Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Agent-Dice: Disentangling Knowledge Updates via Geometric Consensus for Agent Continual Learning

Zheng Wu, Xingyu Lou, Xinbei Ma, Yansi Li, Weiwen Liu, Weinan Zhang, Jun Wang, Zhuosheng Zhang

TL;DR

This work argues that the stability-plasticity dilemma fundamentally arises from the failure to explicitly distinguish between common knowledge shared across tasks and conflicting knowledge introduced by task-specific interference, and proposes Agent-Dice, a parameter fusion framework based on directional consensus evaluation.

Abstract

Large Language Model (LLM)-based agents significantly extend the utility of LLMs by interacting with dynamic environments. However, enabling agents to continually learn new tasks without catastrophic forgetting remains a critical challenge, known as the stability-plasticity dilemma. In this work, we argue that this dilemma fundamentally arises from the failure to explicitly distinguish between common knowledge shared across tasks and conflicting knowledge introduced by task-specific interference. To address this, we propose Agent-Dice, a parameter fusion framework based on directional consensus evaluation. Concretely, Agent-Dice disentangles knowledge updates through a two-stage process: geometric consensus filtering to prune conflicting gradients, and curvature-based importance weighting to amplify shared semantics. We provide a rigorous theoretical analysis that establishes the validity of the proposed fusion scheme and offers insight into the origins of the stability-plasticity dilemma. Extensive experiments on GUI agents and tool-use agent domains demonstrate that Agent-Dice exhibits outstanding continual learning performance with minimal computational overhead and parameter updates. The codes are available at https://github.com/Wuzheng02/Agent-Dice.

Agent-Dice: Disentangling Knowledge Updates via Geometric Consensus for Agent Continual Learning

TL;DR

This work argues that the stability-plasticity dilemma fundamentally arises from the failure to explicitly distinguish between common knowledge shared across tasks and conflicting knowledge introduced by task-specific interference, and proposes Agent-Dice, a parameter fusion framework based on directional consensus evaluation.

Abstract

Large Language Model (LLM)-based agents significantly extend the utility of LLMs by interacting with dynamic environments. However, enabling agents to continually learn new tasks without catastrophic forgetting remains a critical challenge, known as the stability-plasticity dilemma. In this work, we argue that this dilemma fundamentally arises from the failure to explicitly distinguish between common knowledge shared across tasks and conflicting knowledge introduced by task-specific interference. To address this, we propose Agent-Dice, a parameter fusion framework based on directional consensus evaluation. Concretely, Agent-Dice disentangles knowledge updates through a two-stage process: geometric consensus filtering to prune conflicting gradients, and curvature-based importance weighting to amplify shared semantics. We provide a rigorous theoretical analysis that establishes the validity of the proposed fusion scheme and offers insight into the origins of the stability-plasticity dilemma. Extensive experiments on GUI agents and tool-use agent domains demonstrate that Agent-Dice exhibits outstanding continual learning performance with minimal computational overhead and parameter updates. The codes are available at https://github.com/Wuzheng02/Agent-Dice.
Paper Structure (37 sections, 3 theorems, 31 equations, 8 figures, 7 tables, 1 algorithm)

This paper contains 37 sections, 3 theorems, 31 equations, 8 figures, 7 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. LLM Agent
  4. LLM Continual Learning
  5. Agent-Dice
  6. Theoretical Support
  7. Parameter Space Linearization.
  8. Noise Suppression via Geometric Consistency.
  9. Saliency Maximization via Boltzmann Distribution.
  10. Method: Directional Consensus Evaluation for Parameter Fusion
  11. Stage 1: Geometric Consensus Filtering.
  12. Stage 2: Curvature-based Importance Weighting.
  13. Final Fusion.
  14. Experiments
  15. Implementation
  16. ...and 22 more sections

Key Result

Theorem 1

Assume that for a local neighborhood around $\boldsymbol{\theta}_{\text{pre}}$, the loss function $\mathcal{L}_k$ is approximately linear with respect to $\boldsymbol{\tau}_k$. Let $\mathbf{W} \in \mathbb{R}^{d \times K}$ be a weighting matrix where $\sum_{k=1}^K w_{k,i} = 1$. The update rule $\bold

Figures (8)

  • Figure 1: The stability-plasticity dilemma in agent continual learning (from Task A to Task C). Agent-Dice finds a balance between stability and plasticity by learning common knowledge.
  • Figure 2: The Agent-Dice Parameter Fusion Pipeline. Task vectors $\boldsymbol{\tau}_k$ undergo a two-stage aggregation policy: Geometric Consensus Filtering for variance reduction via outlier pruning, and Curvature-based Importance Weighting for entropy maximization based on parameter saliency. The final refined update is added to $\boldsymbol{\theta}_{\text{pre}}$.
  • Figure 3: Ablation study on GUI Agent tasks.
  • Figure 4: Ablation study on Tool-use Agent tasks.
  • Figure 5: Model similarity analysis in the GUI agent domain. The similarity between Agent-Dice and OS-Atlas-Pro-7B is only marginally lower than that of models trained on a single dataset.
  • ...and 3 more figures

Theorems & Definitions (6)

  • Theorem 1: First-Order Manifold Aggregation
  • proof
  • Theorem 2: Consensus-Induced Variance Reduction
  • proof
  • Theorem 3: Optimal Weighting under Saliency Constraints
  • proof