Papers with RoPE
Extending LLM Context Window with Adaptive Grouped Positional Encoding: A Training-Free Method (2025.acl-long)
Copied to clipboard
| Challenge: | Existing long-context training data is scarce and requires substantial GPU resources for training. |
| Approach: | They propose a training-free plug-and-play method to enhance long-context understanding in existing large language models. |
| Outcome: | The proposed method outperforms existing LLMs on various tasks and surpasses baseline methods. |
Resonance RoPE: Improving Context Length Generalization of Large Language Models (2024.findings-acl)
Copied to clipboard
| Challenge: | Recent advances in Large Language Models (LLMs) have demonstrated their potential across a wide spectrum of natural language processing tasks. |
| Approach: | They propose a novel approach to narrow the generalization gap in TSTL scenarios by refining the interpolation of RoPE features for OOD positions. |
| Outcome: | The proposed approach improves performance without additional online computational costs on train-short-test-long scenarios. |
RoBSA: RoPE-based Blockwise Sparse Multi-head Latent Attention (2026.acl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have advanced in recent years, scaling up in both parameter count and context length. |
| Approach: | They propose a method to compute attention over a subset of context tokens and to implement token selection in a blockwise manner. |
| Outcome: | The proposed method reduces end-to-end inference latency by up to 2.55x with minimal accuracy loss compared to full attention in long-context scenarios for very large models. |
LongEmbed: Extending Embedding Models for Long Context Retrieval (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing embedding models support only 512 input tokens, hindering their application in scenarios requiring long inputs. |
| Approach: | They evaluate the performance of existing embedding models by using a new benchmark and a training-free context window extension strategy. |
| Outcome: | The proposed model extends the input window of existing models by several folds. |
mGTE: Generalized Long-Context Text Representation and Reranking Models for Multilingual Text Retrieval (2024.emnlp-industry)
Copied to clipboard
Xin Zhang, Yanzhao Zhang, Dingkun Long, Wen Xie, Ziqi Dai, Jialong Tang, Huan Lin, Baosong Yang, Pengjun Xie, Fei Huang, Meishan Zhang, Wenjie Li, Min Zhang
| Challenge: | Existing models for text retrieval are based on a multi-stage process that involves retrieving documents from a large corpus. |
| Approach: | They propose to build a multilingual text representation model and a cross-encoder reranker from scratch for text retrieval. |
| Outcome: | The proposed models outperform the state-of-the-art models on long-context retrieval benchmarks. |
CoCA: Fusing Position Embedding with Collinear Constrained Attention in Transformers for Long Context Window Extending (2024.acl-long)
Copied to clipboard
| Challenge: | Existing models that use self-attention and position embedding have anomalous behavior that hinder long context window extrapolation. |
| Approach: | They propose a collinear constraint between Q and K to integrate RoPE and self-attention. |
| Outcome: | The proposed model integrates self-attention and position embedding into LLMs without fine-tuning. |
E2-LLM: Efficient and Extreme Length Extension of Large Language Models (2024.findings-acl)
Copied to clipboard
Jiaheng Liu, ZhiqiBai ZhiqiBai, Yuanxing Zhang, Chenchen Zhang, YuangZh YuangZh, Ge Zhang, JiakaiWang JiakaiWang, Haoran Que, Yukang Chen, Wenbo Su, Tiezheng Ge, Jie Fu, Wenhu Chen, Bo Zheng
| Challenge: | Existing techniques for extending context capabilities in LLMs require additional training procedures and access to datasets with long context (e.g., sequences of 32K tokens). |
| Approach: | They propose a solution to extend context capabilities in Large Language Models by training a single process over a sequence of 4K tokens. |
| Outcome: | The proposed solution significantly reduces the cost of continual-pretraining or fine-tuning over short sequences and improves robustness to diverse relative positions. |
Position Encoding with Random Float Sampling Enhances Length Generalization of Transformers (2026.findings-eacl)
Copied to clipboard
| Challenge: | Length generalization is the ability of language models to maintain performance on inputs longer than those seen during pretraining. |
| Approach: | They propose a position encoding strategy that uses random float sampling to generalize to unseen lengths. |
| Outcome: | The proposed strategy can generalize to lengths unseen during training and in benchmarks. |
Beyond Position: the emergence of wavelet-like properties in Transformers (2025.acl-long)
Copied to clipboard
| Challenge: | Despite its widespread adoption, theoretical limitations in positional encodings are resolved by developing emergent, wavelet-like processing strategies. |
| Approach: | They propose to use Rotary Position Embeddings to develop emergent, wavelet-like properties that compensate for the positional encoding’s theoretical limitations. |
| Outcome: | The attention heads evolve to implement multi-resolution processing analogous to wavelet transforms. |
Mitigate Position Bias in LLMs via Scaling a Single Hidden States Channel (2025.findings-acl)
Copied to clipboard
Yijiong Yu, Huiqiang Jiang, Xufang Luo, Qianhui Wu, Chin-Yew Lin, Dongsheng Li, Yuqing Yang, Yongfeng Huang, Lili Qiu
| Challenge: | Long-context language models exhibit position bias, also known as "lost in the middle" research shows that even long-contemporary LLMs fail to utilize all context information effectively . |
| Approach: | They propose a method to mitigate position bias by scaling positional hidden states . they propose to use a channel of hidden states to modify positional Hidden states a LCLM's positional bias . |
| Outcome: | The proposed method can improve performance by 15.2% in a "lost in the middle" benchmark. |
Breaking the Stage Barrier: A Novel Single-Stage Approach to Long Context Extension for Large Language Models (2025.coling-main)
Copied to clipboard
Haoran Lian, Junmin Chen, Wei Huang, Yizhe Xiong, Wenping Hu, Guiguang Ding, Hui Chen, Jianwei Niu, Zijia Lin, Fuzheng Zhang, Di Zhang
| Challenge: | Recent studies show that Large language models struggle with handling long token sequences due to limited training context size. |
| Approach: | They propose a single-stage continual pretraining method to equip LLMs with long context modeling capabilities. |
| Outcome: | The proposed method outperforms existing methods on 4 language modeling benchmarks. |
Advancing General Multimodal Capability of Vision-language Models with Pyramid-descent Visual Position Encoding (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing methods to encode visual positions inhibit the performance of vision-language Models (VLMs) however, language constitutes only one aspect of communication. |
| Approach: | They propose a method to assign visual position indexes from the periphery to the center and expand the central receptive field incrementally to enhance the perception of visual tokens within VLMs. |
| Outcome: | The proposed method reduces the relative distance between interrelated visual elements and instruction tokens, promoting a more rational allocation of attention weights and allowing for a multi-granularity perception of visual elements. |
On the token distance modeling ability of higher RoPE attention dimension (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Existing work on extending the context length of language models based on Rotary position embedding (RoPE) has shown promising results in capturing longer-range contextual information. |
| Approach: | They propose to use a hidden dimension of an attention head to investigate its contribution to capturing long-distance dependencies. |
| Outcome: | The proposed model can capture long-distance dependencies by extending the attention of a particular dimension of an attention head. |
PSC: Extending Context Window of Large Language Models via Phase Shift Calibration (2024.emnlp-main)
Copied to clipboard
| Challenge: | Large-scale language models (LLMs) have shown impressive results across a variety of tasks. |
| Approach: | They propose a module for calibrating the frequencies predefined by existing methods . they conducted extensive experiments across multiple models and tasks . |
| Outcome: | The proposed method reduces perplexity as the context window size is varied from 16k to 32k and up to 64k. |
Towards Inducing Long-Context Abilities in Multilingual Neural Machine Translation Models (2025.naacl-long)
Copied to clipboard
| Challenge: | Neural Machine Translation models traditionally use Sinusoidal Positional Embeddings . retraining with newer methods like ROPE or ALIBI is computationally expensive . |
| Approach: | They propose to transition NMT models from Sinusoidal to Relative PEs without compromising performance. |
| Outcome: | The proposed approach outperforms models trained with Sinusoidal PEs on document-level benchmarks . the results show that parameter-efficient fine-tuning can facilitate the transition . |
Extending Context Window of Large Language Models from a Distributional Perspective (2024.emnlp-main)
Copied to clipboard
Yingsheng Wu, Yuxuan Gu, Xiaocheng Feng, Weihong Zhong, Dongliang Xu, Qing Yang, Hongtao Liu, Bing Qin
| Challenge: | Existing scaling methods for extending context window rely on empirical approaches and lack understanding of the internal distribution within RoPE resulting in suboptimal performance. |
| Approach: | They propose to optimize the context window extending task from the view of rotary angle distribution by minimizing disturbance between rotary angles to maintain consistency with the pre-training phase. |
| Outcome: | The proposed approach reduces by up to 72% of the distributional disturbance when extending LLaMA2’s context window to 8k, and reduces it by up 32% when extending to 16k. |
VEEF-Multi-LLM: Effective Vocabulary Expansion and Parameter Efficient Finetuning Towards Multilingual Large Language Models (2025.coling-main)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have a significant disadvantage for low-resource languages . VEEF-Multi-LLM-8B excels in multilingual instruction-following tasks . |
| Approach: | They propose a low-resource multilingual large language model that expands the vocabulary for multilingual support. |
| Outcome: | The proposed model outperforms existing models in multilingual instruction-following tasks, but lags behind English-centric models in some tasks. |
Understanding the RoPE Extensions of Long-Context LLMs: An Attention Perspective (2025.coling-main)
Copied to clipboard
| Challenge: | Enabling LLMs to handle lengthy context is currently a research hotspot . a notable challenge limiting further customization is the inability of LLM to utilize context beyond pretrained length due to the inherent flaw of rotary position embedding (RoPE). |
| Approach: | They propose to extend the RoPE from an attention perspective and on two benchmarking tasks. |
| Outcome: | The proposed extension of the RoPE improves extrapolation and retrieval errors. |
The Rotary Position Embedding May Cause Dimension Inefficiency in Attention Heads for Long-Distance Retrieval (2025.findings-acl)
Copied to clipboard
| Challenge: | We hypothesize that the wide range of rotation angles may prevent LLMs from utilizing certain dimensions. |
| Approach: | They propose to use the Rotary Position Embedding (RoPE) for long context modeling . they hypothesize that the wide range of rotation angles may prevent LLMs from utilizing those dimensions. |
| Outcome: | The proposed model may not be useful for long-context modeling. |
VRoPE: Rotary Position Embedding for Video Large Language Models (2025.emnlp-main)
Copied to clipboard
Zikang Liu, Longteng Guo, Yepeng Tang, Tongtian Yue, Junxian Cai, Kai Ma, Qingbin Liu, Xi Chen, Jing Liu
| Challenge: | Existing versions of Large Language Models (LLMs) lack a positional encoding strategy for video. |
| Approach: | They propose a new positional encoding method tailored for Video-LLMs that mitigates positional biases and ensures a more uniform distribution of spatial focus. |
| Outcome: | The proposed method outperforms existing versions of RoPE in video understanding and reasoning tasks. |
DCIS: Efficient Length Extrapolation of LLMs via Divide-and-Conquer Scaling Factor Search (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing frameworks for large language models with context length limitations are suboptimal for initialization and fine-tuning. |
| Approach: | They propose a RoPE-based fine-tuning framework that strategically determines the best scaling factors for LLMs by a Divide-and-Conquer Incremental Search algorithm. |
| Outcome: | The proposed framework mitigates performance decay at extended target lengths and can perform effectively without fine-tuning. |
Fixing Semantic Blind Spots in Anchor Tokens of dMLLMs (2026.findings-acl)
Copied to clipboard
| Challenge: | Autoregressive models (ARMs) are prone to hallucinations due to their sequential text generation and high latency. |
| Approach: | They propose a training-free decoding strategy that augments the attention key space with a static, distance-aware matrix to reduce the attention sink effect on semantic anchors. |
| Outcome: | The proposed method reduces the attention sink effect on semantic anchors while enhancing their ability to aggregate global visual information. |
Mitigating Position Bias in Transformers via Layer-Specific Positional Embedding Scaling (2026.findings-acl)
Copied to clipboard
Zhenghua Wang, Yiran Ding, Changze Lv, Yixin Wu, Tianlong Li, Zhibo Xu, Muling Wu, Tianyuan Shi, Shizheng Li, Qi Qian, Xuanjing Huang, Xiaoqing Zheng
| Challenge: | Existing methods to address the "lost-in-the-middle" problem suffer from high latency or suboptimal hand-crafted scaling strategies. |
| Approach: | They propose a layer-specific positional embedding scaling method that assigns distinct scaling factors to each layer. |
| Outcome: | Experiments show that the proposed method mitigates positional attention bias and delivers consistent improvements across multiple long-context benchmarks. |
HoPE: A Novel Positional Encoding Without Long-Term Decay for Enhanced Context Awareness and Extrapolation (2025.acl-long)
Copied to clipboard
| Challenge: | Existing positional encodings exhibit long-term decay, based on an entrenched and long-standing opinion that tokens farther away from the current position carry less relevant information. |
| Approach: | They propose a high-frequency rotary position encoding (HoPE) that replaces specific components in RoPE with position-independent ones, retaining only high- frequency signals. |
| Outcome: | The proposed method exhibits greater robustness to the out-of-distribution behavior in attention patterns during extrapolation. |
PEPE: Long-context Extension for Large Language Models via Periodic Extrapolation Positional Encodings (2025.findings-emnlp)
Copied to clipboard
Jikun Hu, Dongsheng Guo, Yuli Liu, Qingyao Ai, Lixuan Wang, Xuebing Sun, Qilei Zhang, Quan Zhou, Cheng Luo
| Challenge: | Long-context extension attempts to extend contextual window in pre-trained LLMs . primary method involves expanding initial positional encodings, disrupting positional learning . |
| Approach: | They propose a new extension strategy based on Rotary Position Embedding to extend contextual window in pre-trained large language models. |
| Outcome: | The proposed method can extend the contextual window in pre-trained large language models . expansion disrupts positional encodings learned during pre-training, authors show . |
EQUIP: EQUivariant preserving In-Place updates for Efficient Token Pruning (2026.acl-long)
Copied to clipboard
| Challenge: | Token-pruning methods cause "holes" in KV tensors, posing major challenges . equip reduces recomputation of rotation operations through in-place update, caching and re-indexing . |
| Approach: | They propose an EQUIP-based in-place token update mechanism that preserves the equivariance property of the operations performed in the attention computation. |
| Outcome: | EQUIP reduces recomputation of rotation operations and reduces eviction overheads . it achieves geomean speedups of 1.62 (or 1.47) over StreamingLLM and 3.45 ( or 1.86) |
Do Transformers Grok Succinct Algorithms? Mechanistic Evidence for Counting Circuits (2026.findings-acl)
Copied to clipboard
| Challenge: | Recent studies suggest that Transformers are inherently succinct, capable of representing recursive algorithms like binary counting over exponential state spaces. |
| Approach: | They propose to bridge this gap by testing the Succinctness Hypothesis using mechanistic interpretability on a large-scale computation task. |
| Outcome: | The proposed model can represent recursive algorithms over exponential state spaces . the proposed model is able to generalize perfectly, whereas massive LSTM baselines fail completely. |
Towards Economical Inference: Enabling DeepSeek’s Multi-Head Latent Attention in Any Transformer-based LLMs (2025.acl-long)
Copied to clipboard
Tao Ji, Bin Guo, Yuanbin Wu, Qipeng Guo, Shenlixing Shenlixing, Chenzhan Chenzhan, Xipeng Qiu, Qi Zhang, Tao Gui
| Challenge: | Multi-head Latent Attention (MLA) is an innovative architecture designed to ensure efficient and economical inference by significantly compressing the Key-Value (KV) cache into a latent vector. |
| Approach: | They propose a data-efficient fine-tuning method for transitioning from MHA to MLA using a latent vector cache. |
| Outcome: | The proposed architecture reduces the KV cache size of Llama2-7B by 92.19%, with only 1% drop in LongBench performance. |
LaMPE: Length-aware Multi-grained Positional Encoding for Adaptive Long-context Scaling Without Training (2026.findings-acl)
Copied to clipboard
| Challenge: | Large language models (LLMs) experience significant performance degradation when the input exceeds the pretraining context window due to the out-of-distribution (OOD) behavior of Rotary Position Embedding (RoPE). |
| Approach: | They propose a training-free method that remaps out-of-distribution (OOD) positions into the in-distance range with fixed mapping strategies, ignoring the dynamic relationship between input length and effective context window. |
| Outcome: | Experiments on three representative LLMs across five mainstream long-context benchmarks show that the proposed method achieves significant performance improvements compared to existing methods. |
SCOPE: Boosting LLM Efficiency with Scoped Position Encoding (2026.acl-long)
Copied to clipboard
| Challenge: | Positional encodings are fundamental to Transformers, but explicit methods like RoPE can degrade under length extrapolation and incur extra arithmetic and memory-access overhead. |
| Approach: | They propose a framework that reimagines structured sparsity as an intrinsic position encoding mechanism. |
| Outcome: | The proposed framework reduces the number of attention FLOPs by 8x compared to RoPE on LLaMA-3-8B architectures while reducing training and inference latency. |
Paramanu: Compact and Competitive Monolingual Language Models for Low-Resource Morphologically Rich Indian Languages (2026.acl-long)
Copied to clipboard
| Challenge: | Multilingual large language models are expensive to pretrain and suffer from imbalances across languages and datasets. |
| Approach: | They propose a family of Indian language-only autoregressive language models trained on open-source language-specific data for the five most spoken Indian languages. |
| Outcome: | The proposed model outperforms most larger models up to 8B across all five languages. |