Optimizer choice during LLM fine-tuning produces up to 7x variation in emergent misalignment rates, with spectral regularization on LoRA adapters substantially mitigating misalignment for prone optimizers.
International Conference on Learning Representations , year=
9 Pith papers cite this work. Polarity classification is still indexing.
9
Pith papers citing it
verdicts
UNVERDICTED 9representative citing papers
MixTTA equips normalization layers with low-rank cross-channel transformations plus decoupling and spectral projections to correct cross-channel structural changes during test-time adaptation.
The work derives the optimal ratio of dynamics-to-reward samples that minimizes a bound on return error and characterizes the tradeoff between noisy but cheap rewards versus accurate but expensive ones in imagination-based policy optimization.
PHALAR achieves up to 70% relative accuracy gain in stem retrieval over prior art using under half the parameters and 7x faster training by enforcing musical equivariances via spectral pooling and complex heads.
SPACO is a single-loop penalty-based stochastic algorithm for minimax optimization with nonlinear coupled constraints, achieving non-asymptotic complexity bounds and asymptotic convergence to enhanced KKT points.
DAPPr projects a possibilistic posterior over network parameters to predictions using supremum operators and approximates it with learnable Dirichlet functions to yield an efficient training objective for epistemic uncertainty.
Latent diffusion models exhibit geometric decoupling where curvature in out-of-distribution generation is misallocated to unstable semantic boundaries instead of image details, identifying geometric hotspots as the structural cause of editing instability.
A single-objective rectified flow variant uses neural ODEs trained by regression to monotonically decrease a fixed convex transport cost while preserving marginal distributions.
LHSD estimates local intrinsic dimension in high-D spaces by spectral filtering of the log-density Hessian via SLQ to isolate zero-curvature tangent directions.
citing papers explorer
-
Evil Spectra: How Optimisers can Amplify or Suppress Emergent Misalignment
Optimizer choice during LLM fine-tuning produces up to 7x variation in emergent misalignment rates, with spectral regularization on LoRA adapters substantially mitigating misalignment for prone optimizers.
-
MixTTA: Low-Rank Cross-Channel Mixing for Reliable Test-Time Adaptation
MixTTA equips normalization layers with low-rank cross-channel transformations plus decoupling and spectral projections to correct cross-channel structural changes during test-time adaptation.
-
On Training in Imagination
The work derives the optimal ratio of dynamics-to-reward samples that minimizes a bound on return error and characterizes the tradeoff between noisy but cheap rewards versus accurate but expensive ones in imagination-based policy optimization.
-
PHALAR: Phasors for Learned Musical Audio Representations
PHALAR achieves up to 70% relative accuracy gain in stem retrieval over prior art using under half the parameters and 7x faster training by enforcing musical equivariances via spectral pooling and complex heads.
-
A Single-Loop Penalty-based Algorithm for Stochastic Minimax Optimization with Nonlinear Coupled Constraints
SPACO is a single-loop penalty-based stochastic algorithm for minimax optimization with nonlinear coupled constraints, achieving non-asymptotic complexity bounds and asymptotic convergence to enhanced KKT points.
-
Possibilistic Predictive Uncertainty for Deep Learning
DAPPr projects a possibilistic posterior over network parameters to predictions using supremum operators and approximates it with learnable Dirichlet functions to yield an efficient training objective for epistemic uncertainty.
-
Geometric Decoupling: Diagnosing the Structural Instability of Latent
Latent diffusion models exhibit geometric decoupling where curvature in out-of-distribution generation is misallocated to unstable semantic boundaries instead of image details, identifying geometric hotspots as the structural cause of editing instability.
-
Rectified Flow: A Marginal Preserving Approach to Optimal Transport
A single-objective rectified flow variant uses neural ODEs trained by regression to monotonically decrease a fixed convex transport cost while preserving marginal distributions.
-
Local Hessian Spectral Filtering for Robust Intrinsic Dimension Estimation
LHSD estimates local intrinsic dimension in high-D spaces by spectral filtering of the log-density Hessian via SLQ to isolate zero-curvature tangent directions.