acceptodds
Under review as a conference paper at ICLR 2027

Geometric Calibration of Supervised Fine-Tuning

Abstract

Recent studies investigate the weight geometry through its spectral structure, which characterizes how the weights transform different input directions. Within this view, supervised fine-tuning (SFT) is shown to induce dense updates that perturb the spectral geometry of pre-SFT weights along their principal directions, potentially disrupting the model’s prior knowledge. These observations motivate using the geometry of pre-SFT weights as a reference for refining learned SFT updates. We study this perspective as Geometric Calibration and develop SPACE (Subspace Projection for Adapted Checkpoint Enhancement), a lightweight posthoc method that selectively reverts SFT update components within the principal subspace of the pre-SFT weights while preserving those in its orthogonal complement. SPACE is a data-free, training-free method that requires only the pre- and post-SFT checkpoints. We evaluate SPACE on Qwen2.5, Qwen3, Qwen3.5, Llama-3.1, and Olmo-3 models ranging from 2B to 14B, across mathematical, coding, and medical SFT, as well as several SFT variants. Overall, SPACE mitigates out-of-domain (OOD) capability regression while further improving indomain (ID) performance over the uncalibrated SFT checkpoints. For math SFT, SPACE yields average ID/OOD gains of 2.85/2.26 points on Qwen2.5-7B, and 2.41/1.52 points on Qwen3-14B-Base, respectively. These results highlight how the geometric relationship between SFT updates and pre-SFT weights shapes the adapted model’s capabilities and, more broadly, motivate geometry-aware parameter updates for continual model improvement.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Supervised Fine-Tuning, Large Language Models, Geometric Calibration, Post-hoc Model Refinement

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_zp6GOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies whether the geometry of pre-SFT weights can guide improvements to a completed SFT checkpoint. It proposes SPACE, which uses the leading right singular vectors of each pre-SFT weight matrix to define a principal input subspace. SPACE removes the SFT update within this subspace and retains the update in its orthogonal complement. The method requires the pre-SFT and post-SFT checkpoints, without training data or additional optimization. The authors evaluate SPACE across five model families, model sizes from 2B to 14B, and mathematical, coding, and medical fine-tuning. They report improvements in average in-domain and out-of-domain performance over the corresponding uncalibrated checkpoints.

Critical strengths

I find this paper valuable, particularly its method. Selectively reverting SFT update components within the principal subspace of the pre-SFT weights while retaining those in the orthogonal complement is a well-motivated approach. The experimental evaluation is also extensive, covering both older and newer backbones and a range of datasets. It is also reasonable that the OOD performance of the method will be greatly improved.

Critical weaknesses

My main concern is the experimental evidence for improvements in in-domain (ID) performance. On Qwen2.5-7B, SPACE increases the MATH500 score from 44.84 after standard SFT to 49.36. I do not understand why this calibration improves mathematical ability on ID benchmarks, and the paper does not explain this clearly. This leaves me with a strong concern that the improvement may reflect an insufficiently trained or tuned SFT baseline.

I therefore checked results from other papers using the same model and dataset. Anchored Supervised Fine-Tuning (ASFT) also uses the Qwen2.5-7B base model and 100K samples from NuminaMath-CoT:

Paper and methodMATH500
SPACE: Base model28.79
SPACE: Standard SFT44.84
SPACE: SFT + SPACE49.36
ASFT: Base model28.79
ASFT: Standard SFT47.30

ASFT’s standard SFT score is 3 points higher than SPACE’s standard SFT score and fairly close to the result after SPACE calibration. Of course, the two papers may use different sampled training subsets, which could explain some of the difference. Nevertheless, this discrepancy warrants an explanation.

Table 10 raises a related concern. With four training epochs, standard SFT reaches 56.04 on MATH500, and SPACE provides no additional improvement. Although gains reappear at eight epochs, the dependence on training duration needs to be explained. My main concern is whether the ID gains persist with a well-tuned SFT baseline. The current results do not establish this clearly.

Questions for authors

Section 5.1 suggests that reverting updates within the principal subspace improves generalization to unseen ID examples. Could you clarify why this would improve ID performance, and what evidence in the paper supports this explanation?

4 comments
AuthorsOpenReview ↗

Response to Reviewer zp6G (Part 1)

Comment W1

My main concern is the experimental evidence for improvements in in-domain (ID) performance. …

Thank you for raising this concern. We address the explanation for ID gains, the SFT score discrepancy, and the dependence on training duration below.

1. Why can SPACE improve ID performance?

Prior work provides a useful perspective. In Weight Ensembling Improves Reasoning in Language Models, Dang et al. (2025) observe that SFT can improve Pass@1 while reducing generation diversity and degrading Pass@k: repeated samples increasingly concentrate on a limited set of solutions, which can be incorrect. Interpolating early and late SFT checkpoints mitigates this diversity collapse and improves both Pass@1 and Pass@k.

Our results show a similar pattern. We compare Pre-SFT, SFT, and SPACE on mathematical benchmarks using Qwen3-14B-Base:

BenchmarkMetricPre-SFTSFTSPACE
AIME25 · (k=16)Pass@1 (Avg@16)5.21%32.71%35.83%
Pass@1630.00%60.00%80.00%
Answer diversity87.92%50.63%56.88%
No-success questions ↓21126
Minerva Math · (k=3)Pass@1 (Avg@3)22.92%45.22%46.08%
Pass@339.71%52.21%56.25%
Answer diversity88.97%59.07%60.66%
No-success questions ↓164130119
AIME24 · (k=16)Pass@1 (Avg@16)6.25%42.08%43.54%
Pass@1636.67%76.67%73.33%
Answer diversity80.42%47.29%50.42%
No-success questions ↓1978

Here, k denotes samples per question. Pass@1 averages correctness over all samples; answer diversity averages the fraction of distinct final answers among each question’s k samples. No-success questions have no correct sample.

SFT reduces answer diversity relative to Pre-SFT, while SPACE partially restores diversity and improves Pass@1. On AIME25 and Minerva Math, SPACE also improves Pass@k, while reducing the number of no-success questions. On AIME24, Pass@16 decreases slightly despite higher diversity and Pass@1, indicating that increased diversity does not uniformly improve coverage within a finite sampling budget. Overall, these observations are consistent with the findings of Dang et al. (2025) and suggest that SPACE can mitigate diversity collapse, thereby helping improve ID performance.

A complementary explanation comes from Understanding Overadaptation in Supervised Fine-Tuning: The Role of Ensemble Methods. Hao et al. (2025) show that merging pretrained and fine-tuned weights can improve target-domain performance. Their analysis in an overparameterized linear setting explains how merging can balance bias from insufficient adaptation against variance from fitting finite fine-tuning data.

Both studies discussed above (Dang et al. (2025); Hao et al. (2025)) analyze weight ensembling/merging. SPACE can likewise be interpreted as a form of model merging: WSPACE=W0P0+WSFT(I−P0)W_{\mathrm{SPACE}}=W_0P_0+W_{\mathrm{SFT}}(I-P_0), where P0P_0 projects onto the principal input subspace of W0W_0. This provides a plausible connection to the ID improvements observed in weight ensembling.

Finally, ID improvements also appear among the model merging baselines in Table 1. For example, WiSE-FT yields an average ID gain of +0.78 on Qwen2.5-7B, and LiNeS yields +1.70 on Qwen3-14B-Base, consistent with prior observations that weight ensembling can improve ID performance. SPACE achieves larger average ID gains. We hypothesize that useful task adaptation and updates detrimental to generalization may lie in different directions. SPACE’s larger gains over global (WiSE-FT) or layer-wise merging (LiNeS) may therefore stem from selectively reverting detrimental updates while preserving more useful adaptation.

We have incorporated this discussion and analysis into Appendix C.7 of the uploaded revised manuscript, with a brief mention in Section 5.1 of the main text. These changes are highlighted in blue.

References:

Dang et al. (2025), Weight Ensembling Improves Reasoning in Language Models.

Hao et al. (2025), Understanding Overadaptation in Supervised Fine-Tuning: The Role of Ensemble Methods.

AuthorsOpenReview ↗

Response to Reviewer zp6G (Part 2)

Comment W1 (Continued)

2. Why does the ASFT paper report a higher standard SFT score?

Thank you for your careful review. We also noticed this discrepancy during our experiments. After a detailed investigation, we identified a formatting bias in the training data used by ASFT.

Formatting bias. Specifically, around 13K of the 100K responses in the training data begin with a line break character \n, while the rest do not. These responses are approximately twice as long as the others:

Training responsesNumber of examplesMean response length
Beginning with \n~13K~766 tokens
Without a leading \n~87K~375 tokens
All responses100K~426 tokens

Also, ASFT’s evaluation template ends with ### Response:\n. This \n can cue the longer response style associated with the newline-prefixed training examples. We hypothesize that these longer reasoning chains contribute to the higher evaluation scores. However, this advantage partly arises from the interaction between the formatting bias in the training data and the evaluation prompt. Therefore, to better isolate the effect of SFT from this bias, in our experiments, we simply remove leading line break characters \n from the training responses.

Controlled experiments support this explanation. To clarify this effect, we conducted controlled experiments using ASFT’s official open-source code and training data, varying only whether to keep the leading line break character (\n) in training responses and whether to include it at the end of the evaluation prompt. We evaluated all 500 MATH500 questions with 16 samples per question:

Leading \n in training responsesEvaluation prompt endingMean output length (tokens)MATH500 (Avg@16)
Kept### Response:335.7544.61
Kept### Response:\n572.2347.08
Removed### Response:336.1144.66
Removed### Response:\n333.2545.00

When training responses retain their leading \n, evaluation prompts ending with \n produce substantially longer answers and higher accuracy. After removing the leading \n from training responses, answer length and accuracy remain largely unchanged regardless of whether evaluation prompts end with \n. These results support formatting bias as an explanation for the score discrepancy. We hope this addresses your concern.

SPACE also remains effective using ASFT’s official code and data. We applied SPACE to a standard SFT checkpoint trained and evaluated using ASFT’s official code, retaining the original training data and evaluation prompt template:

SetupStandard SFTSFT + SPACE
Our original setup44.8449.36
ASFT official code (our reproduction)47.0855.95

Our reproduced standard SFT score closely matches the result reported in the ASFT paper. SPACE substantially improves MATH500 even with the formatting bias present, supporting its robustness across these experimental settings. We hope this addresses your concern.

3. Training duration

Combining the one-epoch result in Table 1 with the longer training runs in Table 10 gives:

Training epochsSFT MATH500SPACE MATH500Average ID gain
144.8449.36+2.85
253.0653.55+0.74
456.0456.04+0.10
855.4958.43+2.08

SPACE’s gains persist after longer SFT training. Standard SFT improves substantially from one to four epochs, then largely plateaus. SPACE still improves the eight-epoch checkpoint, exceeding the best uncalibrated SFT result among the reported training durations. At four epochs, gains are smaller and benchmark-dependent: MATH500 remains unchanged, while improvements are observed on OlympiadBench and AMC23 (ID), and MMLU-Pro (OOD).

Our training settings follow established recipes. Our main experiments adopt training datasets and hyper-parameters from previous work ("On the generalization of sft: A reinforcement learning perspective with reward rectification" and "Rethinking generalization in reasoning SFT: A conditional analysis on optimization, data, and model capability"). These choices were made for comparability with prior work, rather than selected to favor SPACE.

Overall, these results show that SPACE’s benefits extend beyond a single favorable experimental setting and are not limited to a short-training or weaker SFT baseline.

Comment Q1

Section 5.1 suggests that reverting updates within the principal subspace improves generalization to unseen ID examples. Could you clarify why this would improve ID performance, and what evidence in the paper supports this explanation?

Thank you for raising this question. Please see Section 1 ("Why can SPACE improve ID performance?") of our response to Comment W1 for our explanation and supporting evidence.

Thank you for your detailed reply, which has solved my previous concerns.

Thank you for your follow-up. We are glad that our response has addressed your concerns. Please let us know if you have any other questions. We would also appreciate it if you could reconsider your rating.

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

This paper proposes SPACE, a data-free, training-free method for post-hoc refinement of SFT checkpoints. It removes SFT update components within the principal input subspace of the pre-SFT weights while preserving the update on its orthogonal complement. Experiments across multiple model families, domains, and SFT variants report average improvements in both ID and OOD performance..

Critical strengths
  1. SPACE provides a simple and practical way to improve existing SFT checkpoints without additional training or calibration data.
  2. The evaluation covers multiple model families, adaptation domains, and training objectives. The simultaneous ID and OOD improvements suggest useful applications beyond mitigating forgetting.
Critical weaknesses
  1. The incremental contribution needs clarification. The core idea shares substantial similarities with CSP, proposed in Revisiting the Role of Pretrained Weights in Model Merging: On Near-Optimality within the Core Subspace, and PACT: Preserving Anchored Cores in Task-vectors for Model Merging. These methods also filter parameter updates using subspaces derived from pretrained weights to preserve existing capabilities. While SPACE introduces different design choices and an SFT-specific evaluation, the paper should more clearly establish its new contribution.
  2. The geometric explanation needs stronger evidence. Although the subspace ablations are useful, they do not clearly separate the importance of the selected directions from the amount of update removed. A norm-matched scaling or random-subspace control would help establish whether the gains come from preserving specific pretrained knowledge or from a broader regularization effect.
Questions for authors
  1. What is the main new insight beyond existing pretrained-subspace projection methods?
  2. What evidence specifically supports the proposed geometric explanation? Could a representative comparison with norm-matched update scaling or random-subspace removal help distinguish directional effects from update reduction?
3 comments
AuthorsOpenReview ↗

Response to Reviewer vRpD (Part 1)

Comment W1

The incremental contribution needs clarification. The core idea shares substantial similarities with CSP, proposed in Revisiting the Role of Pretrained Weights in Model Merging: On Near-Optimality within the Core Subspace, and PACT: Preserving Anchored Cores in Task-vectors for Model Merging. These methods also filter parameter updates using subspaces derived from pretrained weights to preserve existing capabilities. While SPACE introduces different design choices and an SFT-specific evaluation, the paper should more clearly establish its new contribution.

Thank you for pointing out the connections to CSP and PACT. We acknowledge SPACE shares a similar idea with these methods in preserving pretrained-weight subspaces. However, we believe its contribution should be considered in terms of the problem it addresses, its concrete calibration design, and the empirical evidence supporting that design.

1. Different problem setting and objective. CSP and PACT consolidate multiple fine-tuned experts while mitigating task interference. In contrast, SPACE studies whether a single completed SFT checkpoint can be further improved using its own pre-SFT initialization as a geometric reference. Our findings show that geometric calibration mitigates OOD capability regression while improving ID performance. CSP and PACT’s multi-expert merging results do not directly establish such an outcome, which requires dedicated design analysis and validation in the post-SFT setting.

2. Relationship to CSP. Let Δ=WSFT−W0\Delta=W_{\mathrm{SFT}}-W_0, and let PUP_U and PVP_V project onto the principal output and input subspaces of W0W_0, respectively. SPACE uses

WSPACE=W0+Δ(I−PV).W_{\mathrm{SPACE}}=W_0+\Delta(I-P_V).

Applied to the same update with unit scaling, CSP’s two-sided projection gives

WCSP=W0+(I−PU)Δ(I−PV),W_{\mathrm{CSP}}=W_0+(I-P_U)\Delta(I-P_V),

imposing an additional output-side constraint that removes update components retained by SPACE.

Importantly, our existing L+R-top ablation in Fig. 2(d) applies CSP’s two-sided projection operator to the same SFT update, using the same subspace threshold as SPACE. SPACE achieves average ID/OOD score changes of +2.85/+2.26 relative to SFT, compared with −0.35/−0.02 for the CSP projection. These results favor SPACE for our objective of refining a single post-SFT model, while CSP may be more appropriate for multi-expert merging scenarios.

3. Relationship to PACT. PACT is designed for multi-expert merging scenarios, where each expert’s update is filtered using a protection subspace constructed from the task updates of all other experts. However, in our setting, only the pre- and post-SFT models are available, yielding a single task update. Thus, despite related motivations, SPACE and PACT have distinct designs. SPACE’s contribution lies in its specific formulation, design analysis, and systematic validation for post-SFT calibration.

4. Distinct experimental contribution. We evaluate modern generative LLMs across five model families, spanning 2B to 14B parameters. Our evaluation covers mathematical reasoning, code generation, and medical question answering, which are widely used tasks for assessing LLM capabilities. In contrast, CSP evaluates image classification, BLIP-based captioning/VQA, and Qwen3 embedding/reranking. PACT primarily evaluates multi-task image classification using CLIP-based vision models. These studies place less emphasis on evaluating modern generative LLMs (such as Qwen and Llama) on mainstream tasks assessing advanced LLM capabilities like mathematical reasoning and code generation. We also observe that SPACE simultaneously improves average ID and OOD performance across different SFT variants. Our experiments therefore provide a systematic and practically relevant assessment of post-SFT calibration on contemporary models and mainstream benchmarks, closely aligned with current LLM research and applications.

Revision. In the revised manuscript we have already uploaded, we have clarified the connections to CSP and PACT and SPACE’s distinct contributions in Related Work, and explained the correspondence between CSP’s two-sided projection and our existing L+R-top ablation in Section 5.2. These changes are highlighted in blue.

AuthorsOpenReview ↗

Response to Reviewer vRpD (Part 2)

Comment W2

The geometric explanation needs stronger evidence. Although the subspace ablations are useful, they do not clearly separate the importance of the selected directions from the amount of update removed. A norm-matched scaling or random-subspace control would help establish whether the gains come from preserving specific pretrained knowledge or from a broader regularization effect.

Thank you for this suggestion. We added two control experiments:

  • Norm-matched Scaling: For each weight matrix, let Δ=WSFT−W0\Delta=W_{\mathrm{SFT}}-W_0 and ΔSPACE=WSPACE−W0\Delta_{\mathrm{SPACE}}=W_{\mathrm{SPACE}}-W_0. We set WScaling=W0+βΔW_{\mathrm{Scaling}}=W_0+\beta\Delta, β=∣∣ΔSPACE∣∣F∣∣Δ∣∣F\beta=\frac{||\Delta_{\mathrm{SPACE}}||_F}{||\Delta||_F}. This scales the original update to match SPACE’s remaining update norm.
  • Random-subspace Control: We remove update components in a random input subspace with the same per-matrix dimension as SPACE’s selected subspace, comparing pre-SFT principal directions with random directions.

As shown in the table, SPACE achieves larger average ID and OOD gains than both controls on both models. These results support the value of SPACE’s geometric calibration beyond simple update scaling or random subspace removal.

In the revised manuscript we have already uploaded, we have added the experimental results and analysis to Appendix C.6, with a brief summary in Section 5.2. These additions are highlighted in blue.

ModelQwen2.5-7BQwen3-14B-Base
BenchmarkSFTScalingRandomSPACESFTScalingRandomSPACE
MATH-50044.8445.7946.4149.3688.2788.4089.6090.27
AIME250.831.040.831.4632.7134.7933.5435.83
AIME241.040.631.041.4642.0843.5443.1343.54
OlympiadBench14.0814.8515.8717.7255.1155.0156.7458.37
Minerva Math12.6813.9012.9615.2645.2246.5745.4746.08
AMC2319.5319.8422.1924.8477.0380.7880.9480.78
ID Avg. (∆ vs. SFT)-+0.51+1.05+2.85-+1.45+1.50+2.41
MMLU-Pro38.0539.3440.1342.0365.9067.7066.3069.40
GPQA-D23.3624.8125.7627.2744.1142.7643.7744.78
LCB v226.2227.7928.1128.7044.8144.0345.4043.84
IFEval42.5744.6045.8044.6048.8050.4852.8854.08
AlpacaEval−0.14−0.16−0.19−0.10−0.92−0.79−0.62−0.36
HaluEval46.7547.0746.8047.8575.4775.9375.6375.53
OOD Avg. (∆ vs. SFT)-+1.11+1.60+2.26-+0.32+0.87+1.52

Comment Q1

What is the main new insight beyond existing pretrained-subspace projection methods?

The key insight is that selectively reverting components of a single SFT update, guided by pre-SFT weight geometry, can mitigate OOD capability regression while improving ID performance.

Compared to CSP and PACT, SPACE addresses a different problem setting and objective. Despite conceptual similarities, its calibration design is distinct from the main methods of CSP and PACT. Our experiments offer broader coverage of modern generative LLMs and more comprehensive evidence for post-SFT refinement. Please see our response to Comment W1 for details.

Comment Q2

What evidence specifically supports the proposed geometric explanation? Could a representative comparison with norm-matched update scaling or random-subspace removal help distinguish directional effects from update reduction?

Please see our response to Comment W2.

Dear Reviewer vRpD,

Thank you for your review and feedback. As the discussion period is nearing its end, we would appreciate it if you could review our response and revised manuscript. If they address your concerns, we would be grateful if you could consider updating your assessment.

Thank you again for your time and consideration.

Best regards,

The Authors

Reviewer_9bdiOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents SPACE, a post-hoc model editing method that combines a pre-trained model with its fine-tuned version by projecting the SFT update onto the orthogonal complement of the subspace spanned by the leading right singular vectors of the pre-trained weights. Experiments show that SPACE improves performance over the SFT checkpoint across several benchmarks and model families (2B–14B), outperforming existing post-hoc baselines.

Critical strengths
  1. The method is very simple and explained clearly.

  2. The evaluation is broad, covering multiple benchmarks, domains, and model sizes.

Critical weaknesses
  1. Experimental rigor can be improved, e.g. multiple seeds and error bars, clarifying hyperparameter search space for all baselines

  2. The paper lacks a systematic characterization of when / how much SPACE is expected to help SFT. For example, does the fraction of the SFT update lying in the principal subspace correlate with SPACE's gains?

Questions for authors

Is it reasonable to assume that the gains from SPACE are mostly correlated with more overfitting from SFT? It would be nice to understand the potential of SPACE to recover from an under-tuned SFT vs. outperform a well-tuned SFT.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 9bdi (Part 1)

Comment W1

Experimental rigor can be improved, e.g. multiple seeds and error bars, clarifying hyperparameter search space for all baselines

Thank you for this suggestion. We address the two points below.

1. Multiple seeds and error bars. Due to limited computational resources, our original experiments used a single training seed, while evaluation averaged multiple sampled responses on the benchmarks specified in Appendix B.2. We acknowledge that repeated evaluation sampling does not capture variability across training seeds. To directly assess this, we now report results from five independent SFT runs with different seeds for Qwen2.5-7B and Qwen3-14B-Base, applying both MonoSoup and SPACE to each resulting checkpoint. The table below compares the original results in Table 1 with the mean ± standard deviation across the five runs. The five-seed averages are generally close to the original single-seed results. SPACE also continues to outperform MonoSoup, our strongest overall baseline. These results support the reliability of our conclusions.

ModelQwen2.5-7BQwen3-14B-Base
BenchmarkSFT (single seed)SFT (5 seeds)MonoSoup (single seed)MonoSoup (5 seeds)SPACE (single seed)SPACE (5 seeds)SFT (single seed)SFT (5 seeds)MonoSoup (single seed)MonoSoup (5 seeds)SPACE (single seed)SPACE (5 seeds)
MATH-50044.8444.94 ± 0.5046.4346.54 ± 0.2749.3648.98 ± 0.4188.2788.56 ± 0.5189.2089.12 ± 0.4490.2790.05 ± 0.30
AIME250.830.79 ± 0.371.040.79 ± 0.171.460.92 ± 0.5032.7132.71 ± 0.4936.6734.79 ± 1.4635.8334.50 ± 0.79
AIME241.040.67 ± 0.471.461.38 ± 0.621.461.50 ± 0.2742.0841.96 ± 0.6841.8842.54 ± 1.2643.5443.58 ± 1.51
OlympiadBench14.0813.95 ± 0.2115.0615.15 ± 0.1517.7217.77 ± 0.2155.1155.02 ± 0.5156.2556.86 ± 0.7458.3758.06 ± 0.42
Minerva Math12.6812.61 ± 0.3012.6813.23 ± 0.5115.2615.23 ± 0.4045.2245.49 ± 0.5744.4945.59 ± 0.9946.0845.37 ± 0.79
AMC2319.5319.69 ± 0.4620.0020.50 ± 1.4324.8423.81 ± 0.8077.0377.37 ± 0.6280.1681.63 ± 1.5380.7881.75 ± 0.80
ID Avg. (∆ vs. SFT)——+0.61+0.82 ± 0.35+2.85+2.60 ± 0.23——+1.37+1.57 ± 0.27+2.41+2.03 ± 0.28
MMLU-Pro38.0538.98 ± 1.3139.9940.82 ± 1.2542.0343.25 ± 1.0965.9067.32 ± 0.9967.8067.66 ± 0.8969.4068.54 ± 0.86
GPQA-D23.3624.07 ± 0.9825.4426.19 ± 1.2527.2727.08 ± 0.5444.1143.40 ± 1.3844.9544.11 ± 1.0244.7845.56 ± 1.53
LCB v226.2227.76 ± 1.5028.2528.39 ± 1.1528.7029.38 ± 1.4044.8144.79 ± 0.5644.8145.47 ± 0.5443.8444.97 ± 1.11
IFEval42.5742.23 ± 1.1245.4445.11 ± 0.9644.6045.16 ± 0.8248.8048.08 ± 0.9851.2050.94 ± 0.6654.0852.93 ± 1.70
AlpacaEval (RM)−0.14−0.15 ± 0.02−0.16−0.11 ± 0.06−0.10−0.06 ± 0.03−0.92−0.84 ± 0.37−0.55−0.71 ± 0.20−0.36−0.27 ± 0.42
HaluEval46.7547.48 ± 1.4247.2348.02 ± 1.1347.8548.59 ± 1.1075.4775.39 ± 0.4775.8375.79 ± 0.2175.5375.92 ± 0.50
OOD Avg. (∆ vs. SFT)——+1.56+1.34 ± 0.57+2.26+2.17 ± 0.55——+0.98+0.85 ± 0.29+1.52+1.58 ± 0.38

2. Hyperparameter settings for the baselines. As described in Appendix B.3, “Implementation details”, we follow the default settings in the original papers or released implementations, with two adaptations. For TIES, our setting contains only a single SFT update, so the method reduces to global magnitude trimming. All other hyperparameters follow the defaults in the original paper. For FAPM, the original retention ratio of 0.1 (90% sparsity) performs poorly in our setting, likely due to excessive pruning. We therefore keep its scoring rule unchanged and evaluate retention ratios of {0.1, 0.2, 0.5, 0.8, 0.9}. We select 0.9, retaining 90% of entries per tensor, which provides good overall performance.

We have incorporated these results and clarifications into Section 5.1 and Appendices B.3 and C.8 of the uploaded revised manuscript, with the corresponding revisions marked in blue.

AuthorsOpenReview ↗

Response to Reviewer 9bdi (Part 2)

Comment W2

The paper lacks a systematic characterization of when / how much SPACE is expected to help SFT. For example, does the fraction of the SFT update lying in the principal subspace correlate with SPACE's gains?

Thank you for this suggestion. To address your question, we examine how SPACE’s gains vary across SFT checkpoints trained with different hyperparameters, which yield different weight updates. Specifically, we analyze five final Qwen2.5-7B SFT checkpoints with different learning rates and batch sizes, comparing each SPACE model with its corresponding SFT model.

We measure three norm-based statistics: RR, the ratio of the principal-subspace update norm to the total update norm; and AA and UU, the principal-subspace and total update norms, respectively, normalized by the base-weight norm. All three are expressed as percentages. ID and OOD gains are the mean score improvements over the ID benchmarks and OOD benchmarks, respectively, following the same aggregation used in the paper.

SFT learning rateBatch sizeCheckpoint stepID gain ↑OOD gain ↑RR (%)AA (%)UU (%)
2e-5256390+0.05+1.7347.010.2850.607
5e-5256390+2.85+2.2646.570.8011.719
1e-4256390+4.92+3.0348.132.0914.345
5e-5128781+4.05+2.0146.401.0882.344
5e-5512195+2.19+1.8746.940.5991.275

Across these five checkpoints, RR shows little rank association with ID/OOD gains (Spearman correlations: 0.00/0.10), whereas both AA and UU show strong positive associations (Spearman correlations: 1.00/0.90). Since RR varies only slightly across these checkpoints, AA closely follows UU, and their rankings are identical.

Within these evaluated configurations, our empirical finding is that larger updates are associated with larger SPACE gains, while the principal-subspace fraction alone is less informative.

This correlation does not imply that simply shrinking updates would yield the same gains. SPACE achieves larger average ID and OOD gains than norm-matched scaling (see our response to Reviewer vRpD’s Comment W2), supporting the importance of direction selection.

We have incorporated this analysis into Section 5.6 and Appendix C.9 of the uploaded revised manuscript, with the corresponding revisions marked in blue.

Comment Q1

Is it reasonable to assume that the gains from SPACE are mostly correlated with more overfitting from SFT? It would be nice to understand the potential of SPACE to recover from an under-tuned SFT vs. outperform a well-tuned SFT.

Thank you for raising this question. Mitigating overfitting may contribute to SPACE’s gains. To examine whether its benefits extend beyond a short-training SFT baseline, we compare checkpoints trained for different numbers of epochs.

Combining the one-epoch result in Table 1 with the longer training runs in Table 10 gives the following results on Qwen2.5-7B.

Training epochsSFT MATH500SPACE MATH500Average ID gain
144.8449.36+2.85
253.0653.55+0.74
456.0456.04+0.10
855.4958.43+2.08

SFT performance improves with longer training and then largely plateaus. MATH500 performance improves from one to four epochs, with no further gain at eight epochs.

SPACE’s gains persist beyond this plateau. SPACE provides average ID gains across several training durations. At eight epochs, it surpasses the best uncalibrated SFT result on MATH500 among the evaluated checkpoints.

These results show that SPACE’s benefits are not limited to a short-training or weaker SFT baseline: it can further improve performance even after the gains from longer SFT training have largely plateaued.

Dear Reviewer 9bdi,

Thank you for your review and feedback. As the discussion period is nearing its end, we would appreciate it if you could review our response and revised manuscript. If they address your concerns, we would be grateful if you could consider updating your assessment.

Thank you again for your time and consideration.

Best regards,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal and Discussion

Dear SACs and ACs,

Thank you for handling our submission and for the reviewers’ constructive feedback. We briefly summarize our responses and revisions below.

  • Reviewer 9bdi: Experimental rigor and characterization of gains. We added results across multiple training seeds, clarified baseline settings, and analyzed how update statistics and training configurations relate to SPACE’s gains.
  • Reviewer vRpD: Contribution and geometric evidence. We clarified the distinctions from CSP and PACT and added control experiments supporting the importance of SPACE’s selected calibration directions beyond a general update-reduction effect.
  • Reviewer zp6G: ID improvements and baseline strength. We provided explanations and supporting evidence for ID gains, investigated the ASFT baseline discrepancy, and demonstrated benefits with stronger, longer-trained SFT baselines. The reviewer subsequently confirmed that their previous concerns were resolved.

Key revisions are highlighted in blue in the updated manuscript.

However, we have not yet received further feedback from Reviewers 9bdi and vRpD. We would appreciate it if you could consider our responses and additional evidence in your final assessment.

Finally, we would like to re-emphasize the contributions of this paper:

  • Geometric Calibration. We study the refinement of completed SFT updates using pre-SFT weight geometry as a reference. Building on this idea, we propose SPACE, a lightweight post-hoc method that selectively reverts updates within the principal subspace of the pre-SFT weights.
  • Practical post-hoc refinement. SPACE requires only the pre- and post-SFT checkpoints, without additional data or training.
  • Broad empirical validation. Experiments across model families, sizes, training domains, and SFT variants show improvements in average ID and OOD performance, with ablations and additional controls supporting the calibration design.

Thank you again for your time and consideration.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.