Reviews Review 1 Review. This paper studies the problem conversational recommendations lack collaborative filtering (CF) signals. To solve this problem, this paper proposes a framework that integrates conversational context from LLMs with CF-based item representations. The experimental results are conducted to demonstrate the efficiency of the proposed framework. Overall evaluation 5: (Marginally below acceptance threshold) Reviewer's confidence 4: (I am confident but not absolutely certain that the evaluation is correct) Review 2 Review. The authors propose a new Reddit-ML32M dataset and an LLM-based framework for conversational recommendation. The proposed framework outperforms the best-performing conversational baseline when using only context information. Pros: 1. The paper is well-organized and easy to follow. 2. The proposed method bridges conversational context with collaborative filtering information. 3. The proposed method is technically sound. Cons: The proposed method is essentially a form of item representation enhancement, which is currently limited to movie representations, and its generalizability is unknown. If applied to an e-commerce scenario, it implies that the dataset may need to be reconstructed, which poses a significant limitation. Overall evaluation 6: (Marginally above acceptance threshold) Reviewer's confidence 4: (I am confident but not absolutely certain that the evaluation is correct) Review 3 Review. The manuscript presents integrating collaborative filtering (CF) signals with conversational recommendation systems (CRS). The methods are clearly defined, and the experimental results are relatively comprehensive, demonstrating improvements over the baselines. Including a novel dataset, Reddit-ML32M, and the proposed framework are notable contributions. However, some concerns remain (e.g. the reliance on specific datasets and models, generalization issues) The manuscript is a little bit clear, but certain technical explanations, such as embedding-based item ranking, might benefit from further simplification or illustrative examples for accessibility to a broader audience. Sure, I know there can be some issues with the page limitation, but it is required to present it. In addition, the framework relies heavily on models like SASRec and GPT-3.5-t. This dependency can cause some limitations on its applicability to the real world. Combining LLMs with CF embeddings could be computationally expensive, making it difficult to implement in resource-constrained environments. This issue is not addressed in the paper. Pros. - Experiments are valuable, with multiple baselines and evaluation metrics. - Clear contributions, including the dataset and framework. Cons. - Some sections (e.g. embedding-based ranking) are dense and can be challenging for readers. - The practical integration can be limited (see above). Overall evaluation 6: (Marginally above acceptance threshold) Reviewer's confidence 3: (I am fairly confident that the evaluation is correct) Metareview Metareview for paper 3280 Title Bridging Conversational and Collaborative Signals for Conversational Recommendation Authors Ahmad Bin Rabiah, Nafis Sadeq and Julian McAuley Recommendation accept Text This paper presents a new dataset and proposes an approach for leveraging Collaborative Filtering (CF) signals in Conversational Recommender Systems (CRS). The authors uses a zero shot LLM for initial item recommendation, then link these recommended items (based on "exact match") to a large dataset of candidate items, which has already used CF to learn embedding. Similarity of items in the database to the initial recommendations is used to create a final ranking of recommended items. Using their newly proposed dataset, Reddit-ML32M, the authors compare their CRS + CF approach to CRS-only and CF-only models, showing significant improvements over most baselines. This paper received relatively consistent scores with two marginal accepts (6) and 1 marginal reject (5). All reviewers agreed that the paper tackles a relevant problem, the paper is well written an mostly easy to follow, and appreciate the effort to link historical interaction datasets to conversational recommendation datasets. One reviewer also appreciated that the contribution is very clear, especially based on the dataset and the framework. In terms of the limitations pointed out by the reviewers, the most significant one was that this is little detail given about how the Reddit-ML32M dataset is constructed. Also, one reviewer felt a detailed algorithm description of the method would be useful. Concerns were also raised about the lack of analysis of the computational cost of the method, and some parts of the paper (in particular, the embedding based ranking) being difficult to follow. From my own understanding of the paper, I am definitely concerned about two of the above weaknesses, in particular: - (1) Lack of description of how the Reddit-ML32M dataset is constructed. I would expand on the reviewers criticism here by clarifying that no meaningful details of (a) how the dataset is constructed, especially linking between CRS and CF, or (b) statistics about dataset size, partition between train/test, etc. This seems to be a glaring omission, considering that the dataset is one of the main contributions that the authors claim in the Abstract and Introductions - (2) I also found it difficult to follow the details if the embedding based ranking and, in particular, I struggle to understand the intuition behind how this ranking is supposed to improve on the initial LLM rankings. In addition, I noticed one other potential discrepancy/mistake in the results that was pointed out by the reviewers: in Table 2 H@1 and N@1 have identical results for all models. Unless I'm mistaken I don't believe that H@K and N@K are expected to be identical in the case where k=1. So there is probably an error here where the authors entering the wrong results. Or, if this is expected, the authors need to take not of this and explain why this is expected. Overall, then, I consider this to be a very borderline case. On the one hand, there is clearly an interesting contribution here, but I do consider the two issues highlighted above to be significant. As such, while I will give a preliminary recommendation to accept the paper, I do feel that this paper might warrant additional discussion.