2026· Annual Meeting of the Association for Computational Linguistics· pp. 35311-35323· 0 citations· 27 references
Computer Science
TL;DR
O CTOPUS is proposed, a framework that confers fixed-memory inference onto pretrained Transform-ers without the information loss of linearization and outperforms state-of-the-art linearized baselines on the GSM8K benchmark, demonstrating that learned sparse retention serves as an effective regular-izer for long-horizon reasoning.
Abstract
Transformer inference becomes increasingly memory-bound as the Key–Value (KV) cache grows linearly with sequence length. While subquadratic architectures offer constant-memory inference, they rely on aggressive state compression that degrades performance on complex reasoning tasks. We propose O CTOPUS , a framework that confers fixed-memory inference onto pretrained Transform-ers without the information loss of linearization. O CTOPUS retrofits attention layers with Gated Selective Attention , a learnable module that enforces an adaptive sparsity policy over the context history. By dynamically scoring and retaining only high-utility KV states, this mechanism transforms the unbounded cache into a compact, evolving memory budget that filters out uninformative noise. Empirically, on the GSM8K benchmark, it outperforms state-of-the-art linearized baselines by over 36 points under identical memory constraints. Re-markably, O CTOPUS also surpasses its own full-cache teacher, demonstrating that learned sparse retention serves as an effective regular-izer for long-horizon reasoning.
Transformer-based large language models (LLMs) achieve strong performance across many tasks, but their Key-Value (KV) cache grows linearly with sequence length, creating a severe memory bottleneck for long-context inference. Existing heuristic eviction methods (e.g., H$_2$O and SnapKV) rely on static attention or positional signals that often fail to capture a token's future predictive influence. We propose DistillCache, a reinforcement learning framework that formulates KV-cache eviction as a sequential decision problem. DistillCache learns a lightweight policy network using rich internal model signals (attention statistics, value norms, entropy, and position) and trains it with REINFORCE via a per-step KL-divergence reward to preserve the full-cache output distribution. On a 7B-parameter instruction-tuned Transformer (Mistral-7B-Instruct-v0.3), DistillCache retains 94.2% of full-cache accuracy on LongBench at a 25% cache budget, outperforming both strong heuristic baselines (H$_2$O, SnapKV) by up to 2.7 absolute points and, under our re-implementations, concurrent RL-based methods (ForesightKV, RLKV) by up to 1.4 points on long-context tasks. On reasoning benchmarks, DistillCache is competitive with the best concurrent method and surpasses it under aggressive compression. It also delivers up to 2.1x full-cache throughput while maintaining competitive practical efficiency. These results highlight the effectiveness of learned, distribution-aware policies for memory-efficient long-context LLM inference.
MoNe is a lightweight modular neural memory that attaches to any frozen pretrained Transformer to enable long-context inference without retraining, achieving strong performance on needle-in-a-haystack and word extraction benchmarks from RULER, where ICL degrades sharply.
Won-Yong Cho, Kyubyung Chae, Tribhuvanesh Orekondy et al.· 0 citations
The quadratic cost of attention-based sequence models for long contexts has motivated a growing line of research on memory-based models that can compress context into a compact state. However, most existing memory models expose a static memory throughout the entire sequence. Because early tokens face no compression pressure, they occupy too many degrees of freedom and"pollute"the memory state, leaving little capacity for later context and increasing interference between what is stored and what arrives next. We study a new paradigm of incremental memory activation, where the effective capacity of memory is progressively expanded as the context grows. Imposing an early bottleneck forces the model to compress history more effectively, while unlocking fresh capacity over time reduces interference and improves retention of later context. We instantiate this paradigm in Proteus, a straightforward mechanism that can be incorporated into a broad class of neural memory architectures at no additional cost. We apply Proteus to state-of-the-art models, including SWLA, Comba, Titans, and Hope-Attention, and observe consistent improvements on standard language modeling and reasoning, as well as on long-context retrieval and understanding, with gains that grow at longer context lengths. Overall, our results show that static memory is suboptimal and that scheduling effective capacity is a simple and broadly applicable tool for sequence modeling.
Reza Bayat, Ali Behrouz, V. Mirrokni et al.· 0 citations
Modern language models are built primarily from Transformers, recurrent models, and their hybrid architectures. Transformers rely on token-level attention memories, while recurrent models such as state space models (SSMs) and linear attention maintain compact recurrent states. These architectures are typically instantiated separately or interleaved at the layer level, leaving open whether a shared memory representation can support both recurrent compression and attention-style retrieval. We study this question through the state space duality (SSD) view of Mamba-2, where the SSM state can be interpreted as a compressed associative key--value (KV) cache. We observe that Mamba-2 decodes token-conditioned values from this state but does not decode token-conditioned keys. Based on this observation, we propose DART (Decoded Attention over Recurrent sTates), which retains the chunk state contributions produced by the Mamba-2 chunked scan as chunk state memories, decodes token-conditioned keys and values from these memories, and performs state-memory attention (SMA) over the resulting KV pairs. The retrieved output is then combined with the native Mamba-2 output through a gated residual connection. DART supports practical training by reusing the Mamba-2 chunked scan and implementing SMA as a FlashAttention-style computation. Our analysis and experiments show that DART substantially reduces the length-dependent inference cache compared with a matched attention baseline (e.g., $75\%$ savings when the chunk size is $S=256$ and the state size is $N=128$). Compared with Mamba-2, DART substantially improves associative recall and retrieval while preserving general language-modeling quality.
Yixiao Qian, Song Chen, Pengkai Wang et al.· 0 citations
MemDefrag, a training-free and model-agnostic framework that uses a middle-layer tracing signal to conduct memory defragmentation (rank, reorder, and filter memories), and applies an informativeness-guided proportional forgetting mechanism once capacity is exceeded, is proposed.
A lot of prior work addressed key-value (KV) cache selection and compression by sparse attention to enable long-context inference for transformer language models without excessive hardware budgets. We provide a new method for fine-tuning models with sparse attention. It works for any KV cache policy, runs on a moderate hardware budget (e.g., a single Nvidia A100 GPU with 40 GB RAM), and allows the model to co-adapt with the policy, often outperforming models trained with exact attention (sequence parallelism). We also provide an efficient implementation of H2O sparse attention (the leading policy in our experiments) with dedicated scaled dot product attention kernel support. KeysAndValues (https://github.com/awslabs/keys_values), a new open source library for long-context inference and fine-tuning, provides easy-to-use and performant code for all methods discussed here.
Matthias Seeger, Zeyu Zhang, Vihang Patil et al.· 0 citations