mlsanityData to 14 Oct 2026

Stage-Replay Divergence Follows the KV Cache: Fixed-Prefix Precision Controls and Bidirectional Cache Transplantation

arXiv 2607.28495 ↗First posted 30 Jul 2026cs.LG cs.CLcited by 1

On the map: Efficient LLM inference and serving → Efficient reasoning and agent inference

Abstract (arXiv)

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.

Key facts. Cited by 1 paper in the map (data through 14 Oct 2026). Placed in Efficient reasoning and agent inference, one of 12 directions in Efficient LLM inference and serving. Builds on 14 of its 18 references that are in the map. First posted on arXiv 30 Jul 2026.

Builds on

14 of 18 references.

References · 14

  1. 2024Qwen2.5 Technical Reportcited by 4,942
  2. 2023GPQA: A Graduate-Level Google-Proof Q&A Benchmarkcited by 2,879
  3. 2023Branch-Solve-Merge Improves Large Language Model Evaluation and Generationcited by 106
  4. 2025Recursive Self-Aggregation Unlocks Deep Thinking in Large Language Modelscited by 35
  5. 2025Learning to Reason Across Parallel Samples for LLM Reasoningcited by 32
  6. 2026PaCoRe: Learning to Scale Test-Time Compute with Parallel Coordinated Reasoningcited by 17
  7. 2025Deterministic Inference across Tensor Parallel Sizes That Eliminates Training-Inference Mismatchcited by 14
  8. 2026The Illusion of Equivalence: Systematic FP16 Divergence in KV-Cached Autoregressive Inferencecited by 4
  9. 2026Beyond Consensus: Trace-Level Synthesis in Mixture of Agentscited by 4
  10. 2026MarginGate: Sparse Margin-Triggered Verification for Batch-Invariant LLM Inferencecited by 3
  11. 2026SPIRAL: Learning to Search and Aggregatecited by 3
  12. 2026Amortized Reasoning Tree Search: Decoupling Proposal and Decision in Large Language Modelscited by 1
2 more
  1. 2026Counterfactual Likelihood Tests for Indirect Influence in Private Reasoning Channelscited by 1
  2. 2026Re-feeding Is Not Replaying: Measuring Replay Noise in Counterfactual Token-Credit Estimationcited by 1

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 1.

1 paper.

  1. 2026Same Request, Different Answer: Quantization Amplifies Cache-Induced Divergence in LLM Servingcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026Smarter and Cheaper at Once: Byte-Exact KV-Cache Grafting Turns a Frozen Small Model into a Verified-Knowledge Flywheel
  2. 2026Low-Discrepancy Dither for Quantized Recurrent State Caches
  3. 2025Unlearning at Scale: State-Exact Trace-Preserving Deletion in Billion-Parameter Language Models
  4. 2026CacheReforge: Bounded Recovery for Stale KV Caches under Evolving Adapters
  5. 2026Just Let Linear States Forget the Distant Past: Prefix Caching via Suffix Replay for Hybrid LLMs
  6. 2026Aborted but Not Forgotten: KV-Cache Retention Breaks Rollback Consistency in Language Agents
  7. 2026ReplaySSM: Cache SSM Inputs, Not State
  8. 2026Stop the Flip-Flop: Context-Preserving Verification for Fast Revocable Diffusion Decoding