acceptodds
Desk rejected at ICLR 2027

Mixture of Decoders for Diverse Dialog Response Generation

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Mixture modeling is a long established machine learning technique for learning large sets of multi-modal data. While it is known that sequence-to-sequence models for dialog response generation suffer from the problem of low diversity, we hypothesize that it is because sequence-to-sequence models tend to learn a degenerate uni-modal distribution of responses. We then propose to incorporate a mixture of decoders into sequence-to-sequence models and try to make each decoder learn specialized topics in order to improve the diversity of generated responses. Our model is developed under the framework of conditional variational autoencoder (CVAE). We evaluate our approach on an open domain chat corpus and show improvement over strong baselines in quantitative measures and human evaluation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
dialog response generation, mixture of decoders

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_wZcXOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors propose a mixture of decoders for dialogue generation, where each decoder will correspond to a specialised topics. In particular, the authors propose a conditional VAE, where a shared encoder encodes the input into a latent z. During inference, they perform beam search on each decoder, and vote based on the conditional probabilities.

Critical strengths
  1. The authors' method seems intuitive.
  2. The authors provide qualitative study on their method.
Critical weaknesses
  1. The authors miss a highly related paper [1] as a comparison. The ideas are highly related, as they also used a mixture of decoders for dialogue generation. The main difference is that this paper propose CVAE, whereas they proposed an EM variant.

  2. Outdated and poor experimental setup. The authors mainly use BLEU and ROUGE score for dialogue generation, which is quite outdated. It would be nice if the authors can study more recent tasks such as reasoning or instruction following.

[1] An equal-size hard EM algorithm for diverse dialogue generation. [ICLR'23]: https://openreview.net/pdf?id=k5PEHHY4spM

Questions for authors

N/A

Reviewer_X3W7OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a Mixture of Decoders (MoD) for diverse dialogue response generation within a conditional variational autoencoder framework. The model combines a shared encoder with multiple LSTM decoders, using a discrete latent variable to represent different response modes and continuous latent variables for token-level variation. To encourage decoder specialization, the authors introduce pseudo-label supervision derived from response clustering and assign generic responses to a dedicated mixture. At inference, each decoder generates candidates through beam search, which are then ranked using mixture-weighted probabilities. Experiments on the Persona dataset show improvements in lexical diversity and favorable human preferences against the evaluated baselines.

Critical strengths

The approach provides an intuitive mechanism for modeling the one-to-many nature of dialogue, where a single context can support several valid responses. The paper identifies practical challenges in mixture training, including decoder dominance and uninformative mixture assignments, and introduces supervision to encourage specialization.

Critical weaknesses

The main weakness is the limited contemporary relevance of the experimental setting. The evaluation relies entirely on small LSTM encoder–decoder models and older baselines, without establishing whether the proposed method benefits pretrained Transformer models. The approach combines established mixture modeling, variational inference, and clustering techniques, but its methodological novelty is not clearly distinguished from prior work. Furthermore, diversity is measured primarily through lexical overlap and raw n-gram counts, which are affected by the longer responses produced by MoD. Ablations do not fully isolate specialization from additional parameters, special treatment of generic responses, or decoding changes. Finally, the single-dataset evaluation, small human study, and increased computational cost limit the strength and generalizability of the conclusions.

Questions for authors
  1. How could we transfer this idea to more recent model architectures, such as transformers, SSM
  2. How could we use the idea of increasing diversity in which generation diversity matters, such as when creating roll-out in GRPO
  3. How could we reduce the training inefficiency of the proposed approach
Reviewer_yZCROpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a method that incorporates a mixture of decoders to improve the diversity of generated dialogue responses. Each decoder is developed within a conditional variational autoencoder framework. The authors evaluate their method on the Persona dataset and compare the mixture-of-decoders method with a mixture-of-softmaxes baseline.

Critical strengths

The paper addresses the important problem of generating diverse dialogue responses. However, the overall contribution appears limited, as the method is developed on LSTM-based decoders and does not adequately engage with recent advances in diverse dialogue generation. Consequently, I find it difficult to identify substantial strengths in the current work.

Critical weaknesses
  1. The paper's related work appears outdated, with no references published after 2020. Furthermore, the proposed method is developed on LSTM-based decoders, which are less competitive than more recent architectures. The paper does not sufficiently discuss how its approach relates to or advances the state of the art in dialogue response generation.

  2. Given that the primary goal of the paper is to improve dialogue response diversity, the evaluation should include appropriate diversity metrics and a more thorough analysis of response diversity. In particular, the reported Distinct uni-gram, bi-gram, and tri-gram metrics appear to differ from the standard Distinct-1/2/3 metrics, which normalize the number of unique n-grams by the total number of generated n-grams to avoid favoring long sentences (Li et al., 2016).

  3. This paper uses a single dataset for evaluation, and raises the generalizability issue.

Questions for authors

Please see the weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.