Research Papers 论文研究 5h ago Updated 1h ago 更新于 1小时前 48

Below the Noise Floor: Bimodal Seed Collapse and Distinct Failure Modes in Small-Model Knowledge Distillation 噪声 floor 之下:小模型知识蒸馏中的双峰种子崩溃与不同失败模式

Knowledge distillation (KD) gains below 5 percentage points in small-model function routing are statistically indistinguishable from seed variance, as per-seed standard deviations range from 2.8 to 48.7 percentage points Three of seven KD variants exhibit bimodal collapse, with 1 in 3–5 seeds dropping below 55% accuracy while others train normally; a fourth shows elevated variance Collapse manifests in distinct failure modes: wrong-function selection (ce_kd, ce_paraphrase) and a novel output-tru 小模型知识蒸馏中单种子评估存在严重缺陷,无法检测核心失败模式 740实例医疗API路由任务中,8种KD变体有3种出现双峰种子崩溃,标准差高达48.7个百分点 不同KD方法崩溃模式各异:ce_kd和ce_paraphrase表现为错误函数选择,reasoning_kd出现输出截断(准确率仅0.9%) 仅progressive_kd和rank_kd在多种子测试中保持稳定(sigma ≤ 3.9 pp) 输入增强的性能增益在控制变量后逆转,揭示单种子评估的误导性

62
Hot 热度
76
Quality 质量
68
Impact 影响力

Analysis 深度分析

TL;DR

  • Knowledge distillation (KD) gains below 5 percentage points in small-model function routing are statistically indistinguishable from seed variance, as per-seed standard deviations range from 2.8 to 48.7 percentage points
  • Three of seven KD variants exhibit bimodal collapse, with 1 in 3–5 seeds dropping below 55% accuracy while others train normally; a fourth shows elevated variance
  • Collapse manifests in distinct failure modes: wrong-function selection (ce_kd, ce_paraphrase) and a novel output-truncation mode (reasoning_kd), where the model generates reasoning but terminates before emitting a function name, achieving only 0.9% accuracy
  • Only progressive_kd and rank_kd avoid collapse across observed seeds, maintaining sigma ≤ 3.9 percentage points
  • Single-seed evaluation is fundamentally inadequate for detecting central failure modes in small-model KD; input enrichment gains also reverse under controlled multi-seed testing

Why It Matters

This work exposes a critical validity gap in small-model knowledge distillation research: single-seed evaluations routinely overstate KD gains and miss catastrophic bimodal failure modes entirely. For practitioners deploying distilled models in production function-routing scenarios, these findings mean that reported improvements may be noise rather than signal, and that certain KD variants carry hidden collapse risk that only multi-seed evaluation can surface.

Technical Details

  • Setup: 740-instance healthcare API routing task using a 1.5B Qwen student model distilled from a 20B teacher, comparing eight KD variants against supervised cross-entropy baseline with 3–6 seeds per key configuration
  • Bimodal collapse: Three KD variants (ce_kd, ce_paraphrase, reasoning_kd) show bimodal performance distributions; ce_kd and ce_paraphrase fail via wrong-function selection, while reasoning_kd exhibits a previously undocumented output-truncation failure mode (model emits reasoning tokens but terminates before function name output, yielding 0.9% accuracy)
  • Stable variants: progressive_kd and rank_kd are the only configurations avoiding collapse across all observed seeds, with standard deviations capped at 3.9 percentage points
  • Input enrichment reversal: A naive cross-split input enrichment strategy showed a +3.78 pp gain but reversed to −2.70 pp under controlled within-split multi-seed re-testing, demonstrating evaluation artifact susceptibility
  • Variance magnitude: Per-seed standard deviations of 2.8–48.7 pp completely engulf any claimed KD gain below 5 pp, rendering most published small gains statistically meaningless

Industry Insight

  • Multi-seed evaluation should be treated as a minimum standard for any small-model KD paper or production deployment claim; single-seed results are insufficient to distinguish signal from variance and may conceal catastrophic failure modes
  • When selecting KD strategies for production function-routing systems, progressive_kd and rank_kd offer the only empirically verified robustness against bimodal collapse among the variants tested
  • Input enrichment and other preprocessing tricks should be validated under within-split multi-seed protocols rather than cross-split designs, as evaluation methodology alone can reverse apparent performance gains

TL;DR

  • 小模型知识蒸馏中单种子评估存在严重缺陷,无法检测核心失败模式
  • 740实例医疗API路由任务中,8种KD变体有3种出现双峰种子崩溃,标准差高达48.7个百分点
  • 不同KD方法崩溃模式各异:ce_kd和ce_paraphrase表现为错误函数选择,reasoning_kd出现输出截断(准确率仅0.9%)
  • 仅progressive_kd和rank_kd在多种子测试中保持稳定(sigma ≤ 3.9 pp)
  • 输入增强的性能增益在控制变量后逆转,揭示单种子评估的误导性

为什么值得看

这篇论文揭示了小模型知识蒸馏领域长期被忽视的评估缺陷——单种子实验无法捕捉种子方差导致的崩溃风险,对AI从业者优化小模型部署具有重要警示意义。

技术解析

  • 实验设置:740实例医疗API路由任务,1.5B Qwen学生模型,20B教师模型,8种KD变体对比,3-6个种子重复实验
  • 双峰崩溃现象:3/7 KD变体出现双峰分布,部分种子准确率低于55%,其他正常训练
  • 崩溃模式分类:ce_kd和ce_paraphrase表现为错误函数选择,reasoning_kd出现输出截断(模型推理后未输出函数名即终止)
  • 稳定方法:仅progressive_kd和rank_kd在多种子测试中保持稳定(sigma ≤ 3.9 pp)
  • 评估方法学:输入增强的+3.78 pp增益在受控多种子测试下逆转至-2.70 pp,证明单种子评估的不可靠性

行业启示

  • 小模型蒸馏评估必须采用多种子重复实验,单种子结果不可信
  • 选择KD方法时应优先考虑progressive_kd和rank_kd等稳定性高的变体
  • 输入增强等数据策略的效果需要严格的多种子验证,避免虚假增益

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Research 科学研究 Training 训练 Deployment 部署 Healthcare AI 医疗AI