Challenge: Subquadratic architectures rely on aggressive state compression that degrades performance on complex reasoning tasks.
Approach: They propose a framework that confers fixed-memory inference onto pretrained Transformers . they use a learnable module that enforces an adaptive sparsity policy over the context history .
Outcome: The proposed framework outperforms state-of-the-art linearized baselines on the GSM8K benchmark by over 36 points under identical memory constraints.

Similar Papers

ABC: Attention with Bounded-memory Control (2022.acl-long)

Copied to clipboard

Challenge: Existing approaches to attention with bounded-memory control (ABC) have a quadratic complexity in sequence lengths, making it prohibitive for long sequences.
Approach: They propose a new abstraction that bounds memory size to improve efficiency . they propose bounded-memory control, which connects several efficient attention variants .
Outcome: The proposed approach outperforms existing approaches on language modeling, machine translation, and masked language model finetuning.
Lizard: An Efficient Linearization Framework for Large Language Models (2026.acl-long)

Copied to clipboard

Challenge: Existing linearization frameworks that rely on softmax attention with quadratic time and memory complexity pose significant computational and memory bottlenecks for long-context applications.
Approach: They propose a linearization framework that transforms pretrained Transformer-based Large Language Models into subquadratic architectures that closely approximate softmax attention while preserving model quality.
Outcome: Experiments show that the proposed framework outperforms existing methods by 9.4 - 24.5 points on the 5-shot MMLU benchmark and demonstrating superior associative recall.
Context Compression for Auto-regressive Transformers with Sentinel Tokens (2023.emnlp-main)

Copied to clipboard

Challenge: Existing Transformer-based LLMs have limited performance due to complexity of attention module . key-value cache is the major memory footprint and inference latency problem .
Approach: They propose a plug-and-play approach that incrementally compresses token activation into compact ones . they also profile the benefit of context compression on improving the system throughout .
Outcome: The proposed approach reduces memory footprint and inference latency by compressing tokens into compact ones.
DELTA: Dynamic Layer-Aware Token Attention for Efficient Long-Context Reasoning (2026.findings-acl)

Copied to clipboard

Challenge: Large reasoning models generate long chains of intermediate steps, but their inference cost is dominated by decoding, where each new token must attend to the entire growing sequence.
Approach: They propose a training-free sparse attention mechanism that reduces inference cost by evicting entries from the key-value cache.
Outcome: The proposed model matches or surpasses full attention on reasoning benchmarks . it reduces the number of attended tokens by up to 4.25 and delivers 1.54 speedup .
Sparsifying Transformer Models with Trainable Representation Pooling (2022.acl-long)

Copied to clipboard

Challenge: Existing approaches to sparsify attention in the Transformer model are based on quadratic memory complexity and a lack of information for each word.
Approach: They propose a method to sparsify attention in a Transformer model by learning to select the most-informative token representations during the training process.
Outcome: The proposed model performs better than the current SOTA model while being 1.8 faster during training, 4.5 faster inference and 13 more efficient in the decoder.
The Sparse Frontier: Sparse Attention Trade-offs in Transformer LLMs (2026.findings-acl)

Copied to clipboard

Challenge: Sparse attention is a promising strategy to extend long-context capabilities in LLMs . but its efficiency–accuracy trade-offs remain unclear due to the lack of comprehensive evaluation .
Approach: They evaluate sparse attention methods across multiple model families and sizes . they find larger sparser models outperform smaller dense ones at equivalent cost .
Outcome: The proposed methods outperform smaller sparse models at equivalent cost and improve the Pareto frontier.
SCA: Selective Compression Attention for Efficiently Extending the Context Window of Large Language Models (2024.findings-emnlp)

Copied to clipboard

Challenge: Existing methods to compress the KV cache of large language models are expensive and limited in their context window and cost.
Approach: They propose a method to expand the context window and reduce memory footprint by compressing the KV cache of large language models.
Outcome: The proposed method can reduce memory footprint and expand context window of large language models without training.
Latent-Condensed Transformer for Efficient Long Context Modeling (2026.acl-long)

Copied to clipboard

Challenge: Existing approaches address these bottlenecks separately: Multi-head Latent Attention (MLA) reduces the KV cache by projecting tokens into a low-dimensional latent space, while sparse attention reduces computation.
Approach: They propose a Latent-Condensed Attention mechanism that performs structured context condensation directly within MLA's latent space.
Outcome: The proposed approach reduces KV cache size and attention cost without adding parameters.
Layer-Condensed KV Cache for Efficient Inference of Large Language Models (2024.acl-long)

Copied to clipboard

Challenge: Using a key-value cache, memory consumption is a bottleneck for high-throughput language models.
Approach: They propose a method that only computes and caches the KVs of a small number of layers, thus saving memory consumption and improving inference throughput.
Outcome: The proposed method achieves higher throughput and competitive performance than standard transformers and is orthogonal to existing transformer memory-saving techniques.
ReGLA: Refining Gated Linear Attention (2025.naacl-long)

Copied to clipboard

Challenge: Recent advances in Large Language Models (LLMs) are known for their computational and storage requirements due to the quadratic computation complexity of softmax attention.
Approach: They propose to reduce the quadratic computation complexity of softmax attention by using feature maps, normalization and the gating mechanism to improve performance.
Outcome: The proposed model outperforms existing gated linear attention models in extensive tasks including training from scratch and post-linearization with continual pre-training.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations