Beyond State Capacity: Why Retrieval Breaks in Mamba

cs.CL

Submitted: 2026-10-14

Updated: 2026-10-14

Comments: 15 pages, 6 figures. Under review

Code: https://github.com/schwartz-lab-NLP/beyond_state_capacity

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Sources

• Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality
• Hymba: A Hybrid-head Architecture for Small Language Models
• Zamba: A Compact 7B SSM Hybrid Model
• Efficiently Modeling Long Sequences with Structured State Spaces
• How to use and interpret activation patching
• Contextual Representation Flow Across Depth in State-Space and Transformer Architectures
• In-context Learning and Induction Heads
• Probing the Probing Paradigm: Does Probing Accuracy Entail Task Relevance?
• The unreasonable effectiveness of the forget gate
• An Empirical Study of Mamba-based Language Models
• LongMamba: Enhancing Mamba's Long Context Capabilities via Training-Free Receptive Field Enlargement
• Context-Selective State Space Models: Feedback is All You Need

Related papers

• Exploring Solution Divergence and Its Effect on Large Language Model Problem Solving
• Ishigaki-IDS-Bench: A Benchmark for Generating Information Delivery Specification from BIM Information Requirements
• Subliminal Steering: Stronger Encoding of Hidden Signals
• MedStruct-S: A Benchmark for Key Discovery, Key-Conditioned QA and Semi-Structured Extraction from OCR Clinical Reports
• The End of Transformers? On Challenging Attention and the Rise of Sub-Quadratic Architectures
• Untangling the Mechanisms of Misleading Context in Medical Question Answering