Rethinking RoPE Scaling in Quantized LLM: Theory, Outlier, and Channel-Band Analysis with Weight Rescaling

• URL: http://arxiv.org/abs/2510.00028v1
• Date: Fri, 26 Sep 2025 01:23:32 GMT
• Title: Rethinking RoPE Scaling in Quantized LLM: Theory, Outlier, and Channel-Band Analysis with Weight Rescaling
• Authors: Ye Qiao, Haocheng Xu, Xiaofan Zhang, Sitao Huang,
• Abstract summary: We show that combining RoPE position-aware (PI) with PTQ degrades accuracy due to effects including long-context aliasing, dynamic-range dilation, anisotropy from axis-aligned quantizers vs rotated RoPE pairs.<n>We propose Q-ROAR (Quantization, RoPE-interpolation, and Outlier Aware Rescaling), a weight-only, awareness-aware stabilization of PI for quantized LLMs.
• Score: 3.7391437252721698
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Extending the context window support of large language models (LLMs) is crucial for tasks with long-distance dependencies. RoPE-based interpolation and extrapolation methods, such as linear scaling and frequency-aware schemes, enable longer input length support without retraining, while post-training quantization (PTQ) makes deployment practical. However, we show that combining RoPE position interpolation (PI) with PTQ degrades accuracy due to coupled effects including long-context aliasing, dynamic-range dilation, anisotropy from axis-aligned quantizers vs. rotated RoPE pairs, and outlier shifting that produces position-dependent logit noise. We provide, to the best of our knowledge, the first systematic analysis of the PI+PTQ approach and introduce two practical diagnostics: interpolation pressure (per-band sensitivity to phase scaling) and tail-inflation ratios (outlier shift from short to long contexts). Following the analysis results, we propose Q-ROAR (Quantization, RoPE-interpolation, and Outlier Aware Rescaling), a weight-only, interpolation-aware stabilization of PI for quantized LLMs. Q-ROAR groups RoPE dimensions into a small number of frequency bands and performs a lightweight search over per-band scales for Key and Query weights (with an optional symmetric variant to preserve logit scale). The search is guided by our diagnostics and uses a tiny long-context development dataset, requiring no fine-tuning to the model, no architecture or kernel changes, and no additional deployment overhead. Empirically, Q-ROAR reduces the model's perplexity on long-context workloads by more than 14%, while preserving short-context performance, inference throughput, and compatibility with existing LLM system stacks.

Related papers

• Beyond Real: Imaginary Extension of Rotary Position Embeddings for Long-Context LLMs [72.8830548005884]
  Rotary Position Embeddings (RoPE) have become a standard for encoding sequence order in Large Language Models.<n>Standard implementations utilize only the real component of the complex-valued dot product for attention score calculation.<n>We propose an extension that re-incorporates this imaginary component.
  arXiv  Detail & Related papers  (2025-12-08T12:59:54Z)
• ParoQuant: Pairwise Rotation Quantization for Efficient Reasoning LLM Inference [13.283581083797484]
  Post-training quantization (PTQ) compresses the weights of Large Language Models (LLMs) into low-precision representations to reduce memory footprint and accelerate inference.<n>The presence of outliers in weights and activations often leads to large quantization errors and severe accuracy degradation.<n>We propose Pairwise Rotation Quantization (ParoQuant) to suppress outliers and introduce significant overhead during inference.<n>ParoQuant achieves an average 2.4% accuracy improvement over AWQ on reasoning tasks with less than 10% overhead.
  arXiv  Detail & Related papers  (2025-11-13T18:59:24Z)
• DoPE: Denoising Rotary Position Embedding [60.779039511252584]
  Rotary Position Embedding (RoPE) in Transformer models has inherent limits that weaken length.<n>We reinterpret the attention map with positional encoding as a noisy feature map, and propose Denoising Positional extrapolation page (DoPE)<n>DoPE is a training-free method based on truncated matrix entropy to detect outlier frequency bands in the feature map.
  arXiv  Detail & Related papers  (2025-11-12T09:32:35Z)
• Q-ROAR: Outlier-Aware Rescaling for RoPE Position Interpolation in Quantized Long-Context LLMs [0.9510848451801044]
  We show that combining PI with PTQ degrades accuracy due to coupled effects long context aliasing, dynamic range dilation, axis grid anisotropy, and shifting that induce position-dependent logit noise.<n>We propose Q-ROAR, a RoPE-aware, weight-only stabilization that groups RoPE dimensions into a few frequency bands and performs a small search over per-band scales for W_Q,W_K, with an optional symmetric variant to preserve logit scale.
  arXiv  Detail & Related papers  (2025-09-17T19:50:16Z)
• HoPE: Hyperbolic Rotary Positional Encoding for Stable Long-Range Dependency Modeling in Large Language Models [19.3827288035483]
  We propose Hyperbolic Rotary Positional.<n>(HoPE) which leverages hyperbolic functions to implement Lorentz rotations on token representations.<n>Tests show HoPE consistently exceeds existing positional encoding methods.
  arXiv  Detail & Related papers  (2025-09-05T16:20:48Z)
• Decoupling the "What" and "Where" With Polar Coordinate Positional Embeddings [29.421443764865003]
  We present an analysis indicating that what and where are entangled in the popular RoPE rotary position embedding.<n>We propose an improvement to RoPE, which we call Polar Coordinate Position Embeddings or PoPE, that eliminates the what-where confound.
  arXiv  Detail & Related papers  (2025-09-05T14:22:27Z)
• QuantVSR: Low-Bit Post-Training Quantization for Real-World Video Super-Resolution [53.13952833016505]
  We propose a low-bit quantization model for real-world video super-resolution (VSR)<n>We use a calibration dataset to measure both spatial and temporal complexity for each layer.<n>We refine the FP and low-bit branches to achieve simultaneous optimization.
  arXiv  Detail & Related papers  (2025-08-06T14:35:59Z)
• LaMPE: Length-aware Multi-grained Positional Encoding for Adaptive Long-context Scaling Without Training [45.74983991122073]
  Large language models (LLMs) experience significant performance degradation when the input exceeds the pretraining context window.<n>Recent studies mitigate this problem by remapping OOD positions into the in-distribution range with fixed mapping strategies.<n>We propose Length-aware Multi-grained Positional Scaling (LaMPE), a training-free method that fully utilizes the model's effective context window.
  arXiv  Detail & Related papers  (2025-08-04T11:22:13Z)
• Agentic Reinforced Policy Optimization [66.96989268893932]
  Large-scale reinforcement learning with verifiable rewards (RLVR) has demonstrated its effectiveness in harnessing the potential of large language models (LLMs) for single-turn reasoning tasks.<n>Current RL algorithms inadequately balance the models' intrinsic long-horizon reasoning capabilities and their proficiency in multi-turn tool interactions.<n>We propose Agentic Reinforced Policy Optimization (ARPO), a novel agentic RL algorithm tailored for training multi-turn LLM-based agents.
  arXiv  Detail & Related papers  (2025-07-26T07:53:11Z)
• RoSTE: An Efficient Quantization-Aware Supervised Fine-Tuning Approach for Large Language Models [53.571195477043496]
  We propose an algorithm named Rotated Straight-Through-Estimator (RoSTE)<n>RoSTE combines quantization-aware supervised fine-tuning (QA-SFT) with an adaptive rotation strategy to reduce activation outliers.<n>Our findings reveal that the prediction error is directly proportional to the quantization error of the converged weights, which can be effectively managed through an optimized rotation configuration.
  arXiv  Detail & Related papers  (2025-02-13T06:44:33Z)
• Scaling Laws of RoPE-based Extrapolation [103.33995311915864]
  We propose textbftextitScaling Laws of RoPE-based Extrapolation to describe the relationship between the extrapolation performance and base value. We achieve extrapolation up to 1 million context length within only 16K training length on LLaMA2 7B and 13B.
  arXiv  Detail & Related papers  (2023-10-08T15:50:36Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.