Q-ROAR: Outlier-Aware Rescaling for RoPE Position Interpolation in Quantized Long-Context LLMs

2025-09-17

AI Summary

Position interpolation (PI) in RoPE coupled with post-training quantization (PTQ) induces position-dependent logit noise—arising from long-context aliasing, dynamic range expansion, axial anisotropy, and outlier shifts—degrading long-context performance. Method: We systematically characterize this coupling effect and propose a fine-tuning-free, architecture-agnostic stabilization framework. We introduce two diagnostic metrics—Interpolation Pressure Ratio and Tail Inflation Ratio—and design a RoPE-aware band-wise rescaling mechanism: grouping dimensions by RoPE frequency and independently optimizing per-group Q/K weight scaling factors under symmetric scaling to preserve logit magnitude. The method requires only a small long-context development set and integrates seamlessly into existing inference pipelines. Results: Experiments show up to 0.7% accuracy recovery on standard benchmarks, >10% perplexity reduction on GovReport, and zero degradation on short-context tasks—demonstrating robustness and practicality.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Natural Language Processing: Safety and Robustness
• Planning, Routing, and Scheduling: Planning with Language Models

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Security and Privacy: Large-scale security measurements
• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work

Abstract

Extending LLM context windows is crucial for long range tasks. RoPE-based position interpolation (PI) methods like linear and frequency-aware scaling extend input lengths without retraining, while post-training quantization (PTQ) enables practical deployment. We show that combining PI with PTQ degrades accuracy due to coupled effects long context aliasing, dynamic range dilation, axis grid anisotropy, and outlier shifting that induce position-dependent logit noise. We provide the first systematic analysis of PI plus PTQ and introduce two diagnostics: Interpolation Pressure (per-band phase scaling sensitivity) and Tail Inflation Ratios (outlier shift from short to long contexts). To address this, 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. The diagnostics guided search uses a tiny long-context dev set and requires no fine-tuning, kernel, or architecture changes. Empirically, Q-ROAR recovers up to 0.7% accuracy on standard tasks and reduces GovReport perplexity by more than 10%, while preserving short-context performance and compatibility with existing inference stacks.

Problem

• Combining position interpolation with quantization degrades LLM accuracy
• Addresses long-context aliasing and outlier shifting in quantized models
• Solves position-dependent logit noise in extended context windows

Innovation

• Per-band frequency scaling for RoPE dimensions
• Diagnostics-guided search without fine-tuning
• Weight-only stabilization preserving logit scale