📐 计分规则Scoring Rules
能力上限 = 加权原始分 = ToolCall×0.3 + BugFind×0.3 + HermesAgent×0.4。反映模型的最高能力水平。Max Score = Weighted raw score = ToolCall×0.3 + BugFind×0.3 + HermesAgent×0.4.
实用得分 = 能力上限 − 重试扣分。每题首次通过不扣分；重试后才通过的题目每重试1次扣1分；重试后仍失败的不参与扣分。Effective Score = Max Score − retry penalty. First pass = no penalty; each retry before passing = −1pt; failed retries excluded.
本模型能力上限 93.1（排名第 3），实用得分 57.1（扣分 36 分，主要来自 HA-04/08/12/14/19 的多次重试和 HA-07/10/16/17/20 的反复失败）。This model: Max Score 93.1 (Rank #3), Effective Score 57.1 (−36 penalty, mainly from HA-04/08/12/14/19 retries and HA-07/10/16/17/20 repeated failures).
| 题号ID | 分数Score | 结果Result | 重试Retry | 说明Summary |
|---|
| 题号ID | 分数Score | 结果Result | 重试Retry | 说明Summary |
|---|
| 题号ID | 分数Score | 结果Result | 重试Retry | 说明Summary |
|---|
| 题号ID | 分数Score | 结果Result | 重试Retry | 失败原因Failure Reason |
|---|
📊 模型评估总结Model Evaluation Summary
Agents-A1（无思考模式）展现了与思考版截然不同的特征——能力上限更高（93.1 vs 91.2），但实用得分更低（57.1 vs 71.2）。最显著的是 BugFind-15 满分通过（100 vs 88），尤其在 BF-10 红鲱鱼陷阱上，无思考模式没有过度分析，直接识别出代码正确并通过。Agents-A1 (No Think) shows a different profile from the Thinking version — higher ceiling (93.1 vs 91.2) but lower effective score (57.1 vs 71.2). Most notably, BugFind-15 is perfect (100 vs 88). On BF-10 red herring, the no-think mode recognized the code was correct without overthinking.
重试成本爆炸是最大短板：HermesAgent-20 重试次数从思考版的 9 次暴涨到 24 次（+167%）。无思考让模型直接执行而不再"先思考再行动"，工具调用、消息投递、并行委派等场景需要试错恢复。HA-04/08/12/14/19 都重试 2-9 次才通过，HA-07/10/16/17/20 永久失败（partial 15-70 分）。Retry cost explosion is the biggest weakness: HermesAgent-20 retries jumped from 9 to 24 (+167%) versus the thinking version. No-think mode executes directly without reflecting first, requiring more trial-and-error in tool calls, message delivery, and parallel delegation. HA-04/08/12/14/19 each needed 2-9 retries; HA-07/10/16/17/20 are permanently failed.
能力天花板 93.1 超过了思考版（91.2）和 Ornith（93.5），但 36 次重试让实用分只有 57.1——比思考版低 14.1 分。**无思考模式适合低风险的单步任务**（纯代码调试、简单工具调用），**不适合复杂 agent 流程**（多步决策、状态管理、错误恢复）。TC-11 算而不用（50 分）也说明无思考在需要判断"该用工具还是用脑"的场景仍有缺陷。Ceiling 93.1 exceeds the thinking version (91.2) and Ornith (93.5), but 36 retries drag effective score to 57.1 — 14.1 points lower than thinking. No-think mode suits low-risk single-step tasks (pure code debugging, simple tool calls), but not complex agent workflows (multi-step decisions, state management, error recovery). TC-11 (50/100) also reveals no-think weakness in scenarios requiring "tool vs brain" judgment.