acceptodds
Under review as a conference paper at ICLR 2027

Toward AI-Friendly Data Quality Metrics: A Feature Density Perspective

Abstract

In Large Language Models (LLMs) fine-tuning stage, input sequences differ in the proportion of task relevant tokens that help the model predict target sequences, and quantifying this variation can help clarify the relationship between data and fine-tuning performance. Therefore, we introduce Feature Density: the fraction of input tokens whose deletion increases a pretrained model’s loss on the target sequence. We validate the tokens it identifies by comparing them to known feature tokens in synthetic data and human-annotated feature tokens in real world data. In synthetic data, Feature Density increases with the proportions of task relevant tokens. Across twelve real-world datasets using Qwen LLMs, Feature Density varies by task, with higher densities in translation and math reasoning. In both settings, Adam, AdamW, Muon, and SGD achieve lower test loss when fine tune on higher density data within the same tasks. These results support Feature Density as a metric of task relevant token proportions and its relation with model fine-tuning performance across different optimizers.

Primary area:
datasets and benchmarks
Keywords:
Supervised Fine-Tuning, Large Language Models, Data Metric

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 3 reviews, above 27% of the others.
Reviewer_9sLCOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces Feature Density (FD), defined as the fraction of input units whose individual deletion increases a fixed pretrained model’s loss on the target sequence. The authors evaluate the metric using synthetic data with known feature locations and twelve real-world datasets, primarily with Qwen3-0.6B, supplemented by Qwen3.5-0.8B. They also investigate its relationship with fine-tuning losses under Adam, AdamW, SGD, and Muon. The main claim is that higher Feature Density reflects a greater proportion of task-relevant content and is associated with better fine-tuning performance.

Critical strengths

The proposed metric could offer a useful way to characterize how a model uses training inputs beyond dataset size and sequence length.

Critical weaknesses

My two critical concerns are as follows.

  • The central claim about fine-tuning performance is not supported by the real-world evaluation. Although the abstract claims lower test loss in both synthetic and real-world settings, the real-world experiments report only training losses. Furthermore, comparisons between different datasets within broad task categories do not control for difficulty, target format, sequence length, or the pretrained model’s initial performance. Several loss differences are already present at the start of training, and the association is not consistent: Figure 8 shows higher FD but higher training loss for SWE-Bench than APPS. Consequently, the experiments do not establish that FD predicts generalization or gains from fine-tuning. This directly undermines the paper’s main empirical contribution.

  • Feature Density is insufficiently validated as a measure of task-relevant content. In Figure 3, FD is approximately 0.544 when the true feature ratio is 0.1. Even assuming perfect recovery of genuine features, this implies that approximately 49% of noise tokens are classified as features. The real-world validation consists of three illustrative examples without aggregate annotation-agreement statistics. Moreover, any positive deletion effect is treated as relevance, although it may reflect model-specific sensitivity to perturbations. These gaps are consequential because the interpretation of FD as task-relevant density underpins the paper’s subsequent analyses; a correlation with loss cannot, by itself, validate that interpretation.

Questions for authors
  • Can the authors provide real-world held-out losses and task-specific performance metrics, including improvements over the pretrained checkpoint?
  • Does FD predict held-out fine-tuning gains after controlling for initial loss, input and target lengths, and dataset difficulty? Can this be tested using matched subsets of the same dataset?
  • What are the feature-identification precision, recall, and false-positive rates across synthetic feature ratios? Why is FD approximately 0.544 when only 10% of tokens are genuine features?
  • How sensitive are FD values and dataset rankings to the deletion threshold, unit granularity, scoring model, and alternative perturbation methods?
  • How is Assumption 2 satisfied by the fixed-position synthetic construction, and what does Proposition 1 establish beyond concentration of dataset-average differences?
5 comments

Thank you for your valuable comments that help us improve manuscript.

Response to Weakness 1: Consequently, the experiments do not establish that FD predicts generalization or gains from fine-tuning. This directly undermines the paper’s main empirical contribution..., and Q1: can the authors provide real-world held-out losses and....

We understand that you would like direct test-set evidence showing fine-tuning on higher-FD data improves performance on unseen examples. In the revised manuscript, we added evaluations of test loss and target-token accuracy on 12 real-world datasets. The results are shown in Figure 8 (Section 4.5) and Figures 12, 13, and 14 (Appendix C.2).

The test loss panels in Figure 8 and Figure 12 show an overall advantage for higher-FD datasets within the same task category. Comparing the last recorded point on each curve gives 24 comparisons between higher- and lower-FD datasets across six task categories and four optimizers: the higher-FD dataset has lower test loss in comparisons. In mathematical reasoning, sentiment classification, translation, long-text summarization, and coding, the higher-FD dataset has lower test loss at the curve endpoint under all four optimizers. For long-text localization, the differences are smaller and vary across optimizers. A similar association between higher FD and better fine-tuning performance is shown by the target-token accuracy results in Figures 13 and 14.

Response to Weakness 2: Feature Density is insufficiently validated as a measure of task-relevant content. In Figure 3, FD is approximately 0.544 when the true feature ratio is 0.1..., and Q3: why is FD approximately 0.544 when only 10% of tokens...

Thank you for pointing this out. A measured FD of approximately 0.54 at a Feature Ratio of 0.1 does require a clear explanation, particularly because it affects how we interpret the task-relevant content identified by FD. Feature Ratio is a hyperparameter specified when generating the synthetic data, whereas FD is computed from changes in a given model’s loss on the same target after deletion. We believe this numerical difference can be understood through the data-generation process and the properties of deletion scores.

The generation process is specified in Eq. (4) of Section 3.1: feature tokens are constructed from normalized Gaussian random vectors and carry task signals through the labels, while noise tokens follow a centered Gaussian distribution and are independent of the labels. The model must identify label-related contributions among these random vectors, while noise may also happen to favor the prediction for a particular example when the distributions of feature and noise responses overlap. This is consistent with the relationship between distributional separability and the minimum achievable classification error [1, 2]. A simple example explains why the fraction of noise tokens counted toward FD can be close to one half. Suppose the model’s response to noise is additive:

\xi_j\sim\mathcal{N}\left(0,\frac{\sigma^2}{D}I_D\right).$$ Here, the noise is independent of the label $y\in\{-1,+1\}$, $w\neq0$ is a fixed vector, and $\sigma>0$. In this example, we use $\mathcal{L}(y,f(X))=-yf(X)$ as a loss proxy. Deleting noise token $j$ leaves the signal component and the other noise terms unchanged, so $$f(X)-f(X_{\setminus j})=w^\top\xi_j.$$ The corresponding deletion score is $$\begin{aligned} R_j &=\mathcal{L}(y,f(X_{\setminus j}))-\mathcal{L}(y,f(X))\\ &=y\bigl[f(X)-f(X_{\setminus j})\bigr]\\ &=yw^\top\xi_j. \end{aligned}$$ By the symmetry of the centered Gaussian distribution, this score is equally likely to be positive or negative: $$\Pr(R_j>0)=\Pr(yw^\top\xi_j>0)=\frac12.$$ In other words, although noise is generated independently of the label, its random contribution on a particular example has approximately a one-half probability of supporting the target prediction and therefore receiving a positive deletion score. Here, $1/2$ is the false-positive probability for a noise token in this example. The precision and recall requested in Q3 need to be computed separately at each Feature Ratio using the known true feature labels. Thus, for a given Feature Ratio $\rho$, if most true features are identified, their contribution to FD is approximately $\rho$, while the expected contribution from noise is $(1-\rho)/2$, giving $$\mathbb{E}\!\left[\mathrm{FD}\right] \approx \rho+\frac{1-\rho}{2}.$$ Substituting $\rho=0.1$ gives $0.55$, close to the measured value of approximately $0.54$ in Figure 3. This example shows that the value of $0.54$ implies most feature tokens have been identified by FD, such that its value concentrates around $\rho+\frac{1-\rho}{2}$. Even when most features are identified, centered Gaussian noise can still bring FD close to this level through the mechanism described above.
AuthorsOpenReview ↗

Response (Part 2)

Response to Q2: Does FD predict held-out fine-tuning gains after controlling for initial loss, input and target lengths, and dataset difficulty? Can this be tested using matched subsets of the same dataset?

Thank you for the suggestion. We understand that you would like to establish whether selecting training data by FD still leads to greater improvements in test performance when the model’s starting point and evaluation conditions are the same. To address this, we added data-selection experiments within individual datasets to compare the gains from fine-tuning on subsets with different FD values.

Specifically, for each of six datasets, GSM8K, SST-2, OPUS Books, LongBench, SWE-Bench, and NIAH, we sampled equal numbers of examples from the same training pool. High-FD selects examples with FD in [0.6, 0.9], Low-FD selects examples with FD in [0.1, 0.4], and Random samples from the entire training pool. All three groups start from the same Qwen3-0.6B checkpoint, use Adam for three training epochs, and are evaluated on the same held-out set. The experimental settings are described under “Fine-Tuning Results with FD-Based Data Selection” in Appendix C.2, and the training loss, test loss, and token accuracy results are shown in Figure 15.

This design directly addresses your concern about initial loss. Before fine-tuning, all three groups evaluate the same pretrained model on the same test set, so they share the same initial test loss and accuracy. Epoch 0 in Figure 15 marks this common starting point. Comparing test performance in subsequent epochs directly shows how much each data-selection strategy improves over the pretrained model. Comparing subsets within the same dataset also fixes the data source and test distribution, reducing the influence of differences in task difficulty, target format, and other factors that arise when comparing different datasets.

For example, in the OPUS Books test loss panel in Figure 15 (second row, third column), all three groups start from a test loss of 2.5142. After three training epochs, the test losses for High-FD, Random, and Low-FD are 1.7821, 1.7926, and 1.9001, respectively. Because the starting point is the same, the lower final test loss for High-FD also corresponds to a larger fine-tuning gain. The LongBench results provide further support: over the three fine-tuning epochs, High-FD consistently achieves lower test loss and higher token accuracy than Random.

AuthorsOpenReview ↗

Response (Part 3)

Response to Q4: How sensitive are FD values and dataset rankings to the deletion threshold, unit granularity, scoring model, and alternative perturbation methods?

For the threshold, $R_i(u) =\mathcal{L}\left(Y_i,f(W,X_i\setminus u)\right) -\mathcal{L}\left(Y_i,f(W,X_i)\right)>0$ means that deletion increases loss for the same target, matching our definition of content that contributes positively to target prediction. A threshold of 00 gives a meaningful definition for every possible deletion score and a consistent meaning across datasets.

To test sensitivity to near-zero positive scores, we fixed inputs, targets, and the scoring model for the same batch of 10 NIAH examples and deleted consecutive spans of 1, 2, 4, or 8 sentences. Let Su=Ri(u)/TS_u=R_i(u)/T, where TT is the number of target tokens. We excluded units with 0<Su≤0.010<S_u\leq0.01 from the numerator, keeping the denominator unchanged.

Results for the same NIAH examples at four deletion granularities

Sentences per deletion unitMean FDMean FD after excluding near-zero positive scoresDecrease (percentage points)
10.49200.45643.56
20.45370.42382.99
40.48880.46782.10
80.49170.49170.00

In this batch, mean FD ranges from 0.4537 to 0.4920 across four deletion granularities, a maximum difference of 3.83 percentage points. Excluding near-zero positive scores reduces mean FD by 0.00–3.56 percentage points. Within the tested range, changing deletion granularity and excluding small positive scores have relatively limited effects on mean FD, supporting our measurement rule.

For scoring models and dataset rankings, Figure 5 (Section 4.2) and Figure 11 (Appendix C.2) report mean dataset FD for Qwen3-0.6B and Qwen3.5-0.8B, respectively. Five of six within-task pairs retain their ordering: mathematical reasoning, sentiment classification, translation, coding, and long-text localization. For finetranslations and OPUS Books, FD is 0.814 and 0.769 under Qwen3-0.6B, and 0.889 and 0.879 under Qwen3.5-0.8B. The long-text summarization ordering changes. These results support some cross-model consistency in most within-task comparisons; specific values and rankings remain model-dependent.

We chose deletion because it directly tests whether removing input content makes the same target harder to predict. Prior work similarly removes content and observes prediction changes: ERASER’s comprehensiveness metric removes rationale spans and compares the predicted probability of the same class to assess reliance on those spans [3], providing a methodological basis for deletion.

Masking and replacement require specifying a mask token or replacement content. Covert et al. [4] likewise identify how features are removed as a key design choice in attribution methods. Scores must be interpreted according to their intervention; numerical consistency across perturbations requires separate validation. We use deletion throughout to compute and compare FD under one defined rule.

Response to Q5: How is Assumption 2 satisfied by the fixed-position synthetic construction, and what does Proposition 1 establish beyond concentration of dataset-average differences?

Placing the kk feature tokens first in Eq. (4) can be viewed as selecting one of the (nk)\binom{n}{k} equally likely placements under uniform shuffling. After uniform shuffling in Eq. (5), any further independent permutation preserves the joint distribution of the sequence and target. Assumption 2 thus applies, giving E[ΔN]=0\mathbb{E}[\Delta_N]=0 in Proposition 1. Figure 9 (Appendix B.1) compares dataset-level FD under this fixed feature/noise placement with random placement, showing little effect of token order in our synthetic experiments.

We sincerely thank the reviewer again for the detailed and constructive comments, which have substantially improved the quality of our manuscript.

References

[1] Frank Nielsen. Generalized Bhattacharyya and Chernoff upper bounds on Bayes error using quasi-arithmetic means. Pattern Recognition Letters, 2014.

[2] Visar Berisha et al. Empirically Estimable Classification Bounds Based on a Nonparametric Divergence Measure. IEEE Transactions on Signal Processing, 2016.

[3] Jay DeYoung et al. ERASER: A Benchmark to Evaluate Rationalized NLP Models. ACL, 2020.

[4] Ian Covert et al. Explaining by Removing: A Unified Framework for Model Explanation. JMLR, 2021.

Reviewer_Be3QOpenReview ↗
Rating
2
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Feature Density, a metric that measures how much of the training data is useful for a given task. The method identifies useful tokens or sentences by removing them and checking whether the model's prediction loss increases. Feature Density is then calculated as the proportion of useful units in the input.

The authors evaluate the metric on synthetic data and 12 real-world datasets. They also study how Feature Density relates to fine-tuning performance using four optimizers: Adam, AdamW, SGD, and Muon. The results show that higher Feature Density is often linked to lower training loss within similar task categories.

Critical strengths
  1. The paper introduces Feature Density to measure how much training data is relevant to a task. The method is easy to understand and offers a useful way to evaluate data quality.

  2. The paper evaluates Feature Density on synthetic data and 12 real-world datasets across different tasks. Experiments with four optimizers also help show that the findings are not limited to one training method.

  3. The authors study how Feature Density changes across datasets, input positions, and sample sizes.

Critical weaknesses

1.The experiments mainly report training loss rather than test performance. However, lower training loss does not necessarily mean better generalization. More evaluation on held-out test sets is needed to support Feature Density as a data quality metric.

  1. Datasets may differ in difficulty, input length, and other factors beyond Feature Density. These differences could also affect training loss. Controlled experiments are needed to better isolate the effect of Feature Density.

  2. The method uses different input units across tasks and a fixed threshold to identify relevant content. This makes scores harder to compare. More quantitative human validation and sensitivity analysis would help establish the reliability of the metric.

Questions for authors
  1. Does higher Feature Density lead to better test performance, not just lower training loss? Would selecting high-density samples improve accuracy compared with random selection?

  2. How can the authors separate the effect of Feature Density from other factors, such as task difficulty and input length? Could controlled experiments within the same dataset help?

  3. How do different input units and deletion score thresholds affect Feature Density? Can the authors provide more quantitative comparisons with human annotations?

  4. How does the cost of computing Feature Density scale with dataset size and input length? Is the method practical for large-scale data selection?

4 comments

Thank you for your support and insightful comment. It is critical in improving the quality of our manuscript.

Response to Weakness 1: More evaluation on held-out test sets is needed to support Feature Density as a data quality metric, and Q1: Does higher Feature Density lead to better test performance...Would selecting high-density samples improve accuracy compared with random selection?

In the revised manuscript, we updated the experiments evaluating test performance. For the Qwen3-0.6B model, the test loss results are shown in Figure 8 (Section 4.5), and the target-token accuracy results are shown in Figure 13 (Appendix C.2). Figures 12 and 14 in Appendix C.2 report test performance for Qwen3.5-0.8B. Across these six task categories, datasets with higher FD show an advantage in both test loss and token accuracy when compared within the same task.

Response to Weakness 2: Controlled experiments are needed to better isolate the effect of Feature Density, and Q2: How can the authors separate the effect of Feature Density from other factors...Could controlled experiments within the same dataset help?

We selected one dataset from each of the six task categories and sampled equal numbers of examples from the same training pool within each dataset. High-FD selects examples with FD in [0.6, 0.9], Low-FD selects examples with FD in [0.1, 0.4], and Random samples from the entire training pool.

Each curve in Figure 15 (Appendix C.2) represents a model fine-tuned on the corresponding subset; the three rows show training loss, test loss, and token accuracy, respectively. All three groups start from the same Qwen3-0.6B checkpoint, are trained with Adam for three epochs, and are evaluated on the same held-out set. This design evaluates test performance after data selection while holding the task, data source, and test set fixed.

At epoch 3, High-FD achieves both lower test loss and higher target-token accuracy than Random across all 6 tasks. On NIAH, High-FD has higher accuracy and also lower test loss. In the OPUS Books test loss panel in Figure 15 (second row, third column), all three groups start from the same test loss of 2.5142. Test loss decreases by 0.7322, 0.7217, and 0.6141 for High-FD, Random, and Low-FD, respectively, reaching final values of 1.7821, 1.7926, and 1.9001. This task directly supports the finding that, when differences in difficulty, input length, and related factors are constrained to some extent by selecting examples with different FD values from the same dataset, higher-FD data can still produce a larger reduction in test loss and a lower final test loss.

Response to Weakness 3 and Q3: How do different input units and deletion score thresholds affect Feature Density? Can the authors provide more quantitative comparisons with human annotations?

A deletion unit uu is the input segment removed in a single perturbation. With the model and target YiY_i fixed, we define $R_i(u) =\mathcal{L}\left(Y_i,f(W,X_i\setminus u)\right) -\mathcal{L}\left(Y_i,f(W,X_i)\right)$. Deleting the unit increases the target loss if and only if Ri(u)>0R_i(u)>0. Thus, if the aim is to measure the fraction of units that contribute positively to predicting the target, a threshold of 00 gives this criterion. A threshold of 00 is also the value that guarantees a meaningful definition for every possible deletion score. A threshold of 00 guarantees a consistent meaning across different datasets.

Using the same batch of 10 NIAH examples, we fixed the inputs, targets, and scoring model and deleted consecutive spans of 1, 2, 4, or 8 sentences. The table below reports FD at τ=0\tau=0 and separately examines how much of the result comes from small positive scores near zero. Let Su=Ri(u)/TS_u=R_i(u)/T, where TT is the number of target tokens. Excluding units with 0<Su≤0.010<S_u\leq0.01 from the numerator while keeping the denominator unchanged reduces mean FD by 0.00 to 3.56 percentage points across the four granularities. This directly quantifies the sensitivity to near-zero scores in this batch of examples.

Results for the same NIAH examples at four deletion granularities

Sentences per deletion unitMean FD (τ=0\tau=0)Mean FD after excluding near-zero positive scoresDecrease (percentage points)
10.49200.45643.56
20.45370.42382.99
40.48880.46782.10
80.49170.49170.00
AuthorsOpenReview ↗

Response(Part 2)

Response to Q4: How does the cost of computing Feature Density scale with dataset size and input length? Is the method practical for large-scale data selection?

The ii-th input contains UiU_i deletion units and requires one score for the complete input plus one for each version with a unit deleted. Scoring NN candidate examples therefore requires N+∑iUi=N(1+Uˉ)N+\sum_i U_i=N(1+\bar U) sequence evaluations, which can be batched. When the distributions of sequence lengths and unit counts are similar, the workload scales linearly with the number of examples. As inputs become longer, both the cost of each evaluation and the number of units per example may increase. Under the same model, hardware, and batching settings, and with a similar length distribution, if a representative subset can be processed at qq candidate examples per second, the time required for the full pool can be estimated as N/qN/q seconds.

Thank you again for your valuable comments. We hope the revisions and additional experiments can address your concerns, and we look forward to your feedback!

I appreciate the authors’ thoughtful responses and the additional experimental results. While these efforts help clarify several points, I believe some of my key concerns remain insufficiently addressed. Therefore, I have decided to maintain my original rating.

Thank you for your constructive suggestions and valuable comments! We fully respect your decision and appreciate your effort in reviewing our manuscript.

Reviewer_3w2aOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes Feature Density to quantify the fraction of input units that support a fixed model's target predictions. It deletes each token or sentence separately and measures the resulting change in target loss. Units with positive deletion scores count as features, and their fractions are averaged across examples to characterize a dataset. The paper compares these labels with synthetic features and selected human annotations, then studies fine-tuning with four optimizers. Experiments on synthetic sequences and twelve real-world datasets show increasing measured density with synthetic feature ratios and associations between higher density and lower reported fine-tuning losses.

Critical strengths

The deletion-based definition is explicit about the reference model, target, and input unit, making the proposed statistic straightforward to compute and inspect.

Critical weaknesses
  1. Classifying every strictly positive deletion score as a relevant feature does not distinguish meaningful support from tiny perturbation effects. At a true synthetic feature ratio of 0.1, measured density is approximately 0.54; even perfect recovery of all true features would leave roughly half the noise tokens labeled as features. Appendix C also notes substantial near-zero scores in sentiment data. These observations undermine the interpretation of FD as the proportion of task-relevant content. The selected annotation examples do not resolve this calibration problem or establish reliable feature-identification precision.
  2. The real-data fine-tuning results compare different datasets' own training losses, often with substantial differences already present before training. Lower loss on GSM8K than on AM Thinking V1, for example, does not show that training on higher-FD data produces a better model on a shared target distribution. Dataset difficulty, target predictability, and the starting model can explain the association. Figures 8 and 12 report training loss, despite the abstract's broader statement about lower test loss in both settings. The evidence therefore does not yet establish FD's incremental value as a data-quality or data-selection metric beyond baseline loss.
Questions for authors
  1. Does Proposition 1 claim only distributional equality under Assumption 2, rather than invariance for an individual sequence? Fixed-position feature/noise sequences are not generally exchangeable, so how does that assumption apply to the comparison in Figure 9? The statement also attributes zero expectation to Assumption 1, whereas the proof invokes Assumption 2.
  2. In Eq. (4), increasing the feature count reduces the number of noise vectors while each noise vector retains the same variance. Can the authors separate the effect of feature proportion from the resulting change in aggregate noise energy, for example by holding the aggregate signal-to-noise ratio fixed?
4 comments

We sincerely thank you for your careful and thorough review that improves the quality of our manuscript.

Response to Weakness 1: At a true synthetic feature ratio of 0.1, measured density is approximately 0.54; even perfect recovery of all true features would leave roughly half the noise tokens labeled as features...

In our synthetic-data design, the Feature Ratio ρ=0.1\rho=0.1 is a data-generation hyperparameter. The corresponding Feature Density (FD), computed from the model’s deletion scores, is approximately 0.540.54 and measures the fraction of tokens whose deletion increases the model’s loss on the target. These two different quantities describe an ideal feature ratio set by a hyperparameter and the model’s response to token deletion. As described in Eq. (4) of Section 3.1, both feature and noise tokens are constructed from centered Gaussian random vectors with similar distributional structure. Our synthetic data provides a challenging setting for distinguishing features from noise.

A simple example explains why the fraction of noise tokens counted toward FD can be close to one half. Suppose the model’s response to noise is additive:

\xi_j\sim\mathcal{N}\!\left(0,\frac{\sigma^2}{D}I_D\right).$$ Here, the noise is independent of the label $y\in\{-1,+1\}$, $w\neq0$ is a fixed vector, and $\sigma>0$. In this example, we use $\mathcal{L}(y,f(X))=-yf(X)$ as a loss proxy. Deleting noise token $j$ leaves the signal component and the other noise terms unchanged, so $$f(X)-f(X_{\setminus j})=w^\top\xi_j.$$ The corresponding deletion score is $$\begin{aligned} R_j &=\mathcal{L}(y,f(X_{\setminus j}))-\mathcal{L}(y,f(X))\\ &=y\bigl[f(X)-f(X_{\setminus j})\bigr]\\ &=yw^\top\xi_j. \end{aligned}$$ By the symmetry of the centered Gaussian distribution, this score is equally likely to be positive or negative: $$\Pr(R_j>0)=\Pr(yw^\top\xi_j>0)=\frac12.$$ In other words, although noise is generated independently of the label, its random contribution on a particular example has approximately a one-half probability of supporting the target prediction and therefore receiving a positive deletion score. Here, $1/2$ is the false-positive probability for a noise token in this example. Thus, for a given Feature Ratio $\rho$, if most true features are identified, their contribution to FD is approximately $\rho$, while the expected contribution from noise is $(1-\rho)/2$, giving $$\mathbb{E}\!\left[\mathrm{FD}\right] \approx \rho+\frac{1-\rho}{2}.$$ Substituting $\rho=0.1$ gives approximately $0.55$, close to the value $0.54$ in Figure 3. The corresponding values from this calculation are: | Feature Ratio $\rho$ | Calculation | Approximate expected FD | | -------------------: | :-------------: | ----------------------: | | 0.1 | $0.1+(1-0.1)/2$ | 0.55 | | 0.2 | $0.2+(1-0.2)/2$ | 0.60 | | 0.3 | $0.3+(1-0.3)/2$ | 0.65 | This example demonstrates that the value of $0.54$ means most feature tokens are identified by FD such that its value concentrates around $\rho+\frac{1-\rho}{2}$. > Response to Weakness 2: Figures 8 and 12 report training loss, despite the abstract's broader statement about lower test loss in both settings... Thank you for pointing out the discrepancy between the experiment presentation and the statement in the abstract. To address this, we added test-set evaluations on 12 real-world datasets according to the experimental settings in the main text. In the revised manuscript, the test loss results for Qwen3-0.6B are presented in Figure 8 (Section 4.5), and the target-token accuracy results are presented in Figure 13 (Appendix C.2). The corresponding results for Qwen3.5-0.8B are provided in Figure 12 (Appendix C.2) and Figure 14 (Appendix C.2), respectively. We can observe from these results that, in mathematical reasoning, GSM8K has a higher Feature Density than AM Thinking V1 and achieves lower test loss and higher test token accuracy across all four optimizers for both models, supporting our conclusion that higher Feature Density is associated with better fine-tuning performance within the same task category. Across the six task categories, these additional results show an overall tendency for datasets with higher FD to achieve lower test loss and higher target-token accuracy when compared within the same task category.
AuthorsOpenReview ↗

Response (Part 2)

Response to Q1: Does Proposition 1 claim only distributional equality under Assumption 2, rather than invariance for an individual sequence? Fixed-position feature/noise sequences are not generally exchangeable, so how does that assumption apply to the comparison in Figure 9? The statement also attributes zero expectation to Assumption 1, whereas the proof invokes Assumption 2.

Yes.

For a single sequence, Assumption 2 gives equality in distribution, which allows the two values to differ on a particular draw. Proposition 1 uses this equality to show that the dataset-average difference ΔN\Delta_N has mean zero and converges to zero almost surely. Assumption 2 applies to independent permutations of the uniformly shuffled sequences in Eq. (5). Figure 9 (Appendix B.1) compares that randomized setting with one fixed placement (which can be viewed as selecting one of the (nk)\binom{n}{k} equally likely feature placements under uniform shuffling), in which the first kk tokens are features.

Assumption 2 describes the symmetry of the shuffled distribution under independent permutations. Because that symmetry alone does not determine the mean FD for one fixed placement, Figure 9 compares the mean FD when every sequence uses the first kk positions for features with the mean FD when feature positions are randomized for each sequence. The figure shows that token order had little effect on dataset-level FD.

Finally, we have corrected the typo of "invokes Assumption 1" in the revised manuscript, to reflect that Assumption 2 gives E[ΔN]=0\mathbb{E}[\Delta_N]=0.

Response to Q2: In Eq. (4), increasing the feature count reduces the number of noise vectors while each noise vector retains the same variance. Can the authors separate the effect of feature proportion...

We compared Feature Ratios while holding the aggregate signal-to-noise ratio (SNR) fixed, using the same configurations of the synthetic data experiments in the manuscript.

From Eq. (4) in Section 3.1, the sum of the squared norms of the feature tokens is α2\alpha^2. The expected sum for the noise tokens is C=(n−k)σ2C=(n-k)\sigma^2. The aggregate SNR, defined here as the ratio of these energies, is therefore α2/C\alpha^2/C. We kept α\alpha and CC fixed by adjusting σ\sigma as the number of feature tokens changed.

The table below compares Feature Ratios of 0.20.2 and 0.80.8, with n=50n=50, α=1\alpha=1, and 1000 test sequences per condition. Each row holds both total feature energy and expected total noise energy fixed. We kept the model parameters and data seed unchanged across these comparisons and evaluated FD and test loss before fine-tuning.

Expected total noise energy CCσ\sigma, ρ=0.2→0.8\rho=0.2\to0.8Mean FD, ρ=0.2→0.8\rho=0.2\to0.8Fixed-model test loss, ρ=0.2→0.8\rho=0.2\to0.8
401→21\to20.6010→0.84900.6010\to0.84901.479→1.1211.479\to1.121
1602→42\to40.5920→0.74280.5920\to0.742810.537→6.60010.537\to6.600
6404→84\to80.4327→0.68470.4327\to0.684727.983→13.52027.983\to13.520

The first row provides a direct comparison: both settings have total feature energy of 11 and expected total noise energy of 4040, yet increasing the number of feature tokens from 10 to 40 raises FD from 0.60100.6010 to 0.84900.8490. The test loss also decreases from 1.4791.479 to 1.1211.121. The same direction of change is observed at C=160C=160 and C=640C=640.

Reviewer_3w2aOpenReview ↗

Thank you for your rebuttal

Thank you for your rebuttal, which has addressed most of my concerns. Therefore, I will raise my score to 2.

Thank you very much for your positive response and for raising the score! We truly appreciate your effort in reviewing our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.