We sincerely thank the reviewer for their thoughtful feedback and for recognizing the motivation, novelty, and effectiveness of our PANGEA framework. The weaknesses and questions raised were highly valuable, and we are pleased to provide further clarification and internal results.

W1 (Domain Relevance and Filtering of Synthetic Data). Is every general data entry suitable for transformation into any arbitrary domain? In other words, some data may not be appropriate for use in specific domains. The authors should consider designing a procedure to filter out less desirable data for the target domain during both the generation and training stages.

Your insight into maintaining and precisely controlling the quality of domain-specific synthetic data through filtering is highly valuable. It closely reflects the challenges we faced during the early stages of our research, and we believe your comments accurately pinpoint the core issues we encountered at the time.

We agree that maintaining domain relevance and filtering out low-quality data are critical challenges in synthetic data generation, as synthetic data often suffers from significant quality degradation. As you suggested, using reweighting methods for synthetic data and filtering components within the PANGEA process are both promising directions. Among various filtering strategies, we explored an approach that directly filters the general data. Indeed, in the early phase of our research, we explored this direction by training a verifier model to pre-filter general data. Specifically, we trained a model to predict whether a given general data sample ($D_g$), when paired with source domain data ($D_s$), is suitable for generating high-quality synthetic data.

To supervise the model, we used the o1 metric from the PANGEA framework to label data pairs. Samples deemed high-quality were labeled as 1, and those that did not meet the quality threshold were labeled as 0. We manually annotated approximately 20,000 examples, maintaining a balanced 1:1 class ratio. The verifier was trained using linear probing on the Llama-3.1-8B model and produced a probability score between 0 and 1, estimating whether a given (source data, general data) pair, or Synth-Guide Block ($\tau$), would yield high-quality synthetic outputs.

However, in practice, the verifier’s precision was limited. It achieved only 50 to 60 percent accuracy in correctly identifying positive cases, which resulted in a low yield when attempting to collect high-quality samples using this method. We also conducted an additional experiment in which we filtered generated synthetic data based on quality. The results are summarized below:

Interestingly, even when low-quality samples were included, the overall performance improved when the model was trained on the entire dataset of 30,000 samples. This suggests that the inclusion of diverse examples, including negative ones, contributes to generalization. In hindsight, we found that directly utilizing general data was ultimately more effective than filtering or relying only on high-quality subsets. Building on these observations, we shifted our focus from selecting only high-quality out-of-domain data to designing a framework that makes the best use of any available general data. In particular, our proposed three-stage framework, centered around the high-quality Synth-Profile ($\eta$) constructed from source data, has proven more effective at producing useful synthetic data even when the general data is arbitrarily selected. While we recognize that filtering mechanisms can still be beneficial, especially when the proportion of ambiguous or low-quality synthetic samples (cf. Appendix Table 10) becomes too high, we believe that training a dedicated verifier and applying data filtering for every individual domain is neither scalable nor generalizable. Such an approach increases both complexity and cost, and it limits the practicality of applying synthetic data generation across diverse domains.

W2 (Lack of Human Annotation for Quality Measurement). Quality measurement lacks human annotation: While the use of o1 as a judge may provide a strong automatic evaluation...

We randomly sampled 100 synthetic data points from each benchmark (GSM8K, FinQA, MedQA, and CDSL) to compare three generation methods: A (Naive), B (Evol-Instruct), and C (PANGEA). Fifty-seven graduate-level annotators, after a brief pilot, evaluated the three generated synthetic data in random order using the same criteria provided in the supplement, labeling each as ‘Best’, ‘Second’, and ‘Worst’.

Based on quality assessments of 100 items per benchmark, PANGEA (C) consistently received the highest ratings across all four domains. In contrast, Naive (A) was most frequently rated lowest in quality, while Evol-Instruct (B) typically received intermediate evaluations. Inter-annotator agreement was moderate, with a Krippendorff’s α of 0.61.

To determine whether these observed differences in quality ratings were statistically significant, we applied χ² and Friedman tests to the annotation results for each domain.

Across all domains (GSM8K, FinQA, MedQA, CDSL), PANGEA (C) received the highest proportion of “Best” ratings, indicating it consistently produced the highest-quality outputs. In contrast, Naive (A) was most frequently rated as “Worst,” while Evol-Instruct (B) generally ranked in the middle. To assess the statistical significance of these differences, we conducted both χ² (chi-squared) and Friedman tests, which confirmed significant performance differences across all domains (p < 0.01).

Q1. What is the amount of the CoT-Collection after excluding math, medical, and finance-related entries?

After removing entries related to our target domains, the CoT-Collection contained approximately 900k samples. For storage and computational efficiency, we created a working subset of about 500k samples, from which general data instances were randomly drawn for augmentation.

Q2. CoT-Collection is more than 1.8M, but the authors only generated 10k, 30k, or 120k synthetic data. Is this randomly selected?

Arithmetically, PANGEA can generate up to (# general data) × (# source data) instances. However, our objective was not merely to produce a large volume of synthetic data for specific domains, but to develop a universal pipeline applicable to any downstream task as needed. Accordingly, we focused on validating performance trends and scalability across a range of domains, rather than simply increasing the amount of synthetic data.

To this end, we selected three representative scales (10k, 30k, and 120k) that clearly demonstrate how PANGEA (1) consistently outperforms baseline approaches and (2) continues to scale more effectively as additional synthetic data are introduced. All general data were randomly sampled from the CoT-Collection after filtering out domain-related content per task, as our goal is to demonstrate the feasibility of using domain-unrelated data for large-scale synthetic generation.

We sincerely thank the reviewer for their thoughtful and constructive feedback. Your insightful suggestions have significantly helped us improve the clarity, rigor, and scope of our work. Below, we address each of your comments in detail.

W1 (Scalability to Larger Models). Results on larger models, such as 7B variants.

We are sincerely grateful to the reviewer for this excellent suggestion that significantly strengthens our paper. We conducted an experiment using the Llama-3.1-8B model at the 10k setup. Notably, our PANGEA framework is highly effective on the larger 8B model, even outperforms Llama-3.1-8B-Instruct:

W2. The LLMs used in Stage 2 and Stage 3 may possess prior knowledge of downstream benchmarks…

Thank you for your insightful comment. We address your concerns in two parts: (1) fairness in comparison with baselines and (2) the role of prior knowledge.

First, all methods, including baselines, use the same Llama3.3-70B-Instruct model for synthetic data generation, ensuring a fair comparison. The performance gains from PANGEA are not due to differences in prior knowledge but rather stem from its systematic approach. Specifically, PANGEA performs stage-wise profiling to extract key components from seed data, and then leverages diverse general data to enhance variety while maintaining difficulty and thematic consistency.

Second, to assess the role of prior knowledge, we conducted a new experiment on the Korean CSAT exam, data not included in the LLM’s training data. We generated 10,000 synthetic CSAT problems using the 2023 and 2024 official exams as seed data, and evaluated performance on a held-out test set from the 2025 CSAT. As shown in our response to Reviewer Puqb, PANGEA outperformed baselines that use a smaller 1B model across all difficulty levels (Easy, Medium, Hard) and in total score. This further supports that PANGEA’s effectiveness does not rely on prior knowledge.

These results clearly demonstrate that even for tasks unseen during pre-training of the LLM in Stages 2 and 3, PANGEA can still generate useful synthetic data to train the target model.

W3. More downstream datasets are needed, preferably with fine-grained splits; the introduced CDSL dataset lacks representativeness.

More downstream datasets with fine-grained splits. We believe this concern is effectively addressed by the additional experiments we conducted on the Korean CSAT task. Detailed settings and results can be found in our response to Reviewer Puqb.

Representativeness of CDSL. We are fully aware of concerns regarding potential knowledge leakage from widely used benchmarks such as GSM8K, as also noted in your W2. Since such leakage is difficult to eliminate entirely, we deliberately designed CDSL as a novel, out-of-distribution “acid test.” Its symbolic syntax, custom operators, and domain-specific content are entirely absent from any public dataset, and even the 70B model achieves only around 20% accuracy. In this sense, we understand your concern regarding the representativeness of CDSL.

However, CDSL is not meant to reflect typical downstream tasks. Rather, it is intentionally constructed to evaluate the robustness of data augmentation and domain adaptation in truly novel settings, where prior knowledge is unavailable and only a small number of seed examples (e.g., 100) are provided. While standard benchmarks such as GSM8K, MedQA, and FinQA are undoubtedly more representative of real-world tasks, they cannot be entirely free from prior knowledge leakage. CDSL was specifically designed with this issue in mind, providing a controlled setting where such leakage is virtually impossible.

W4 (Performance with Strong Reasoning Models). It remains unclear how the proposed method performs with strong reasoning models such as QWQ or DeepSeek-R1.

We would like to clarify that the primary goal of PANGEA is not merely to achieve state-of-the-art performance on reasoning tasks. Rather, our objective is to develop a universal and robust framework for data augmentation that can be effectively applied to any new, domain-specific task, even when starting from a small set of seed examples (e.g., 100). Owing to its broad applicability, PANGEA also enables strong publicly available reasoning models to be fine-tuned into domain-specific specialists using the same approach.

As an illustration, we conducted an additional experiment using DeepSeek-R1-Distill-Llama-8B; the only difference from our original setup is that we created new labels, including a “reasoning-path,” to align with the model’s intended prompt template:

These results clearly show that fine-tuning with data generated by PANGEA provides substantial gains even for a strong, specialized reasoning model. This validates that our framework is not only effective for general-purpose models but also serves as a powerful tool to further enhance the capabilities of expert models in their respective domains.

W5 (Ablation Study on the Number of Source Samples). An ablation study on the number of domain-specific samples in the source dataset would be valuable for understanding the method’s robustness under varying levels of data scarcity.

PANGEA was originally designed with the assumption of scarce seed data in mind, so exploring its performance under such conditions is certainly a valuable direction. While the current main experiments use 100 seed examples, we additionally tested scenarios with fewer seed samples:

The results demonstrate that PANGEA’s performance remains remarkably robust across varying amounts of domain-specific seed data. Notably, as long as the number of seed examples is not extremely small (e.g., 10), PANGEA outperforms the baseline trained with 100 seed examples even when using significantly fewer seeds (e.g., 80, 40, 20). We further conduct an in-depth analysis to better understand what happens when the number of seed data becomes very limited:

Root Cause: The Impact on Synth-Profiling. We identified that a limited number of seed data negatively impacts the quality of the Synth-Profiling stage. With fewer examples, the generated profile ($\eta$) tends to be less detailed and less specific. The followings are Synth-Profiles for GSM8K with 10/100 seeds; the 10-seed profile asks for fewer equations and lacks details compared to the 100-seed profile (e.g., "kid-friendly," "multi-step solution," specific units):

Synth-Profile from 10 seeds:

- Pick one ordinary scene (errands, snacks, chores, mini-trip, etc.).
- List number with a symbol (A, B, C, D) → value + unit + brief label.
- Write 2–3 equations that use those symbols.
- End by stating which symbol the solver must find.
- Give a bullet-style plan showing the order the equations should be tackled.
- Do not write the full word problem or reveal any answers.

Synth-Profile from 100 seeds (Original):

### Choose a Kid-Friendly Scenario
- Restate one everyday clause students can picture (shopping, chores, snacks, simple travel).
- Remove unnecessary details.
### Select 3-5 Meaningful Numbers
- Prefer numbers in General Question.
- Invent realistic values if needed, supporting clear multi-step solution.
- Use everyday units (dollars, minutes, km, items, °C, simple percents).
### Define Symbols & Units
- Assign symbols A, B, C, (D, E) with brief label and explicit unit.
### Write 4-5 Simple Equations
- Use only +, -, ×, ÷, %, or single-step unit conversions.
- No new numbers; use chosen symbols or intermediate results.
### State the Target
- Specify final symbol to solve and its numeric meaning.
### Outline Step-by-Step Plan
- Provide one bullet per equation, in order (4-5 steps).

The following illustrative examples show that although the synthetic data generated from 10 seed examples remains mathematically valid, it displays qualitative differences. While the 10-seed example is more verbose and contains an unrealistic scenario (a 200°F appetizer), the 100-seed example is clear and contextually coherent.

Example from 10 seeds:

Using identical food thermometers, Sarah notes that her appetizer registers 200 °F, while Tim observes his at 150 °F. Sarah then turns up the heat, raising her appetizer by exactly 20 °F. Tim follows suit, but rather than copying the same degree change, he warms his snack (in the upward direction) by an amount equal to one-half of the percentage that Sarah’s 20 °F rise represents relative to her appetizer’s original temperature. After both of these adjustments have been completed, how many degrees Fahrenheit apart are the two snacks?

Example from 100 seeds (Original):

Sarahn and Timon discovered two planets, where Sarahn’s crust temperature is 48 °C, and Timon’s crust is half as hot due to lower geographic activity. If the temperature difference between their planets is twice Timon’s crust temperature, what is the difference in their crust temperatures?

In conclusion, this ablation study confirms the robustness of PANGEA and highlights that a moderately small set of ~40 high-quality examples is sufficient to generate a strong domain profile. We are grateful to the reviewer for encouraging this valuable investigation.

We sincerely thank the reviewer for the positive feedback on the creativity, practicality, modularity, and effectiveness of our work. We also appreciate the insightful questions, which have prompted us to conduct further analysis that strengthens our paper. We provide our responses below.

W1, Q1 (Feasibility of Domain-Agnostic Prompt Templates). Is it feasible to design domain-agnostic prompt templates that could improve the flexibility of this method?

As noted (cf. Appendices A.2–A.4), different prompts were used for each downstream task. To assess domain-agnosticity, we conducted additional experiments using generalized prompts across all three stages: 1) Domain-agnostic Synth-Profiling: $\eta$ was extracted without domain-specific cues. 2) Domain-agnostic Prompt-Writer: $\eta$ was incorporated without referring to any domain. 3) Domain-agnostic Projection: data was synthesized without domain mentions. The table below summarizes accuracy (%) at the 10k-sample scale:

As shown, although there is a slight performance gap compared to our original domain-specific prompts (Original), the domain-agnostic version of PANGEA (Domain-Agnostic) still outperforms all other baselines by a wide margin. This indicates that the pipeline can operate effectively in a domain-agnostic setting, as the Synth-Profiling stage functions as a domain-adaptive component that guides the rest of the process.

We will add this experiment and the unified template to the supplementary materials in the camera-ready version. Thank you for the constructive feedback.

Q2 (Controlling the Difficulty of Synthetic Data). Is there a way to control or assess the difficulty of generated samples? Can unnecessarily simple or complex examples harm training effectiveness?

This is an excellent point. In our framework, the difficulty distribution of the generated data is primarily controlled and guided by the provided source data. The Synth-Profiling stage does not just extract topics; it analyzes the structural complexity, reasoning depth, and stylistic nuances of the seed examples. For instance, the profile for GSM8K data is given by:

### Choose a Kid-Friendly Scenario
- Restate one everyday clause students can picture (shopping, chores, snacks, simple travel).
- ...
### Select 3-5 Meaningful Numbers
- Prefer numbers in General Question.
- ...
### Define Symbols & Units
- Assign symbols A, B, C, (D, E) with brief label and explicit unit.
- ...
### Write 4-5 Simple Equations
- Use only +, -, ×, ÷, %, or single-step unit conversions.
- ...
### State the Target
- Specify final symbol to solve and its numeric meaning.
- ...
### Outline Step-by-Step Plan
- Provide one bullet per equation, in order (4-5 steps).
- ...

Here, “Outline Step-by-Step Plan“ and “Define Symbols & Units” control the difficulty of the synthetic problems. Thus, our framework adaptively generates synthetic data that aligns with the “difficulty distribution” of the original, expert-curated source data. This prevents the generation process from collapsing into overly “trivial” examples or drifting toward “unnecessarily complex” ones, as it remains anchored to the specifications of the original data. Our primary goal is to efficiently augment the existing source data distribution, and this mechanism ensures we remain faithful to that objective.

Q3 (Ensuring the Quality of Synthetic Data).. How to ensure the quality of the automatically generated data? Is there any human verification or automated quality filtering?

Thank you for raising the insightful question regarding the quality of synthetic data. We also acknowledge the potential risk of generating low-quality data during the synthetic generation process. To address this concern, we conducted an experiment on the GSM8k dataset (10k examples) using an LLM-based self-evaluation to quantitatively measure the quality of each generated sample on a scale from 0 to 100 (cf. Figure 4). The resulting score distributions per method demonstrated that our approach consistently maintained superior data quality compared to baseline methods.

Notably, even our lower-scoring data retained meaningful mathematical contexts despite issues related to realism or situational ambiguity, as illustrated by the following example. This implies that such data, despite their lower scores, could still meaningfully contribute to the learning process.

Low quality example Using identical food thermometers, Sarah notes that her appetizer registers 200 °F, while Tim observes his at 150 °F (…) After both of these adjustments have been completed, how many degrees Fahrenheit apart are the two snacks?

To further validate this hypothesis, we compared models trained exclusively on low-quality data, exclusively on high-quality data, and on the complete dataset. The results clearly showed that even data rated as low-quality under our method were beneficial for model training.

The generation of this "low-quality yet helpful" data is attributable to our Synth-Profiling stage, which analyzes crucial elements from source data required for meaningful synthetic data generation. By extracting these essential components from general data, our approach leverages diverse and valuable information while effectively filtering out unnecessary details, thus ensuring the relevance and usefulness of generated synthetic data.

Q4 (Applicability to Multilingual Tasks). Can the proposed framework be applied on multilingual or cross-lingual tasks? If so, what challenges might arise in adapting the projection and augmentation process to languages with different linguistic structures?

Whether general data in a different language can be effectively projected to generate synthetic data in the target language is both an intriguing and highly practical question. The LLM model (Llama-3.3-70B-Instruct) we utilize for synthetic dataset generation is inherently multilingual. Thus, even if the provided general data and seed data are composed in different languages, the model is fully capable of effectively generating synthetic data through appropriate projection across languages. To verify this, we conducted a new experiment on a Korean-language task: the Korean College Scholastic Ability Test (CSAT). This serves as a test case for whether a primarily English CoT-Collection database can be used to synthesize downstream data in Korean.

More precisely, we generated 10k synthetic Korean CSAT problems with our framework, leveraging official problems from the 2023 and 2024 CSAT exams as source data. The model was then evaluated on a held-out set of official problems from the 2025 CSAT, which were not included in the seed data. Evaluation was based on the accuracy (%) for Easy, Medium, and Hard questions, as well as the overall Total Score. The comparative results of synthetic data generation methods using the Llama-3.2-1B model was as follows:

Notably, PANGEA not only achieves the highest total score but is also the only method that enables the model to solve any “Hard” problems. This demonstrates that the proposed PANGEA framework is both effective and capable of delivering meaningful performance gains, even for languages that are underrepresented in large-scale, domain-unrelated datasets. This result highlights the robustness and generalizability of our approach.

W2 (Risks Arising from the Gap between General and Source Data). This paper lacks analysis of failure cases—particularly when the gap between general and domain-specific data is significant. Is there a risk of semantic drift?

We agree that transferring general data to domain-specific contexts can risk semantic drift. To address this, our experiments include settings where the source data is intentionally selected to be domain-unrelated, creating a substantial gap between general and source data. Despite this, as discussed in Q2 and Q3, we observed no signs of semantic drift or degradation in the difficulty and quality distributions of the generated synthetic data.

This robustness stems from the Synth-Guide Block ($\tau$), obtained via Stages 1-2, which serves as a semantic anchor that guides the projection process toward the structural and topical intent of the original domain. As shown in the example in Figure 3, performing projection without $\tau$ leads to failure cases, whereas the use of $\tau$ effectively resolves this issue.

W3 (Lack of Theoretical Justification). This work is primarily supported by empirical results, with limited theoretical explanation...

We acknowledge that the theoretical foundations of synthetic data generation are still underexplored, with only a few recent efforts addressing this gap [1]. Our work instead builds a strong empirical basis for the practical effectiveness of LLM-driven synthetic data generation. Through ablation studies, we isolate the impact of each component in our projection mechanism, showing consistent gains in diversity, difficulty alignment, and overall quality. While our focus is empirical, we believe these findings can inform future theoretical work.

References
[1] Gan and Liu, “Towards a Theoretical Understanding of Synthetic Data in LLM Post-Training: A Reverse-Bottleneck Perspective.”

We thank the reviewer for their time and effort. However, we must respectfully point out that the content of this review appears to be for a different paper. Our submission, "PANGEA: Projection-Based Augmentation with Non-Relevant General Data for Enhanced Domain Adaptation in LLMs," is a study on synthetic data generation for Large Language Models (LLMs). The review discusses topics entirely unrelated to our work, such as "embodied navigation," "pre-extracted target images," and "CLIP embeddings for navigation." These concepts belong to the fields of robotics and computer vision, not LLM data augmentation. Due to this fundamental mismatch, we are unable to address the specific strengths and weaknesses raised. We believe there has been a mix-up with another submission and kindly request that this be taken into consideration.

We sincerely thank Reviewer for the constructive and insightful feedback. We appreciate the positive evaluation of our work's motivation, extensive evaluations, and clarity. We are grateful for the opportunity to address the questions raised in the "Weaknesses" section, and we believe our responses and new experiments further strengthen our contributions.

W1 (Source data needs & effects). What are the requirements for the initial source dataset (e.g., size, diversity)? How do its characteristics affect the performance of data synthesis and domain adaptation?

Thank you for this important question regarding the sensitivity of PANGEA to the initial seed data. Our method was designed for data-scarce scenarios, and we agree that its performance under varying conditions is a crucial aspect to explore.

Regarding the number of initial seed samples: To address this, we conducted additional experiments during the rebuttal period by varying the number of seed examples from 100 down to 10. These experiments were run on our 10k synthetic data setup.

As the results show, PANGEA remains robust and outperforms baselines that use 100 seed samples, even when the number of seed samples is significantly reduced to as few as 20. The performance degrades gracefully as the seed count decreases, which is expected since the quality of the initial Synth-Profiling stage depends on having a representative sample of the target domain.

Regarding the diversity of the source data and the setting of the general dataset: These are excellent points. We hypothesize that higher diversity in the initial seed data would improve the quality of the generated profile($\eta$) , leading to better synthetic data. Similarly, the breadth and diversity of the general dataset likely set the upper bound for the diversity of the generated data. We will add a detailed discussion of these points to our final manuscript.

W2(saturation with respect to the number of training samples). How about more training data is provided? Will it be saturated given more training data?

This is a very insightful question. We anticipate that performance will eventually saturate. This saturation point would likely be determined by the combined diversity of the fixed source data and the large-scale general data. If the number of source data examples is fixed, the diversity of the general data ultimately bounds the variety of the synthetic data that can be generated. Once the diversity from the general data pool is exhausted, performance gains will likely plateau. While conducting experiments on a much larger scale (e.g., >120k) would be valuable, the time constraints of the discussion period make it challenging to complete them now. We are committed to including a thorough discussion of this expected saturation in the final manuscript. Thank you again for this excellent suggestion.

W3(projection stage). Why does the projection stage lead to a distribution aligned with the source domain?

The key to ensuring the projected data ($D_syn$) aligns with the source domain lies in our

Stage 1: Synth-Profiling. In this stage, we distill a structured domain-specific profile ($/eta$) from the source data ($D_s$). This profile explicitly captures the core characteristics, formats, and structural patterns of the target domain. The subsequent stages use this profile as a strict guide:

Stage 2 (Prompt-Writer) extracts relevant information from general data ($D_g$) that maps directly to the elements defined in the profile($\eta$), creating a Synth-Guide Block($\tau$).

Stage 3 (Projection) then uses this structured block ($\tau$) to generate a new data point, ensuring it conforms to the format and style of a source data example ($d_s$). In essence, the Synth-Profiling stage acts as a "compass," ensuring that even though we are drawing diversity from an unrelated general dataset, the final output is "projected" back into the specific structure and context of the target domain.

W4(Scalability to larger LLMs). How would the performance change if larger models (e.g., 7B or 14B) were used?

We are sincerely grateful to the reviewer for this excellent suggestion, which significantly strengthens our paper. To test PANGEA's effectiveness on larger models, we conducted a new experiment using the Llama-3.1-8B model on our 10k data setup. The results are highly encouraging. PANGEA not only delivers substantial improvements over the pre-trained model but also outperforms the heavily-resourced Llama-3.1-8B-Instruct model on three of the four benchmarks.his demonstrates that our framework is not only applicable but highly effective for larger, more capable models, confirming its scalability and generalizability.

Once again, we sincerely thank the reviewer for their valuable time and insightful suggestions. The new experimental results you recommended will be included in the final camera-ready version, and the points discussed during the rebuttal will also be reflected in the final manuscript. We hope our responses have fully addressed your concerns.

Additionally, we were not able to view the score that was given. If possible, we would greatly appreciate it if you could kindly leave a brief comment indicating your initial score for our reference. Lastly, if the main concerns have been resolved, we would be grateful if the score could be updated to reflect that.