发表机构
Openhagen（奥彭哈根）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
该研究基于Qwen2.5系统，通过实验证实阶段重放分歧由KV缓存携带，固定前缀精度会影响分歧表现，双向缓存移植可复现分歧轨迹，精确令牌重放无需保留实时状态保真度。
AI 中文摘要
阶段重放诊断会重构中间令牌前缀，并将新鲜预填充的续段视为从最初到达该前缀的解码器状态开始的续段。我们在基于Qwen2.5的系统中，于整个推理阶段边界处检验了该假设。一项包含200个样本的匹配实验，比较了保留的实时缓存与相同整数令牌的一次性预填充，并在两侧放置了完全相同的副本。在BF16精度下，副本完全一致，而两种构造在166个后缀和20个正确性标签上存在差异；准确率差异仅为1个百分点（配对95%置信区间[-3.5, +5.5]）。固定前缀2x2设置使所有200个令牌状态保持恒定，同时跨越构造和精度进行对比。BF16下的分歧再次出现，而FP32下未产生解码分歧（95%威尔逊上界为1.88%）。一项前瞻性桥接使逐令牌增量缓存与保留的实时缓存在12/12行上逐位精确匹配；对全部200个保存的轨迹账本的审计，复现了所有保留的轨迹和比较指纹。对全部48个键/值层的双向移植，使所有测试的分歧续段均遵循其缓存供体，无论是主检查点的选定集合（24/24）还是后续检查点的盲结果复现（43/43）。因此，精确令牌重放可在不保留实时状态保真度的情况下实现可重复性。在测试状态下，边界KV缓存是分歧轨迹的因果充分载体，而数值精度会调节其行为表现。
英文摘要
Stage-replay diagnostics reconstruct intermediate token prefixes and treat fresh-prefill continuation as continuation from the decoder state that originally reached the prefix. We audit that assumption at a whole reasoning-stage boundary in a Qwen2.5-derived system. A matched 200-item experiment compares retained live cache with one-shot prefill of identical integer tokens and places an exact replica on both sides. In BF16, replicas remain exact while the constructions differ on 166 suffixes and 20 correctness labels; the accuracy difference is only one point (paired 95% CI [-3.5, +5.5]). A fixed-prefix 2x2 holds all 200 token states constant while crossing construction and precision. The BF16 disagreements recur, whereas FP32 produces no decoded disagreement (95% Wilson upper bound 1.88%). A prospective bridge makes token-by-token incremental and retained live caches bit-exact on 12/12 rows; an all-200 saved-ledger audit reproduces every retained trajectory and comparison fingerprint. Bidirectional transplantation of all 48 key/value layers makes every tested divergent continuation follow its cache donor, both on a selected set at the primary checkpoint (24/24) and an outcome-blind replication at a later checkpoint (43/43). Exact-token replay can therefore be repeatable without preserving live-state fidelity. On the tested states, boundary K/V cache is a causally sufficient carrier of the divergent trajectory, while numerical precision moderates its behavioral expression.
Comments15 pages, 1 figure, 6 tables. Reproducibility artifacts (frozen manifests, token IDs, per-item scores, analysis harnesses) described in Section 3.9
