FlowCompile: An Optimizing Compiler for Structured LLM Workflows
Summary
FlowCompile is a compiler for structured LLM workflows that performs compile-time exploration of configurations to balance accuracy and latency, achieving up to 6.4x speedup without retraining.
View Cached Full Text
Cached at: 05/15/26, 12:21 AM
Paper page - FlowCompile: An Optimizing Compiler for Structured LLM Workflows
Source: https://huggingface.co/papers/2605.13647
Abstract
FlowCompile is a structured LLM workflow compiler that optimizes complex multi-agent tasks by performing compile-time exploration of workflow configurations to balance accuracy and latency without retraining.
Structured LLM workflows, where specialized LLMsub-agentsexecute according to a predefined graph, have become a powerful abstraction for solving complex tasks. Optimizing such workflows, i.e., selecting configurations for each sub-agent to balance accuracy and latency, is challenging due to the combinatorial design space over model choices, reasoning budgets, and workflow structures. Existing cost-aware methods largely treatworkflow optimizationas a routing problem, selecting a configuration at inference time for each query according to the accuracy-latency objective used during training. We argue thatstructured LLM workflowscan also be optimized from a compilation perspective: before deployment, the system can globally explore the workflow design space and construct a reusable set ofworkflow-level configurationsspanning diverseaccuracy-latency trade-offs. Drawing inspiration frommachine learning compilers, we introduce FlowCompile, a structured LLM workflow compiler that performscompile-time design space explorationto identify a high-quality, reusable trade-off set. FlowCompile decomposes a workflow intosub-agents, profiles each sub-agent under diverse configurations, and composes these measurements through astructure-aware proxyto estimate workflow-level accuracy and latency. It then identifies diverse high-quality configurations in a single compile-time pass, without retraining or online adaptation. Experiments across diverse workflows and challenging benchmarks show that FlowCompile consistently outperforms heuristically optimized workflow configurations and routing-based baselines, delivering up to 6.4x speedup. The compiled configuration set further serves as a reusable optimization artifact, enabling flexible deployment under varyingruntime preferencesand supporting downstream selection or routing.
View arXiv pageView PDFGitHub0Add to collection
Get this paper in your agent:
hf papers read 2605\.13647
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2605.13647 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2605.13647 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2605.13647 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
LeanFlow: A Case Study in Workflow-Driven Lean Autoformalization
LeanFlow is an LLM agent system for translating mathematical papers into formalized Lean projects, evaluated through case studies and benchmarks with Kimi-K2.6 and GPT-5.5, achieving high completion rates within budget constraints.
[R] Compiling Agentic Workflows into LLM Weights: Near-Frontier Quality at Two Orders of Magnitude Less Cost
This paper demonstrates that compiling agentic workflow procedures into the weights of a small fine-tuned model achieves near-frontier quality at 128–462× cost reduction compared to in-context baselines, addressing perceived barriers of quality, cost, and flexibility.
Schedules Are Solvable Symbols: Tuning-Free Compilation of Tile Programs on Dataflow Architectures
A paper from NUS introducing Loom, a tuning-free symbolic compiler framework that formulates tile-based SPMD compilation as CP-SAT optimization over symbolic schedules, matching or exceeding vendor-optimized TTNN library performance on Tenstorrent Wormhole and Blackhole for GEMM, Flash Attention, and Flash Decode without per-shape profiling.
ReaComp: Compiling LLM Reasoning into Symbolic Solvers for Efficient Program Synthesis
ReaComp compiles LLM reasoning traces into reusable symbolic program synthesizers that achieve strong accuracy on program synthesis benchmarks while eliminating LLM calls at test time, significantly reducing computational cost.
FlowEdit: Information-Theoretic Control of LLM Reasoning Flows for Ill-posed Problems Involving Conflicts
FlowEdit is a novel framework that uses information-theoretic principles to regulate internal reasoning flows in LLMs, enabling them to generate multiple alternative responses in a single pass for ill-posed problems with conflicting conditions. Experiments show 68% improvement in exact-set-match accuracy and 24% boost in response informativeness over leading proprietary models.