Detecting a Route Flip Is Easier Than Knowing Whether to Fix It: Causal Route-Mediated Damage in Quantized Mixture-of-Experts

检测路由翻转比判断是否需要修复更容易：量化混合专家模型中的因果路由介导损伤

Abstract: Top-k Mixture-of-Experts (MoE) routing is discontinuous, so a deployment-motivated numerical disturbance — simulated 4-bit KV-cache quantization read by a protected BF16 gate — pushes tokens across decision boundaries and flips which experts fire. This paper proposes no new mitigation; it supplies a causal apparatus, empirical findings, and a detection-limit result.

摘要： Top-k 混合专家模型（MoE）的路由机制是不连续的，因此由部署需求引发的数值扰动（即由受保护的 BF16 门控读取模拟的 4-bit KV 缓存量化）会将 Token 推过决策边界，从而改变激活的专家模型。本文并未提出新的缓解方案，而是提供了一套因果分析装置、实证研究结果以及关于检测极限的结论。

A four-run apparatus prices the route-mediated fraction (RMF) of quantization damage, a token-level attribution decomposes it by mechanism, and pre-registered probes carry the findings across three architectures. On OLMoE-1B-7B at 4-bit KV (pilot), about a third of the damage is routing-mediated: RMF ~ 0.31 (discovery 0.31 [0.20, 0.41]; process-replicated mean 0.313 +/- 0.020; pre-registered re-execution 0.231).

一套四轮运行的装置评估了量化损伤中路由介导的部分（RMF），通过 Token 级别的归因分析将其按机制分解，并利用预注册的探针在三种架构上验证了这些发现。在 OLMoE-1B-7B 模型（4-bit KV 试点）中，约三分之一的损伤是由路由介导的：RMF 约为 0.31（发现值为 0.31 [0.20, 0.41]；过程重复均值为 0.313 +/- 0.020；预注册重执行结果为 0.231）。

The deployable router margin detects that a flip occurred (AUC 0.772) but cannot tell a harmful flip from a helpful one (at chance): among the tested local, inference-observable router statistics we find no predictor of a flip’s loss sign above chance — an empirical benefit-detection barrier bounding selective repair restricted to this feature family.

可部署的路由器边际（Router Margin）能够检测到翻转的发生（AUC 0.772），但无法区分翻转是有害还是有益（准确率仅为随机水平）：在测试的本地及推理可观测的路由器统计数据中，我们未发现任何能以高于随机概率预测翻转损失符号的指标——这构成了一个实证上的“收益检测壁垒”，限制了该特征族内选择性修复的可行性。

The signed-flip tax and sign-inseparability carry cross-model; the clean-reference remedy’s payout is architecture-modulated; a controlled same-checkpoint flag-swap re-scopes the gate’s normalization convention to a damage-magnitude moderator, not a route-recoverability mechanism.

这种“翻转代价”及其符号不可分性在不同模型间具有通用性；“干净参考修复法”的收益受架构调节；通过受控的同一检查点标志交换（flag-swap），我们将门控归一化约定的作用重新界定为损伤幅度的调节器，而非路由可恢复性机制。

A real int4 KV kernel yields a fraction compatible with the fake-quant dose curve but underpowered (95% CI [-0.111, 0.394] includes zero) — ruling out gross disagreement, not an independent replication. Hypotheses, thresholds, and evaluations were pre-registered before measurement, with misses reported; a pre-registered held-out read replicates the partition and the near-cancelling tax out of sample, while the strict impossibility exclusion narrowly misses.

真实的 int4 KV 内核产生的比例与伪量化剂量曲线相符，但统计效能不足（95% 置信区间 [-0.111, 0.394] 包含零）——这排除了严重偏差的可能性，但尚不能视为独立的重复验证。研究假设、阈值和评估方法均在测量前进行了预注册，并报告了未达标项；一项预注册的留出集读取验证在样本外重复了该分区结果及近乎抵消的代价，而关于“严格不可能”的排除结论则以微小差距未被完全证实。