acceptodds
Under review as a conference paper at ICLR 2027

Gradient Skill Decomposition: Interpretable LLM Skill Vectors for Continual Learning

Abstract

Humans describe learning in terms of named and often compositional skills, whereas LLM training records model change through high-dimensional gradient updates. Without a correspondence between these views, a gradient does not directly reveal what an example teaches or which capabilities later training should preserve. We introduce Gradient Skill Decomposition (GSD), a framework for human–machine skill alignment in model gradient update space. GSD extracts compact layer-wise structure from per-example gradients, aligns this structure with multi-label skill annotations, and maps the recovered skill directions back to the parameter coordinates used for fine-tuning. The resulting shared representation supports both skill interpretation and skill-guided intervention. Across SmolLM3-3B, Qwen3.5-2B, and MiniCPM5-1B, GSD predicts human-defined skill sets with Jaccard scores of 64.2%–87.1%. During continual learning, we use the aligned directions as preservation constraints by projecting fine-tuning updates away from the selected skill subspace. Compared with standard LoRA, GSD-guided projection improves protected-task accuracy across 6 independently fine-tuned tasks and 3 models, with broadly comparable new continual-task performance. After 6 sequential tasks on MiniCPM5-1B, the projection improves protected-task accuracy by 12.5–18.0 percentage points over standard LoRA across two task orders. A targeted subset ablation provides initial evidence that selecting particular skill directions can concentrate protection on the corresponding behaviors. Across both independent and sequential continual-learning runs, once the skill basis is constructed, GSD processes the same number of fine-tuning tokens as standard LoRA, with only small increases in adaptation time and peak memory. Together, these results show that GSD connects an interpretable account of what examples teach to actionable control over continual learning, turning human-interpretable skill directions into a practical mechanism for reducing forgetting across successive tasks while maintaining new-task performance.

Primary area:
interpretability and explainable AI
Keywords:
LLM Interpretability

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_UcVKOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Gradient Skill Decomposition (GSD), which connects human-named skills with parameter-update directions in large language models. GSD extracts per-example LoRA-B gradients at a fixed original checkpoint, learns compact layer-wise representations, aligns them with skill annotations through a multilabel predictor, and maps the recovered directions back to parameter space. During subsequent supervised fine-tuning, it removes components of parameter displacement within selected skill subspaces to mitigate forgetting. Experiments include skill prediction on a synthetic algebra dataset, independent fine-tuning of three models on six continuation tasks, and sequential training under two task orders. The reported results show improved protected-task performance with broadly comparable continuation-task metrics, while a restricted skill-subset ablation provides preliminary evidence of selective protection.

Critical strengths

S1: Connecting skill interpretation with training intervention: The paper connects human-interpretable skill descriptions with the model’s parameter-update space and uses this correspondence for practical training interventions. The recovered skill directions are used not only to characterize the skills associated with individual examples, but also to define preservation constraints during subsequent fine-tuning. The method therefore goes beyond skill prediction or representation visualization by making skill interpretations actionable during optimization, providing a concrete and testable framework that connects training interpretability with capability preservation.

S2: Supporting skill-selective preservation and complementing replay: Compared with the replay baseline used in the paper, which preserves prior capabilities by rehearsing old-task examples, GSD provides a parameter-space mechanism for selecting preservation targets by skill name. It thus supports finer-grained capability preservation while mitigating catastrophic forgetting, with the skill-subset ablation providing preliminary empirical evidence of selectivity. Moreover, GSD can be combined with replay, and the combined method improves six-task-average retention over replay alone across all three evaluated models. This is consequential because the method offers more controllable preservation while adding value to existing replay-based training pipelines, rather than requiring replay to be replaced.

Critical weaknesses

Although GSD improves capability retention, the main experiments do not sufficiently isolate the additional contribution of skill-label alignment. Since Mean-Grad and GSD differ in protected subspace dimensionality, it remains unclear whether the gains arise from semantically guided direction selection or from protecting richer gradient structure. Comparisons with matched-dimensional gradient-subspace methods that do not use skill labels would help clarify this distinction.

Questions for authors

Both skill-basis construction and preservation evaluation currently rely on different splits of Math Gen. Do the authors have evidence that skill directions extracted from one dataset can preserve the same skill on an independently sourced dataset?

1 comment
AuthorsOpenReview ↗

Response to weakness and question

Response summary

Thank you for recognizing the connection between skill interpretation and training intervention, the ability to select preservation targets by skill name, and the complementarity with replay. These strengths capture the paper's central contribution: a shared representation that describes the skills associated with examples and turns those descriptions into explicit constraints on later model updates. The original experiments support its practical value through improved retention in all 18 independent settings without replay and higher average retention from adding GSD to replay across all three models. Your questions help distinguish this operational contribution from the additional evidence needed to isolate the effect of skill-label alignment and establish protection across datasets.

  1. Alignment contribution (W1). We add PCA and Random comparisons: GSD exceeds PCA on five of six tasks and Random on all six, with the highest average protected accuracy (Section 5.2, Table 4; Appendix I). Selected-skill studies separately test controllability (Section 5.1, Figure 3; Appendix H.5, Figure 8). Isolating label effects at matched per-layer rank remains open.

  2. Cross-dataset protection (Q1). We add original-MATH evaluation using the unchanged Math Gen-derived basis and continual-task training. GSD improves accuracy over Vanilla by 6.5 and 11.0 percentage points after FinQA and SQuAD, respectively (Appendix K, Table 16), demonstrating protection beyond the generated evaluation corpus.

Critical Weaknesses

W1

Although GSD improves capability retention, the main experiments do not sufficiently isolate the additional contribution of skill-label alignment. Since Mean-Grad and GSD differ in protected subspace dimensionality, it remains unclear whether the gains arise from semantically guided direction selection or from protecting richer gradient structure. Comparisons with matched-dimensional gradient-subspace methods that do not use skill labels would help clarify this distinction.

Response location: Respond here; this weakness is not covered by a separate explicit question.

Response: GSD outperforms PCA on five of six tasks and Random on all six, achieving the highest average protected accuracy under the shared data and no-replay adaptation protocol (Section 5.2, Table 4).

PCA retains ten global components, so this comparison evaluates complete basis constructions rather than label effects at matched per-layer rank; Appendix I provides the basis definitions and continual task results. Alongside stronger retention, GSD offers explicit control over what to protect: users can select and combine targets by skill name through its learned skill-to-direction mapping (Section 5.1, Figure 3).

Paper references: Section 5.2, Table 4; Section 5.1, Figure 3; Appendix I, Table 15.

Questions for Authors

Q1

Both skill-basis construction and preservation evaluation currently rely on different splits of Math Gen. Do the authors have evidence that skill directions extracted from one dataset can preserve the same skill on an independently sourced dataset?

Response: Yes. Appendix K, Table 16 evaluates original-MATH algebra retention on MiniCPM5-1B after FinQA and SQuAD adaptation using the unchanged Math Gen-derived basis and continual-task training. Only the protected evaluation dataset changes; no basis is refitted on original MATH.

GSD improves accuracy over Vanilla by 6.5 percentage points after FinQA and 11.0 after SQuAD. A GSD variant achieves the highest original-MATH accuracy in both settings: Replay+Ours after FinQA and projection alone after SQuAD. The same learned directions therefore preserve algebra capability on independently sourced problems, extending protection beyond the generated Math Gen evaluation corpus.

Paper references: Appendix K, Table 16.

Reviewer_BgHXOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Gradient Skill Decomposition (GSD), which links human-named skills to directions in LoRA-B gradient space and uses them as protected subspaces during continual fine-tuning. A two-stage pipeline (unsupervised layer-wise factor estimation, then a multilabel MLP) recovers skill directions, and updates are projected orthogonally to them. On a synthetic algebra dataset across three models, the authors report Jaccard scores of 64.2–87.1% for skill recovery and improved protected-task accuracy over LoRA in 18 independent settings and in six-task sequences.

Critical strengths
  • The question is well motivated: expressing gradient updates in a vocabulary people can inspect, and using that vocabulary for control, is a worthwhile direction.

  • The method is carefully built. Fixing the LoRA A factor to make per-example signatures comparable is well justified.

Critical weaknesses
  • Narrow data. All skills come from high-school algebra, generated by one teacher model (Qwen3.5-27B) and labeled at construction time. Models are small (1B–3B). Transfer to other domains is untested.
  • Dimensionality not analyzed. Factor rank r, subsample size, and basis rank have no sensitivity analysis. The protection-vs-plasticity trade-off is not shown as a curve.
  • Fixed skill vocabulary. GSD is supervised with a predefined skill set. It cannot produce directions for new skills, and the effect of vocabulary granularity is not studied.
Questions for authors
  • How does new-task performance change as more skills or rank are protected?
  • Does the method work with non-mathematical domains?
6 comments
AuthorsOpenReview ↗

Response summary

Response summary

Thank you for recognizing the value of expressing gradient updates in an inspectable skill vocabulary and using that vocabulary for control. We also appreciate your assessment of the method's construction, particularly the fixed LoRA-A coordinates that keep extracted directions comparable to later updates. This connection is central to GSD: the representation used to interpret an example also specifies which update directions to protect. Your concerns motivate broader domain evidence and a clearer account of the method's dimensions, vocabulary, and protection–plasticity trade-off.

  1. Domain breadth (W1, Q2). We add coding experiments demonstrating skill-recovery Jaccard and the strongest average independent and final sequential coding retention among the evaluated methods (Section 4.4, Table 3; Appendix H). Original-MATH evaluation additionally tests transfer of the unchanged mathematical basis (Appendix K, Table 16).

  2. Dimensionality (W2). We add a controlled factor-dimension sweep showing monotonically improving validation and test skill recovery (Appendix J, Figure 9). This tests representation capacity; subsample-size and end-to-end protection–plasticity sweeps remain open.

  3. Vocabulary (W3). Our W3 response outlines factor reuse or expansion for new skills, motivated by the transfer-learning and subspace-tracking work cited there. This is a proposed extension; the paper evaluates fixed vocabularies.

  4. Plasticity (Q1). We make the no-protection versus all-skill-protection comparison explicit in Section 5.1. Continual-task performance remains broadly comparable alongside stronger retention in mathematics and coding (Appendix F, Table 8; Appendix H.4, Table 14).

AuthorsOpenReview ↗

Response to weakness 1 and 2

Critical Weaknesses

W1

Narrow data. All skills come from high-school algebra, generated by one teacher model (Qwen3.5-27B) and labeled at construction time. Models are small (1B–3B). Transfer to other domains is untested.

Response: We extend the same GSD pipeline to Generated Coding-10Skills on MiniCPM5-1B, testing skill recovery and protection beyond algebra (Section 4.4). The vocabulary is derived from MBPP programs, and generation uses Qwen3.8-27B rather than Qwen3.5-27B, adding a different teacher model. Related problem families are grouped within splits (Appendix H.1). GSD reaches 92.3% non-oracle Jaccard and achieves the highest independent-average coding retention and final retention in both sequential orders (Tables 3 and 11–13), with broadly comparable continual task performance (Table 14). These results extend the method to another domain within the paper's evaluated 1B–3B model scale.

We also test mathematical protection on algebra problems from the original MATH dataset (Appendix K, Table 16). Keeping the learned protection basis and continual-task training unchanged, we replace only the evaluation dataset. GSD improves accuracy over Vanilla by 6.5 and 11.0 percentage points after FinQA and SQuAD, respectively. The learned mathematical directions therefore preserve capability on independently sourced benchmark problems beyond the teacher-generated evaluation corpus.

Paper references: Sections 4.1 and 4.4; Appendix H.1–H.4, Tables 10–14; Appendix K, Table 16.

W2

Dimensionality not analyzed. Factor rank r, subsample size, and basis rank have no sensitivity analysis. The protection-vs-plasticity trade-off is not shown as a curve.

Response location: Q1 addresses protection and plasticity by comparing the no-protection and all-skill-protection endpoints using the existing continual task results. Factor-dimension and subsample-size concerns are addressed below.

Response:

  1. Intermediate subspace size. In GSD, rr controls the representation passed between the two stages. Our controlled sweep uses identical gradient examples and a fixed split; validation and test Jaccard improve monotonically across the tested dimensions (Appendix J, Figure 9). A full end-to-end sweep would additionally require refitting alignment, rebuilding the skill basis, and repeating independent and sequential continual-learning runs for every dimension, at substantial computational cost.

  2. Estimation subsample. The subsample supplies the data used to estimate the Stage-1 gradient subspace; it is an estimation budget, not a separate protection-strength parameter (Appendix B.3). We fix the estimation data across dimension settings to isolate the effect of rr.

  3. Protected-basis rank. The protected basis is specified by the selected skills, with its numerical rank determined by their directions. Arbitrary rank truncation would remove part of the prescribed skill span. We have an evaluation section the choice of skills directly through selected-skill ablations in Math Gen and Generated Coding-10Skills (Section 5.1, Figure 3; Appendix H.5, Figure 8), demonstrating improved performance on the corresponding skills.

Paper references: Appendix B.3; Sections 3.1–3.2; Appendix J, Figure 9; Section 5.1, Figure 3; Appendix H.5, Figure 8.

AuthorsOpenReview ↗

Response to weakness 3

W3

Fixed skill vocabulary. GSD is supervised with a predefined skill set. It cannot produce directions for new skills, and the effect of vocabulary granularity is not studied.

Response location: Respond here; this weakness is not covered by a separate explicit question.

Response: Although GSD is supervised with a predefined skill list, the skill list can be extended when new data with new skills come in. We can first construct a training set with new labeled skills. We do not need to redo the entire supervision procedure, but can instead extend the skill list by adding only the new skills with the new training set. To incorporate new skills, we would first update the underlying factor representation and then extend the skill-label mapping. Following the principle of adaptive factor transfer, we would assess whether incoming data share the existing factor structure before deciding to pool information or estimate additional structure see [1–3] for this idea.

Let Λ^old\widehat{\Lambda}_{\mathrm{old}} and Λ^new\widehat{\Lambda}_{\mathrm{new}} be estimated loading matrices in the same observed feature space, with latent factors normalized to identity covariance. Their common covariance structures can be compared through the rotation-invariant discrepancy

\Delta=\frac{1}{p}\left\\|\widehat{\Lambda}_{\mathrm{new}}\widehat{\Lambda}_{\mathrm{new}}^\top-\widehat{\Lambda}_{\mathrm{old}}\widehat{\Lambda}_{\mathrm{old}}^\top\right\\|_F,

where pp is the feature dimension. If Δ\Delta is below a threshold reflecting estimation uncertainty, we would pool old and new data to refine the shared factor basis and incorporate the new skills through updated alignment. Otherwise, we would estimate the new structure separately and augment the existing factor basis when residual variation outside its span supports additional factors. Thus, compatible skills reuse existing structure, while genuinely new factor components introduce additional dimensions.

Paper references: Section 3; Appendix B.3, factor estimation.

External references:

[1] Fan, Jianqing, Cheng Gao, and Jason M. Klusowski. "Robust transfer learning with unreliable source data." The Annals of Statistics 53, no. 4 (2025): 1728-1752.

[2] Chai, Jinhang, Jianqing Fan, Cheng Gao, and Qishuo Yin. "SMART Fine-tuning Factor Augmented Neural Lasso." arXiv preprint arXiv:2604.12288 (2026).

[3] Narayanamurthy, Praneeth, and Namrata Vaswani. "Nearly optimal robust subspace tracking." In International Conference on Machine Learning, pp. 3701-3709. PMLR, 2018.

Reviewer_TdiHOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a method that extracts skill directions using per-example gradients and human-defined skill labels. Based on these directions, the authors perform skill-orthogonal supervised fine-tuning, which projects parameter updates onto the orthogonal complement of the protected skill subspace. Experimental results show that the knowledge from the Math Gen dataset is protected after training on other datasets.

Critical strengths
  1. The authors try to connect skill directions with gradient-based continual learning. The motivation is intuitive and easy to understand.
  2. Experiments in table 2 show that the knowledge learned from Math Gen is well retained compared to baselines.
  3. Once the skill basis is constructed, the additional computational cost is marginal.
Critical weaknesses
  1. The proposed method lacks general continual learning ability. In continual learning, new knowledge is continuously acquired and should be protected during subsequent learning. However, in this paper, only a fixed, predefined skill set is protected, while newly acquired knowledge from subsequent tasks is not considered for protection. This design differs from general continual learning, and I think it would be better to evaluate the stability and plasticity of all tasks (e.g., FinQA, LIMA, and UltraChat) during sequential fine-tuning, instead of only showing Math Gen retention.
  2. There are many existing orthogonal subspace-based methods in continual learning, like O-LoRA and Inflora. The authors should compare with them experimentally to demonstrate their superiority.
  3. This method depends on predefined, human-annotated skill labels to train the MLP and obtain the skill direction. However, in real-world scenarios, reliable skill annotations can be difficult to obtain, which may limit the practicality of the method.
  4. The experiments are limited to a synthetic math benchmark, where skill labels are relatively easy to obtain. The performance on more general benchmarks remains unclear.
  5. I also have concerns about the method design, where the lora-A is frozen after random initialization. I think this may make the results sensitive to random initialization and also limit the representation capacity of the model.
  6. In Table 5, the reported computational time excludes basis construction. The basis construction time should also be reported to enable a fair cost comparison with other baselines.
  7. The title in the PDF is inconsistent with the title in the openreview portal. Please check it carefully.
  8. Figure 1 is of low quality and difficult to understand.
Questions for authors

See weaknesses, ranked by importance.

2 comments
AuthorsOpenReview ↗

Response summary

Response summary

Thank you for recognizing the value of connecting skill directions with gradient-based continual learning, the stronger retention, and the small recurring computational cost. We summarize our responses in the order of your concerns:

  1. Continual-learning objective (W1). We respectfully disagree the narrow interpretation that every newly learned task must become a protection target. Continual learning can also test whether a model retains a chosen capability as it learns successive tasks [1]. This setting lets us test our interpretation of the skill basis directly without distraction. GSD links human-defined skills to directions in the model’s updates. By choosing those directions once and reusing them across six tasks in two orders, we test whether they still help preserve the corresponding skills as training progresses (Sections 1 and 3, Tables 9 and 13). [details see below]

  2. Skill interpretability and compute (W2). GSD maps named skills to update directions and reuses the selected basis across tasks. O-LoRA and InfLoRA incur substantial recurring costs to protect task subspaces. In our sixth-task example, O-LoRA’s penalty costs 32 times as many FLOPs per step as GSD’s projection. InfLoRA’s original implementation repeats a dense decomposition for each task at a cost cubic in layer width. Neither method lets users select a named skill within a task, so neither directly tests skill-level control (Sections 3, 5.1 and 5.3, Appendix L). [details see below]

  3. Skill annotation and reliability (W3). Our pipeline requires no manual skill annotation and can be applied across domains. For Math Gen and our new coding experiment, we use LLMs to extract skills from sample dataset, and cluster them into a vocabulary for each respective domain. Each synthetic example inherits the skill labels requested during generation. In a blinded audit of 600 Math Gen solutions, a separate model identified all requested skills in 94.17% of solutions and 97.00% of requested skill instances overall (Section 4.1 and Appendices C.1 and M.3), suggesting the validity of our approach.

  4. Evaluation beyond synthetic mathematics (W4). We add Generated Coding-10Skills to test GSD beyond the mathematical dataset to be protected. GSD achieves 92.3% skill-recovery Jaccard and the highest average independent and final sequential coding retention among the five evaluated methods (Section 4.4, Table 3; Appendix H).

  5. Fixed LoRA-A design (W5). Freezing LoRA-A keeps extracted skill directions consistent with later updates while making gradient analysis tractable. All compared methods share this parameterization, with effective new-task learning in mathematics and coding (Section 2; Appendix F, Table 8; Appendix H.4, Table 14).

  6. Basis-construction cost (W6). We add separate timings for factorization, skill-map fitting, and vector mapping: 57–117 seconds across models, incurred once and amortized across tasks (Section 5.3; Appendix L, Tables 17–18).

AuthorsOpenReview ↗

Details of response to weakness 1 and 2

W1 details

The proposed method lacks general continual learning ability. In continual learning, new knowledge is continuously acquired and should be protected during subsequent learning. However, in this paper, only a fixed, predefined skill set is protected, while newly acquired knowledge from subsequent tasks is not considered for protection. This design differs from general continual learning, and I think it would be better to evaluate the stability and plasticity of all tasks (e.g., FinQA, LIMA, and UltraChat) during sequential fine-tuning, instead of only showing Math Gen retention.

Response: We respectfully disagree with the reviewer’s narrow interpretation of “general continual learning.” Preserving a specified set of skills through successive training tasks is also a valid objective, as illustrated by Chen et al.'s [1] evaluation of fact retention through subsequent training.

Our paper studies human–model skill alignment and interpretability by linking human-understandable skill labels to model-update directions. Protecting the corresponding skills as a model learns new tasks is both a concrete application of this alignment and a meaningful continual-learning objective in its own right. Keeping the skills and their learned basis fixed across six sequential tasks in two orders tests whether those directions remain useful for protection after repeated model updates.

GSD improves Math Gen retention over Vanilla in all 18 independent comparisons (Section 4.3, Table 2). After six sequential tasks, final retention gains are 12.5–18.0 percentage points in mathematics and 3.2–13.2 points in coding across the two orders (Appendix G, Table 9; Appendix H.3, Table 13). We also add final scores on all six continual tasks after both coding sequences, including earlier tasks after later training; performance remains broadly comparable to Vanilla (Appendix H.4, Table 14).

Extending protection to newly acquired tasks could use additional skill directions estimated after each task, with the compatibility and effectiveness of the expanded basis requiring separate evaluation.

Paper references: Sections 1 and 3; Section 4.3, Table 2; Section 4.4; Appendix F, Tables 7–8; Appendix G, Figure 7 and Table 9; Appendix H.3–H.4, Tables 13–14.

External reference:

[1] Howard Chen, Jiayi Geng, Adithya Bhaskar, Dan Friedman, and Danqi Chen. "Continual Memorization of Factoids in Language Models." Transactions on Machine Learning Research, 2026.

W2 details

There are many existing orthogonal subspace-based methods in continual learning, like O-LoRA and Inflora. The authors should compare with them experimentally to demonstrate their superiority.

Response: GSD uses labeled reference examples to map named skills to update directions and reuses the selected directions across tasks. O-LoRA and InfLoRA incur substantial recurring computational costs. O-LoRA evaluates a penalty against every earlier adapter at each micro-batch, and InfLoRA’s original implementation repeats a dense subspace decomposition for each task.

The methods also differ in what users can protect. GSD lets users select skills the model already has, without first training a separate adapter on those skills. O-LoRA and InfLoRA derive their protected subspaces from previously learned tasks. Their standard formulations are therefore not directly applicable to our experiment, which varies the selected skills within a task.

Recurring compute.

  • O-LoRA evaluates an orthogonality penalty against each earlier adapter on every micro-batch. For one illustrative 2,048-by-2,048 weight matrix, rank 16, four micro-batches per step, and ten protected GSD directions, the sixth-task penalty costs about 41.94 million FLOPs per optimizer step versus 1.31 million for GSD’s projection. That is 40.63 million extra FLOPs per matrix per step for the constraint alone.
  • InfLoRA collects representations before and after training on each task. In the original implementation, each adapted attention layer of width dd accumulates a d×dd \times d matrix and computes a dense singular value decomposition. The decomposition costs O(d3)O(d^3) per layer per task. Doubling the width from 2,048 to 4,096 therefore increases its leading arithmetic cost eightfold, in addition to the cost of collecting representations. GSD reuses its fitted skill basis across tasks and avoids this repeated decomposition. Section 5.3 and Appendix L report GSD’s measured adaptation costs.
Reviewer_SpZaOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies whether named mathematical skills can be linked to model-update directions and used to protect existing abilities during fine-tuning.

GSD trains a skill classifier on compressed per-example LoRA-B gradients. Derivatives of the classifier define skill directions, which are mapped back to adapter coordinates and used to constrain parameter changes after each optimizer step.

The reported average Math Gen retention improves from 27.6 to 32.1, 46.8 to 49.5, and 46.7 to 52.7 across the three models. The paper also tests two six-task orders on MiniCPM5-1B.

Critical strengths
  1. Letting users choose which named skills to protect is a useful idea. The paper tests this through label prediction, retention, and selected-skill interventions. Keeping LoRA-A fixed also gives the extracted directions a consistent coordinate system for later updates.

  2. Math Gen retention improves over Vanilla in all 18 reported independent settings without replay. The two task orders and the replay combinations broaden the evaluation. I also appreciate the continuation-task results and the storage and adaptation costs in the appendix.

Critical weaknesses
  1. The main comparison does not tell me whether the skill labels matter. Mean-Grad protects one direction per layer, while GSD can protect up to ten. A larger protected subspace could already explain the better retention. I would want to see a comparison with an unlabeled gradient or PCA basis of the same rank, using the same examples.

  2. The skill labels come from the generation requests, filtering relies on the generator's own quality flags, and train and test questions share generation groups. Independent checks of the labels and answers would help establish what the classifier has learned. The targeted-protection study is also narrow: one model and continuation task, with only 17 two-skill questions in Figure 3d. The increase from 35.3% to 41.2% is one net additional correct answer. The authors call this initial evidence, which seems fair.

  3. I find it hard to judge how much new learning is achieved at the reported level of retention. Table 6 has no unadapted-model scores. FinQA F1 is low, and dialogue losses often barely change at the reported precision. Similar final scores are not enough to show that the new tasks were learned well. The sequential tests track Math Gen but do not show what happens to earlier continuation tasks.

Questions for authors
  1. With the protected data and per-layer rank matched, does a PCA basis perform as well as GSD? How much does GSD gain from the skill labels?

  2. Can you independently check some generated answers and skill labels, test combinations left out of training, and add a simple prediction baseline? The selected-skill study would also benefit from larger relevant subsets and another continuation task.

  3. Please add the base-checkpoint scores to Table 6 so we can see how much the new tasks improve. A small behavioral evaluation would make the instruction-tuning results easier to interpret. Please also report performance on earlier continuation tasks after later training.

5 comments
AuthorsOpenReview ↗

Response summary

Response summary

Thank you for recognizing the value of choosing protection targets by skill name, the consistent update coordinates provided by fixed LoRA-A, and the retention improvements across all 18 independent settings without replay. We also appreciate your attention to the two sequential orders, compatibility with replay, and adaptation costs. These strengths support the paper's central contribution: the same learned correspondence between named skills and model updates supports both interpretation and control during training. Your concerns identify where further controls and validation are needed to establish what this correspondence contributes. Our responses distinguish the evidence added in the revision from the questions it leaves open.

  1. Basis comparison (W1, Q1). We add PCA and Random controls, with GSD achieving the highest average protected accuracy (Section 5.2, Table 4; Appendix I). These compare complete basis constructions; isolating label effects at matched per-layer rank remains open.

  2. Validation and selectivity (W2, Q2). We add a blinded cross-model audit supporting all requested skills in 94.2% of 600 Math Gen solutions (Section 4.1; Appendix C.1; protocol in Appendix M.3). Coding-data checks (Appendix H.1) and original-MATH transfer (Appendix K, Table 16) broaden validation. Selected-skill coding protection adds another model and task (Section 5.1; Appendix H.5, Figure 8); joint-skill evidence remains preliminary.

  3. Continued learning (W3, Q3). Our original paper already reports broadly comparable continual-task performance alongside stronger retention (original Table 6, now Appendix F, Table 8; sequential results in Appendix G, Figure 7). We add base references (Appendix F, Table 7) and final all-task scores after both coding sequences (Appendix H.4, Table 14) to clarify learning and earlier-task outcomes.

AuthorsOpenReview ↗

Response to weakness 1 and 2

Critical Weaknesses

W1

The main comparison does not tell me whether the skill labels matter. Mean-Grad protects one direction per layer, while GSD can protect up to ten. A larger protected subspace could already explain the better retention. I would want to see a comparison with an unlabeled gradient or PCA basis of the same rank, using the same examples.

Response: GSD outperforms PCA and Random mostly on the six continual tasks, achieving the highest average protected accuracy (Section 5.2, Table 4). This comparison exactly use the main-experiment GSD basis under the same data and no-replay adaptation/evaluation setup.

PCA uses ten global components, so the comparison evaluates complete basis constructions. GSD additionally provides named skill directions that users can better imterpretation about what does the direction mean (Section 5.1, Figure 3).

Paper references: Section 5.2, Table 4; Appendix I. Q1 gives the numerical comparison and rank details.

W2

The skill labels come from the generation requests, filtering relies on the generator's own quality flags, and train and test questions share generation groups. Independent checks of the labels and answers would help establish what the classifier has learned. The targeted-protection study is also narrow: one model and continuation task, with only 17 two-skill questions in Figure 3d. The increase from 35.3% to 41.2% is one net additional correct answer. The authors call this initial evidence, which seems fair.

Response:

  1. Checking the labels and generated data. We asked a separate model, GPT-6.1 Terra, to identify the skills used in 600 Math Gen test solutions, with 20 questions from each generation group (Appendix C.1). It saw the question, the saved Qwen3.5-27B solution, and the ten skill definitions. It did not see the requested labels, their number, the generation group, or student results. Skills used in the main reasoning and in supporting steps both counted, and each confirmed label required a quoted passage from the solution. The model found evidence for every requested skill in 565 of 600 solutions (94.17%). Across all requested labels, 1,164 of 1,200 were supported (97.00%). These percentages measure support for the requested labels; they do not establish overall annotation accuracy or independently verify the answers. The protocol is in Appendix M.3.

    For the coding dataset, we also execute reference programs on their tests, check for overlap with source problems, remove duplicates and problems expressing the same task, and independently review a sample (Appendix H.1). Related questions and templates are kept in the same split. Skill combinations are still shared across splits, so generalization to combinations absent from training remains untested.

  2. Checking what the protection preserves. We first test whether the learned mathematical basis also protects performance on the original MATH benchmark. The basis and continual-task training stay unchanged; only the evaluation dataset changes. GSD improves accuracy over Vanilla by 6.5 and 11.0 percentage points after FinQA and SQuAD, respectively (Appendix K, Table 16).

    We then extend the experiment that protects selected skills to MiniCPM5-1B during UltraChat fine-tuning, adding another model, task, and domain (Section 5.1; Appendix H.5, Figure 8). Protecting frequency counting or literal string operations improves accuracy on problems involving the chosen skill by 5.0 and 10.0 points, respectively. Accuracy on the other problems in each panel falls by 0.6 points. Because the panels contain many skill combinations, only a small subset involves both chosen skills. That joint result remains preliminary: the gain is one additional correct program out of six, and all six also require regular expressions (C07).

Paper references: Section 4.1; Appendix C.1, skill audit; Appendix M.3, audit protocol; Section 5.1; Appendix H.1 and H.5, Figure 8; Appendix K, Table 16.

AuthorsOpenReview ↗

Response to weakness 3

W3

I find it hard to judge how much new learning is achieved at the reported level of retention. Table 6 has no unadapted-model scores. FinQA F1 is low, and dialogue losses often barely change at the reported precision. Similar final scores are not enough to show that the new tasks were learned well. The sequential tests track Math Gen but do not show what happens to earlier continuation tasks.

Response: The original submission compared continual task performance across all five SFT methods (original Table 6, now Appendix F, Table 8) and during sequential adaptation (Appendix G, Figure 7). GSD retains the adaptation performance of these methods while improving protected-skill retention. Its continual task scores are comparable, with improvements in some settings, including the highest SQuAD F1 on SmolLM3-3B.

Appendix F now places the base references (Table 7) before the post-SFT results (Table 8). The comparison shows substantial reading-comprehension gains and lower dialogue loss across the SFT methods; GSD retains these learning benefits. The base references are contextual because four evaluation panels differ, and FinQA remains difficult.

The coding extension also reports all six continual tasks after the complete sequence in both orders (Appendix H.4, Table 14), including earlier tasks after later training. These final scores remain broadly competitive alongside stronger coding retention (Appendix H.3, Table 13). Thus the protection advantage coexists with effective learning of continual tasks.

Paper references: Section 4.3, Table 2; Appendix F, Tables 7–8; Appendix G, Figure 7; Appendix H.3–H.4, Tables 13–14.

Comments on the paper

1 comment

Dear Reviewers,

As the discussion phase comes to an end, we'd appreciate it if you could take a moment to review the authors' responses. If you have any remaining concerns or questions, please raise them with the authors before the discussion closes.

Best,

Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.