract reports relative gains up to 5.26% over base models on ARCs, OBOA, and HellaSwag, and parity with directly trained LoRA adapters on other commonsense reasoning benchmarks. The full text supplied, however, is arXiv:2508.05220v2, a mathematics paper on parabolic abstract evolution equations in uniformly local Sobolev spaces; it contains no mention of LoRA, LLMs, subspace alignment, or any of the experiments. The only in-scope evidence for the claimed contribution is the abstract.","tokens_in":15335,"tokens_out":5754,"duration_ms":62960,"significance":"If the claimed transfer method works, the contribution could be practically useful: the data-free and training-free property is attractive, and the high-level construction (rank-truncated SVD plus optimal linear map) is concrete enough to be falsifiable. The paper should receive credit for proposing a clearly specified, parameterizable mechanism and for not relying on target-task labels. Nevertheless, as submitted, the manuscript cannot be technically evaluated. There are no derivations, algorithm definitions, benchmark tables, baselines, variance estimates, or artifact links. The body is an unrelated mathematics paper, so the central claim is unsupported by any in-scope evidence. This is a submission-integrity problem, not a normal scientific disagreement with the field's consensus.","major_comments":[{"comment":"The body of the submission is arXiv:2508.05220v2, 'Parabolic abstract evolution equations in cylindrical domains and uniformly local Sobolev spaces' by Romain Joly. This text has no overlap with the abstract's topic: there is no LoRA, no SVD-based subspace alignment, no LoRA-Shift projection, and no LLM experiments. The manuscript therefore contains no definitions, equations, pseudo-code, or experimental protocol for the claimed Cross-LoRA framework. The central claim cannot be checked from the submitted material.","section":"Full Text (entire document)"},{"comment":"The headline result, 'relative gains of up to 5.26%', is reported as a selected maximum rather than as a distribution over benchmarks and model pairs. No per-benchmark results, model architectures, LoRA ranks, base-model pairs, or baselines are provided. A single 'up to' figure is not an evaluable empirical claim, especially when the sentence covers three different benchmarks (ARCs, OBOA, HellaSwag) with presumably different gains.","section":"Abstract"},{"comment":"The statement that on 'other commonsense reasoning benchmarks' Cross-LoRA 'maintains performance comparable to that of directly trained LoRA adapters' is made without naming those benchmarks, giving their metrics, or defining 'comparable.' No confidence intervals, standard deviations, or significance tests are reported. Because the method is data-free, there is also no discussion of how a silent transfer failure would be detected or bounded; this is especially important given the unsupervised nature of the projection.","section":"Abstract (parity claim)"},{"comment":"The method rests on the assumption that the top-r singular subspaces of source and target base-model weight matrices are semantically comparable. The abstract provides no evidence for this assumption across heterogeneous architectures, tokenizers, or pretraining runs. The Frobenius-optimal linear map is optimal in a matrix-norm sense, but optimality in that sense does not imply preservation of task-relevant directions. With no ablations over the rank-truncation level r and no layer-wise analysis, this load-bearing premise remains unsupported.","section":"Abstract (LoRA-Align)"}],"minor_comments":[{"comment":"The first page of the supplied full text displays arXiv:2508.05220v2 with the title of the mathematics paper, while the submission is identified as arXiv:2508.05232 and titled 'Cross-LoRA'. This identifier/title mismatch must be resolved before the manuscript can be handled as the intended submission.","section":"Title/identifier"},{"comment":"The benchmark names ARCs and OBOA are not expanded or cited, and the 'other commonsense reasoning benchmarks' are not enumerated. Full benchmark names, licenses, and evaluation settings would be needed even in a short abstract.","section":"Abstract"},{"comment":"The abstract mentions 'lightweight adaptation on a commodity GPU in 20 minutes' but gives no hardware, framework, code, or seed information. A reproducibility statement and an artifact link are needed for the claimed experiments.","section":"Reproducibility"}],"recommendation":"reject","confidential_remarks":"This looks like the wrong PDF was supplied for review. As a reviewer I can only judge the submitted text, and on that text the manuscript cannot be reviewed: the body is an unrelated mathematics paper and the abstract alone is not sufficient evidence for any of the empirical claims. I recommend returning the submission; if the authors intended to submit the actual Cross-LoRA paper, it should be submitted as a new manuscript with full methods and results, including per-benchmark tables, baselines, variance measures, and an artifact link."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: we can't review this paper as submitted. The full text attached to arXiv:2508.05232 is actually arXiv:2508.05220v2, a math.AP paper by Romain Joly on parabolic evolution equations in uniformly local Sobolev spaces. It has nothing to do with LoRA, adapters, or LLMs. So the only in-scope evidence is the abstract.\n\nWhat the abstract describes is coherent and worth taking seriously as an idea. Transferring a LoRA adapter across heterogeneous base models via rank-truncated SVD subspace alignment plus a Frobenius-optimal (Procrustes-style) linear map is a clean combination of standard linear algebra tools. It is genuinely data-free and training-free, which is a meaningful practical target for PEFT and model serving. The construction does not fit to task labels, so circularity is not a concern. The proposal is not a restatement of a named prior result, as far as that can be judged from the abstract.\n\nThe problems start with evidence. The headline number is \"up to 5.26%\" — that is a best case, not a spread. The claim that Cross-LoRA \"maintains performance comparable to directly trained LoRA adapters\" comes without per-benchmark numbers, baselines, standard deviations, or even the model pairs used. Only three benchmarks are named, all English multiple-choice commonsense tasks. No code or artifact is referenced. Most importantly, the load-bearing assumption is unexamined: that the top-r singular subspaces of different base models are semantically comparable, so a Frobenius-optimal map computed from base-model geometry can transfer task-relevant LoRA directions. That assumption can silently fail across different tokenizers, training objectives, and model scales, and because the method is data-free there is no signal to detect the failure.\n\nSo the honest verdict is: a plausible abstract, but nothing that can be checked. The missing body is not a minor referee concern; it is the entire paper. I would not send this artifact to peer review, because there is no manuscript to review. If the authors' actual paper exists, it may well deserve a look — the idea is worth a rigorous test. But from what we were given, I can't recommend engagement, citation, or reading-group time.\n\nMy recommendation: desk reject this artifact, and if the authors resubmit the correct full text, do a proper eval then.","headline":"Can't actually review this paper — the body we got is an unrelated math.AP article, so the only reviewable evidence is an abstract that describes a plausible but unverified LoRA transfer method.","tokens_in":15950,"tokens_out":2467,"would_cite":false,"duration_ms":29990,"reading_group":"no","serious_thinker":"unclear","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Cross-LoRA claims to transfer LoRA adapters between different large language models without any task data, using only the geometry of the base models' weight matrices.","keywords":["LoRA transfer","parameter-efficient fine-tuning","subspace alignment","singular value decomposition","data-free","cross-model adaptation","commonsense reasoning"],"falsifier":"Take a source model with a task-trained LoRA, compute Cross-LoRA's alignment to a target model, and compare the transferred adapter against the same LoRA projected onto random subspaces of the target base model. If the gain over the random projection falls to zero on held-out examples, then the singular-subspace alignment carries no task signal.","tokens_in":14955,"feed_emoji":"🧩","tokens_out":4859,"duration_ms":48730,"temperature":0.7,"pith_summary":"Cross-LoRA aims to make LoRA adapters portable across different large language models—models that differ in size, architecture, or tokenizer—without using any task data. It does this by rewriting a source model's LoRA update in the target model's parameter space, after aligning the two base models' dominant singular subspaces. The paper reports that this transferred adapter improves over the target base model by up to 5.26% on ARCs, OBOA, and HellaSwag, and matches directly trained LoRA adapters on other commonsense reasoning benchmarks. The whole procedure is data-free and training-free, running in about 20 minutes on a commodity GPU.","feed_headline":"Data-free LoRA transfer lifts LLM scores by 5.26%","feed_subtitle":"A subspace-alignment trick ports fine-tuned adapters across model architectures in 20 minutes, with no target data.","key_machinery":"Two components form the method. LoRA-Align performs rank-truncated singular value decomposition on the source and target base-model weight matrices and finds the Frobenius-optimal linear map between their dominant singular subspaces, so the two models' update directions become comparable despite dimension mismatch. LoRA-Shift applies that aligned basis to the source LoRA weight updates, projecting each low-rank factor into the target model's parameter space. Together they convert a source adapter into a target adapter with no training and no data.","core_discovery":"The central claim is that task behavior encoded in a LoRA adapter is not tied to the specific base model it was trained on. Cross-LoRA decomposes the source base model's weight matrix by rank-truncated SVD, computes a Frobenius-optimal linear transformation that maps the source's top singular subspaces onto the target's top singular subspaces, and then uses that map (LoRA-Align) to project the source LoRA update into the target parameter space (LoRA-Shift). The result is an adapter that can be applied directly to the target model. The paper argues this works without any target-task training data, and the experiments on commonsense reasoning benchmarks support the claim that the transferred a","pith_inferences":["The paper reports gains on multiple benchmarks, but a systematic study of failure cases—pairs of models whose singular subspaces are semantically far apart—would map the boundary of the assumption.","If the singular-subspace alignment genuinely captures shared feature structure, the same LoRA-Align map could be reused to transfer several adapters from the same source model, amortizing the one-time alignment cost.","A harder test would transfer to a model with a different tokenizer while controlling for architecture and size, isolating whether vocabulary mismatch or geometric mismatch is the bigger obstacle."],"forward_implications":["Adapters tuned on one model can be dropped onto a different model, so task knowledge survives model upgrades.","Organizations can transfer fine-tuned behavior without sharing or generating data, which matters for privacy and licensing.","The 20-minute, single-GPU budget makes the transfer usable as a routine operation.","If the alignment works across architectures, the same idea could apply to other low-rank adapters or to submodules beyond attention weights."],"supporting_citations":[],"fun_headline_variants":["LoRA adapters jump models with no retraining","Transfer LoRA across LLMs in 20 min, no data","Subspace trick ports LoRA to new LLMs instantly","5.26% boost: LoRA transfers without target data","Cross-model LoRA: no data, no training, gains"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The method assumes that the dominant singular directions of the source and target base models encode the same learned features, so that a linear map computed from base-model geometry can faithfully transplant a task adapter; if the directions do not match, the transfer silently distorts the adapter.","fun_headline_variants_meta":{"raw":{"variants":["LoRA adapters jump models with no retraining","Transfer LoRA across LLMs in 20 min, no data","Subspace trick ports LoRA to new LLMs instantly","5.26% boost: LoRA transfers without target data","Cross-model LoRA: no data, no training, gains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000151,"raw_usage":{"total_tokens":1038,"prompt_tokens":745,"completion_tokens":293,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":489,"completion_tokens_details":{"reasoning_tokens":209}},"tokens_in":489,"tokens_out":293,"duration_ms":3635,"temperature":1.0,"reasoning_tokens":209,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T23:29:02.221783+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a source model with a task-trained LoRA, compute Cross-LoRA's alignment to a target model, and compare the transferred adapter against the same LoRA projected onto random subspaces of the target base model. If the gain over the random projection falls to zero on held-out examples, then the singular-subspace alignment carries no task signal.","supporting_citations":[],"review_version":1}