cs.CLNov 6, 2026

Retrofitting Unbounded Context onto Stock Language Models

Authors: Mark Oskin

Organizations: School of Computer Science and Engineering, University of Washington

Abstract

A language model that never had to forget would keep every document it has read and every turn it has spoken, and answer from any of it. This paper shows that two stock models can do that on unmodified weights by manipulating their geometry. Rotary attention turns each pair of key dimensions, a plane, at its own rate. The slow planes turn once in thousands of tokens and carry the address a query matches when it reads a far key; the fast planes turn once in a few tokens, and the order of the tokens inside a span is already written into its keys and values by the time the span leaves the near window. At distance the position blurs, and what attention reads is a bag rather than a sequence. A far span is read by its words: under competition it is read only if it is visible at the query of the question's own handle word, and a read through a description rather than a word fails at distance in every model measured, a 70B included. These facts allow a trick. What sits in the far part of attention does not have to be the history as it was; it can be installed ahead of the prediction. The design keeps a near window of 8K tokens at their true positions and changes what lies beyond it with the prompt: every far key and value goes to the host exact, a word index finds the units the request's rare words name, a far window of slots inside the trained range holds them while the request is encoded, and a rare word the model generates fetches its units mid-generation with one exact re-encode. Implemented in llama.cpp on Mistral-7B and Llama-3.1-8B, the design reads planted single needles at 99 to 100 of 100 and three-key discrimination at 88 to 100 of 100 from 32K to one million tokens, scores 70.8 on RULER against 69.9 for dense attention on Llama-3.1-8B at 128K and 67.8 against 75.5 on Mistral-7B at 32K, and at 128K prefills at 1.85 times and decodes at 1.65 to 1.68 times stock speed.

Figures & tables

Explore similar work

Mar 18, 2026

Learning When to Attend: Conditional Memory Access for Long-Context LLMs

Language models struggle to generalize beyond pretraining context lengths, limiting long-horizon reasoning and retrieval. Continued pretraining on long-context data can help but is expensive due to the quadratic scaling of Attention. We observe that most tokens do not require (Global) Attention over the entire sequence and can rely on local context. Based on this, we propose L2A (Learning To Attend), a layer that enables conditional (token-wise) long-range memory access by deciding when to invoke global attention. We evaluate L2A on Qwen 2.5 and Qwen 3 models, extending their effective context length from 32K to 128K tokens. L2A matches the performance of standard long-context training to within 3% while skipping Global Attention for ∼\sim80% of tokens, outperforming prior baselines. We also design custom Triton kernels to efficiently implement this token-wise conditional Attention on GPUs, achieving up to ∼\sim2×\times improvements in training throughput and time-to-first-token over FlashAttention. Moreover, L2A enables post-training pruning of highly sparse Global Attention layers, reducing KV cache memory by up to 50% with negligible performance loss. Our code is released under Apache 2.0 at https://github.com/awslabs/hybrid-model-factory/tree/main/examples/research/L2A.
Oct 6, 2026

Hybrid Latent Attention for Looped Language Models

Looped language models apply the same stack of layers T times to each token, which deepens the model without adding parameters but multiplies its key-value (KV) cache by T. The larger cache limits how many sequences a GPU can decode at once and slows each decoding step, which reads the whole cache. We propose Hybrid Latent Attention (HLA), which keeps exact keys and values within a sliding window of W recent tokens and stores each older token as a compact latent that the query of each loop reads directly, without reconstructing keys and values. We uptrain HLA on Ouro looped models (T=4) with 1.4B and 2.6B parameters, keeping the pretrained weights frozen and training only the added parameters to reproduce the original attention. The cache shrinks by 10.7x per token, fitting 4.0-8.8x as many concurrent sequences per GPU, and decoding throughput improves by 2.5x at 1K-token contexts and by up to 7.4x at 16K. HLA retains over 97% of the original accuracy on math, knowledge and reasoning benchmarks, and 96-100% on long-context retrieval up to 16K tokens. After supervised fine-tuning, it performs on par with the fine-tuned original model on competition-level math.
Nov 3, 2026

Length Is Cheap, Competition Is Not: Attention Predicts Context Rot in Retrieval

Why do language models fail on long inputs? The common answer is length. We show that, for retrieval, length is cheap and competition is not. On five open models (0.5B to 3B parameters, four families), neutral filler costs nothing: accuracy stays at 0.97 to 1.00, up to 32K tokens on Qwen2.5-1.5B. Statements in the format of the evidence, placed at the same positions, drop it to 0.16 to 0.33. A single statement that gives the queried key a rival value multiplies the loss by 9 to 110, and every additional rival adds to it. Inside the model, one quantity tracks the failure: the attention mass left on the evidence. Across 126,473 trials it explains up to 58% of the variance in loss; log length explains at most 8%. The masked-filler protocol behind the claim that length alone hurts reproduces on MMLU (11 to 18 points lost at 30K tokens), but fails a visible control in retrieval: masking costs 45 accuracy points at 32K tokens, while the same filler left visible costs none. In retrieval, context rot is a competition for attention. We release the probe and all 126K trial-level measurements.