S$^3$-Attention:Attention-Aligned Endogenous Retrieval for Memory-Bounded Long-Context Inference

Large language models are increasingly applied to multi-document and long-form inputs, yet long-context inference remains memory- and noise-inefficient. Key-value (KV) caching scales linearly with context length, while external retrieval methods often return lexically similar but causally irrelevant passages. We present S3-Attention, a memory-first inference-time framework that treats long-context processing as attention-aligned endogenous retrieval. S3-Attention decodes transient key and query projections into top-k sparse feature identifiers using lightweight sparse autoencoders, and constructs a CPU-based inverted index mapping features to token positions or spans during a single streaming scan. This design allows the KV cache to be discarded entirely and bounds GPU memory usage by the scan chunk size. At generation time, feature co-activation is used to retrieve compact evidence spans, optionally fused with BM25 for exact lexical matching. Under a unified LongBench evaluation protocol with fixed prompting, decoding, and matched token budgets, S3-Hybrid closely matches full-context inference across multiple model families and improves robustness in several information-dense settings. We also report an engineering limitation of the current prototype, which incurs higher wall-clock latency than optimized full-KV baselines, motivating future kernel-level optimization.

Interpretable Safety Alignment via SAE-Constructed Low-Rank Subspace Adaptation

Parameter-efficient fine-tuning has become the dominant paradigm for adapting large language models to downstream tasks. Low-rank adaptation methods such as LoRA operate under the assumption that task-relevant weight updates reside in a low-rank subspace, yet this subspace is learned implicitly from data in a black-box manner, offering no interpretability or direct control. We hypothesize that this difficulty stems from polysemanticity--individual dimensions encoding multiple entangled concepts. To address this, we leverage pre-trained Sparse Autoencoders (SAEs) to identify task-relevant features in a disentangled feature space, then construct an explicit, interpretable low-rank subspace to guide adapter initialization. We provide theoretical analysis proving that under monosemanticity assumptions, SAE-based subspace identification achieves arbitrarily small recovery error, while direct identification in polysemantic space suffers an irreducible error floor. On safety alignment, our method achieves up to 99.6% safety rate--exceeding full fine-tuning by 7.4 percentage points and approaching RLHF-based methods--while updating only 0.19-0.24% of parameters. Crucially, our method provides interpretable insights into the learned alignment subspace through the semantic grounding of SAE features. Our work demonstrates that incorporating mechanistic interpretability into the fine-tuning process can simultaneously improve both performance and transparency.

Unlocking the Address Book: Dissecting the Sparse Semantic Structure of LLM Key-Value Caches via Sparse Autoencoders

The Key-Value (KV) cache is the primary memory bottleneck in long-context Large Language Models, yet it is typically treated as an opaque numerical tensor. In this work, we propose \textbf{STA-Attention}, a framework that utilizes Top-K Sparse Autoencoders (SAEs) to decompose the KV cache into interpretable ``semantic atoms.'' Unlike standard $L_1$-regularized SAEs, our Top-K approach eliminates shrinkage bias, preserving the precise dot-product geometry required for attention. Our analysis uncovers a fundamental \textbf{Key-Value Asymmetry}: while Key vectors serve as highly sparse routers dominated by a ``Semantic Elbow,'' deep Value vectors carry dense content payloads requiring a larger budget. Based on this structure, we introduce a Dual-Budget Strategy that selectively preserves the most informative semantic components while filtering representational noise. Experiments on Yi-6B, Mistral-7B, Qwen2.5-32B, and others show that our semantic reconstructions maintain perplexity and zero-shot performance comparable to the original models, effectively bridging the gap between mechanistic interpretability and faithful attention modeling.