Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Reinforcement Tuning for Detecting Stances and Debunking Rumors Jointly with Large Language Models

Ruichao Yang, Wei Gao, Jing Ma, Hongzhan Lin, Bo Wang

TL;DR

The paper tackles the data-scarce problem of jointly detecting stance and verifying rumors by introducing JSDRV, a reinforcement-tuned framework that uses LLMs as annotators for both tasks and a two-level selector policy to curate high-quality training data from a small seed set. It combines an LLM-based stance detector, an LLM-based rumor verifier, and a reinforcement learning–driven selector to iteratively improve labeling quality and model performance, allowing fine-tuning via PEFT techniques such as LoRA. Across multiple rumor-stance datasets, JSDRV outperforms strong baselines, including fully supervised and weakly supervised approaches, and generalizes to non-LLM task models. The work highlights the potential of reinforcement-guided data selection and prompt-based LLM fine-tuning for data-efficient, joint social-media understanding tasks, while acknowledging limitations related to distribution-based rewards and training overhead.

Abstract

Learning multi-task models for jointly detecting stance and verifying rumors poses challenges due to the need for training data of stance at post level and rumor veracity at claim level, which are difficult to obtain. To address this issue, we leverage large language models (LLMs) as the foundation annotators for the joint stance detection (SD) and rumor verification (RV) tasks, dubbed as JSDRV. We introduce a novel reinforcement tuning framework to enhance the joint predictive capabilities of LLM-based SD and RV components. Specifically, we devise a policy for selecting LLM-annotated data at the two levels, employing a hybrid reward mechanism to choose high-quality labels for effective LLM fine-tuning on both tasks. Results demonstrate that JSDRV improves the capabilities of LLMs in the joint tasks, not only outperforming state-of-the-art methods but also generalizing to non-LLMs accommodated as task models.

Reinforcement Tuning for Detecting Stances and Debunking Rumors Jointly with Large Language Models

TL;DR

The paper tackles the data-scarce problem of jointly detecting stance and verifying rumors by introducing JSDRV, a reinforcement-tuned framework that uses LLMs as annotators for both tasks and a two-level selector policy to curate high-quality training data from a small seed set. It combines an LLM-based stance detector, an LLM-based rumor verifier, and a reinforcement learning–driven selector to iteratively improve labeling quality and model performance, allowing fine-tuning via PEFT techniques such as LoRA. Across multiple rumor-stance datasets, JSDRV outperforms strong baselines, including fully supervised and weakly supervised approaches, and generalizes to non-LLM task models. The work highlights the potential of reinforcement-guided data selection and prompt-based LLM fine-tuning for data-efficient, joint social-media understanding tasks, while acknowledging limitations related to distribution-based rewards and training overhead.

Abstract

Learning multi-task models for jointly detecting stance and verifying rumors poses challenges due to the need for training data of stance at post level and rumor veracity at claim level, which are difficult to obtain. To address this issue, we leverage large language models (LLMs) as the foundation annotators for the joint stance detection (SD) and rumor verification (RV) tasks, dubbed as JSDRV. We introduce a novel reinforcement tuning framework to enhance the joint predictive capabilities of LLM-based SD and RV components. Specifically, we devise a policy for selecting LLM-annotated data at the two levels, employing a hybrid reward mechanism to choose high-quality labels for effective LLM fine-tuning on both tasks. Results demonstrate that JSDRV improves the capabilities of LLMs in the joint tasks, not only outperforming state-of-the-art methods but also generalizing to non-LLMs accommodated as task models.
Paper Structure (40 sections, 5 equations, 8 figures, 6 tables)

This paper contains 40 sections, 5 equations, 8 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Problem Statement
  4. Our JSDRV Framework
  5. LLM Stance Detection Network
  6. $\epsilon$-greedy Controller for Post Pre-Selection
  7. Stance Prompt Learning
  8. Pretraining
  9. LLM Rumor Verification Network
  10. Pretraining
  11. Selector Policy
  12. State
  13. Action
  14. Reward
  15. Model Training
  16. ...and 25 more sections

Figures (8)

  • Figure 1: A false rumor claim and the stances of responding posts. The arrow lines indicate the direction of inference for claim veracity and posts stances.
  • Figure 2: Our reinforcement fine-tuning framework for joint stance detection and rumor verification.
  • Figure 3: Visualization of stance distributions of original and selected posts.
  • Figure 4: An example of stance prompt.
  • Figure 5: An example of rumor prompt.
  • ...and 3 more figures