acceptodds
Under review as a conference paper at ICLR 2027

Structured Algebraic Adapters for Fine-Tuning of Large Language Models

Abstract

Parameter-efficient fine-tuning restricts the trainable update space of a pretrained model, commonly through low-rank factorization. We study Structured Algebraic Adapters (SAA), which instead learn coefficients over a fixed family of blockwise signed-permutation operators. This parameterization supports residual-state and weight-space adaptation without imposing an explicit low-rank bottleneck, while keeping the pretrained backbone frozen. We evaluate SAA against LoRA at approximately matched trainable-parameter budgets on mathematical supervised fine-tuning, data-to-text generation, and reinforcement learning with verifiable rewards. On Ministral-3B with canonical-boxed training targets, weight-space SAA reaches 76.6% GSM8K accuracy under a MetaMathQA-aligned evaluation prompt, compared with 74.2% for projection-level LoRA. However, improvements in answer formatting do not imply uniform gains in mathematical correctness, and neither method improves MATH performance over the frozen backbone in this setting. On GPT-2 Medium, selected SAA and LoRA configurations achieve comparable E2E generation scores. In our Qwen3-0.6B RL experiments, SAA obtains higher answer accuracy than LoRA and the frozen backbone on all four evaluations. These single-seed results motivate structured operator parameterizations as an alternative to low-rank adaptation, while revealing task- and placement-dependent trade-offs.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
PEFT, LoRA, Operator Representations, Efficient LLMs

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_e9M2OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

SAA learns coefficients over fixed blockwise signed-permutation operators, supporting residual-state and weight-space adaptation without an explicit low-rank constraint. Comparisons with LoRA cover mathematical SFT, E2E generation, and mathematical RL. Weight-space SAA improves some SFT results, E2E scores are comparable, and residual SAA improves all four reported RL evaluations.

Critical strengths

Table 4 reports consistent gains in answer accuracy over both LoRA and the frozen backbone across four RL evaluations. The paper also distinguishes improvements in output formatting from gains in mathematical correctness and explicitly reports negative transfer under SFT, making its empirical findings easier to interpret.

Critical weaknesses

W1. The method is insufficiently specified to reproduce or assess its central contribution. Section 3 presents a generic linear combination of fixed operators but omits the explicit XOR/Walsh construction, block size, basis-selection rule, coefficient-sharing scheme, and mapping to rectangular weight matrices. The exact RL adapter parameter budget and key SFT training settings are also missing. Section 4.3 refers to an appendix containing the full configuration sweep, but the supplied PDF ends with the references. These omissions prevent verification of the implemented method and the claimed parameter matching. W2. The evaluation does not isolate the contribution of the proposed algebraic structure. All trained baselines are LoRA variants; the experiments include neither FourierFT, which also learns coefficients in a fixed representation, nor a random-basis control under a matched parameter budget. FourierFT is already discussed in the paper, so the missing comparison represents an experimental gap. Moreover, Section 4.3 selects configurations based on the highest observed test BLEU, meaning that the reported comparison of 66.77 versus 66.60 does not provide an unbiased estimate of relative test performance. Each comparison also uses a single training seed (Section 4.1), and potential overlap between the RL training data and evaluation datasets remains unassessed (lines 419–420). Taken together, these limitations leave the reported gains insufficient to establish a reliable advantage of the proposed operator family.

Questions for authors
  1. Can you provide the explicit operator construction, rectangular mapping, complete training configurations, exact parameter counts, and the referenced appendix?
  2. How does SAA compare with FourierFT and a random-basis control under matched budgets and placements? Since the reported experiments use one seed, do the key gains persist across training seeds?
  3. Can you provide E2E results with configurations selected solely on validation data, and document Math10K provenance and overlap with the RL evaluation sets?
1 comment
AuthorsOpenReview ↗

Addressing Methodological and Experimental Concerns Fixes

We thank the reviewer for the detailed and actionable feedback. The revised manuscript directly addresses the central methodological and experimental concerns in W1 and W2.

W1: Method specification and reproducibility.

Section 3 now distinguishes the full-rank elementary signed-permutation directions from their learned linear combinations, specifies square-block updates for rectangular projections, and describes both residual-state and weight-space insertion sites.

Appendix A provides the complete SFT configurations. Appendix C adds the previously missing Qwen RL configurations, including exact parameter counts, initialization, learning rates, and group sizes. For example, residual SAA and residual LoRA each use 1,048,576 trainable parameters, while the matched 131K SAA/FourierFT configurations each use 131,072 parameters ( Table 9). The referenced appendices are now included in the PDF.

The residual mask-bank selection algorithm is not yet fully specified, which we acknowledge as a remaining reproducibility limitation.

W2: Alternative bases and the contribution of algebraic structure.

We added FourierFT and WaveFT baselines at matched placements and budgets, as well as random signed-permutation, random sparse, random dense, and single-entry controls (Appendix E.1).

At 393K on E2E, random signed permutations achieve 66.41 BLEU versus 66.35 for SAA. We therefore do not claim that the XOR/Walsh structure itself produces a demonstrated performance advantage on this benchmark.

E2E selection bias and statistical reliability.

The original exploratory test-BLEU selection has been replaced by validation-loss-based selection. All reported principal SFT and E2E results now use three seeds.

At 48K, SAA obtains 65.52 +/- 0.89 BLEU compared with FourierFT's 64.34 +/- 0.82 under greedy decoding. The paired test-input bootstrap interval for the difference is [0.63, 1.69], although this interval does not measure training-seed variability and the advantage disappears under beam search.

RL evaluation and dataset overlap.

The revised RL experiments now use three seeds, independently tuned learning rates, multiple adaptation placements, and two completion-group sizes where applicable ( Tables 4 and 9-12).

The Math10K preparation removes duplicates and screens for exact matches and word 6-gram near-duplicates against GSM8K, SVAMP, and MATH-500. Of 8,731 numeric records, 6,021 remain; we reserve 500 validation problems and sample 1,200 training examples per seed from the remaining 5,521.

The expanded RL results are more informative than the original single-seed comparison. SAA is competitive, but the strongest baseline depends on the setting, and most test-set differences remain within seed variability.

We have revised our claims to reflect these results rather than asserting a universal or algebra-specific advantage.

Reviewer_1N51OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes structured algebraic adapters to avoid low-rank adapters. For given operators, only the coefficients of the operators are trained. The experiments are performed on weight-side adaptation and residual adaptation.

Critical strengths

A method that only trains coefficients of fixed operators.

Critical weaknesses
  1. The training is done on a fixed set of operators. To my understanding, the operator selection should be the key innovation and contribution of the paper. But the paper only mentions, "The operators are real signed permutations constructed from XOR shifts and Walsh sign patterns". It does not clearly specify which operators are used and the motivation/reason for using them.

  2. The paper does not compare the proposed method with existing SOTA PEFT methods. The experimental results don't demonstrate the effectiveness of the proposed method. In the reported results, the proposed method even often loses to vanilla LoRA.

  3. The discussion and citations of existing related PEFT methods and fine-tuning are not sufficient at all.

  4. Both the motivation and the novelty of the paper are unclear.

Questions for authors

See weakness

1 comment
AuthorsOpenReview ↗

Clarified Novelty, Operator Construction, and Stronger Baselines fixes

We thank the reviewer for highlighting the need for a clearer motivation, stronger related-work positioning, and a direct assessment of the proposed operator family.

Motivation and novelty. We agree that full-rank parameter-efficient adaptation is not a totally new concept. The revised Introduction and Related Work now discuss RandLoRA, MoRA, HiRA, FourierFT, WaveFT, BOFT, QuanTA, and Quantum-PEFT.

The specific contribution of SAA is a different choice of elementary adaptation directions: fixed, structured matrices that are individually full rank on local square blocks, with only their scalar coefficients optimized. Unlike RandLoRA's factorized terms or selected Fourier/wavelet coefficients, the elementary SAA directions themselves are full rank. We do not claim that this property guarantees superior performance.

Operator construction. Section 3 explains the blockwise update, orthogonality of the elementary matrices, treatment of rectangular projections, and the two adaptation placements. Appendix A reports the experimental block sizes, while Appendix F provides an illustrative Walsh-based dictionary. The precise residual mask-bank selection remains a limitation of the current specification.

Evaluation against stronger baselines. We added tuned FourierFT and WaveFT comparisons and random-basis ablations, using three seeds throughout the main experiments.

On GPT-2 Medium E2E at 393K parameters, SAA achieves 66.35 BLEU, compared with 66.46 for LoRA, 66.28 for FourierFT, and 66.51 for WaveFT. At 48K, SAA reaches 65.52 versus 64.34 for FourierFT under greedy decoding. Random sparse and signed-permutation controls are also competitive, showing that the particular XOR/Walsh organization is not uniquely advantageous in this setting.

The expanded Qwen RL evaluation likewise demonstrates competitive behavior across placements and parameter budgets, although most differences remain within training-seed variability.

We have therefore narrowed the central claim. The paper evaluates an alternative fixed-operator PEFT parameterization and characterizes its performance and limitations, rather than claiming general superiority or a novel theory of full-rank adaptation.

We hope these revisions clarify both the intended contribution and the empirical evidence supporting it.

Reviewer_bjMUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SAA. It fixes a set of block-wise signed permutation matrices in advance, learns only their combination coefficients, and uses the combined result to modify the weights or hidden states, with the base model kept frozen. This lets it represent high-rank updates with few parameters and avoids LoRA’s explicit low-rank constraint. Under a similar parameter budget, one set of math SFT experiments raises GSM8K accuracy from 74.2% with LoRA to 76.6%, but MATH performance is still below the model without fine-tuning. On text generation tasks, SAA performs close to LoRA. In the RL experiments, SAA outperforms LoRA on all four math benchmarks, but these are all single-seed results, which are not enough to show that it can reliably replace LoRA.

Critical strengths

(1) A worthwhile alternative to low-rank adaptation. By learning combination coefficients over fixed signed permutation operators, SAA can represent high-rank updates with a compact parameterization, so the parameter budget is no longer tied directly to an explicit rank constraint.

(2) The weight-space and residual-stream versions are tested against matching LoRA baseline methods.

(3) Careful separation of evaluation signals. The paper distinguishes validation loss, answer-format compliance, and the correctness of generated answers. This makes the results more informative, and in particular shows that better supervised fitting or format following does not necessarily lead to higher math accuracy.

(4) Promising RL results. In the reported Qwen3-0.6B experiments, SAA outperforms both LoRA and the frozen base model on all four math benchmarks, which suggests this parameterization is worth further study.

Critical weaknesses

(1) No details on how the operators in the paper are built, which makes the paper hard to reproduce. The authors also need to provide more details about the experimental settings and hyperparameter selections.

(2) The value of the proposed algebraic basis is not verified on its own. The experimental comparison in this paper mainly focuses on LoRA. The authors also need to provide the comparison with other baseline methods, such as FourierFT, MoRA. It's not very clear about the contribution of the proposed signed permutation construction.

(3) All comparisons use a single training seed, with no error bars. Some of the differences are small, so repeated runs are needed to confirm that the gains are consistent.

(4) The E2E experiments pick the configuration with the highest test BLEU from an exploratory search. This introduces test-set selection bias. A better approach would be to select the configuration based on validation performance and then evaluate it separately on the test set.

Questions for authors

N/A

Comments on the paper

1 comment
AuthorsOpenReview ↗

Major Revision: Crucial Updates: Expanded Experiments, Reproducibility, and Clarified Contributions Fixes.

Dear Area Chair,

We thank the reviewers for their careful assessments. We have substantially revised Submission 47016 to address the principal concerns about method specification, baseline coverage, single-seed reliability, test-set selection, and RL dataset overlap.

Method and reproducibility. Section 3 clarifies the use of fixed, locally full-rank signed-permutation operators, blockwise updates to rectangular projections, and residual-state versus weight-space placement. Appendices A and C provide detailed SFT and RL configurations, exact parameter counts, learning rates, and selection protocols. Appendix F gives an illustrative nonredundant Walsh-based operator dictionary. We acknowledge that the residual mask-bank selection procedure is not yet completely specified.

Expanded experimental evidence. All principal experiments now report three training seeds. We added parameter-matched FourierFT and WaveFT baselines, random signed-permutation and other random-basis controls, and validation-based hyperparameter selection. The revised Qwen3-0.6B RL study covers multiple adapter placements, budgets, and completion-group sizes ( Tables 4, 9-10 and 12).

Dataset verification. We reduced 8,731 numeric Math 10K records to 6,021 after deduplication and overlap screening against GSM8K, SVAMP, and MATH-500 using exact matching and word 6-gram similarity. We use a separate 500-example validation set and sample 1,200 training problems per seed.

Revised scientific claims. We no longer claim universal superiority over LoRA or an established advantage of the XOR/Walsh construction. The results instead demonstrate that fixed structured full-rank elementary directions provide a viable, competitive PEFT parameterization. Random signed permutations achieve comparable E2E performance, and most RL differences remain within training-seed variability.

We hope the reviewers will consider the substantially expanded experimental evidence and the clarified contribution when reassessing the submission. Our initial objective was to demonstrate our method in the context of the LoRA. We intended to detail the differences between our approach and various Fourier-based variants in the next part of our research. We hope that our primary thesis now appears more clearly substantiated.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.