[2402.09221] Spectral Filters, Dark Signals, and Attention Sinks

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs. arXiv Operational Status Get status notifications via email or slack

[2402.09221] Spectral Filters, Dark Signals, and Attention Sinks Skip to main content arXiv is now an independent nonprofit! Learn more × Search arXiv Press Enter to search · Advanced search --> Computer Science > Artificial Intelligence arXiv:2402.09221 (cs) [Submitted on 14 Feb 2024] Title: Spectral Filters, Dark Signals, and Attention Sinks Authors: Nicola Cancedda View a PDF of the paper titled Spectral Filters, Dark Signals, and Attention Sinks, by Nicola Cancedda View PDF HTML (experimental) Abstract: Projecting intermediate representations onto the vocabulary is an increasi

related reading

• [2401.05566] Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Trainingarxiv.org
• [2512.19428] Attention Is Not What You Needarxiv.org
• [2406.11717] Refusal in Language Models Is Mediated by a Single Directionarxiv.org
• [2012.09852] SpAtten: Efficient Sparse Attention Architecture with Cascade Token and Head Pruningarxiv.org
• 2502.11089arxiv.org
• Sparse Autoencoders Find Highly Interpretable Features in Language Modelsarxiv.org
• [2312.15159] Understanding the Potential of FPGA-Based Spatial Acceleration for Large Language Model Inferencearxiv.org
• Spotting LLMs With Binoculars: Zero-Shot Detection of Machine-Generated Textarxiv.org
• [2406.02061] Alice in Wonderland: Simple Tasks Showing Complete Reasoning Breakdown in State-Of-the-Art Large Language Modelsarxiv.org
• [2309.01809] Are Emergent Abilities in Large Language Models just In-Context Learning?arxiv.org
• Sparse Attention Post-Training for Mechanistic Interpretabilityarxiv.org
• [2401.06102] Patchscopes: A Unifying Framework for Inspecting Hidden Representations of Language Modelsarxiv.org