Qingsen Ma

Qingsen Ma, Dianyun Wang, Jiaming Lyu et al.

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.

Dianyun Wang, Qingsen Ma, Yuhu Shang et al.

Safety alignment -- training large language models (LLMs) to refuse harmful requests while remaining helpful -- is critical for responsible deployment. Prior work established that safety behaviors are governed by low-rank structures, suggesting parameter-efficient fine-tuning (PEFT) should be well-suited for alignment. However, Low-Rank Adaptation (LoRA) consistently underperforms full fine-tuning and reinforcement learning on safety benchmarks. We attribute this gap to semantic entanglement: safety-relevant directions are intertwined with unrelated concepts due to polysemanticity, impeding implicit subspace identification. To address this, we propose SAILS (Safety Alignment via Interpretable Low-rank Subspace), which leverages Sparse Autoencoders (SAEs) to disentangle representations into monosemantic features, constructs an interpretable safety subspace from SAE decoder directions, and uses it to initialize LoRA adapters. Theoretically, we prove that SAE-based identification achieves arbitrarily small recovery error under monosemanticity assumptions, while direct identification suffers an irreducible error floor. Empirically, SAILS achieves up to 99.6% safety rate on Gemma-2-9B -- exceeding full fine-tuning by 7.4 points and matching RLHF-based models -- while updating only 0.19% of parameters and providing interpretability.

Qingsen Ma, Dianyun Wang, Ran Jing et al.

Large language models often hallucinate when processing long and noisy retrieval contexts because they rely on spurious correlations rather than genuine causal relationships. We propose CIP, a lightweight and plug-and-play causal prompting framework that mitigates hallucinations at the input stage. CIP constructs a causal relation sequence among entities, actions, and events and injects it into the prompt to guide reasoning toward causally relevant evidence. Through causal intervention and counterfactual reasoning, CIP suppresses non causal reasoning paths, improving factual grounding and interpretability. Experiments across seven mainstream language models, including GPT-4o, Gemini 2.0 Flash, and Llama 3.1, show that CIP consistently enhances reasoning quality and reliability, achieving 2.6 points improvement in Attributable Rate, 0.38 improvement in Causal Consistency Score, and a fourfold increase in effective information density. API level profiling further shows that CIP accelerates contextual understanding and reduces end to end response latency by up to 55.1 percent. These results suggest that causal reasoning may serve as a promising paradigm for improving the explainability, stability, and efficiency of large language models.

Qingsen Ma, Dianyun Wang, Yaoye Wang et al.

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.

Qingsen Ma, Chen Zou, Dianyun Wang et al.

Under extremely low-light conditions, novel view synthesis (NVS) faces severe degradation in terms of geometry, color consistency, and radiometric stability. Standard 3D Gaussian Splatting (3DGS) pipelines fail when applied directly to underexposed inputs, as independent enhancement across views causes illumination inconsistencies and geometric distortion. To address this, we present DTGS, a unified framework that tightly couples Retinex-inspired illumination decomposition with thermal-guided 3D Gaussian Splatting for illumination-invariant reconstruction. Unlike prior approaches that treat enhancement as a pre-processing step, DTGS performs joint optimization across enhancement, geometry, and thermal supervision through a cyclic enhancement-reconstruction mechanism. A thermal supervisory branch stabilizes both color restoration and geometry learning by dynamically balancing enhancement, structural, and thermal losses. Moreover, a Retinex-based decomposition module embedded within the 3DGS loop provides physically interpretable reflectance-illumination separation, ensuring consistent color and texture across viewpoints. To evaluate our method, we construct RGBT-LOW, a new multi-view low-light thermal dataset capturing severe illumination degradation. Extensive experiments show that DTGS significantly outperforms existing low-light enhancement and 3D reconstruction baselines, achieving superior radiometric consistency, geometric fidelity, and color stability under extreme illumination.