Thank you for your review and for acknowledging the interesting phenomenon we identified.

W1: The cognitive experts seem to be domain-specific. I wonder whether different datasets in the same domain would result in different cognitive experts? For example, will MATH-500 and AIME have the same cognitive experts?

You are correct that the exact set of top experts can vary between datasets, even within the same domain. For instance, comparing the top-5 cognitive experts for AIME24, MATH, and GSM8K, we find both shared and dataset-specific experts. For example, the 182nd expert at layer 39 appear in the top-5 across all three math datasets, suggesting they support general mathematical reasoning. In contrast, other experts may focus on patterns unique to each dataset. We'll add this observation to the paper.

W2&Q1: The reinforcing method seems to only work in math, instead of in any other domains. Whether the proposed method still outperform other steering methods?

• Performance in Other Domains: Applying RICE sometimes degraded performance in non-math dmains (e.g., Physics, Biology in Table 4). This was the result of a specific experimental choice to test generalization. In the original manuscript, we used hyperparameters (steering multiplier = 64) tuned only on the math domain (AIME) and applied them directly to all other domains.

  This setup was designed to test the raw generalization of the math-derived setting. However, different domains may require different intensities of cognitive steering. When we use a simple, domain-specific steering multiplier (selected from a small set of {16, 32, 64}), RICE delivers consistent and significant improvements across all domains.

DeepSeek-R1 with Domain-Tuned Multipliers

Qwen3-235B with Domain-Tuned Multipliers

Our hypothesis is that math problems demand more intensive reasoning, hence the higher optimal multiplier. We will clarify this in the paper to avoid misinterpretation.

• Comparison with Other Steering Methods: We ran a new comparison against a strong baseline, Contrastive Activation Addition (CAA), on AIME 24. RICE not only performs better but also avoids the negative side effects on general instruction-following in Arena that CAA exhibits.

Lower is better for Side Effects; RICE improves general capabilities.

W3: Whether reinforcing cognitive experts could help the model to reason more effectively is in doubt. The experiment in Section 3.2 mainly supports this claim. However, as shown in Table 5, all models using the reinforce trick generate more tokens compared with the vanilla model. This contradicts the claim.

This is a misunderstanding of the purpose of Table 5.

• The experiment in Table 5 (Arena-Hard) is NOT meant to measure reasoning efficiency. Its purpose is to check for negative side effects on general instruction-following capabilities.
• Arena-Hard includes tasks like creative writing, where longer, more detailed answers are often better. The slight increase in token count here is not a sign of inefficiency but reflects the nature of the tasks.
• Our efficiency claims are based on reasoning benchmarks like AIME, where RICE often reduces token count while improving accuracy (see table above).

Reference:

[1] Steering Llama 2 via Contrastive Activation Addition, ACL 2023.

Cognitive experts reflect the model’s reasoning ability, and their steering multipliers modulate the degree of reasoning enhancement. We hypothesize that the optimal multiplier varies across domains because each domain requires a different level of reasoning intensity.

In this paper, we recommend setting the optimal multiplier for each domain: 64 for the math domain; 32 or 16 for the biology, physics, and chemistry domains. Dynamic adaptation of the multiplier based on the difficulty of each input is left for future work.

1. Solving biology problems generally requires moderate reasoning combined with factual recall. Some errors in this domain result from missing core factual knowledge in the base model (e.g., “What protein is encoded by gene X?”). Our RICE aims to enhance reasoning given existing facts, rather than to supply missing knowledge.

2. In contrast, solving challenging math problems generally demands stronger reasoning ability than biology questions. As a result, the optimal steering multiplier in the mathematics domain tends to be higher than that in the biology domain.

Experimental results consistent with this claim can be found in the first two tables of our previous response. The optimal steering multiplier for the math domain is 64, while for other domains it is lower than 64.

3. Additionally, we find that in the same domain, using a moderate multiplier leads to better performance. For example, in the math domain, both overly large and overly small multipliers fail to achieve optimal results. The impact of the steer multiplier on performance is detailed in Table 2 of our paper.

We have further elaborated on the above analyses and results in the revised manuscript.

We sincerely appreciate your valuable suggestions and thoughtful feedback. Should you have any further questions or require additional clarification, please do not hesitate to contact us.

Thanks for your responses. I have one more question about W2&Q1: why is the proposed method so sensitive to the hyperparameters (steering multiplier)?

Thank you for your very positive and constructive review. We are delighted you found the paper enjoyable and the idea clear and well-executed.

W1&Q1: The distribution of nPMI distribution

This is an excellent question. We list the nPMI of top 10 experts for Deepseek-R1. The distribution is sharply peaked, which strongly supports our hypothesis that a small number of experts are highly specialized for cognitive functions. The group of top five “thinking-specialized” experts have significantly higher nPMI scores than the rest.

W2: Paper structure

Your suggestions for restructuring the research questions are excellent and create a more logical narrative flow. We will adopt this structure for the camera-ready version. Thank you!

W3: No uncertainty estimates

This is a fair point. Our results are currently based on single, deterministic runs due to the high computational cost. For the final version, we will perform multiple runs with different seeds to report standard deviations, adding to the rigor of our findings.

Q2: Why does stimulating the “biology experts” fail to improve performance in the biology domain?

This is an excellent question with a two-part answer:

• Hyperparameter Mismatch:

  As you noted, applying RICE sometimes degraded performance in non-math dmains (e.g., Biology in Table 4). This was the result of a specific experimental choice to test generalization. In the original manuscript, we used hyperparameters (steering multiplier = 64) tuned only on the math domain (AIME) and applied them directly to all other domains.

  This setup was designed to test the raw generalization of the math-derived setting. However, different domains may require different intensities of cognitive steering. When we use a simple, domain-specific steering multiplier (selected from a small set of {16, 32, 64}), RICE delivers consistent and significant improvements across all domains.

DeepSeek-R1 with Domain-Tuned Multipliers

Qwen3-235B with Domain-Tuned Multipliers

Our hypothesis is that math problems demand more intensive reasoning, hence the higher optimal multiplier. We will clarify this in the paper to avoid misinterpretation.

• Failure Mode Analysis: We also found that many errors in the biology domain stem from a lack of core factual knowledge in the base model (e.g., "What protein is encoded by gene X?"). Our method, RICE, is designed to enhance the reasoning process given a set of facts, not to inject missing knowledge. We will add this important clarification about the scope of our method to the paper.

Thank you for your encouraging review and for highlighting our method's interpretability and strong performance on math problems.

W1&W3: Limited Evaluation We address concerns about the number of models and benchmarks used.

• Models: Our work focuses on identifying cognitive experts in large reasoning models with MoE architectures. Concurrent work [1] has shown that these models exhibit semantic expert specialization, which is distinct from the more syntactic specialization found in general-purpose MoEs [2]. At present, DeepSeek-R1 and Qwen3-235B are the only publicly available, state-of-the-art LRMs with an MoE architecture. We are committed to extending our evaluation as more such models become open-source.
• Benchmarks: We acknowledge the request for broader evaluation. To supplement our original results on AIME and GPQA Diamond, we have run new experiments on three additional diverse benchmarks: GSM8K (grade-school math), MATH (competition math), and HLE (Humanity's Last Exam, covering social science, CS, etc.). As shown below, RICE consistently provides benefits. These results will be added to the revised paper.

The number in parentheses after each dataset indicates the number of test examples. For HLE, we randomly sampled 100 examples as the test set. Due to resource and time constraints, we will conduct more extensive evaluations for the camera-ready version.

Reference

[1] Semantic Specialization in MoE Appears with Scale: A Study of DeepSeek R1 Expert Specialization, arxiv 2025.

[2] OpenMoE: An Early Effort on Open Mixture-of-Experts Language Models, ICML 2024.

W2: Insignificant/misleading non-math results. As you noted, applying RICE sometimes degraded performance in non-math dmains (e.g., Physics, Biology in Table 4). This was the result of a specific experimental choice to test generalization. In the original manuscript, we used hyperparameters (steering multiplier = 64) tuned only on the math domain (AIME) and applied them directly to all other domains.

This setup was designed to test the raw generalization of the math-derived setting. However, different domains may require different intensities of cognitive steering. When we use a simple, domain-specific steering multiplier (selected from a small set of {16, 32, 64}), RICE delivers consistent and significant improvements across all domains.

DeepSeek-R1 with Domain-Tuned Multipliers

Qwen3-235B with Domain-Tuned Multipliers

Our hypothesis is that math problems demand more intensive reasoning, hence the higher optimal multiplier. We will clarify this in the paper to avoid misinterpretation.

W4: It is not clear if the method generalizes to other reasoning tasks beyond QA.

To address this, we evaluated on GSM8K (math word problems), MATH (competition math), and HLE (multi-domain problem-solving), which are not simple QA tasks. Our method shows clear benefits on these diverse reasoning formats.

The number in parentheses after each dataset indicates the number of test examples. For HLE, we randomly sampled 100 examples as the test set due to the limited time available for author response.

If "beyond QA" refers to other task families (e.g., code generation, summarization), we believe this is an exciting direction for future work, but our current focus is on complex reasoning.

Q1: Could you provide insights into the distribution of the nPMI for experts?

This is an excellent question. The distribution is sharply peaked, which strongly supports our hypothesis that a small number of experts are highly specialized for cognitive functions. We list the nPMI of top 10 experts for Deepseek-R1. The group of top five “thinking-specialized” experts have significantly higher nPMI scores than the rest.

Q2: There is a qualitative example of how reasoning changes. What are some other ways to evaluate if the reasoning/thinking tokens improved?

We evaluate the improvement in reasoning through a combination of methods:

1.  Final Accuracy (Quantitative): The most critical metric, as improved reasoning should lead to more correct answers (Tables in main paper and General Response).

2.  Qualitative Analysis: We provide examples (Table 7) showing how RICE leads to more structured and effective reasoning paths.

3.  Efficiency (#Tokens): On reasoning tasks, RICE often achieves higher accuracy with fewer tokens, suggesting more direct and less wasteful thinking.

4.  Future Analysis: As promised to reviewer R-RpVC, we will add analysis of activation patterns in the final version for a deeper look at the mechanism.

Thanks for your response.

Our study investigates the reasoning capabilities of large-scale Mixture of Experts (MoE) models, such as DeepSeek-R1 (671B) and Qwen3-235B. Due to resource constraints and the limited response period, we have not yet identified cognitive experts for advanced reasoning tasks beyond QA. We will evaluate the generalization of our RICE framework to these reasoning tasks beyond QA in future work when resources are available.

Thank you for your valuable feedback and positive assessment of our method's novelty and clarity.

W1: The results rely on only two benchmarks (AIME and GPQA Diamond) and two models (DeepSeek-R1 and Qwen3-235B), which limits the generalizability of the findings.

We address concerns about the number of models and benchmarks used.

• Models: Our work focuses on identifying cognitive experts in large reasoning models with MoE architectures. Concurrent work [1] has shown that these models exhibit semantic expert specialization, which is distinct from the more syntactic specialization found in general-purpose MoEs [2]. At present, DeepSeek-R1 and Qwen3-235B are the only publicly available, state-of-the-art LRMs with an MoE architecture. We are committed to extending our evaluation as more such models become open-source.
• Benchmarks: We acknowledge the request for broader evaluation. To supplement our original results on AIME and GPQA Diamond, we have run new experiments on three additional diverse benchmarks: GSM8K (grade-school math), MATH (competition math), and HLE (Humanity's Last Exam, covering social science, CS, etc.). As shown below, RICE consistently provides benefits. These results will be added to the revised paper.

The number in parentheses after each dataset indicates the number of test examples. For HLE, we randomly sampled 100 examples as the test set. Due to resource and time constraints, we will conduct more extensive evaluations for the camera-ready version.

W2: As shown in Table 4, activating domain-specific experts does not consistently improve performance. As you noted, applying RICE sometimes degraded performance in non-math domains (e.g., Physics, Biology in Table 4). This was the result of a specific experimental choice to test generalization. In the original manuscript, we used hyperparameters (steering multiplier = 64) tuned only on the math domain (AIME) and applied them directly to all other domains.

This setup was designed to test the raw generalization of the math-derived setting. However, different domains may require different intensities of cognitive steering. When we use a simple, domain-specific steering multiplier (selected from a small set of {16, 32, 64}), RICE delivers consistent and significant improvements across all domains.

DeepSeek-R1 with Domain-Tuned Multipliers

Qwen3-235B with Domain-Tuned Multipliers

Our hypothesis is that math problems demand more intensive reasoning, hence the higher optimal multiplier. We will clarify this in the paper to avoid misinterpretation.

W3: Despite impressive performance gains, the paper lacks in-depth explanations for why RICE works.

We agree that a deeper analysis is valuable. Visualizing feature distributions to understand how re-weighting steers the model is an excellent suggestion. Given the scale of these models and the short rebuttal period, we were unable to complete this analysis in time. We have begun this investigation and commit to including visualizations of activation patterns in the camera-ready version to provide these crucial insights.

W4: What is the data for identifying the cognitive experts, and could there be some data leakage? Could the author adopt some other benchmarks are evaluate the activated experts?

This is a critical point that we are happy to clarify.

• No Data Leakage: - Our method for identifying cognitive experts is fully unsupervised. It relies only on the co-occurrence of expert activations with reasoning tokens (<think>) from input queries, not on any ground-truth labels. Therefore, the concept of data leakage in the supervised sense does not apply.
• Expert Generalization: To explicitly test generalization, we identified experts on AIME24 and applied them to new domains. As shown in our General Response (Sec 1), these same experts (identified on math problems) improve performance on GSM8K, MATH, and HLE, demonstrating strong task generalization.

The number in parentheses after each dataset indicates the number of test examples. For HLE, we randomly sampled 100 examples as the test set due to the limited time available for author response.

W5: Modification of formula tokens in Eq. (6) Thank you for catching this. We have corrected the typo and have performed another thorough proofread of the manuscript.

Reference

[1] Semantic Specialization in MoE Appears with Scale: A Study of DeepSeek R1 Expert Specialization, arxiv 2025.

[2] OpenMoE: An Early Effort on Open Mixture-of-Experts Language Models, ICML 2024.