Resolving Gradient Conflicts in Machine Unlearning

This direction is merged into another direction: catastrophic forgetting in unlearning, gradient conflict in unlearning. Its last version stays readable and citable here.

Version 1 of 132 papers40 conceptsfrozen 2026-10-04

Machine unlearning is often framed as a simple deletion task, but in deep neural networks, knowledge is not stored in discrete files; it is entangled within millions of shared parameters. When we attempt to "forget" a specific subset of data, we are essentially performing a second round of optimization that seeks to maximize the loss on the "forget set" while maintaining a low loss on the "retain set."

This creates a fundamental tug-of-war. Recent research identifies gradient conflict as the primary reason why unlearning often leads to "catastrophic collapse," where a model successfully forgets the target data but loses its general intelligence or reasoning capabilities in the process.

The Geometry of Optimization Conflict

In a standard training setting, gradients from different data points generally point in a similar direction toward a shared local minimum. However, unlearning introduces a contradictory objective. To forget, the model must move away from the patterns it previously learned.

As Hadi Reisizadeh and colleagues observe in BLUR, the gradients for forgetting and retaining often exhibit negative cosine similarity. This means the mathematical direction required to erase sensitive data points directly opposes the direction required to keep the model useful. When these gradients are combined using a simple weighted sum, the model oscillates between competing goals, leading to unstable updates and poor convergence.

Yue Wang and Qizhou Wang describe this in GRU as a geometric conflict. If the unlearning update is not "rectified," the signal intended to remove harmful information overwhelms the signal for retaining useful capabilities, effectively "lobotomizing" the model.

Gradient Surgery and Orthogonalization

To resolve this conflict, researchers have moved toward "gradient surgery"—techniques that modify the update vector before it is applied to the weights.

The intuition is to find a "safe" direction for forgetting that lies in the null space of the retention objective. Aviv Shamsian and the team behind OrthoGrad achieve this through per-sample gradient orthogonalization. By projecting the unlearning gradients into a subspace that is orthogonal to the gradients of the retain samples, they ensure that the "forgetting" update has zero mathematical impact on the model's performance on the data it is supposed to keep.

Similarly, in the context of diffusion models, Jing Wu and Mehrtash Harandi propose EraseDiff, which uses closed-form solutions to resolve these conflicts efficiently without the massive cost of retraining.

Beyond Weighted Sums: Bi-Level and Game-Theoretic Frameworks

If forgetting and retaining are inherently at odds, perhaps they shouldn't be treated as equal parts of a single loss function. Instead, researchers are framing unlearning as a hierarchical optimization problem.

  1. Bi-Level Optimization: Frameworks like OFMU and BLUR treat forgetting as a non-negotiable constraint. They prioritize the "forgetting" objective in an inner loop and only allow the model to optimize for utility in the outer loop, provided those updates do not interfere with the erasure of sensitive data.
  2. Nash Bargaining: Jing Wu and Mehrtash Harandi take this further in MUNBa by treating the two objectives as players in a cooperative game. By applying Nash bargaining theory, they derive a solution that balances the two goals such that neither objective can be improved without disproportionately harming the other, reaching a "Pareto optimal" state of unlearning.

Ensuring Robustness through Smoothness

A final challenge is that even if a model "forgets" a concept, that knowledge can often be recovered through adversarial attacks or "relearning." This happens when the unlearning process creates a sharp, narrow valley in the loss landscape—the data is hidden, but the underlying weights haven't fundamentally changed.

To combat this, Han Yan and Zheyuan Liu introduced PRISM, which enforces dual-space smoothness. By smoothing both the internal representations and the parameter space, they widen the "margin" between harmful and safe content. This ensures that the unlearning is not just a superficial shift in weights but a robust structural change that resists being undone.

By treating unlearning as a sophisticated conflict-resolution problem rather than a simple fine-tuning task, these methods allow models to adhere to the "right to be forgotten" without sacrificing the intelligence that makes them useful.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_resolving_gradient_conflicts_in_machine__117f7c87,
  title = {Resolving Gradient Conflicts in Machine Unlearning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/resolving-gradient-conflicts-in-machine-unlearning/rdr_117f7c87356d4923bc32c3f02bbee5a7},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_resolving_gradient_conflicts_in_machine__117f7c87_v1,
  title = {Resolving Gradient Conflicts in Machine Unlearning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/resolving-gradient-conflicts-in-machine-unlearning/rdr_117f7c87356d4923bc32c3f02bbee5a7/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}