Beyond the Surface: Enhancing LLM-as-a-Judge Alignment with Human via Internal Representations

We sincerely appreciate the time and effort you have devoted to offering insightful and constructive comments on the paper. Below, we present our main responses to the key concerns raised:

Q1: The method lacks novelty. As the authors themselves acknowledge, the idea of using internal hidden states has been explored extensively since the introduction of BERT.

A1:

• We would like to clarify that there might be a misunderstanding, as we did not make such a claim in our paper. In lines 77–79, we specifically pointed out that “hidden representations across different layers have been widely observed to exhibit distinct characteristics”. We will revise and clarify this description in future work to prevent potential misunderstandings.

• This work introduces a new perspective to enhance point-wise evaluation by leveraging intermediate layer representations of the model. Unlike traditional methods that rely solely on final layer outputs, we build a distinct evaluation framework based on an in-depth analysis of internal layers, laying the groundwork for further improving evaluation capabilities through deeper exploration of internal representations.

Q2: My key concern is whether the lightweight training of layer weights can truly generalize across diverse evaluation benchmarks without additional fine-tuning. Fig. 2 illustrates that the correlation with human scores fluctuates significantly across layers, suggesting potential dataset dependency. This argument would be stronger if tested on a broader and more diverse set of benchmarks. The current evaluation, limited to three datasets, does not convincingly establish generalizability.

A2:

• In Section 6, we conduct experiments across diverse scenarios, including sentiment understanding, data selection, and large model knowledge boundary recognition (without any additional training), and found that LAGER effectively generalizes to these scenarios while outperforming the baselines.

• Regarding the fluctuations in layer-wise correlations shown in Figure 2, the datasets (FLASK, BiGGen Bench, HelpSteer) indeed differ significantly in their design and objectives, yet we observe similar fluctuation trends across them. This suggests that while dataset-specific characteristics may have some influence, the inherent properties and representational differences of the model's intermediate layers exert a more dominant effect on evaluation performance.

• Our experiments cover multiple widely used benchmark datasets spanning diverse domains and evaluation criteria [1–6,10]. Broader validation could enhance generalizability, but our current results already demonstrate strong performance.

Q3: Increasingly, modern evaluation benchmarks are label-scarce, more complex, or dynamically generated without human-annotated scores. It is unclear whether the proposed method is applicable or effective in such settings where scalar labels are unavailable or ill-defined.

A3：

• In this paper, we focus on the point-wise evaluation setting. Our method requires only a small amount of annotated data, demonstrates strong transferability, and incurs an acceptable cost for obtaining a limited set of high-quality annotations. While we appreciate your suggestion, it is beyond the scope of this paper, and we will address this issue in future work.

• In fact, this relates closely to recent studies [7][8][9], which have leveraged LLMs to automatically generate scoring rubrics and annotations, effectively addressing challenges related to scarce, unavailable, or ill-defined scalar labels. Meanwhile, our method focuses on enhancing the generalization and efficiency of structured scalar scoring tasks, ensuring stability and reliability. These directions are complementary: the former mitigates label scarcity, while the latter strengthens the robustness of evaluation methodology.

Q4: The work focuses exclusively on evaluations that assign discrete integer scores within a predefined range. It does not explore more general evaluation scenarios, such as those involving natural language feedback or descriptive labels. It remains uncertain whether the proposed method would be robust and adaptable in these more open-ended evaluation settings.

A4:

• Unstructured scoring performs worse than structured scalar scoring. In our early experiments, we also explored unstructured scoring methods (i.e., not confined to predefined discrete integer ranges), but found that the model's performance degraded significantly compared to structured scalar scoring.

• Point-wise evaluation suits open-ended tasks. Previous studies [1][10][11] have shown that point-wise scoring is widely adopted in LLM-as-a-judge research, being well-suited for open-ended tasks without unified standards and demonstrating strong generality and scalability. Notably, models can first perform explicit reasoning before providing a score, thereby enhancing the transparency and rationality of the evaluation.

• This work leverages the core idea of utilizing intermediate layer representations and lightweight weight training, aiming to capture a better mapping relationship between the model's internal states and the evaluation criteria. While the current focus is on structured scalar scoring to establish a solid foundation, this core idea is not confined to integer scoring tasks. This lays the groundwork for future extensions to handle more complex and open-ended feedback forms (e.g., natural language descriptions).

Q5: As noted in the weaknesses, have you explored the method’s performance in more general or open-ended evaluation settings beyond predefined integer scores?

A5:

• We have also explored more general evaluation settings, as presented in Section 6, including sentiment understanding, data selection, and large model knowledge boundary recognition (all without any additional training). These tasks typically involve greater task variability and weaker supervision signals, yet our method consistently demonstrates strong robustness in such scenarios. While we have not yet extended the method to fully open-ended evaluation formats, its strong generalization across diverse tasks suggests promising potential for application in more open-ended evaluation settings, which we plan to pursue in future work.

Q6: Do you have additional experimental results on recent benchmarks that are very different from the 1,000 samples of the HelpSteer dataset used for training?

A6:

• We emphasize that our experimental benchmarks cover diverse evaluation scenarios and exhibit significant heterogeneity in task design, evaluation dimensions, and target scenarios, rather than being highly similar. Specifically, the FLASK dataset includes 12 fine-grained scoring dimensions, covering tasks from multiple different domains (e.g., Culture, Health, History); BiGGen Bench encompasses 77 tasks (e.g., Semantic Constraints and Code Revision) that evaluate various capabilities such as reasoning and instruction following; HelpSteer contains data annotated with multiple evaluation criteria, including helpfulness, correctness, coherence, complexity, and verbosity. And the results on downstream tasks (as presented in Section 6) are also sufficient to comprehensively evaluate the generalization ability of our method.

---

Thank you again for your time and effort in the review process. We will carefully revise the manuscript based on your suggestions and hope that our responses adequately address your concerns. We remain open to any further suggestions or requests for clarification you may have.

Reference

[1] Prometheus: Inducing Fine-grained Evaluation Capability in Language Models (ICLR 2024)

[2] Prometheus 2: An Open Source Language Model Specialized in Evaluating Other Language Models (EMNLP 2024)

[3] Aligning to Thousands of Preferences via System Message Generalization (NeurIPS 2024)

[4] Evaluating the Evaluator: Measuring LLMs' Adherence to Task Evaluation Instructions (AAAI 2024)

[5] Evaluating Scoring Bias in LLM-as-a-Judge

[6] RewardAnything: Generalizable Principle-Following Reward Models

[7] G-EVAL: NLG Evaluation using GPT-4 with Better Human Alignment (ACL 2023)

[8] MCRanker: Generating Diverse Criteria On-the-Fly to Improve Point-wise LLM Rankers (WSDM 2025)

[9] Rationale Behind Essay Scores: Enhancing S-LLM’s Multi-Trait Essay Scoring with Rationale Generated by LLMs (NAACL 2025)

[10] FLASK: Fine-grained Language Model Evaluation based on Alignment Skill Sets (ICLR 2024 Spotlight)

[11] CODEJUDGE:Evaluating Code Generation with Large Language Models (EMNLP 2024)

Dear Reviewer 8jsL,

We have noted that you have re-evaluated and updated the ratings for our paper regarding quality, clarity, significance, and originality. The adjustment of these scores may reflect that there are still some issues in the paper that require further clarification. As the discussion period is coming to a close, although you mentioned in your initial feedback that we have partially addressed your concerns, we would still like to take this opportunity to engage in in-depth discussion and provide comprehensive responses to any remaining questions you may have. Your further feedback will be of significant value to the refinement and improvement of our paper.

We look forward to your response or acknowledgment once you have read our message, as your support is very important to us.

Best regards,

The Authors

We are truly grateful for your insightful feedback on our manuscript. Our key responses are summarized below:

Q1: In spite of the discussed in Section 3.4, the approach is mainly restricted to open-sourced models with accessible intermediate layers for idealistic performance, so its applicability is still limited.

A1:

• We acknowledge that our method relies on access to intermediate layer logits. However, closed-source models could provide access to internal representations (e.g., via enterprise APIs or special access programs), our method could be readily applied without modification, and would perform similarly.

Q2: The evaluation is limited to point-wise benchmarks, while the pairwise/listwise evaluation are not involved.

A2:

• In our study, we focus on point-wise evaluation, because pairwise and listwise methods have high computational complexity and limited scalability, which restricts their use in large-scale data synthesis and filtering. In contrast, point-wise evaluation is computationally simpler, adaptable to various standards, and thus better suited for practical applications.

• Following your suggestion, we evaluated our method on the RewardBench dataset, which is designed for pairwise comparisons. The results in the table indicate that although our approach is primarily designed to enhance performance in point-wise evaluation, it also leads to certain improvements in pair-wise judgment. We will include these experiments and provide further analysis in the next version of the paper.

Q3: Apart from the weaknesses, I want to see a bit more insights in why reasoning leads to weaker results for LAGER (with supports from concrete examples if possible).

A3:

• Through experimental analysis of the model’s internal attention scores, we found that the weaker performance of LAGER under the reasoning setting is closely related to the drift in attention distribution. This is consistent with the findings of previous work[1][2][3]: As the reasoning steps progress, the model’s focus on the original input and the text to be evaluated gradually diminishes, shifting more toward its self-generated reasoning trajectory. This attention shift weakens the model’s direct perception of the source text, causing it to rely on its own reasoning chain when information is insufficient, thereby exhibiting overconfidence. As a result, the model tends to exaggerate supporting evidence while overlooking potential errors or shortcomings, leading to inflated scores. Below, we present real evaluation examples (assessed by the llama3.1-8b-instruct model)：

---

A Real Evaluation Example(assessed by the LLaMA3.1-8B-Instruct model)：

• Instruction: Solve this equation. Input: x^3 - 4x^2 + 6x - 24 = 0

• Response: Sure, I can help you with that. The equation x^3 - 4x^2 + 6x - 24 = 0 can be solved using the following steps: Factor the equation. Set each factor equal to 0 and solve for x. The factors of the equation are (x - 4)(x - 3)(x + 2) = 0. Setting each factor equal to 0 and solving for x, we get the following solutions: x = 4, x = 3, x = -2. Therefore, the solutions to the equation x^3 - 4x^2 + 6x - 24 = 0 are x = 4, x = 3, and x = -2.

• Evaluation Feedback: The response provides a clear step-by-step approach and attempts to factor the cubic equation. However, the factorization presented, (x - 4)(x - 3)(x + 2), is incorrect as it does not satisfy the original equation upon substitution. This is a significant flaw since correct factorization is critical for solving the equation properly. Despite this, the explanation is well-structured and easy to follow, which demonstrates solid reasoning skills. The response presents the solutions clearly and in an organized manner. Based on the rubric, considering the clarity and structure of the explanation alongside the attempt at solving, I would rate this response as highly accurate and logical. Score: 5

---

Thank you again for your time and effort in the review process. We will carefully revise the manuscript based on your suggestions and hope that our responses adequately address your concerns. We remain open to any further suggestions or requests for clarification you may have.

---

Reference：

[1] When Backdoors Speak: Understanding LLM Backdoor Attacks Through Model-Generated Explanations (ACL 2025)

[2] Mechanistically Demystifying Extreme-Token Phenomena in LLMs (NeurIPS 2024)

[3] Measuring and Controlling Instruction (In)Stability in Language Model Dialogs (COLM)

Dear Reviewer Yh62,

We hope this message finds you well. As the discussion period is nearing its end with less than two days remaining, we would like to confirm whether we have adequately addressed all your concerns. If these clarifications help to resolve your doubts and improve your understanding of our work, we would be very grateful if you could consider giving our work a higher evaluation and support.

If you have any further comments or feedback, please feel free to let us know. Your insights are very important to us, and we are more than willing to address any remaining issues to further improve our work.

Best regards,

The Authors

We appreciate the reviewers’ time and effort in evaluating our work. Below, we present our main responses to the key concerns raised:

Q1: The proposed method relies on accessing the internal representations of the LLM; thus, this limits applicability to open-source models. Closed-source or API-only LLMs cannot benefit.　The proposed method relies on accessing the internal representations of the LLM; thus, this limits applicability to open-source models. Closed-source or API-only LLMs cannot benefit.

A1:

• We acknowledge that our method relies on access to intermediate layer logits. However, closed-source models could provide access to internal representations (e.g., via enterprise APIs or special access programs), our method could be readily applied without modification, and perform similarly.

Q2: The supervised weight-tuning step still requires human-annotated validation data.

A2:

• This is a common requirement for most supervised training methods and should not be viewed as a weakness, especially for approaches focused on human alignment, as evidenced by the paper [1] cited by Reviewer HC1G.

• Minimal human-annotated data and lightweight training parameters. Our method trains only dozens of lightweight parameters using a few hundred annotated data points (about 5% of the data required by traditional methods), avoiding the need to fine-tune billions of backbone parameters.

• Strong generalization ability. Furthermore, as demonstrated in Section 6, our method demonstrates strong generalization ability, effectively transferring to other domains while requiring substantially fewer annotations and incurring no significant additional costs, thus offering clear advantages over traditional approaches.

Q3: The experimental results are not thorough, missing important baselines, for example RewardBench and DPO rewarding. Also, LLM-as-a-Judge heavily rely on the prompt templates, so it would be better to test on different popular prompts which you can find in recent papers such as self-rewarding LLM.

A3:

• It is our view that RewardBench is fundamentally an evaluation benchmark intended to measure the performance of reward models, rather than a specific evaluation method. Moreover, we are unable to find a directly comparable baseline method reported in the RewardBench study.

• Similarly, DPO rewards are typically obtained by fine-tuning the base language model itself, which differs from our approach that only introduces lightweight aggregation parameters without altering the backbone model. So we believe that the DPO reward cannot be used as a directly comparable baseline in our setting. If our understanding is incorrect, we would appreciate clarification on the role of RewardBench and DPO rewarding in the comparisons.

• Our primary prompt is adapted from templates widely adopted in prior evaluation studies[2][3][4], which have demonstrated strong and reliable baseline performance. While prompt design can influence performance, it is not the primary focus of our work. Nevertheless, we also tested two popular prompt variants (T1 refers to the template used in Self-Rewarding LLM, and T2 refers to the template used in G-Eval), and the conclusions were consistent with those obtained using our primary prompt, further confirming the robustness of our approach.

Q4: Will LAGER become better if scale more training data and params?

A4:

• Increasing the model parameters with an MLP and enlarging the dataset did not result in significantly better performance. In preliminary experiments, replacing simple weighting with an MLP led to unstable training, likely due to increased complexity and limited data. Our current L+1 parameter approach is simpler and achieves great performance. We also observed performance saturates as data scales, with limited gains from additional data. In future work, we will explore larger datasets and more complex aggregations to evaluate scalability.

---

Thank you again for your time and effort in the review process. We will carefully revise the manuscript based on your suggestions and hope that our responses adequately address your concerns. We remain open to any further suggestions or requests for clarification you may have.

Reference

[1] HD-Eval: Aligning Large Language Model Evaluators Through Hierarchical Criteria Decomposition (ACL 2024)

[2] G-Eval: NLG Evaluation using Gpt-4 with Better Human Alignment (EMNLP 2023)

[3] Prometheus: Inducing Fine-grained Evaluation Capability in Language Models (ICLR 2024)

[4] Prometheus 2: An Open Source Language Model Specialized in Evaluating Other Language Models (EMNLP 2024)

Dear Reviewer DQpf,

We hope this message finds you well. As the discussion period is nearing its end with less than two days remaining, we would like to confirm whether we have adequately addressed all your concerns. If these clarifications help to resolve your doubts and improve your understanding of our work, we would be very grateful if you could consider giving our work a higher evaluation and support.

If you have any further comments or feedback, please feel free to let us know. Your insights are very important to us, and we are more than willing to address any remaining issues to further improve our work.

Best regards,

The Authors

Thank you for your thoughtful feedback and constructive suggestions. Our key responses are summarized below:

Q1: The paper does not compare LAGER against several relevant evaluators that also target human alignment, such as G-Eval (reasoning-based) or HD-Eval (which also leverages multi-layer information). These comparisons are important to contextualise LAGER’s gains.

A1:

• G-Eval and E-Score: Essentially the Same. We would like to clarify that G-Eval is essentially the same as the E-score baseline included in our experiments. Moreover, the G-Eval(reasoning-based) evaluation results you mentioned are also incorporated within our experimental comparisons (see the performance of E-score under the reasoning setting on each dataset in Tables 1 and 6).

• LAGER outperforms HD-Eval. Regarding the HD-Eval baseline, we would like to clarify: HD-Eval enhances the alignment between LLM evaluators and human judgments by decomposing evaluation into fine-grained criteria ("layers" refer to fine-grained decomposition stages, not model layers) and employing human-guided aggregation. This is fundamentally different from our approach, which utilizes intermediate layer representations of the model. We compared HD-Eval (chat task setting) with LAGER in evaluation performance. As shown in the table below, HD-Eval's performance is lower than our method, primarily because it requires manually customizing fine-grained criteria for different tasks, making it difficult to generalize effectively across the diverse evaluation datasets used in our paper.

Q2: GPT-4o-mini achieved very strong results, raising questions about the impact of model scale. However, the current experiments focus only on cross-family comparisons. It would strengthen the work to test LAGER across different sizes within the same family (e.g., Qwen 2.5B, 7B, 14B).

A2:

• LAGER brings improvements across model scales. In fact, we have already conducted such an analysis in Section 5.3. This section includes extensive experiments across multiple sizes within the same Qwen2.5 family, ranging from 0.5B to 72B. These results demonstrate that LAGER consistently improves evaluation performance across a wide range of model scales. Moreover, larger models generally tend to achieve better performance.

Q3: While LAGER with tuning performs best overall, the tuning appears to succeed on some backbones (e.g., Qwen) but fails or regresses on others (e.g., Mistral, LLaMA). Yet the discussion (p. 6) focuses only on LAGER w/ tuning (without comparing to the ones w/ tuning), making the discussion a bit insufficient.

A3：

• We would like to clarify that across all evaluation datasets, LAGER(w. tuning) outperforms E-score by an average of over 6% on both backbones. The number of instances in which LAGER with tuning underperforms E-Score is 3 out of 36. It is worth noting that in these cases, the performance of LAGER with tuning is comparable to that of E-Score, with differences of less than 0.5%.

• We would like to emphasize that our tuning-based method demonstrates strong transferability, as validated by the multi-task experiments in Section 6. Even without tuning, our method outperforms E-Score in 5 out of 6 backbone models, with an average improvement of 4.5%. In the few cases where it underperforms, the performance gap is within 3%. Moreover, it consistently outperforms V-Score. Therefore, we primarily recommend the LAGER with tuning approach and conduct subsequent analyses based on it.

Q4: Compared to E-Score, LAGER without tuning performs worse (or on-par) on several backbones (e.g., InternLM, Qwen), suggesting that the observed gains largely stem from fine-tuning. This makes the comparisons not quite fair, comparing a tuned model to unturned baselines.

A4:

• Our method does not involve fine-tuning the backbone language model. LAGER only introduces L+1 lightweight parameters (where L is the number of transformer layers) to aggregate multi-layer logits for more accurate scoring. This approach is fundamentally different from traditional fine-tuning: we do not update any of the original backbone model's parameters.

• Except for InternLM, our method outperforms E-Score by an average of 4.5 points across all other model backbones, even without introducing any additional weight parameters. As a result, LAGER introduces minimal computational overhead and is highly scalable. The performance gains of LAGER stem not from large-scale fine-tuning, but from a more efficient and structured utilization of the model’s internal representations.

---

Thank you again for your time and effort in the review process. We will carefully revise the manuscript based on your suggestions and hope that our responses adequately address your concerns. We remain open to any further suggestions or requests for clarification you may have.