华为论文 EvoTrias:搜索、写码、评分三权分立的代码进化智能体更省更强
华为这篇论文做了组很干净的消融:把搜索、写码、打分拆给三个角色,打分器还对 agent 隐藏。结果比 4-agent 的 CORAL 省 64% token,7 个数学任务赢了 6 个。做代码优化 agent 循环的工程师直接能抄结论。
华为发布 EvoTrias 论文,核心结论是代码进化系统把搜索规划、代码编写、评分三个环节交由独立角色承担,效果更好且成本更低。现有 AlphaEvolve 式系统要么限制代码编写者自由度,要么让单个编码智能体主导搜索,代价是 token 消耗高,且没有任何系统对智能体隐藏评分器。
EvoTrias 的机制是:controller 负责引导搜索方向,Claude Code 这类编码智能体只负责写代码,evaluator 打分且智能体无法读取其内容,防止“刷分”行为。在 7 个数学任务上,EvoTrias 在 6 个任务领先,相比 4-agent 方案 CORAL,每次运行的 token 成本降低 64%。
在 OR-Tools 这个 68 万行 C++ 代码库的实测中,EvoTrias 在 300 个实例上将求解器质量提升 13.2%,而同一智能体的 1-shot 运行平均成绩为 -0.1%。对做代码优化智能体循环的开发者,论文给出的操作建议是:对智能体隐藏评分器,并避免让编码智能体自己规划搜索。
New Huawei paper finds that code-evolution agents do better and spend less when search planning, code writing, and scoring are handled by 3 separate owners.
AlphaEvolve-style systems improve programs by writing code, scoring it, and picking what to try next. Some give the code writer little freedom, others let 1 coding agent steer the search at high token cost, and none hides the scorer.
EvoTrias gives each job its own owner. A controller steers the search, a coding agent like Claude Code writes code, and an evaluator the agent cannot read scores it.
On 7 math tasks it led 6 and cut token cost per run by 64% versus 4-agent CORAL. On OR-Tools, a 680K-line C++ repository, it lifted solver quality 13.2% on 300 instances, while 1-shot runs of the same agent averaged -0.1%.
For code-optimizing agent loops, hide the scorer from the agent and keep search planning out of the coding agent.
This report is categorized under AITOP's "Papers" desk, focusing on China's AI industry and domestic foundation model ecosystem. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include EvoTrias, AlphaEvolve, Claude Code, OR-Tools.
Source attribution: curated from rohanpaul_ai (@rohanpaul_ai).