Haiquan Lu, Zigeng Chen, Gongfan Fang, Xinyin Ma, Xinchao Wang
LLM agents have recently emerged as a powerful paradigm for solving complex tasks through planning, tool use, memory retrieval, and multi-step interaction. However, these agentic workflows often introduce substantial input-side overhead, making the compute-intensive prefilling stage a key bottleneck in long-context, multi-turn inference. In this work, we propose Mix-Quant, a simple and effective phase-aware quantizat ...
ion framework for fast agentic inference. We first investigate FP4 quantization in agentic LLM workflows and observe that quantizing the entire inference process can incur significant performance degradation. In contrast, the prefilling stage exhibits substantial quantization redundancy and can therefore be quantized with minimal accuracy loss, despite being the dominant source of computation. Based on this insight, we apply high-throughput NVFP4 quantization to the prefilling phase while preserving BF16 precision for decoding. By decoupling prefilling acceleration from decoding quality, Mix-Quant combines phase-aware algorithmic quantization with hardware-efficient NVFP4 execution to alleviate the inference bottleneck in LLM agents. Extensive experiments across long-context and agentic benchmarks demonstrate that Mix-Quant largely preserves task performance while delivering significant efficiency improvements, achieving up to a 3x speedup during prefilling.
Audit each benchmark finding before selecting an implementation path. Evidence refs map to the disclosure section below.
| Task | Dataset | Metric | Value | Source | Evidence refs |
|---|---|---|---|---|---|
| Agentic tool use | Qwen3-8B | LongMemEval | 57.00 | paper-derived | No explicit refs |
| Agentic tool use | Qwen3-8B-NVFP4 | LongMemEval | 49.82 | paper-derived | No explicit refs |
| Agentic tool use | Mix-Quant | LongMemEval | 54.85 | paper-derived | No explicit refs |
LLM agents have recently emerged as a powerful paradigm for solving complex tasks through planning, tool use, memory retrieval, and multi-step interaction.
Recommendation evidence is currently too limited for a maintained-repo choice. Use Implementation Status and Reproduction Path for a practical baseline plan.
Hardware Notes
Expect multi-day setup/compute for meaningful reproduction based on current guidance.
Evidence graph: 2 refs, 1 links.
Utility signals: depth 95/100, grounding 68/100, status medium.
Compare maintenance quality, reproducibility coverage, and evidence confidence before choosing a reproduction baseline.
Matched via arXiv identifier search
Risk flags
Matched via arXiv identifier search · Partial overlap with paper title keywords
Risk flags
Matched via arXiv identifier search
Risk flags
There is no verified maintained implementation yet. Use this baseline plan to decide whether to prototype now or defer.
Hardware requirements
No verified implementation available
No additional verified repositories beyond the primary recommendation.
These repositories had low-confidence matching signals and are hidden by default.
No trustworthy direct or curated related Hugging Face artifacts were found yet.
Continue with targeted Hugging Face searches derived from the paper title and method context:
Models
Tip: start with models, then check datasets/spaces if you need evaluation data or demos.
Direct artifact matches are currently sparse. Use targeted Hugging Face searches to quickly locate candidate models, datasets, and demos.
Tasks
Agentic tool use, Retrieval / indexing
Methods
Quantization
Domains
Large Language Models, AI Agents, Information Retrieval
Evaluation & Human Feedback Data
Open this paper in HFEPX to review benchmark signals, evaluation modes, and human-feedback protocol context.
Open in HFEPXExplore Similar Papers
Jump to Paper2Code search queries derived from this paper's research context.
Need human evaluators for your AI research? Scale annotation with expert AI Trainers.