acceptodds
Under review as a conference paper at ICLR 2027

FamiFT: Familiarity-Adaptive Supervised Fine-Tuning

Abstract

Supervised fine-tuning (SFT) typically applies the same training objective across examples, while recent probability-based methods such as Dynamic Fine-Tuning (DFT) adapt token supervision using the model's online confidence. However, the appropriate strength of such token weighting may itself depend on how well the pretrained model already fits each training example. We first conduct a controlled study and find a consistent familiarity-dependent pattern: relatively familiar examples benefit from sharper probability-based weighting, whereas relatively unfamiliar examples benefit from flatter weighting. Motivated by this observation, we propose Familiarity-Adaptive Fine-Tuning (FamiFT), a simple extension of probability-weighted SFT that assigns each example a sample-specific exponent. FamiFT measures response familiarity once using the frozen initial model, centers it by the dataset median, and uses the resulting score to modulate online token probabilities throughout fine-tuning. Across mathematical reasoning and code generation, FamiFT consistently improves average performance over standard SFT and competitive token-weighting baselines, providing a simple way to incorporate pretrained-model familiarity into supervised fine-tuning.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Supervised Fine-Tuning, LLM

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_gT91OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies whether probability-based token-weighting strength should depend on a pretrained model’s familiarity with each demonstration. It identifies a gap between token-local weighting and sample-level information that can also reshape within-response weights. FamiFT computes each response’s geometric-mean target probability under the frozen initial model and sets a fixed exponent on detached online token probabilities. Experiments fine-tune two math models and one code model on task-specific 10,000-example datasets. The authors report the highest benchmark-averaged accuracy among the compared methods and use fixed, shuffled, reversed, and matched-weight controls to investigate the mechanism.

Critical strengths
  1. Concrete sample-adaptive extension: The technical delta is replacing DFT’s shared probability exponent with an initial-familiarity-dependent exponent. This is a novel combination, not a new problem or theorem: unlike DFT it changes weighting strength per example; unlike TALR it does not share a batch temperature; unlike CADFT it changes token ratios rather than only sample scale; unlike PriFT it aggregates frozen-model probabilities to modulate online weights instead of directly weighting tokens. My judgment is that this incremental delta could meet ICLR’s bar if the proposed insight is reliably established.
  2. Controls test the correspondence, not merely the full system: Shuffling preserves the exponent distribution while breaking example-score matching, and reversal tests assignment direction. Reported averages are 37.20 for FamiFT, 35.88 for shuffling, and 35.64 for reversal. These are more diagnostic than comparing only against SFT, although the protocol concern below remains.
  3. Low-complexity training modification with matched backbones/data: All methods share checkpoints, task data, and training configurations, reducing backbone and extra-data confounds. Scoring is performed once; the training rule uses stored scalars and current probabilities, with no added inference mechanism. Reported gains over the strongest weighting baseline are 0.44 and 0.98 math points and 0.51 code points. Modest gains are not intrinsically a weakness.
Critical weaknesses
  1. Sampling provenance prevents validating controlled comparisons: Table 1 labels results Avg@16, whereas Table 8 labels identical per-benchmark FamiFT/DFT rows Avg@8; Table 9 likewise repeats every FamiFT/DFT Pass@16 entry from Table 5 as Pass@8. Appendix B.2 explicitly distinguishes these budgets. Equality is not proof of an error, but the text does not establish whether controls and references used matched samples. This affects the mechanism evidence, not merely presentation. Provide generation-file provenance and recompute all four matched controls at one budget; if this is only a labeling error, the concern becomes minor.
  2. Reliability and tuning fairness remain unestablished: Training uses seed 1, and uncertainty across training runs is not reported in the main text or appendix. The principal incremental margins are 0.44–0.98 points. Baseline thresholds/coefficients are fixed, while FamiFT receives exponent sweeps; matched tuning budgets and a held-out selection protocol are not reported in the main text or appendix. Shared configurations do not establish equal optimization effort. Report at least three paired training seeds, uncertainty, and equal-budget baseline selection on held-out data.
  3. Attribution supports a joint rule more than an independent shape benefit: Shape-only scores 36.48 and scale-only 35.44, both below DFT’s 36.76; only their combination reaches 37.20. The conditional comparison favors shape, but isolated changes do not improve DFT; the mechanistic interpretation should emphasize interaction rather than an independently beneficial component. The controls also recompute weights from their own trajectories.
  4. Exact reproduction remains underspecified — source code is missing, and data, baselines, and metrics are partially specified.
Questions for authors

(W1) Were all matched controls evaluated with the same eight or sixteen generations? Matched provenance would raise my score; unequal budgets requiring materially changed results would lower it. (W2) Do three paired seeds and equal-budget held-out baseline tuning preserve the incremental gains? Consistent gains would move me toward acceptance; disappearing gains would reinforce rejection. (W3) Does the shape-by-scale interaction remain positive across seeds, and does the familiarity-split effect survive response-weight normalization? (W4) Can you provide exact subset IDs, optimizer configs, trainer revision, and a minimal gradient-reduction example reproducing Appendix A.1?

4 comments

We thank the reviewer for the detailed and constructive feedback. We appreciate the recognition of our method and controlled analyses. We conducted additional experiments to address the concerns about evaluation consistency, robustness, hyperparameter fairness, and the shape–scale interaction. We also added reproduction materials to the supplementary material.

1. Evaluation Consistency (W1)

We apologize for an error in reporting the evaluation budget. The Qwen2.5-Math-7B results in Table 1 were mistakenly reported using 8 samples instead of 16. The corresponding Pass@n results in Table 5 also need correction. Tables 8 and 9, which report the matched-weight ablations, correctly use 8 samples for all methods. We re-evaluated the main Qwen2.5-Math-7B results using 16 samples. The corrected Avg@16 scores are:

MethodMATH500MinervaOlympiadAIME24AMC23Mean
Base41.1013.5017.608.1028.4021.74
SFT48.9015.8015.902.9021.6021.02
DFT69.4029.1033.1011.0040.8036.68
ASFT68.9026.2032.808.8041.4035.62
FamiFT70.7029.2034.207.5044.8037.28
TALR68.7027.4032.205.8042.2035.26
ProFit67.7024.9031.907.5038.8034.16

All matched-weight controls in Tables 8 and 9 were evaluated using the same sampling budget (n=8). Thus, the controlled comparisons themselves used matched evaluation budgets. We have also re-evaluated the main Qwen2.5-Math-7B experiments using n=16. FamiFT still achieves the highest average accuracy (37.28), outperforming DFT (36.68) by 0.60 percentage points. We will correct Tables 1 and 5 and clarify the evaluation budgets.

2. Multi-Seed Robustness (W2)

We ran two additional training seeds for all three original models. The math models were evaluated using Avg@8 and Pass@8, while the code model was evaluated using pass@1. The tables below report the results averaged across three training seeds.

Qwen2.5-Math-7B

MethodMATH500MinervaOlympiadBenchAIME24AMC23Mean
Base41.20 / 84.0013.40 / 41.5017.70 / 48.006.20 / 26.7030.30 / 77.5021.76 / 55.54
SFT49.67 / 81.6716.60 / 43.1316.30 / 42.603.20 / 14.4324.30 / 65.0022.01 / 49.37
DFT69.30 / 85.2028.70 / 44.8732.77 / 53.038.60 / 20.0041.70 / 70.8336.21 / 54.79
ASFT69.37 / 89.6027.40 / 49.4033.03 / 58.638.47 / 20.0043.73 / 80.0036.40 / 59.53
TALR69.27 / 86.9328.63 / 45.3732.53 / 53.536.53 / 17.7742.07 / 69.1735.81 / 54.55
ProFit67.00 / 87.2025.57 / 46.4731.00 / 55.407.20 / 22.2339.30 / 77.5034.01 / 57.76
FamiFT70.20 / 87.4029.80 / 47.1033.30 / 56.507.90 / 17.7744.80 / 77.5037.20 / 57.25

Qwen3-8B-Base

MethodMATH500MinervaOlympiadBenchAIME24AMC23Mean
Base57.60 / 89.4020.80 / 50.7027.30 / 60.106.70 / 23.3035.30 / 70.0029.54 / 58.70
SFT50.07 / 81.2717.17 / 46.3317.17 / 45.032.50 / 13.3324.30 / 58.3322.24 / 48.86
DFT69.33 / 85.2726.83 / 39.5732.33 / 51.577.93 / 16.6738.63 / 68.3335.01 / 52.28
ASFT66.17 / 87.7326.53 / 50.6332.03 / 57.677.53 / 16.6738.87 / 77.5034.23 / 58.04
TALR69.30 / 85.9327.53 / 43.6031.83 / 53.337.37 / 18.9038.27 / 70.8334.86 / 54.52
ProFit67.73 / 87.3327.13 / 47.6731.77 / 55.806.97 / 17.7740.23 / 73.3334.77 / 56.38
FamiFT69.13 / 87.3327.93 / 45.8332.63 / 54.609.33 / 18.9041.43 / 73.3336.09 / 56.00

Qwen2.5-Coder-3B-Instruct

MethodHumanEval+MBPP+LCB v5BigCodeBench HardAvg.
Base76.2265.089.5812.7740.91
SFT78.8666.4010.9814.8942.78
DFT79.2766.0512.1814.6543.04
ASFT78.8666.5813.1711.1142.43
TALR79.4766.5811.9813.9542.99
ProFit79.4765.8713.1713.9543.12
FamiFT79.8867.0213.5715.3643.96

FamiFT retains the highest benchmark-averaged accuracy on all three models. It also achieves the highest mean accuracy on four of five benchmarks for each math model and all four code benchmarks.

3. Hyperparameter fairness (W2).

FamiFT uses the fixed rule Bi=1+gi−MB_i=1+g_i-M in all main experiments. We did not tune this rule for individual benchmarks. The default FamiFT rule was fixed across all models and benchmarks. The exponent sweeps were conducted only to analyze sensitivity and were not used to select the main experimental results.

For the tunable baselines, we followed their recommended settings and further evaluated five candidate values for each method. We used a fixed 500-problem development set from the official 7,500-problem MATH training split. Each configuration was trained with three seeds and evaluated with six samples per question. We report the average across seeds. The recommended configurations achieved the highest Avg@6 among the tested values:

MethodSelected ParameterAvg@6Pass@6
ASFTλ = 0.0168.987.4
ASFTλ = 0.0372.587.8
ASFTλ = 0.05(default)74.789.6
ASFTλ = 0.0772.288.8
ASFTλ = 0.1072.389.0
TALRwmin⁡w_{\min} = 0.00368.079.8
TALRwmin⁡w_{\min} = 0.00771.485.2
TALRwmin⁡w_{\min}= 0.01(default)71.884.6
TALRwmin⁡w_{\min} = 0.0269.880.2
TALRwmin⁡w_{\min} = 0.0369.785.8
ProFitτ\tau = 0.0568.886.6
ProFitτ\tau = 0.10(default)73.887.0
ProFitτ\tau = 0.3072.182.6
ProFitτ\tau = 0.2070.182.0
ProFitτ\tau = 0.5069.382.0

These results support our choice of baseline hyperparameters. We will include the full validation results in the appendix.

4. Shape–Scale Interaction and Familiarity-Split Analysis (W3)

We repeated the matched-weight ablation across three training seeds.

MethodMeanRatiosAvg@8 (Mean ± Std)Pass@8 (Mean ± Std)
DFTDD36.21 ± 0.5054.79 ± 1.02
Scale-onlyFD34.82 ± 0.6953.40 ± 1.61
Shape-onlyDF36.81 ± 0.3155.72 ± 0.33
FamiFTFF37.20 ± 0.0857.25 ± 1.53

Shape-only now slightly outperforms DFT in the three-seed average, while Scale-only performs worse. Combining both effects in FamiFT gives the best performance.

We also computed the shape–scale interaction:

$ I=\mathrm{FamiFT}-\mathrm{Shape}-\mathrm{Scale}+\mathrm{DFT}. $

The interaction is positive for all three seeds, with an average of +1.78 ± 0.24 points in Avg@8 and +2.92 ± 1.89 points in Pass@8.

This supports the reviewer's observation that the two components interact. We will revise the mechanism discussion to emphasize this interaction.

5. Familiarity-Split Experiment with Response-Weight Normalization (W3)

To further investigate whether the familiarity-split effect comes from response-level scaling or relative token weighting, we extended the motivating experiment in Figure 2 with Shape-only and Scale-only controls across three seeds.

Shape-only preserves the exponent-dependent token ratios while matching each response's mean weight to DFT. Scale-only retains the exponent-dependent response-level mean weights while using DFT's token ratios. Each entry reports Avg@8 ± Std / Pass@8 ± Std (%).

SettingOriginalShape-onlyScale-only
DFT33.79±1.70 / 55.81±2.0933.80±2.58 / 54.71±1.9932.11±1.34 / 52.33±0.85
Fixed 0.932.98±0.39 / 55.48±2.6833.15±2.58 / 55.13±2.6035.01±0.44 / 56.64±1.79
Fixed 1.132.75±1.65 / 53.46±3.0532.51±2.25 / 53.31±2.4234.05±2.20 / 56.60±3.31
Mixed35.31±0.62 / 58.15±0.1134.49±0.99 / 57.43±0.1333.45±2.07 / 55.29±2.12
Reversed31.95±0.36 / 52.21±1.5033.33±0.10 / 55.08±0.6132.91±1.87 / 55.76±3.30

Mixed achieves the highest mean Avg@8 and Pass@8 in both the original and Shape-only settings. In contrast, under Scale-only weighting, Fixed 0.9 achieves the highest average accuracy, while Mixed no longer performs best.

These results suggest that response-level scaling alone does not explain the familiarity-split effect. The advantage of Mixed remains when relative token weights are preserved and response-level mean weights are normalized.

Together with the positive shape–scale interaction observed in our multi-seed ablation, these findings support the importance of familiarity-adaptive token ratios and their interaction with response-level scaling.

6. Reproducibility (W4)

We have added the training code, trainer source snapshot, and reproduction scripts for Tables 1 and 2 to the supplementary material.

Both math models use the same 10,000-example NuminaMath-CoT subset. The code model uses a 10,000-example OpenCodeInstruct subset. The supplement provides the exact training files, ordered sample IDs, and trainer checksums.

We use AdamW with β1=0.9\beta_1=0.9, β2=0.95\beta_2=0.95, weight decay 0.01, gradient clipping at 1.0, cosine learning-rate scheduling, and a global batch size of 256. The math models train for one epoch at 5×10−55\times10^{-5} with 10% warmup. The code model trains for three epochs at 5×10−65\times10^{-6} with two warmup steps.

We also added a minimal CPU example for Appendix A.1. It verifies response-token normalization, data-parallel gradient averaging, and microbatch gradient accumulation. The division by the number of microsteps affects only the logged loss, as it occurs after backward().

We thank the reviewer again for the suggestions. These additional experiments and reproduction materials strengthen the evidence for FamiFT and clarify the scope of our conclusions.

Dear Reviewer gT91,

Thank you again for your detailed and constructive feedback. We have now addressed your concerns (W1–W4) with additional experiments, and reproducibility materials.

We would greatly appreciate it if you could take another look at our responses and reconsider your assessment in light of the new evidence. Please let us know if any concerns remain. We would be happy to provide further clarification.

Best regards,
The Authors

Reviewer_D6WFOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2(+1)
Read review
Paper summary

The paper builds on "Dynamic Fine-Tuning", a method that weights cross entropy loss during SFT, based on token probability of the model. The authors make a hypothesis that familiar and unfamiliar samples benefit from different exponents applied to these token probabilities. Their method precomputes familiarity scores for each sample with frozen base model, which determine the sample-specific exponent during training.

Critical strengths

The method is simple.

Critical weaknesses

The core issue is the robustness and generality of the claims. The motivating experiment, suggesting that samples with different familiarity benefit from different exponents, uses a single training seed and a single model (Figure 2).

The main experiments comparing authors' method and baselines include only a single model family (Qwen) and are run on a single seed as well. Combined with modest gains over the baselines (+0.44pp for Qwen2.5-Math-7B, +0.98pp for Qwen3-8B-Base, +0.51pp for Qwen2.5-Coder-3B-Instruct) on avg. accuracy, it makes it unclear whether the improvements are robust, and generalize to other model families.

Moreover, in terms of pass@n metric, the author's method leads to worse performance than baselines -- ASFT outperforms FamiFT on both math models (Tables 5 and 6).

Questions for authors

Could you provide results for Figure 2 and Table 1-2 over multiple training seeds, reporting means and variability?

Could you evaluate an additional model family, preferably a model with open pretraining recipe/data, e.g. newest Olmo base?

3 comments

We thank the reviewer for the constructive feedback. We conducted additional multi-seed experiments for both the motivating analysis and the main evaluations, and extended our experiments to the OLMo model family.

1. Robustness of the Motivating Experiment (Figure 2)

To examine whether the familiarity-dependent pattern in Figure 2 is sensitive to training randomness, we repeated the experiment using three independent training seeds, following the same familiarity split and exponent assignments.

Exponent AssignmentAvg@8 (Mean ± Std)Pass@8 (Mean ± Std)
DFT (B = 1.0)33.79 ± 1.7055.81 ± 2.09
Fixed B = 0.932.98 ± 0.3955.48 ± 2.68
Fixed B = 1.132.75 ± 1.6553.46 ± 3.05
Mixed (Familiar: 1.1 / Unfamiliar: 0.9)35.31 ± 0.6258.15 ± 0.11
Reversed (Familiar: 0.9 / Unfamiliar: 1.1)31.95 ± 0.3652.21 ± 1.50

The mixed assignment retains the highest mean performance on both Avg@8 and Pass@8 across three training seeds. Specifically, it improves over DFT by 1.52 percentage points in Avg@8 and 2.34 points in Pass@8. Moreover, the mixed assignment exhibits relatively low variability, with standard deviations of 0.62 for Avg@8 and 0.11 for Pass@8.

These findings strengthen our original motivating observation: assigning a larger exponent to familiar examples and a smaller exponent to unfamiliar examples is more effective than using a shared exponent, and this advantage persists in the multi-seed averages.

2. Multi-Seed Evaluation on the Original Models (Tables 1–2)

To further investigate the robustness of FamiFT, we repeated the main experiments across 3 training seeds for all three original models, including Qwen2.5-Math-7B, Qwen3-8B-Base, and Qwen2.5-Coder-3B-Instruct with n=8. All methods were evaluated under the same training and evaluation configurations within each setting.

Qwen2.5-Math-7B

Each entry reports Avg/Pass (%), averaged across training seeds.

MethodMATH500MinervaOlympiadBenchAIME24AMC23Mean
Base41.20 / 84.0013.40 / 41.5017.70 / 48.006.20 / 26.7030.30 / 77.5021.76 / 55.54
SFT49.67 / 81.6716.60 / 43.1316.30 / 42.603.20 / 14.4324.30 / 65.0022.01 / 49.37
DFT69.30 / 85.2028.70 / 44.8732.77 / 53.038.60 / 20.0041.70 / 70.8336.21 / 54.79
ASFT69.37 / 89.6027.40 / 49.4033.03 / 58.638.47 / 20.0043.73 / 80.0036.40 / 59.53
TALR69.27 / 86.9328.63 / 45.3732.53 / 53.536.53 / 17.7742.07 / 69.1735.81 / 54.55
ProFit67.00 / 87.2025.57 / 46.4731.00 / 55.407.20 / 22.2339.30 / 77.5034.01 / 57.76
FamiFT70.20 / 87.4029.80 / 47.1033.30 / 56.507.90 / 17.7744.80 / 77.5037.20 / 57.25

Qwen3-8B-Base

Each entry reports Avg@n / Pass@n (%), averaged across training seeds.

MethodMATH500MinervaOlympiadBenchAIME24AMC23Mean
Base57.60 / 89.4020.80 / 50.7027.30 / 60.106.70 / 23.3035.30 / 70.0029.54 / 58.70
SFT50.07 / 81.2717.17 / 46.3317.17 / 45.032.50 / 13.3324.30 / 58.3322.24 / 48.86
DFT69.33 / 85.2726.83 / 39.5732.33 / 51.577.93 / 16.6738.63 / 68.3335.01 / 52.28
ASFT66.17 / 87.7326.53 / 50.6332.03 / 57.677.53 / 16.6738.87 / 77.5034.23 / 58.04
TALR69.30 / 85.9327.53 / 43.6031.83 / 53.337.37 / 18.9038.27 / 70.8334.86 / 54.52
ProFit67.73 / 87.3327.13 / 47.6731.77 / 55.806.97 / 17.7740.23 / 73.3334.77 / 56.38
FamiFT69.13 / 87.3327.93 / 45.8332.63 / 54.609.33 / 18.9041.43 / 73.3336.09 / 56.00

Qwen2.5-Coder-3B-Instruct

All values are pass@1 (%), averaged across training seeds.

MethodHumanEval+MBPP+LCB v5BigCodeBench HardAvg.
Base76.2265.089.5812.7740.91
SFT78.8666.4010.9814.8942.78
DFT79.2766.0512.1814.6543.04
ASFT78.8666.5813.1711.1142.43
TALR79.4766.5811.9813.9542.99
ProFit79.4765.8713.1713.9543.12
FamiFT79.8867.0213.5715.3643.96

FamiFT achieves the highest benchmark-averaged accuracy on all three models in the multi-seed evaluations. Specifically, FamiFT achieves average accuracies of 37.20, 36.09, and 43.96 on Qwen2.5-Math-7B, Qwen3-8B-Base, and Qwen2.5-Coder-3B-Instruct, outperforming their strongest competing baselines by 0.80, 1.08, and 0.84 percentage points, respectively. These multi-seed results strengthen the original single-seed comparisons and provide additional evidence that FamiFT's improvements in average accuracy persist across different training seeds.

3. Generalization Beyond the Qwen Family: OLMo Experiments

Following the reviewer's suggestion, we additionally evaluated FamiFT on the OLMo model family to investigate whether familiarity-adaptive weighting generalizes beyond Qwen. We compared FamiFT with SFT, DFT, ASFT-KL, TALR, and ProFit on the same five mathematical reasoning benchmarks, repeated 3 training seeds with n=8.

Each entry reports Avg / Pass (%).

MethodMATH500MinervaOlympiadBenchAIME24AMC23Mean
SFT53.63 / 80.4718.20 / 39.9320.47 / 46.403.17 / 15.5728.03 / 60.0024.70 / 48.47
DFT54.20 / 73.2021.43 / 35.6722.47 / 42.202.33 / 11.1029.17 / 54.1725.92 / 43.27
ASFT55.60 / 79.0019.00 / 38.4723.43 / 47.602.63 / 10.0326.87 / 59.1725.51 / 46.85
TALR56.27 / 78.0021.33 / 36.5322.63 / 41.032.63 / 8.8728.57 / 51.6726.29 / 43.22
ProFit54.60 / 77.8720.07 / 38.3721.80 / 42.631.77 / 7.8030.00 / 59.1725.65 / 45.17
FamiFT55.40 / 76.4022.23 / 35.3323.63 / 44.232.23 / 12.2032.00 / 59.1727.10 / 45.47

FamiFT achieves the highest average accuracy of 27.10, outperforming the strongest baseline, TALR, by 0.81 percentage points and DFT by 1.18 points.

These results extend our findings to a different model family, providing additional evidence that the benefits of familiarity-adaptive token weighting is robust.

Thank you for the additional experiments. They make the results more robust.

The method itself have mixed results overall -- it often improves average performance over baselines for specific model/benchmark combinations, but underperforms on other (e.g. Olmo/MATH500 0.87pp under TALR baseline). It's also noticable that in most cases the method underperforms baselines on pass@n metric. This pattern could be potentially explained by the method making the model's token distribution more spiky -- similarly to sampling with lower temperature that usually leads to higher avg. performance and lower pass@n.

I will increase my score, but I'm still leaning towards rejection.

Reviewer_222uOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FamiFT, which assesses the familiarity of each training sample by calculating the average token likelihood under a frozen initial model, and assigns a sample-specific exponent to the token probability weights.

Critical strengths
  1. This paper proposes a simple and well-motivated extension. The approach is easy to follow.
  2. The experiments provide a sound demonstration of the proposed mechanism.
Critical weaknesses
  1. Lack of comparison with closely related methods. Although the paper discusses PriFT and CADFT in the Related Work, neither is included in the main experiments. Furthermore, RankTuner is not included as a baseline. Could the authors provide a direct comparison with these baselines, or offer a reasonable explanation for their exclusion?

  2. Insufficient validation of the core motivation. The motivation experiments in Section 3 spporting this motivation primarily report overall performance. However, they do not directly examine the specific impact of the proposed method on sample groups with varying levels of familiarity. A more detailed analysis of performance across familiarity groups would better support the central motivation.

Questions for authors

See Weaknesses.

2 comments

We thank the reviewer for the constructive feedback. We conducted additional experiments to compare FamiFT with closely related methods and further validate our familiarity-dependent motivation.

1. Comparison with Related Methods

Following the reviewer's suggestion, we added CADFT, RankTuner, and PriFT-prob as baselines on Qwen2.5-Math-7B. All results are averaged across three training seeds (mean ± std).

MethodMATH500Minerva MathOlympiadBenchAIME24AMC23Mean
CADFT68.60(0.92) / 85.47(1.67)28.77(0.50) / 43.27(1.52)31.63(1.01) / 52.30(0.89)4.73(1.90) / 18.90(6.97)42.50(1.47) / 71.67(5.20)35.25(0.47) / 54.32(2.74)
RankTuner68.57(0.42) / 87.87(1.47)27.83(1.31) / 47.80(0.96)31.57(0.31) / 55.90(0.78)7.20(1.57) / 18.87(7.68)39.60(2.95) / 73.33(5.20)34.95(0.86) / 56.75(1.64)
PriFT-prob68.73(0.61) / 89.13(1.10)27.50(0.96) / 49.40(1.11)33.07(0.06) / 59.00(0.35)9.87(0.92) / 25.57(1.96)43.97(1.31) / 83.33(5.20)36.63(0.38) / 61.29(0.93)
FamiFT70.20(0.87) / 87.40(1.22)29.80(1.18) / 47.10(2.91)33.30(1.05) / 56.50(2.00)7.90(0.69) / 17.77(5.08)44.80(0.46) / 77.50(2.50)37.20(0.08) / 57.25(1.53)

FamiFT achieves the highest average accuracy, outperforming CADFT, RankTuner, and PriFT-prob by 1.95, 2.25, and 0.57 percentage points, respectively. These results strengthen our comparisons with closely related methods.

2. Further Validation of the Familiarity-Dependent Motivation

We agree that a more detailed analysis of different familiarity groups would help validate our motivation.

Group-specific exponent interventions.

We repeated the motivating experiment in Section 3 across three training seeds, using the 1,000 most familiar and 1,000 least familiar training examples.

The four exponent assignments allow us to examine the impact of changing each training group's exponent while keeping the other group's exponent fixed.

For the familarity group, keeping the unfamiliar exponent at 0.9, increasing the familiar exponent from 0.9 to 1.1 improves Avg@8 from 32.98 to 35.31 (+2.33 points). For unfamiliar group, keeping the familiar exponent at 1.1, decreasing the unfamiliar exponent from 1.1 to 0.9 improves Avg@8 from 32.75 to 35.31 (+2.56 points). These comparisons provide more targeted evidence for assigning larger exponents to familiar training examples and smaller exponents to unfamiliar ones.

Shape–Scale decomposition.

We further examined whether this pattern comes from response-level scaling or relative token weighting. We added Shape-only and Scale-only controls to the same familiarity-split experiment across three seeds.

Shape-only preserves exponent-dependent token ratios while matching the response-level mean weights to DFT. Scale-only preserves the response-level mean weights while retaining DFT's token ratios.

Each entry reports Avg@8 (mean ± std) / Pass@8 (mean ± std).

SettingOriginalShape-onlyScale-only
DFT33.79±1.70 / 55.81±2.0933.80±2.58 / 54.71±1.9932.11±1.34 / 52.33±0.85
Fixed 0.932.98±0.39 / 55.48±2.6833.15±2.58 / 55.13±2.6035.01±0.44 / 56.64±1.79
Fixed 1.132.75±1.65 / 53.46±3.0532.51±2.25 / 53.31±2.4234.05±2.20 / 56.60±3.31
Mixed35.31±0.62 / 58.15±0.1134.49±0.99 / 57.43±0.1333.45±2.07 / 55.29±2.12
Reversed31.95±0.36 / 52.21±1.5033.33±0.10 / 55.08±0.6132.91±1.87 / 55.76±3.30

Mixed achieves the highest mean Avg@8 and Pass@8 in both Original and Shape-only. Even after response-weight normalization, Mixed outperforms Reversed by 1.16 points in Avg@8 and 2.35 points in Pass@8. In contrast, under Scale-only weighting, Fixed 0.9 performs best, while Mixed no longer achieves the highest accuracy. The advantage of Mixed remains when relative token weights are preserved and response-level mean weights are normalized.

Together, the group-specific interventions and normalization controls provide further evidence for our central motivation: the preferred token-weighting strength depends on training-sample familiarity, with relative token weighting playing an important role.

Thank you for the new experiments, they address my concerns to a reasonable extent. The new results do, however, raise one point I would like to flag. In the shape/scale table, the three-seed standard deviations are considerably larger than most of the differences being compared. I recognise that the setting and metrics differ from the main experiments, but I would encourage the authors to report results under the same three-seed protocol, since that is the comparison the main claim rests on. I will maintain my positive score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.