Requiring LICQ/SCS/SOSC everywhere in bilevel optimization is non-prevalent and rigid, while holding almost everywhere is prevalent, but the distinction introduces fundamental difficulties.
SEUF: Is unlearning one expert enough for mixture-of-experts LLMs?
5 Pith papers cite this work. Polarity classification is still indexing.
5
Pith papers citing it
citation-role summary
method 1
citation-polarity summary
roles
method 1polarities
use method 1representative citing papers
A deterministic single-loop cubic regularized Newton method for NCSC bilevel optimization that attains the optimal O(ε^{-1.5}) SOSP rate without repeated lower-level solves.
CURE disentangles LLM recommendation circuits into forget-specific, retain-specific, and task-shared modules with tailored update rules to achieve more effective unlearning than weighted baselines.
RQ-TTSA achieves O(T^{-(p-1)/(3p-2)}) convergence for nonconvex-strongly convex bilevel optimization under heavy-tailed noise (p in (1,2]) via quantile-guided Huber clipping and shows empirical gains on vision, games, and RL tasks.
Downgrading optimizers to lower-information variants during LLM unlearning yields more robust forgetting on MUSE and WMDP benchmarks by converging to harder-to-perturb loss basins.
citing papers explorer
-
On the Nature of Regularity Assumptions in Bilevel Optimization with Constrained Lower-level Problem
Requiring LICQ/SCS/SOSC everywhere in bilevel optimization is non-prevalent and rigid, while holding almost everywhere is prevalent, but the distinction introduces fundamental difficulties.
-
On Second-Order Methods for Bilevel Optimization
A deterministic single-loop cubic regularized Newton method for NCSC bilevel optimization that attains the optimal O(ε^{-1.5}) SOSP rate without repeated lower-level solves.
-
CURE:Circuit-Aware Unlearning for LLM-based Recommendation
CURE disentangles LLM recommendation circuits into forget-specific, retain-specific, and task-shared modules with tailored update rules to achieve more effective unlearning than weighted baselines.
-
Distribution-Aware Robust Bilevel Optimization: Quantile-Guided Huber Updates in Two-Timescale Stochastic Approximation
RQ-TTSA achieves O(T^{-(p-1)/(3p-2)}) convergence for nonconvex-strongly convex bilevel optimization under heavy-tailed noise (p in (1,2]) via quantile-guided Huber clipping and shows empirical gains on vision, games, and RL tasks.
-
Downgrade to Upgrade: Optimizer Simplification Enhances Robustness in LLM Unlearning
Downgrading optimizers to lower-information variants during LLM unlearning yields more robust forgetting on MUSE and WMDP benchmarks by converging to harder-to-perturb loss basins.