[
    {
        "id": "osp-19607",
        "type": "article-journal",
        "title": "BitNest: Bit-Nested Speculative Decoding for Memory-Efficient LLM Inference Acceleration",
        "author": [
            {
                "family": "Yang",
                "given": "Chence"
            },
            {
                "family": "Cheng",
                "given": "Ningxi"
            },
            {
                "family": "Akbari",
                "given": "Arash"
            },
            {
                "family": "Tan",
                "given": "Qitao"
            },
            {
                "family": "Zhu",
                "given": "Qingchan"
            },
            {
                "family": "Zhang",
                "given": "Ci"
            },
            {
                "family": "Yang",
                "given": "Changdi"
            },
            {
                "family": "Wang",
                "given": "Yanzhi"
            },
            {
                "family": "Niu",
                "given": "Wei"
            },
            {
                "family": "Wang",
                "given": "Jinhui"
            },
            {
                "family": "Lu",
                "given": "Jin"
            },
            {
                "family": "Yuan",
                "given": "Geng"
            }
        ],
        "URL": "https://omanscience.com/ar/articles/bitnest-bit-nested-speculative-decoding-for-memory-efficient-llm-inference-acceleration",
        "language": "en",
        "issued": {
            "date-parts": [
                [
                    2026
                ]
            ]
        },
        "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."
    }
]