Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Affective Flow Language Model for Emotional Support Conversation

Chenghui Zou, Ning Wang, Tiesunlong Shen, Luwei Xiao, Chuan Ma, Xiangpeng Li, Rui Mao, Erik Cambria

TL;DR

This work tackles the problem of long-horizon emotional support conversation where traditional alignment signals are sparse and poorly credit intermediate strategy decisions. It introduces AFlow, a framework that injects dense prefix-level supervision through affective flow signals derived from search-distilled trajectories and a Generative Flow Network perspective. AFlow couples AFPO, a subpath flow-balance objective, with a policy and value heads so that downstream value propagates coherently across dialogue prefixes, enabling stable and diverse strategy progression. Empirical results on two ESC datasets show consistent improvements in strategy alignment and response diversity, with robust performance across open-source backbones and even competitive standing against strong proprietary LLMs, underscoring the practical value of flow-shaped process supervision for long-horizon emotional support.

Abstract

Large language models (LLMs) have been widely applied to emotional support conversation (ESC). However, complex multi-turn support remains challenging.This is because existing alignment schemes rely on sparse outcome-level signals, thus offering limited supervision for intermediate strategy decisions. To fill this gap, this paper proposes affective flow language model for emotional support conversation (AFlow), a framework that introduces fine-grained supervision on dialogue prefixes by modeling a continuous affective flow along multi-turn trajectories. AFlow can estimate intermediate utility over searched trajectories and learn preference-consistent strategy transitions. To improve strategy coherence and empathetic response quality, a subpath-level flow-balance objective is presented to propagate preference signals to intermediate states. Experiment results show consistent and significant improvements over competitive baselines in diverse emotional contexts. Remarkably, AFlow with a compact open-source backbone outperforms proprietary LMMs such as GPT-4o and Claude-3.5 on major ESC metrics. Our code is available at https://github.com/chzou25-lgtm/AffectiveFlow.

Affective Flow Language Model for Emotional Support Conversation

TL;DR

This work tackles the problem of long-horizon emotional support conversation where traditional alignment signals are sparse and poorly credit intermediate strategy decisions. It introduces AFlow, a framework that injects dense prefix-level supervision through affective flow signals derived from search-distilled trajectories and a Generative Flow Network perspective. AFlow couples AFPO, a subpath flow-balance objective, with a policy and value heads so that downstream value propagates coherently across dialogue prefixes, enabling stable and diverse strategy progression. Empirical results on two ESC datasets show consistent improvements in strategy alignment and response diversity, with robust performance across open-source backbones and even competitive standing against strong proprietary LLMs, underscoring the practical value of flow-shaped process supervision for long-horizon emotional support.

Abstract

Large language models (LLMs) have been widely applied to emotional support conversation (ESC). However, complex multi-turn support remains challenging.This is because existing alignment schemes rely on sparse outcome-level signals, thus offering limited supervision for intermediate strategy decisions. To fill this gap, this paper proposes affective flow language model for emotional support conversation (AFlow), a framework that introduces fine-grained supervision on dialogue prefixes by modeling a continuous affective flow along multi-turn trajectories. AFlow can estimate intermediate utility over searched trajectories and learn preference-consistent strategy transitions. To improve strategy coherence and empathetic response quality, a subpath-level flow-balance objective is presented to propagate preference signals to intermediate states. Experiment results show consistent and significant improvements over competitive baselines in diverse emotional contexts. Remarkably, AFlow with a compact open-source backbone outperforms proprietary LMMs such as GPT-4o and Claude-3.5 on major ESC metrics. Our code is available at https://github.com/chzou25-lgtm/AffectiveFlow.
Paper Structure (110 sections, 4 theorems, 36 equations, 8 figures, 5 tables)

This paper contains 110 sections, 4 theorems, 36 equations, 8 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Preliminaries
  4. Methodology
  5. Affective Flow Reward Acquisition
  6. Role-Based Dialogue Environment.
  7. Affective Flow Rewards on Dialogue States.
  8. Selection.
  9. Expansion.
  10. Simulation/rollout.
  11. Backpropagation.
  12. Affective Flow Preference Optimization
  13. Optimization variables.
  14. Affective Flow parameterization.
  15. Flow Construction in Search.
  16. ...and 95 more sections

Key Result

Theorem 1

If the subpath flow-balance loss is zero and the terminal flow is matched to the terminal reward, then the induced policy samples trajectories with probability proportional to their terminal affective reward:

Figures (8)

  • Figure 1: Comparison of Emotional Support Conversation approaches
  • Figure 2: Detailed diagram of the AFlow framework for emotional support conversation.
  • Figure 3: (a) Value-gap distribution. AFlow exhibits a stronger positive shift than DPO. (b) Turn-wise separation. AFlow provides steadier guidance across turns.
  • Figure 4: Stage-wise effectiveness scores. AFlow achieves the strongest overall performance.
  • Figure 5: Trajectory analysis of reward scores. AFlow shows robust long-horizon improvement, with competitive late-turn performance.
  • ...and 3 more figures

Theorems & Definitions (4)

  • Theorem 1: Distribution Matching Property of AFPO
  • Theorem 2: Entropy Lower Bound
  • Theorem 3: Prefix Consistency of Affective Flow
  • Theorem 4: Consistency Supports Long-Horizon Stability