Geometry Conflict: Explaining and Controlling Forgetting in LLM Continual Post-Training

Technical Summary: Geometry Conflict: Explaining and Controlling Forgetting in LLM Continual Post-Training

Problem Statement
Continual post-training aims to extend Large Language Models (LLMs) with new knowledge, skills, and behaviors through a sequence of updates. However, this process faces the challenge of catastrophic forgetting, where learning a new task undermines previously acquired capabilities. Existing mitigation strategies—such as sequential fine-tuning, replay-based methods, regularization, and model merging—often lack principled criteria for determining when a new parameter update should be integrated. Specifically, it remains unclear when sequential updates enable capability transfer versus when they cause interference. Current methods often rely on coarse metrics like update magnitude or isolated pairwise compatibility, which fail to fully explain or control forgetting in heterogeneous, sequential settings.

Methodology: Task Geometry and GCWM
The authors propose a framework based on task geometry to analyze and control continual post-training.

1. Task Geometry Representation: Instead of viewing updates solely by their norm, the method represents each task by its parameter update matrix $\Delta_t$. For each layer, the "task geometry" is defined as the covariance matrix $C_t = \Delta_t^\top \Delta_t$, which captures the dominant directions and spectral structure of the update.
2. Geometry Conflict: The paper introduces geometry conflict as a normalized Bures–Wasserstein distance between the covariance geometries of tasks. Crucially, the authors define a state-relative variant, measuring the discrepancy between an incoming task's geometry and the geometry of the evolving model state (which already incorporates previous updates).
3. Key Insight: Through empirical analysis across Qwen3 models (0.6B–14B), the authors find that forgetting is best explained as a state-relative update-integration failure. Forgetting occurs when the covariance geometry induced by a new task misaligns with the geometry of the current model state. Transfer occurs when the new update remains compatible with this state. This state-relative signal outperforms raw update norms, subspace alignment ratios (SAR), and pairwise gradient conflicts in predicting retention loss.
4. Geometry-Conflict Wasserstein Merging (GCWM): Motivated by these findings, the authors propose GCWM, a data-free update-integration method.
   • Shared Metric Construction: GCWM constructs a shared Wasserstein metric for the current set of active tasks using Gaussian Wasserstein barycenters.
   • Conflict-Gated Correction: It computes a layer-wise geometry conflict score. This score gates a geometry-aware correction branch. If the conflict is high (indicating incompatibility), the method restrains the integration of the geometry-aware update; if conflict is low, it allows stronger integration.
   • Incremental Update: The method applies only the incremental change relative to the previous merged state, ensuring the model evolves based on newly induced compatibility-controlled changes.
   • Theoretical Support: The authors provide theoretical bounds showing that the additional loss incurred by GCWM relative to a plain merge is controlled by the geometry conflict and the gated merge displacement.

Key Contributions

1. Analytical Framework: The paper develops a task-geometry analysis showing that forgetting is a state-relative phenomenon, better explained by the misalignment of task-induced covariance geometries with the evolving model state than by update magnitude or isolated pairwise compatibility.
2. New Signal: It introduces geometry conflict (a Bures–Wasserstein distance over task-induced covariance geometries) as both an explanatory signal for forgetting and a practical control signal for update integration.
3. Algorithm: It proposes GCWM, a data-free method that constructs a shared Wasserstein metric and uses geometry conflict to gate geometry-aware corrections layer-wise.
4. Empirical Validation: The method is validated on Qwen3 models (0.6B to 14B) across domain-continual (14 MMLU-Pro sub-domains) and capability-continual (math and code) settings.

Results

• Performance: GCWM consistently outperforms data-free baselines (including Localize-and-Stitch, AIMMerging, and OPCM) without using replay data.
  • In domain-continual settings, GCWM achieves the strongest non-MTL (Multi-Task Learning) overall performance across all scales, improving over the best data-free baseline by up to +1.61 points on Qwen3-1.7B.
  • In capability-continual settings, GCWM leads data-free baselines on average across 1.7B and 14B models, showing significant gains in retention and final performance on benchmarks like GSM8K, HumanEval, and MMLU-Pro.
• Ablation: Removing the conflict gate or replacing the Wasserstein barycenter with a simple mean metric degrades performance, confirming the necessity of both the shared metric and the conflict-controlled gating mechanism.
• Scalability: The method remains effective across model scales, with the state-relative geometry signal becoming a stronger predictor of forgetting as model size increases (e.g., correlation with retention loss reaches -0.86 for Qwen3-14B).

Significance and Claims
The paper claims that geometry conflict serves a dual role:

1. Explanatory: It provides a mechanistic explanation for catastrophic forgetting, identifying it as a failure of state-relative update integration rather than simple parameter drift.
2. Control: It acts as a practical, data-free signal to control how strongly sequential updates are integrated.

The authors position GCWM as a method that makes continual LLM adaptation more practical by reducing reliance on replay data. They emphasize that while the method improves retention and performance, it operates in parameter-update space and does not inherently provide safety guarantees; thus, it should be combined with standard safety checks. The work suggests that understanding the spectral and geometric structure of updates is critical for managing the trade-off between learning new capabilities and retaining old ones in large language models.