Image de couverture de livre

Shared Parameter Subspaces and Cross-Task Linearity in Emergently Misaligned Behavior

Fuente: arXiv

Enregistré dans:

Détails bibliographiques
Auteurs principaux:	Arturi, Daniel Aarao Reis, Zhang, Eric, Ansah, Andrew, Zhu, Kevin, Panda, Ashwinee, Balwani, Aishwarya
Format:	Preprint
Publié:	2025
Sujets:	Machine Learning Artificial Intelligence
Accès en ligne:	Acceder al recurso
Tags:	Ajouter un tag Pas de tags, Soyez le premier à ajouter un tag!

Documents similaires

LoRI: Reducing Cross-Task Interference in Multi-Task Low-Rank Adaptation
par: Zhang, Juzheng, et autres
Publié: (2025)

Decomposing Behavioral Phase Transitions in LLMs: Order Parameters for Emergent Misalignment
par: Arnold, Julian, et autres
Publié: (2025)

Convergent Linear Representations of Emergent Misalignment
par: Soligo, Anna, et autres
Publié: (2025)

A New Linear Scaling Rule for Private Adaptive Hyperparameter Optimization
par: Panda, Ashwinee, et autres
Publié: (2022)

Model Organisms for Emergent Misalignment
par: Turner, Edward, et autres
Publié: (2025)

Amortized Latent Steering: Low-Cost Alternative to Test-Time Optimization
par: Egbuna, Nathan, et autres
Publié: (2025)

Speculating Experts Accelerates Inference for Mixture-of-Experts
par: Madan, Vivan, et autres
Publié: (2026)

Understanding Emergent Misalignment via Feature Superposition Geometry
par: Minegishi, Gouki, et autres
Publié: (2026)

In-Training Defenses against Emergent Misalignment in Language Models
par: Kaczér, David, et autres
Publié: (2025)

Peek-a-Boo Reasoning: Contrastive Region Masking in MLLMs
par: Chaturvedi, Isha, et autres
Publié: (2025)

CommonKV: Compressing KV Cache with Cross-layer Parameter Sharing
par: Wang, Yixuan, et autres
Publié: (2025)

Alignment-Constrained Dynamic Pruning for LLMs: Identifying and Preserving Alignment-Critical Circuits
par: Patel, Dev, et autres
Publié: (2025)

From Narrow Unlearning to Emergent Misalignment: Causes, Consequences, and Containment in LLMs
par: Mushtaq, Erum, et autres
Publié: (2025)

Persona-Model Collapse in Emergent Misalignment
par: Costa, Davi Bastos, et autres
Publié: (2026)

Thought Crime: Backdoors and Emergent Misalignment in Reasoning Models
par: Chua, James, et autres
Publié: (2025)

FRIT: Using Causal Importance to Improve Chain-of-Thought Faithfulness
par: Swaroop, Anand, et autres
Publié: (2025)

Shared Lexical Task Representations Explain Behavioral Variability In LLMs
par: Yang, Zhuonan, et autres
Publié: (2026)

Emergent and Subliminal Misalignment Through the Lens of Data-Mediated Transfer
par: Askin, Baris, et autres
Publié: (2026)

Persona Features Control Emergent Misalignment
par: Wang, Miles, et autres
Publié: (2025)

Dense Backpropagation Improves Training for Sparse Mixture-of-Experts
par: Panda, Ashwinee, et autres
Publié: (2025)

The Devil in the Details: Emergent Misalignment, Format and Coherence in Open-Weights LLMs
par: Dickson, Craig
Publié: (2025)

The Master Key Hypothesis: Unlocking Cross-Model Capability Transfer via Linear Subspace Alignment
par: Balasubramanian, Rishab, et autres
Publié: (2026)

Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs
par: Betley, Jan, et autres
Publié: (2025)

Privacy Auditing of Large Language Models
par: Panda, Ashwinee, et autres
Publié: (2025)

Distilling Linearized Behavior into Non-Linear Fine-Tuning for Effective Task Arithmetic
par: Sommariva, Thomas, et autres
Publié: (2026)

On the Emergence of Cross-Task Linearity in the Pretraining-Finetuning Paradigm
par: Zhou, Zhanpeng, et autres
Publié: (2024)

ROGUE: Misaligned Agent Behavior Arising from Ordinary Computer Use
par: Tien, Jeremy, et autres
Publié: (2026)

Emergent Misalignment is Easy, Narrow Misalignment is Hard
par: Soligo, Anna, et autres
Publié: (2026)

QMP: Q-switch Mixture of Policies for Multi-Task Behavior Sharing
par: Zhang, Grace, et autres
Publié: (2023)

SALT: Steering Activations towards Leakage-free Thinking in Chain of Thought
par: Batra, Shourya, et autres
Publié: (2025)

PASs-MoE: Mitigating Misaligned Co-drift among Router and Experts via Pathway Activation Subspaces for Continual Learning
par: Hou, Zhiyan, et autres
Publié: (2026)

CoPS: Empowering LLM Agents with Provable Cross-Task Experience Sharing
par: Yang, Chen, et autres
Publié: (2024)

Re-Emergent Misalignment: How Narrow Fine-Tuning Erodes Safety Alignment in LLMs
par: Giordani, Jeremiah
Publié: (2025)

Moloch's Bargain: Emergent Misalignment When LLMs Compete for Audiences
par: El, Batu, et autres
Publié: (2025)

Teach LLMs to Phish: Stealing Private Information from Language Models
par: Panda, Ashwinee, et autres
Publié: (2024)

Overtrained, Not Misaligned
par: Schreiber, Joel, et autres
Publié: (2026)

Characterizing the Consistency of the Emergent Misalignment Persona
par: Weckauff, Anietta, et autres
Publié: (2026)

Gemstones: A Model Suite for Multi-Faceted Scaling Laws
par: McLeish, Sean, et autres
Publié: (2025)

BLOCK-EM: Preventing Emergent Misalignment via Latent Blocking
par: Ustaomeroglu, Muhammed, et autres
Publié: (2026)

The Latent Color Subspace: Emergent Order in High-Dimensional Chaos
par: Pach, Mateusz, et autres
Publié: (2026)