arXiv:2602. 02599v4 Announce Type: replace Abstract: Long-context inference in large language models (LLMs) is bottlenecked by the memory and compute of the key-value (KV) cache.
arXiv:2608. 02947v1 Announce Type: new Abstract: The attention score with rotary position embeddings (RoPE) decomposes exactly into a sum over its 2D-rotation frequency pairs, and each pair's wavelength limits how far it can discriminate position.
The attention score with rotary position embeddings (RoPE) decomposes exactly into a sum over its 2D-rotation frequency pairs, and each pair's wavelength limits how far it can discriminate position. Aligned with this structure, we propose the per-RoPE-wavelength distance window: it prunes the query--key inner-product terms beyond a wavelength-proportional distance.
arXiv:2607. 19363v1 Announce Type: new Abstract: Rotary Position Embedding (RoPE) is widely adopted in Transformers to encode positional information, yet standard implementations enforce a uniform frequency schedule and scaling across all attention heads.
arXiv:2606. 09508v1 Announce Type: new Abstract: Existing sparse attention and KV cache compression methods for long-context LLM inference typically apply fixed sparsity patterns or uniform budgets across all attention heads, overlooking the substantial variation in attention behavior among heads and contexts.
The paper investigates the use of partial Rotary Positional Embedding (RoPE) in transformer models, showing that applying RoPE to only about 10% of hidden dimensions can save up to ten times the memory of full RoPE while achieving comparable final loss. The authors systematically analyze training dynamics and convergence across various architectures, model sizes, sequence lengths, and datasets, finding consistent performance trends. They also note that completely omitting positional encoding (NoPE) can cause loss spikes, which are mitigated by minimal RoPE usage or QK‑Norm application.
arXiv:2606. 24975v1 Announce Type: new Abstract: PaTH Attention showed that replacing RoPE's position-indexed rotations with accumulated data-dependent Householder reflections yields strong length extrapolation, though performance degrades at extreme context lengths.
arXiv:2610.02953v1 Announce Type: new Abstract: Long-context LLM serving is increasingly bottlenecked by KV-cache memory, especially in resource-constrained scenarios. Among existing KV-cache compres...
arXiv:2509.12635v4 Announce Type: replace-cross Abstract: We prove under practical assumptions that Rotary Positional Embedding (RoPE) introduces an intrinsic distance-dependent bias in attention sco...
arXiv:2607. 06519v1 Announce Type: new Abstract: Long-context LLM inference is increasingly limited by the memory and bandwidth cost of KV caches, yet aggressive compression can remove the layer-specific evidence needed for retrieval and multi-step reasoning.
arXiv:2609.39929v1 Announce Type: cross Abstract: Long-context failures of RoPE-based language models can arise from RoPE's intrinsic tradeoff between maintaining stable token preferences and disting...
The paper proposes a content‑based addressing scheme for long‑context models that replaces the growing token counter in Rotary Position Embedding (RoPE) with unit‑level addresses derived from the content of each unit. By dividing the token stream into units, the method preserves local RoPE behavior while allowing new units to be addressed via learned content maps, avoiding positional mismatches when extending context length. Experiments on character‑level Tiny Shakespeare show that a model trained on 256‑character contexts achieves lower perplexity at 4096 characters using this scheme, and a second diagnostic demonstrates retrieval of multiple serialized facts.