acceptodds
Under review as a conference paper at ICLR 2027

Introspecting Alignment Shifts Beyond Behaviors Implanted Through Fine-Tuning

Abstract

Fine-tuning enables a source model to acquire desired capabilities and behaviors in a target domain while retaining much of its general-purpose competence. However, this adaptation process can also degrade alignment properties that were present in the source model. Recent work has shown that large language models can be trained using LoRA-based modules known as introspection adapters (IAs) to describe behavioral changes induced by fine-tuning. However, existing studies primarily consider settings in which the model is fine-tuned on datasets explicitly designed to implant a specific behavior and is then asked to explain the implanted behavior. This differs from practical deployment scenarios, where the misalignment is not necessarily deliberately implanted, but arises only as a side effect. To bridge this gap, we formulate a novel problem setting, in which the target of introspection is not necessarily a behavior explicitly implanted through fine-tuning, but rather alignment shifts that may emerge as unintended side effects, and we construct a dataset for this setting. Furthermore, to enhance sensitivity to internal model changes, we propose the Delta-Aware Introspection Adapter (DAIA), a novel mechanism designed to explicitly process both base-model activations and activation differences induced by fine-tuning. Our empirical evaluation shows that introspection learning generalizes to unseen fine-tuned models and safety categories, and that DAIA generally outperforms existing IAs.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Introspection, Interpretability, Alignment, Safety, Fine-tuning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_5T1cOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates whether fine-tuned language models can report unintended safety-related behavioral changes relative to their base models. The authors introduce a practical introspection setting in which the behavior being explained is not deliberately implanted during fine-tuning but instead emerges as a side effect. The paper further proposes the Delta-Aware Introspection Adapter (DAIA), which separately processes base-model activations and fine-tuning-induced activation differences to generate self-reports. Experiments on safety-shift classification and free-form reporting show that introspection learning generalizes to unseen models and safety categories, while DAIA generally outperforms a standard LoRA-based introspection adapter.

Critical strengths

Compared with the existing work, the questions raised by the author are closer to the real model deployment scenarios and have research significance. The design concept of DAIA is straightforward, the motivation is clear, and the mechanism analysis clearly demonstrates how DAIA operates. The experimental setup is comprehensive, demonstrating that DAIA has good generalization ability. The ablation experiments are relatively sufficient, verifying the necessity of the two branches. The mechanism analysis details how DAIA forms the judgment of safe changes.

Critical weaknesses
  1. The presentation of the proposed method should be improved. Although DAIA is mathematically defined, the paper lacks an overall framework diagram or pseudocode illustrating the complete workflow, including dataset construction, adapter training, and inference.
  2. The paper motivates introspection as a potentially inexpensive first-pass safety diagnostic, but it does not provide an end-to-end cost analysis of DAIA. Constructing the training data requires evaluating 213 models on 1,523 safety prompts, performing LLM-based scoring and reference-report generation, and subsequently conducting SFT and DPO. The authors should compare both the one-time training cost and the per-model inference cost of DAIA with those of direct safety evaluation, and clarify under what deployment scale DAIA becomes cost-effective.
  3. The performance improvement of DAIA is not stable. Some indicators are lower than the baselines, especially in the OOD experiments, and the superiority of the algorithm is not demonstrated.
  4. The baseline comparison is insufficient. Although the authors cite two recent introspection methods in the Introduction, they do not clearly identify which method the “LoRA” baseline corresponds to or provide separate comparisons with these methods.
  5. The same LLM is responsible for generating the reference report and evaluating the generated report, which makes the credibility of the report lower. The author did not prove that the internal audit report output by DAIA truly reflects the safety changes, or if it merely outputs text that is more in line with the LLM's preference.
Questions for authors
  1. Which prior method does the LoRA-based IA baseline correspond to? The authors cite the introspection methods of Shenoy et al. and Goel et al., but it is unclear whether the reported baseline faithfully reproduces either method or combines elements from both. The authors should clarify its provenance and explain why the two methods are not evaluated separately.
  2. The authors should provide a detailed analysis of the cost of constructing the dataset and training DAIA. Collecting 213 models, evaluating them on 1,523 safety prompts, and using LLM-based scoring and reference-report generation appear computationally expensive.
  3. How reliable are the reference reports and their associated evaluation scores? Were the ten LLM-generated reference reports manually validated? Did the authors assess agreement across different judge models? Moreover, can DAIA’s self-reports be corroborated using independent behavioral evidence to ensure that they reflect genuine safety changes rather than the preferences of a particular judge model?
5 comments
AuthorsOpenReview ↗

Response to reviewer 5T1c

We thank reviewer 5T1c for taking the time to provide insightful feedback. Below are our responses to each of the concerns.

W1 (The presentation of the proposed method)

We agree that the original manuscript lacked a visual overview of the proposed framework. To address this, we have added Figure 1 to the revised manuscript, illustrating the complete workflow from collecting fine-tuned models and constructing supervision data to training DAIA and performing inference.

W2&Q2 (Cost analysis)

We conducted an end-to-end cost analysis for the Qwen3-14B free-form reporting task, including dataset construction, reference report generation, SFT, and DPO.

Cost ComponentEstimated Cost
Direct benchmark evaluation (per model-category pair)$0.0642
DAIA construction and training (one-time)$197
DAIA inference (per model-category pair)$0.0037

All computational costs were estimated using an NVIDIA A100-SXM4 (80 GB) GPU at RunPod's rate of $1.79/hour.

DAIA reduces per-model evaluation costs by approximately 94.2%, requiring only 5.8% of the cost of direct benchmark evaluation at inference time.

To account for the one-time cost, we derive the break-even point:

N∗=FK(b−d)N^* = \frac{F}{K(b-d)}

where:

  • FF: One-time dataset construction and training cost
  • KK: Number of evaluated safety categories
  • NN: Number of evaluated fine-tuned models
  • bb: Direct evaluation cost per model-category pair
  • dd: DAIA inference cost per model-category pair

Assuming a fixed one-time cost and constant per-category costs, the break-even point decreases inversely with the number of evaluated categories.

For example, with 30 safety categories:

N∗=19730(0.0642−0.0037)≈109N^* = \frac{197}{30(0.0642-0.0037)} \approx 109

Thus, DAIA becomes more cost-effective after evaluating approximately 109 additional fine-tuned models under these assumptions.

Importantly, large-scale collections of over 1,000 fine-tuned models have been observed on model hubs such as Hugging Face and in real-world production environments[1][2].

These findings suggest that DAIA could offer substantial cost savings in large-scale, repeated safety auditing by amortizing its one-time training cost across many fine-tuned models.

We have added this analysis to Appendix D.6 of the revised manuscript.

[1] Liu, Hongyi, et al. "LoRATK: LoRA Once, Backdoor Everywhere in the Share-and-Play Ecosystem.", EMNLP (Findings), 2025.

[2] Jaiswal, Shashwat, et al. "Serving Heterogeneous LoRA Adapters in Distributed LLM Inference Systems.", arXiv, 2025.

W3 (The performance improvement of DAIA)

We acknowledge that DAIA does not consistently outperform the LoRA-based IA across all classification metrics. However, DAIA achieves higher accuracy in 11 out of 12 cases, although improvements in macro-F1 are less consistent. In particular, DAIA achieves a higher average score in OOD-C, where both fine-tuned models and safety categories are unseen during training. More importantly, DAIA consistently achieves higher mean scores than the LoRA-based IA in the free-form reporting task across all cases.

One possible explanation for this difference is the granularity of the output task. Classification requires predicting only a coarse safety-shift label, potentially limiting the extent to which detailed information about fine-tuning-induced differences can be reflected in the output. In contrast, free-form reporting requires generating detailed behavioral descriptions over multiple tokens, allowing DAIA's explicit processing of fine-tuning-induced activation differences to influence a broader range of generated content.

W4&Q1 (The baseline comparison)

Both prior methods correspond to the "LoRA" baseline in our setting. Both studies adopt a standard LoRA architecture for their introspection adapters, and we therefore denote the baseline representing both methods as "LoRA". To clarify this point, we have added an explanation to the Baselines paragraph in Section 4.1 of the revised manuscript.

AuthorsOpenReview ↗

Response to reviewer 5T1c

W5&Q3 (LLM judge reliability)

We agree that the reviewer has raised an important concern. To address this issue, we re-evaluated the free-form reporting task on Qwen3-14B using Gemini 3.8 Flash and Claude Haiku 5.5, following the same evaluation protocol as in the original experiments. The results are presented in Table 11 and Figure 11 of Appendix D.5. For convenience, we reproduce Table 11 below.

・OOD-A

MethodGPTGeminiClaude
Raw1.73 ± 0.021.26 ± 0.011.43 ± 0.02
LoRA3.41 ± 0.123.56 ± 0.093.01 ± 0.08
DAIA3.57 ± 0.103.70 ± 0.093.15 ± 0.07

・OOD-B

MethodGPTGeminiClaude
Raw1.53 ± 0.021.26 ± 0.001.38 ± 0.01
LoRA2.56 ± 0.082.24 ± 0.052.12 ± 0.04
DAIA2.74 ± 0.012.47 ± 0.022.22 ± 0.02

・OOD-C

MethodGPTGeminiClaude
Raw1.41 ± 0.041.19 ± 0.001.26 ± 0.02
LoRA2.49 ± 0.102.18 ± 0.082.06 ± 0.05
DAIA2.75 ± 0.012.47 ± 0.082.21 ± 0.06

The results show that the ranking Raw < LoRA < DAIA is preserved across all OOD settings and all three LLM judges, consistently supporting our main findings. Furthermore, Figure 11 presents the pairwise correlations of evaluation scores among the three LLM judges. Both Pearson and Spearman correlation coefficients are consistently high, ranging approximately from 0.8 to 0.9 across all settings.

These results suggest that our main findings are robust to the choice of LLM judge and are not solely attributable to the preferences of the model used to generate the reference reports. In particular, the improvements of introspection-trained IAs over raw fine-tuned models, as well as the performance advantage of DAIA over the conventional LoRA-based IA, are consistently observed across different LLM judges.

AuthorsOpenReview ↗

Response to reviewer 5T1c

We hope that the above explanations address your concerns. If any concerns remain, please let us know.

Otherwise, we would greatly appreciate it if you could reconsider your score in light of our responses and revisions. Thank you for your time and consideration.

Reviewer_Gr3FOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors extend IA (introspection adapters), which are trained to describe / explain finetuned model behaviour using natural language. Previous work on IAs has relied on explicit datasets that are used in finetuning the model itself to also train the IA. The authors propose a technique that sidesteps this requirement by conducting their own dataset that recovers the finetuning differences by evaluating both the base and finetuned model on a benchmark of interest, and then using the evaluation results to train the introspection adapter. The authors then leverage this technique to explain safety relevant behaviour changes in models, as opposed to recovering the finetuning objectives.

Critical strengths

The paper provides a natural extension to an emerging explanation method, motivates it sufficiently, and provides enough context about the problem and its relevance. The solution proposed is starightforward and lucidly explained, and the evaluation is complete.

Critical weaknesses

I fail to understand why the analysis is conducted at the level of safety "categories". How are these constructed (human review? which humans? what protocols?) Why are the DAIA results analysed at the category level? The model is trained at with a dataset of individual prompts, I don't see why the evaluation is being done this way.

Similarly, the motivation behind the free form evaluation was unclear to me. Why synthesize all existing data using a "verbalizer" into a single report? I would have liked to instead see a baseline where DAIA is compared with introspection adapters (such as Goel et al's DIT) which are explicitly trained against a misalignment objective in addition to IAs trained against innocuous finetuning goals. It would not be necessary to beat this "oracle" baseline (as the DAIA does not optimise for this objective, while the DiT does), but it would be a useful comparison. This is missing from the paper, as far as I understand.

Questions for authors

Line 074: what benchmark are the safety examples coming from? Line 116: what is j? (seems to be labelled QA pairs about i, but this does not seem to detailed)

why restrict to misalignment? it seems that DAIA is a generic technique, where detecting "inadvertent" misalignment is just one possible use case.

Table 1: how are prompts mapped into categories?

lines 301-304: why does experiment protocol first discretise the scores and then aggregate them? Why convert the numbers into categories, and then report classification metrics? why not report the residuals on the numbers directly?

While these are some limited questions, and more fundamental weaknesses are listed above, I still think the work is unique and creative, and therefore lean towards acceptance. However, it would be beneficial if the authors could fix the issues about evaluation (why use categories? why discretize?) and show performance wrt an "oracle" baseline IA.

3 comments
AuthorsOpenReview ↗

Response to reviewer Gr3F

Q1 (What benchmarks?)

As already described at the beginning of Section 4.1, the 1,523 safety-related prompts were collected from four benchmarks: StrongREJECT, HarmBench, HEx-PHI, and AdvBench. For clarity, we have also added this information to Line 074 in the Introduction.

Q2 (What is jj?)

The index jj denotes the introspection question–answer pairs associated with the jj-th fine-tuned model. We acknowledge that the definitions of NN and jj were missing. We have clarified these notations immediately after Eq. (2) in the revised manuscript.

Q3 (Why misalignment alone?)

We agree with the reviewer that our framework and DAIA are, in principle, applicable to introspecting a broader range of behavioral changes induced by fine-tuning, beyond misalignment. However, we focus on alignment-relevant changes because we consider them one of the most practically important use cases, particularly given the growing concern about unintended safety degradation caused by fine-tuning (Qi et al., 2024; Betley et al., 2025). We therefore chose to focus on this setting, building on the motivations of existing research. Evaluating the effectiveness of our approach in explaining more general fine-tuning-induced changes is beyond the scope of this paper and remains an important direction for future work. We have clarified this point in Section 7 (Limitations and Future Work) of the revised manuscript.

Q4 (How are prompts mapped into categories?)

We agree that the original manuscript did not sufficiently explain the categorization procedure. Specifically, we used an LLM (GPT-5.6 Luna) to assign each benchmark prompt to one of the 13 safety categories. We have now clarified this procedure in Section 4.1 and provided the complete categorization prompt and additional details in Appendix B.3.

Q5 (Why discretize the scores rather than directly predicting numerical differences?)

We agree that directly training the model to predict numerical alignment shifts is possible. However, when training an LLM to perform regression over continuous numerical values, the standard token-level cross-entropy loss treats numerical tokens as categorical symbols and does not explicitly account for numerical proximity or distance between them [1]. Therefore, appropriate regression-oriented training may require additional mechanisms, such as a regression head.

However, our ultimate goal is to enable LLMs to generate open-ended, natural-language reports of fine-tuning-induced behavioral changes, for which we adopt standard autoregressive language-model training. Given the aforementioned limitations of numerical prediction under token-level cross-entropy, we chose not to pursue direct numerical regression. Instead, we discretized the continuous score differences into five semantically interpretable labels and trained the model to generate these labels using the standard language-modeling objective.

[1] Zausinger, Jonas, et al. "Regress, Don't Guess--A Regression-like Loss on Number Tokens for Language Models.", ICML, 2025.

AuthorsOpenReview ↗

Response to reviewer Gr3F

We hope that the above explanations address your concerns. If any concerns remain, please let us know. Otherwise, we would greatly appreciate it if you could reconsider your score in light of our responses and revisions. Thank you for your time and consideration.

As the discussion period is approaching the deadline, we would like to kindly follow up and ask whether you might be able to share any further comments or considerations regarding our paper.

Thank you again for all your time and effort in reviewing our work.

Reviewer_WQfTOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

Recent work has shown that language models can be trained using introspection adapters to describe behavioral changes induced by fine-tuning. This paper extends that approach to unintended side effects of fine-tuning, particularly changes in safety alignment. The authors introduce a new training dataset and propose a novel introspection adapter architecture that explicitly processes fine-tuning-induced activation differences. Through experiments involving 213 fine-tuned models and four safety benchmarks, the authors demonstrate that their approach can identify and describe alignment shifts, including those in previously unseen models and safety categories.

Critical strengths
  1. Clear writing. In an age of AI-generated writing, I particularly appreciate papers that are well written and have a strong, distinctive voice. I found this paper clear and engaging throughout, with a comprehensive and accessible discussion of related work, methodology and results.

  2. Insightful ablation study. The ablation study in Section 5.2 is focused and informative, systematically examining the contribution of the different architectural choices. In particular, the comparisons between using base activations, fine-tuning-induced differences, and their combination provide useful evidence for the benefits of processing these signals separately.

  3. Breadth of experimental evaluation. The breadth of fine-tuned models and tasks considered is another strength. Evaluating 213 fine-tuned models across four safety benchmarks and 13 safety categories provides a reasonably diverse experimental setting. I also appreciate the evaluation across multiple OOD settings, which helps assess whether the learned introspection capabilities generalize beyond the specific models and safety categories encountered during training.

  4. Novel problem formulation and methodological contribution. The extension of introspection learning from deliberately implanted behaviors to unintended alignment shifts is a meaningful and novel contribution. This moves the problem closer to realistic fine-tuning scenarios, where developers may not know in advance which behaviors have changed. I also appreciate the architectural intuition behind DAIA, which explicitly separates base-model activations from fine-tuning-induced differences rather than treating them as a single combined representation. Together, these contributions offer an interesting new direction for studying and diagnosing alignment changes.

Critical weaknesses
  1. Inconsistent improvements relative to added complexity. Given the additional architectural complexity introduced by DAIA, the performance gains over the standard LoRA-based introspection adapter are less consistent than one might hope. This is particularly apparent in OOD-A (unseen models, seen categories), where DAIA essentially ties with LoRA on average (52.1 vs. 51.9). The improvements are more substantial in OOD-C, but DAIA does not consistently outperform LoRA across all metrics, particularly macro-F1. This raises questions about whether the added architectural complexity is sufficiently justified by the empirical gains.

  2. Limited diversity of base models. Although the study considers an impressive 213 fine-tuned variants, these are all derived from just two base models, Qwen3-14B and Gemma3-12B-it. Consequently, the diversity of underlying architectures, model sizes, and pretraining regimes is relatively limited. Moreover, the introspection adapters are trained separately for each base-model family, so the experiments do not establish cross-family transfer. Evaluating additional model families and scales would strengthen the generalizability of the conclusions.

  3. Potential confound from modifying the model's forward computation. DAIA does not merely observe the fine-tuned model's internal activations; it actively modifies the forward computation by replacing the original fine-tuning contribution with a learned transformation of the base activations and fine-tuning-induced differences. This raises a conceptual concern about whether the resulting introspection reflects the behavior of the original fine-tuned model or that of the modified model equipped with DAIA. Although the authors investigate adding the original fine-tuning contribution back through a residual connection in Section 5.2, this distinction between observing and modifying the model being evaluated deserves further discussion.

  4. Lack of human evaluation. The evaluation relies heavily on LLM-based judgments, both for constructing safety-shift labels and for assessing the quality of free-form introspection reports. While this is a reasonable approach at the scale of the study, the absence of human validation makes it harder to assess how reliably these judgments reflect actual behavioral changes. Even a modest human evaluation of the safety-shift labels and the faithfulness of generated reports would increase confidence in the findings.

Questions for authors

Can you address the weaknesses I've raised during rebuttal, whether by experiments or clarifications?

5 comments
AuthorsOpenReview ↗

Response to reviewer WQfT

We thank reviewer WQfT for taking the time to provide insightful feedback. Below are our responses to each of the concerns.

W1 (Inconsistent improvements relative to added complexity.)

We acknowledge that DAIA does not consistently outperform the LoRA-based IA across all classification metrics. However, DAIA achieves higher accuracy in 11 out of 12 cases, although improvements in macro-F1 are less consistent. In particular, DAIA achieves a higher average score in OOD-C, where both fine-tuned models and safety categories are unseen during training. More importantly, DAIA consistently achieves higher mean scores than the LoRA-based IA in the free-form reporting task across all cases.

One possible explanation for this difference is the granularity of the output task. Classification requires predicting only a coarse safety-shift label, potentially limiting the extent to which detailed information about fine-tuning-induced differences can be reflected in the output. In contrast, free-form reporting requires generating detailed behavioral descriptions over multiple tokens, allowing DAIA's explicit processing of fine-tuning-induced activation differences to influence a broader range of generated content.

Furthermore, regarding complexity, we conducted a Pareto analysis of training complexity and performance by varying the adapter rank. In addition to rank 16 used in the main experiments, we evaluated ranks 8 and 32, examining the relationship between the number of trainable parameters and macro-F1 scores on the five-class classification task using Qwen3-14B.

The results are shown below. Except for OOD-A, DAIA consistently achieves a more favorable Pareto frontier than LoRA. Notably, DAIA with fewer trainable parameters often outperforms LoRA with more parameters, suggesting greater parameter efficiency. In OOD-A, LoRA performs better at larger parameter budgets, but its performance declines as the number of trainable parameters decreases. In contrast, DAIA exhibits greater robustness to parameter reduction, suggesting a more favorable performance–parameter trade-off in the low-parameter regime.

These results further demonstrate that DAIA's fine-tuning difference-aware design enables more parameter-efficient introspection.

We have added these results in Appendix D.8 of the revised manuscript.

Rank 8Rank 16Rank 32
OOD-ALoRA38.62 ± 7.1942.56 ± 9.3451.03 ± 4.08
DAIA41.49 ± 3.4642.12 ± 2.0142.87 ± 4.32
OOD-BLoRA33.56 ± 2.9234.63 ± 3.1435.75 ± 2.89
DAIA38.18 ± 3.2038.44 ± 0.7740.44 ± 1.65
OOD-CLoRA26.10 ± 7.8226.36 ± 9.8233.38 ± 1.04
DAIA33.13 ± 1.9532.53 ± 1.4233.95 ± 4.29

W2 (Limited diversity of base models.)

We acknowledge that our experiments are restricted to two base-model families with relatively similar model sizes, and that cross-family transfer of a trained DAIA has not been evaluated.

However, we would like to clarify that our primary target setting is generalization across fine-tuned variants derived from the same base model, rather than transferring a single introspection adapter across different base-model families. This setting is motivated by practical deployment scenarios in which numerous fine-tuned adapters are developed and maintained for a common base model. Furthermore, transferring adapters such as DAIA or LoRA-based IAs across base models with different architectures is challenging due to differences in their internal dimensions and structures, and would require further adapter development. We therefore consider this beyond the scope of the present work. Although DAIA is trained separately for Qwen3-14B and Gemma3-12B-it, the same architectural design achieves improvements over the conventional LoRA-based IA across both model families. This provides evidence that the proposed approach is not specific to a single base-model architecture.

Nevertheless, we agree that evaluating DAIA on additional model families and different model scales would strengthen the generalizability of our conclusions. Investigating cross-family transfer is also an important direction for future work. We have added this discussion to Section 7 of the revised manuscript.

AuthorsOpenReview ↗

Response to reviewer WQfT

W3 (Potential confound from modifying the model's forward computation. )

We agree that DAIA does not observe the exact activations of the original fine-tuned model, as its forward computation is influenced by DAIA in preceding layers. However, we emphasize that DAIA can explicitly access the fine-tuning-induced computational difference for the same input at each module. Specifically, for any given module input xx, the base weights produce hbase(x)h_{\text{base}}(x), while the fine-tuned weights produce hbase(x)h_{\text{base}}(x)+Δh(x)\Delta h(x). DAIA directly accesses this difference Δh(x)\Delta h(x), allowing it to capture how fine-tuning changes the module's computation for a fixed input, without requiring the original fine-tuned model's exact activation trajectory.

Furthermore, our activation-patching results in Section 5.1 show that DAIA's introspective judgments depend more strongly on Δh(x)\Delta h(x) than those of the conventional LoRA-based IA, providing empirical support for this interpretation.

W4 (Lack of human evaluation.)

We thank the reviewer for this suggestion. We investigated the extent to which LLM-based judgments align with human evaluations. Specifically, we randomly sampled 150 responses from the 121 fine-tuned Qwen3-14B models evaluated on our benchmark covering 13 safety categories. We then conducted a human evaluation involving three annotators who were not involved in this paper. The evaluation protocol was identical to that used for the LLM judge in the main experiments, where annotators assessed the harmfulness of each prompt–response pair on a five-point scale from 1 to 5.

The agreement between human evaluations and the LLM judge is summarized below.

MetricValue
Pearson correlation (r)0.51
Spearman correlation (ρ)0.52
Agreement within ±1 point72%

First, the Pearson and Spearman correlation coefficients were 0.51 and 0.52, respectively, indicating moderate positive correlations. Furthermore, 72% of the evaluations differed by no more than one point between the human annotators and the LLM judge. These results indicate a moderate, although not strong, agreement between human evaluations and LLM-based judgments.

We would like to emphasize that this imperfect agreement is not a limitation specific to our framework, but rather a broader challenge associated with existing direct benchmark-based evaluations. Developing scalable yet more reliable evaluation protocols using LLM judges remains an important direction for future research. We have added these results and the corresponding discussion to Appendix D.7 of the revised manuscript, highlighted in red.

Furthermore, we sought to strengthen the reliability of our evaluation by examining agreement across different LLM judges. Specifically, we re-evaluated the free-form reporting task on Qwen3-14B using Gemini 3.8 Flash and Claude Haiku 5.5, following the same evaluation protocol as in the original experiments. The results are presented in Table 11 and Figure 11 of Appendix D.5. For convenience, we reproduce Table 11 below.

・OOD-A

MethodGPTGeminiClaude
Raw1.73 ± 0.021.26 ± 0.011.43 ± 0.02
LoRA3.41 ± 0.123.56 ± 0.093.01 ± 0.08
DAIA3.57 ± 0.103.70 ± 0.093.15 ± 0.07

・OOD-B

MethodGPTGeminiClaude
Raw1.53 ± 0.021.26 ± 0.001.38 ± 0.01
LoRA2.56 ± 0.082.24 ± 0.052.12 ± 0.04
DAIA2.74 ± 0.012.47 ± 0.022.22 ± 0.02

・OOD-C

MethodGPTGeminiClaude
Raw1.41 ± 0.041.19 ± 0.001.26 ± 0.02
LoRA2.49 ± 0.102.18 ± 0.082.06 ± 0.05
DAIA2.75 ± 0.012.47 ± 0.082.21 ± 0.06

The results show that the ranking Raw < LoRA < DAIA is preserved across all OOD settings and all three LLM judges, consistently supporting our main findings. Furthermore, Figure 11 presents the pairwise correlations of evaluation scores among the three LLM judges. Both Pearson and Spearman correlation coefficients are consistently high, ranging approximately from 0.8 to 0.9 across all settings.

We believe that the high agreement observed across different LLM judges further strengthens the reliability of our evaluation results.

AuthorsOpenReview ↗

Response to reviewer WQfT

We hope that the above explanations address your concerns. If any concerns remain, please let us know.

Otherwise, we would greatly appreciate it if you could reconsider your score in light of our responses and revisions. Thank you for your time and consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to All Reviewers

Dear reviewers,

Thank you all for your insightful and constructive feedback on our paper.

We believe that we have addressed all of your concerns in our detailed responses below. We would greatly appreciate it if you could revisit our paper in light of these clarifications and revisions and consider updating your scores accordingly. As the discussion period is approaching its deadline, we would also be grateful if you could share any comments or questions to facilitate continued discussion.

Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.