当内置思考既有帮助又有害：指令遵循中的约束级错误转移
When Built-in Thinking Helps and Hurts: Constraint-Level Error Shifts in Instruction Following
AI总结：
研究大型推理模型（LRM）的思考模式对指令遵循的影响，发现思考会改变错误模式而非统一降低性能，其中规划类约束改善而精确类约束恶化，并通过分析思考轨迹和激活修补揭示了机制。
摘要：
大型推理模型（LRM）通常能提升数学和编码性能，但其对指令遵循的影响尚不明确。我们使用 Qwen3 模型（1.7B-32B）研究 IFEval，采用同权重的思考开启/关闭控制；四个 Hunyuan 模型提供跨家族方向性支持。总体通过率变化很小（-0.55 到 -3.52 个百分点），但 10-20% 的提示在两种模式间在通过和失败之间切换，表明思考改变了错误模式——某些提示改善而另一些恶化——而非统一降低性能。在事后 Qwen3 导出的分组下，约束类型分为规划类（全局计数、结构、协调）和精确类（精确局部形式）；规划类在思考下类别层面改善，而精确类持续恶化；尽管 Hunyuan 的总体方向相反，但所有四个 Hunyuan 模型在类别层面的规划/精确符号模式方向一致。思考还改变了最终答案长度；匹配长度分析大幅减少了精确类的下降，但仍有残余惩罚。使用交叉编码器相关性指标分析思考轨迹揭示了三种模式：中性模式显示正的相关性-合规性关联（r ≈ 0.15）；规划模式显示接近零的预测相关性（r ≈ 0.02），尽管有可测量的轨迹参与，这与 CE 测量的轨迹相关性和最终答案合规性之间的执行差距一致；精确模式显示小的负相关性（r ≈ -0.05），失败实例的平均相关性高于通过实例。跨四个模型大小（1.7B-14B）的激活修补显示，精确类翻转实例比规划类翻转实例更常被恢复（32-58% 对 14-40% 的平均层恢复），最大差距在 14B 处（约 30 个百分点）。
英文摘要：
Large reasoning models (LRMs) often improve math and coding performance, but their effect on instruction following is unclear. We study IFEval with Qwen3 models (1.7B-32B), using same-weights Thinking ON/OFF controls; four Hunyuan models provide directional cross-family support. Aggregate pass-rate changes are small (-0.55 to -3.52 pp), yet 10-20% of prompts switch between pass and fail across modes, suggesting that thinking changes the pattern of errors--some prompts improve while others worsen--rather than uniformly degrading performance. Under a post-hoc Qwen3-derived grouping, constraint types separate into Planning (global counting, structure, coordination), which improves at the class level under thinking, and Precision (exact local form), which consistently worsens; the class-level Planning/Precision sign pattern holds directionally for all four Hunyuan models despite Hunyuan's opposite aggregate direction. Thinking also changes final-answer length; matched-length analyses substantially reduce the Precision drop, but a residual penalty remains. Analyzing thinking traces with a cross-encoder relevance metric reveals three patterns: Neutral shows a positive relevance-compliance link (r approximately 0.15); Planning shows near-zero predictive correlation (r approximately 0.02) despite measurable trace engagement, consistent with an execution gap between CE-measured trace relevance and final-answer compliance; Precision shows a small negative correlation (r approximately -0.05), with failing instances having higher mean relevance than passing ones. Activation patching across four model sizes (1.7B-14B) shows that Precision flip instances are more often restored than Planning flip instances (32-58% vs. 14-40% mean layer-restoration), with the largest gap at 14B (about 30 pp).
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
