LoKiFormer: Locality-aware Attention with Decoupled Knowledge Memory for Efficient Large Language Model Pretraining

Representation-Aware Modularity: Efficient Cross-Task Generalization for LLMs

GLA-LoRA: Parameter-efficient LLM fine-tuning with global-local knowledge alignment.

The rapid advancement of large language models (LLMs) has revolutionized natural language processing, yet their enormous parameter size presents significant challenges for fine-tuning on downstream tasks. While parameter-efficient fine-tuning (PEFT) methods such as Low-Rank Adaptation (LoRA) and Quantized LoRA (QLoRA) have substantially reduced computational requirements, they often introduce knowledge misalignment and representation degradation during adaptation. To bridge this gap, we propose GLA-LoRA, a novel PEFT framework centered on Global-Local Knowledge Alignment. Our approach establishes a unified learning strategy that synergistically integrates multi-granular contrastive learning with knowledge distillation. GLA-LoRA explicitly preserves dataset-wide semantic structures through global contrastive learning and maintains sample-level feature consistency via local invariance learning, while a teacher model guides the adaptation to retain the original model's knowledge. Extensive evaluations across eight benchmarks, spanning six GLUE tasks and two intent detection datasets, demonstrate that GLA-LoRA consistently outperforms strong baselines including standard LoRA and QLoRA. Under identical parameter budgets, GLA-LoRA achieves consistent and significant improvements over standard LoRA across three base models and eight benchmark datasets. On sentence-level classification and regression tasks, our method yields average accuracy gains of 1.4%, 1.1%, and 1.2% on LLaMA2-7B, RoBERTabase, and BERTLarge, respectively. For intent detection (CLINC and HWU), it achieves even more pronounced improvements, with up to 2.3% accuracy gain on BERTLarge. These results establish that explicit global-local knowledge alignment is essential for achieving high-fidelity, parameter-efficient fine-tuning across diverse language tasks.

Attention-Free and Lightweight Token Reduction for Efficient Vision-Language Models

Vision-Language Models (VLMs) have achieved strong performance in multimodal understanding, yet remain challenging to deploy on resource-constrained edge devices due to the substantial computational overhead of processing numerous visual tokens. Token reduction is a promising direction for accelerating VLMs inference, but existing approaches either rely on attention maps that are incompatible with modern acceleration frameworks or depend on computationally intensive pairwise similarity comparisons, which undermine scalability and negate their practical benefits in deployment. In this paper, we propose an attention-free and lightweight token reduction framework as a plug-and-play module for VLMs, which preserves both important and diverse tokens to produce a compact visual representation. First, to enable attention-free importance estimation, we adopt an information-theoretic perspective and quantify token information using a novel entropy-based criterion, retaining those with more expressive and less degenerate feature representations. Second, to ensure diverse visual coverage in a lightweight manner, we introduce a transformation-induced consistency signal where similar tokens yield similar signals, such that sorting by this signal places similar tokens close to each other and enables stride-based selection to produce a diverse token set. Extensive experiments across multiple VLMs benchmarks demonstrate that our framework achieves a favorable accuracy-efficiency trade-off, maintaining competitive performance under aggressive compression.

Hierarchical Sparse Attention Done Right: Toward Infinite Context Modeling

Scaling modern large language models (LLMs) to long contexts is limited by the quadratic computation cost, and poor length extrapolation of dense attention. Chunk-wise sparse attention offers a promising alternative, but all existing methods fall short of full attention because of their inaccurate chunk selection. We propose Hierarchical Landmark Sparse (HiLS) Attention, a chunk-wise sparse attention mechanism that learns chunk selection end-to-end under the language-modeling (LM) loss. HiLS factorizes attention hierarchically: each query performs attention independently with each retrieved chunk to extract chunk-specific information, and the resulting outputs are fused according to chunk retrieval scores. By incorporating retrieval scores into the forward attention computation, HiLS optimizes them directly with the LM loss, enabling end-to-end retrieval learning and native sparse training. Experimental results show that HiLS-Attention achieves performance comparable to, and in some cases better than, full attention at in-domain context lengths. Meanwhile, HiLS-Attention extrapolates more than $64\times$ the training context length with 90% retrieval accuracy, far beyond full attention. Moreover, existing full-attention models can be converted to HiLS-Attention with lightweight continued pretraining, preserving in-domain performance while acquiring ultra-long-context extrapolation. Together with its sparse KV access and computation, HiLS-Attention breaks the usual efficiency-performance trade-off, enabling long-context LLMs that are both more efficient and more effective on general long-context tasks than their full-attention counterparts.

Co-LMLM: Continuous-Query Limited Memory Language Models

Limited memory language models (LMLMs) externalize factual knowledge during pretraining to a knowledge base (KB), rather than memorizing it in their weights. During generation, the model then fetches knowledge from the KB as needed. This recently introduced paradigm provides multiple advantages, including knowledge control capabilities that remain beyond conventional LLMs. We propose continuous-query LMLM (CO-LMLM), where the KB pairs continuous keys with textual knowledge values, a significant departure from prior reliance on relational KB and queries. CO-LMLM generates flexible vector queries at minimal cost, while still integrating human-readable and attributable retrieved knowledge into its generation. We pair this design with an annotation pipeline that tags free-form factual spans in arbitrary text, removing prior work's restriction to Wikipedia. Across pretraining on Wikipedia and FineWeb-Edu and at multiple model scales, CO-LMLM outperforms prior LMLMs and vanilla LLMs in both perplexity and factual precision. At 360M scale, this includes lower perplexity than models pretrained on 40x more data, and SimpleQA-verified performance that is in line with gpt-4o-mini and higher than Claude Sonnet 4.5.

MemDefrag: Latent Memory Defragmentation for Large Language Models

Latent memory, which stores past knowledge fragments as per-layer hidden states, has emerged as a promising paradigm (e.g., MemoryLLM and M+) for long-term memory in large language models (LLMs). However, the paradigm suffers from significant performance degradation during memory updates, due to positional encoding misalignment and the absence of any tracing mechanism to distinguish target memory fragments from irrelevant ones. To discover such a tracing mechanism, we probe the layer-wise attention density over stored memory fragments, and find that a small set of middle transformer layers consistently concentrates the highest density on the target fragment - exposing an inherent tracing signal. In light of this, we propose MemDefrag, a training-free and model-agnostic framework that (1) uses a middle-layer tracing signal to conduct memory defragmentation (rank, reorder, and filter memories), and (2) applies an informativeness-guided proportional forgetting mechanism once capacity is exceeded. Experiments show that MemDefrag substantially outperforms MemoryLLM and M+ on knowledge retention (e.g., 43.0% vs. 17.4%/17.6% after 50 memory updates) and long-context benchmarks, and generalizes well across various LLMs and latent-memory variants. The code is available at github.com/ryehr/MemDefrag.