Review 1
Strong Points	1. This paper aims to address the problems in conversational recommender systems, focusing on the abundant semantic information during dialogues.
2. It proposes a reasonable hypothesis that the most frequently used requestions have been addressed, which can be used as a knowledge base for the current request. Therefore, it converts the context of the current request into dense representations and retrieves related experiences from past dialogues to solve them.

3. The results of experiences have verified the improvement of the proposed method, which significantly outperforms current baselines. The author also provides standard variance to show the reliability of the method.
Weak Points	1. It seems that the baselines are out-of-date. Could you conduct the experiments on some recent baselines?
2. I am also wondering about the distribution of the requests in the training set. This distribution could further verify your hypothesis at the beginning of the paper.
3. It seems that the proposed method ignores the personal preference of the current user, where the dialogue history could be helpful.
Overall Evaluation	
1: (weak accept)
This paper proposes a practical method for conversational recommender systems, which could greatly improve the performance, compared with previous works. The motivation and the method are reasonable, and the experiment results are significant. Therefore, I prefer to give a weak acceptance for this paper.
Review 2
Strong Points	- The impact of the proposed solutions is validated on three datasets.
- The paper presents a complete analysis of the results.
- Performance are compared to some LLMs.
Weak Points	- Low details about Related Works: The discussion of related works is very basic, with a list of citations without entering into the details. This very short description hides the true problem that the paper aims to tackle. For a more comprehensive overview, related works should be discussed in greater detail to provide context and highlight the novelty and relevance of the proposed approach.

- The code is not included: The absence of code makes it difficult to reproduce the results and verify the claims made in the paper. Including the code is essential for transparency and for enabling other researchers to build upon this work.

- Missing baselines: Some baselines in the field of Conversational Recommendation are missing. Baselines such as ReDial [1], KBRD [2], and UNICRS [3] are some of the new methods proposed in this field that need to be included to get a comprehensive understanding of the impact of the outcomes. The inclusion of these baselines is crucial for demonstrating the relative performance of the proposed methods.

- Results are not clearly reported and commented: In Table 2, models are reported in two settings (zero-shot and SFT). However, the differences between the two settings are not reported or explained. A clearer explanation of these settings and a detailed discussion of the results would help in understanding the strengths and weaknesses of the proposed methods.

[1] Towards Deep Conversational Recommendations
[2] Towards Knowledge-Based Recommender Dialog System
[3] Towards Unified Conversational Recommender Systems via Knowledge-Enhanced Prompt Learning
Overall Evaluation	
-1: (weak reject)
The paper presents a collaborative (and hybrid) approach to conversational recommendation. The main idea of the authors is to retrieve similar dialogs to the current one in the training set and exploit collaborative information coming from this additional context to perform the recommendation. The results demonstrate competitive performance, but the gap between the baselines and the proposed model(s) is still too high to represent a significant contribution. Moreover, the paper lacks detailed discussion on related works, does not include code for reproducibility, misses important baselines, and fails to clearly explain the results and some methodological components.

Additional
- In section 2, equation 1, it is reported the \( w_i \) factor, however it seems not explained.
Review 3
Strong Points	Interesting and novel idea
Combining a case-based reasoning method in addition to another component for CRS
Weak Points	It is not clear why the only used measure for comparison is R@K
The experiment section need more discussion as it is not clear why the proposed method works better for higher ks and worse for lower ks
Overall Evaluation	
1: (weak accept)
The author proposed a simple method for generating recommendations in conversational recommender systems. Based on their main hypothesis that “many inference time user requests can be answered by reusing popular crowd-written answers associated with similar training queries,” the author introduced a novel approach that employs a case-based reasoning method. This method leverages similar previous discussions to generate a score for the test item, which is then combined with an optional model-based component using a consensus method. Their experiments indicated that for higher values of recall@k, the system outperforms the more complex method (Gemma-2B) and competes with another sophisticated method (Vicuna-7B).

I appreciate this simple idea that clearly combines two different perspectives and achieves better results. However, there are some concerns. Firstly, the authors only reported Recall@k, and it is unclear why they did not discuss Precision@k, for example. Additionally, the discussion in the Experiment section could be improved. Specifically, there is no explanation for why the methods perform differently at various values of k. Since the results are not entirely consistent, more discussion is required to justify them.

Overall, the paper is well-written and clear. I would have higher confidence in suggesting acceptance of the paper if the authors could address the concerns mentioned above.
Review 4
Strong Points	- Interesting complementary strategy for conversational recommendation that reuses of answers associated with similar training queries.

- Conceptually straightforward method to exploit neighborhood-based information in conversational recommendation.

- The experimental analysis guided by general performance, neighborhood importance and size. Ablation studies were conducted considering dataset size and neighborhood weight.
Weak Points	- The efficiency aspect was not measured. The lazy neighborhood-based strategy incurs an additional cost during inference, which cannot be ignored for large datasets. I believe the execution time is an important aspect that should be analyzed.

- There are no statistical tests to support some of the authors' claims regarding the effectiveness of the proposed NBCRS in Table 2.

- The effectiveness is measured using only Recall@k.

- In the conclusion, the authors claim that the proposal is computationally efficient, but there is no experimental analysis of computational time during inference.
Overall Evaluation	
1: (weak accept)
The conceptually straightforward method of exploiting neighborhood-based information in conversational recommendation is an interesting strategy for reusing answers associated with similar training queries. The experimental analysis for this approach presents competitive results, especially when comparing the use of BERT as a sentence encoder (for the proposed NBCRS) against LLMs in terms of Recall@k.

However, the performance should be better addressed in terms of computational efficiency, particularly considering the additional time required to compute the neighborhood. Additionally, I think the abstract is misleading about the use of 170 times less GPU memory compared to LLMs, since the proposal can be seen as a complement to such approaches.
Metareview
Metareview for paper 577
Title	Neighborhood-Based Collaborative Filtering for Conversational Recommendation
Authors	Zhouhang Xie, Junda Wu, Hyunsik Jeon, Zhankui He, Harald Steck, Rahul Jha, Dawen Liang, Nathan Kallus and Julian Mcauley
Text	To better understand users’ preferences in CRS, this work proposes a neighborhood-based approach to consider historical dialogues from the training data to serve as a knowledge base. The authors compared with LLM-based methods and demonstrates their method's outperforming effectiveness in terms of both accuracy and efficiency. Most reviewers agree that the proposed method is reasonable and the experimental results are convincing, but there are some concerns about its baselines, data distribution, reproducibility, related work discussion, and results explanation. The authors have tried to address the reviewers’ comments in their rebuttal. Overall, this is an interesting work and has its technical merit. “Accept” is recommended if the revisions can be well accommodated in their final version.