acceptodds
Withdrawn from ICLR 2027

E-NRPA: Emotion-Aware Nested Rollout Policy Adaptation on Large Language Models for Goal-Oriented Dialogue Planning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

User emotion influences strategy selection in goal-oriented dialogue, but existing online planning methods typically do not explicitly model changes in emotion. This paper proposes Emotion-Aware Nested Rollout Policy Adaptation for Goal-Oriented Dialogue Planning (E-NRPA), which incorporates current emotion and historical emotion trajectories identified by large language models into Monte Carlo simulation and policy update processes, enabling the planner to adjust dialogue strategies based on changes in user state. We conduct experiments on ESConv, ExTES, CraigslistBargain, and P4G. Results show that E-NRPA achieves stable performance on automatic metrics, with advantages in task success rate and negotiation outcome quality. Human evaluations indicate that E-NRPA is preferred in negotiation tasks and receives better ratings for conciseness, goal progress, and overall efficiency in emotional support dialogues. The incorporation of emotion trajectories further enhances the planner’s sensitivity to changes in user state and improves its ability to adapt strategy selection dynamically. Compared with methods that rely only on the current dialogue state, E-NRPA can better capture how emotional evolution influences user goals and interaction intentions. These findings suggest that emotion-aware planning facilitates effective, goal-oriented dialogue strategy selection across different tasks.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Dialog Systems

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_z9CtOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This work presents E-NRPA, an emotion-aware nested rollout policy adaptation framework for goal-oriented dialogue planning. It explicitly integrates user’s current emotion and historical emotion trajectories from LLMs into Monte Carlo simulation and policy update, allowing dynamic adjustment of dialogue strategies to respond to evolving user emotional states.

Critical strengths
  1. This work innovatively incorporates evolving user emotion trajectories into nested Monte Carlo rollout planning, enabling dynamic dialogue strategy adaptation to emotional shifts overlooked by prior planners.
  2. The proposed E-NRPA offers a unified emotion-aware planning formulation that generalizes across distinct goal-oriented dialogue tasks including negotiation and emotional support.
Critical weaknesses

There exists a considerable amount of related work similar to this study. The authors could further highlight the necessity of their proposed method. The paper lacks sufficient explanatory details and necessary validation experiments. Specific comments are as follows:

  1. The image lacks clarity, please draw a vector diagram.
  2. Large language models can already recognize emotions well. Why is it still necessary to introduce Rollout in the Emotion detector?
  3. How to define the action space?
  4. What are the nodes of Monte Carlo Rollout?
  5. The state space contains dialogue history and emotion history, which leads to an enormous state space. How does Monte Carlo Rollout work? Please explain and provide experiments.
  6. How is the reward function defined?
  7. Evaluating trajectories based on rewards introduces additional costs. Please demonstrate this in experiments.
Questions for authors

No

Reviewer_5ptLOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes Emotion-Aware Nested Rollout Policy Adaptation, an online planning framework for goal-oriented dialogue systems. The approach extends Nested Rollout Policy Adaptation for Goal-Oriented Dialogue by explicitly incorporating user emotion into the dialogue-planning state. The authors evaluate the approach on four datasets and experiments use GPT-4o-mini and Qwen2.5-14B as backbone models. The authors report high task success rates, improvements in negotiation outcomes and favorable human preferences compared with NRPA-GD.

Critical strengths
  1. The paper addresses an important limitation in goal-oriented dialogue planning. The motivation is clear and the research question is meaningful. The proposed method builds on NRPA-GD rather than introducing an entirely new planning algorithm.

  2. The evaluation includes both collaborative and non-collaborative tasks. The inclusion of two backbone models also provides some evidence that the approach is not entirely model-specific. The inclusion of human evaluation is also good as task-success metrics alone are insufficient for assessing dialogue quality.

Critical weaknesses
  1. The main contribution is incremental and not reasonably distinguished from NRPA-GD. The main methodological difference between NRPA-GD and E-NRPA is the incorporation of an LLM-inferred emotion trajectory into the dialogue state. Thus, it appears closer to emotion-augmented state representation for NRPA than a fundamentally new emotion-aware policy adaptation algorithm.

  2. The paper mentions that modeling historical emotional trajectories improves strategy selection. However, the ablation study analyses only the number of NRPA iterations. Increasing the number of rollouts is an analysis of search budget, not an ablation of the proposed emotion mechanism. At minimum the authors should be having the following ablation experiments: (i) NRPA-GD without emotion, (ii) NRPA + current emotion only, (iii) NRPA + historical emotion trajectory, (iv) NRPA + shuffled emotion trajectory, (v) NRPA + randomly assigned emotions, (vi) NRPA + textual emotion summary. The lack of these ablations significantly weakens the main contribution.

  3. The paper reports success rates of 1.0000 on several tasks and configurations. While this plausible, it raises a lot of concerns which should be addressed as follows: (i) emotional support does not necessarily have a clearly verifiable terminal condition. A dialogue ending successfully according to a simulator is not equivalent to genuinely helping a user, (ii) the paper does not sufficiently explain the independence of the user simulator, reward function and evaluation procedure, (iii) When most methods approach 100% success, success rate becomes a weak discriminator of planning quality. So the question is: are the tasks too easy or the evaluation criteria too permissive?

  4. The entire method depends on an LLM accurately identifying user emotions across dialogue turns. However, the paper does not evaluate the emotion detector itself. There is no reported emotion-recognition accuracy, agreement with human emotion annotations, confusion matrix or error analysis.

  5. The paper repeatedly claims that E-NRPA dynamically adapts strategy selection in response to changes in user emotion. However, the experiments do not show this mechanism.

Questions for authors
  1. What is the precise contribution of historical emotion trajectories? Can the authors compare full E-NRPA against current-emotion-only, shuffled-history and no-emotion variants under identical experimental conditions?

  2. How is success defined and independently verified? Why do several methods achieve 100% success on emotionally complex tasks such as ESConv and ExTES?

  3. How accurate is the emotion detector? What taxonomy is used and how sensitive is planning performance to incorrect or ambiguous emotion labels?

  4. What changes in the policy adaptation algorithm? Is E-NRPA fundamentally different from NRPA-GD or is it NRPA-GD operating on an augmented state representation?

  5. How do the authors establish that emotional transitions cause changes in strategy selection? Can you provide quantitative transition-conditioned strategy analyses and controlled counterfactual examples?

  6. Are the baseline comparisons computationally fair? What are the relative rollout counts, LLM calls, latency and token costs for E-NRPA and NRPA-GD?

  7. Are the human preference improvements statistically significant? Can the authors report agreement, tie rates and dialogue-level confidence intervals?

Reviewer_iyhNOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes E-NRPA, an extension of NRPA-GD that incorporates LLM-inferred user emotions and historical emotion trajectories into online dialogue planning. The method is evaluated on four goal-oriented dialogue benchmarks.

Critical strengths
  1. Modeling evolving user emotions is a relevant direction for goal-oriented dialogue planning.

  2. The method is evaluated across multiple tasks and achieves encouraging results on negotiation outcomes and some human preference measures.

Critical weaknesses
  1. The central emotion mechanism lacks direct ablations. The reported analysis mainly varies the NRPA iteration count, without comparing no-emotion, current-emotion-only, and full-trajectory variants. Thus, the contribution of historical emotion modeling is not established.

  2. Improvements over NRPA-GD are mixed. On CraigslistBargain with GPT-4o-mini at Level 2, E-NRPA improves the Sale-to-List Ratio from 0.6422 to 0.9345, but increases average dialogue turns from 2.72 to 4.29. The paper provides insufficient statistical evidence to establish the robustness of these trade-offs.

  3. The methodological novelty over NRPA-GD is limited, mainly involving emotion-aware state augmentation. The reliability of inferred emotion labels and the additional computational cost are not adequately evaluated, making the practical benefit difficult to assess.

Questions for authors
  1. Can the authors compare no-emotion, current-emotion-only, and full-emotion-trajectory variants under matched settings?

  2. Can the authors report statistical uncertainty and additional LLM inference costs relative to NRPA-GD?

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

Related papers

Discussion (0)

Sign in to comment.