BitNest: Bit-Nested Speculative Decoding for Memory-Efficient LLM Inference Acceleration

2026-10-02

AI Summary

Existing speculative decoding methods incur substantial memory overhead due to auxiliary draft models, hindering deployment on resource-constrained devices. This work proposes BitNest, a framework that embeds low-precision drafts directly into high-precision target representations via a bit-level nested architecture, enabling single physical weight sharing. By employing progressive precision design and residual refinement strategies, BitNest first constructs a robust low-precision base before recovering the high-precision target, and further extends this mechanism to KV caching to accelerate long-context inference. Evaluated on 7B–8B edge models, the proposed method achieves an average speculative acceptance rate of 95.2%, delivering 1.48× to 1.61× end-to-end speedups over FP16 autoregressive decoding while preserving model quality.

Abstract

Speculative decoding accelerates autoregressive generation by using a lightweight draft to propose multiple tokens for parallel verification. However, existing methods often require an additional draft model or weight representation, introducing non-negligible memory overhead on resource-constrained devices. Self-speculative approaches reduce this overhead, yet still face trade-offs between draft quality, target quality, and storage efficiency. We propose BitNest, a bit-nested speculative decoding framework that embeds a low-precision draft directly into the higher-precision target representation. Instead of deriving a draft from a predefined target, BitNest first constructs a strong low-precision base and then recovers the higher-precision target through residual refinement, enabling both models to share a single physical weight representation. BitNest further extends this progressive-precision design to the KV cache for long-context inference. Across multiple 7B--8B edge-friendly LLMs and diverse workloads, BitNest achieves an average speculative acceptance rate of 95.2% while closely preserving higher-precision model quality, and delivers 1.48--1.61x end-to-end speedup over FP16 autoregressive decoding. On the LLaMA models supported by all representative self-speculative baselines, BitNest also achieves consistently competitive or higher decoding speedup.

Problem

• Speculative Decoding
• Memory Efficiency
• LLM Inference
• Self-Speculative Decoding
• Resource-Constrained Devices

Innovation

• Speculative Decoding
• Bit-Nested Representation
• Memory-Efficient Inference
• Progressive Precision
• KV Cache Optimization