Search papers, labs, and topics across Lattice.
The paper introduces a parallel execution paradigm, "Eager," for LLM-based code generation that overlaps code generation with execution. By chunking code based on Abstract Syntax Trees (ASTs), dynamically batching chunks, and implementing early error interruption, Eager minimizes idle time for both the generator and executor. Experiments across various benchmarks and LLMs demonstrate that Eager achieves up to 99.9% reduction in non-overlapped execution latency and up to 55% reduction in end-to-end latency.
LLMs can generate code 55% faster by executing code *while* generating it, challenging the traditional generate-then-execute paradigm.
Current LLM-based coding agents follow a serial execution paradigm: the model first generates the complete code, then invokes an interpreter to execute it. This sequential workflow leaves the executor idle during generation and the generator idle during execution, resulting in unnecessary end-to-end latency. We observe that, unlike human developers, LLMs produce code tokens sequentially without revision, making it possible to execute code as it is being generated. We formalize this parallel execution paradigm, modeling it as a three-stage pipeline of generation, detection, and execution, and derive closed-form latency bounds that characterize its speedup potential and operating regimes. We then present Eager, a concrete implementation featuring AST-based chunking, dynamic batching with gated execution, and early error interruption. We evaluate Eager across four benchmarks, seven LLMs, and three execution environments. Results show that Eager reduces the non-overlapped execution latency by up to 99.9% and the end-to-end latency by up to 55% across seven LLMs and four benchmarks.
