Pith. sign in

REVIEW 2 major objections 7 minor 36 references

A retained KV-cache event can answer from a source that was never served, because prefill already wrote compact state into its rows.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-30 15:43 UTC pith:OQSONFHI

load-bearing objection Solid causal measurement of source-omitted state in downstream KV; the 99:0 direction holds, the “memory contract” branding runs a bit ahead of the binary-synthetic envelope. the 2 major comments →

arxiv 2607.23693 v1 pith:OQSONFHI submitted 2026-07-26 cs.AI

Compute Globally, Materialize Locally: The Memory Contract of Sparse Event-KV

Zefeng Cai , Zerui Cai This is my paper
classification cs.AI
keywords KV cachesemantic materializationsparse event servinglong-horizon agentsmemory contractevictiondonor paircarrier events
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Long-horizon agents treat the KV cache as memory and drop most events at serve time. This paper shows that a kept downstream event can still carry the value of an omitted upstream observation: when two histories are byte-identical in every served token and position and differ only in the dropped source, answers follow the hidden donor almost perfectly on recent open models. The authors call this semantic materialization—an independently servable view of computation whose inputs are gone. The effect is programmable with deliberately phrased, answer-free carrier events, but passive harvesting of natural mentions from real dialogs gives no detected gain, and only compact state survives. The practical upshot is a memory contract for sparse event-KV serving: what to write, where it lands, and what remains after the source is dropped—plus the warning that an eviction ablation with no accuracy loss has not proved the dropped source was unnecessary.

Core claim

Among donor-sensitive items, with the source event omitted and every served token and position held fixed, answers follow the omitted value (99:0 on Qwen3-8B and directionally across other recent checkpoints and serving cells). A downstream event’s cached rows therefore act as an independently servable carrier of compact upstream state, not merely of their visible text. Deliberate answer-free carriers can raise donor-aligned recovery substantially; passive natural mentions do not.

What carries the argument

Semantic materialization, isolated by the donor-pair test: two histories identical in every served token and position, differing only in the omitted source, so any systematic answer flip must come from state already written into the retained rows at prefill.

Load-bearing premise

That the calibrated synthetic register and mirror constructions, and the per-model carrier phrasings that work in controlled trajectories, will transfer to messy real agent dialogs where passive natural mentions already fail to write reliably.

What would settle it

Run the same donor-pair protocol on production agent traces: omit a source fact, serve only natural downstream mentions at original positions, and check whether free-generation or menu answers still follow the omitted donor above chance after decoy controls; a null that matches the paper’s REALTALK/LoCoMo result would bound the claim to synthetic carriers.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Eviction and episodic-memory policies that keep rows are not selecting only visible token content; retained rows can already hold the answer.
  • Systems can deliberately emit compact, answer-free carrier events for state needed after its operands leave the serve set, with an explicit textual record as fallback.
  • Dropping a source and seeing no accuracy drop does not prove the source was unnecessary—it may have been materialized into a kept row.
  • Updates should be appended as new versioned events and routed by query, not applied as global cache edits, because patches hijack historical queries.
  • Access is a compact-state envelope: binary state recovers well; larger categorical and exact numeric payloads fall toward chance.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Memory-shell designs that only retrieve original-history episodes may be leaving an unused write path on the table if they never emit calibrated carriers.
  • Hybrid or sliding-window models will need position-preserving serve geometry, not only row selection, or materialization readability will diverge from full-attention results.
  • If trigger is surface form rather than meaning, automatic carrier synthesis will need model-specific write-through probes, not a single universal prompt template.
  • The same donor-pair test could audit whether cross-chunk KV fusion and compression pipelines preserve or destroy source-omitted materialized state.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 7 minor

Summary. The paper tests whether a retained KV-cache span remains informative after the observation that produced it is evicted. Using donor pairs — two agent histories byte-identical in every served token and position, differing only in an omitted source event — the authors show that among donor-sensitive items the answer follows the omitted source 99:0 on Qwen3-8B (130:0 in a 256-pair replication), though no served span states the value. They term this semantic materialization, and map its boundary: a deliberately phrased answer-free "carrier" event raises donor-aligned recovery from 6% to 51% (§7); write-through is a surface-form × model property with no construction class-W on all three recent checkpoints (Table 2); roots carry the value while reference edges mostly route (§5); access is bounded to compact state — binary recovers at .934, four/eight-way decay toward chance, three-digit payloads fail entirely (§6). Passive harvesting of natural mentions on REALTALK/LoCoMo yields no detected advantage (Table 6). The stated corollary for eviction research: dropping a source and observing no accuracy loss does not show the source was unnecessary.

Significance. If the result holds, it is a genuinely useful substrate-level measurement for the KV-cache-as-memory literature (SnapKV, H2O, EpiCache, CacheBlend): eviction validations that drop a source event and see no accuracy loss have not isolated necessity, and this paper supplies both the counterexample mechanism and a clean instrument for detecting it. The experimental hygiene is unusually strong for an empirical systems/ML paper: byte-identity and leakage assertions enforced in code, independently sampled decoys with a decoy-stratified re-audit of the binary case, exact sign tests, Wilson/TOST intervals, position-controlled serve-set ablations (X10/X11), a sliding-window native-geometry gate for Gemma-4 (App. D), ITT sensitivity for conditioned readouts, and a seed-pinned reproducibility release with revision SHAs. The negative results (passive-harvest null, no universal construction, Ministral-3 abstention profile) are reported with the same care as the positives, which raises confidence in the whole.

major comments (2)
  1. [App. A, 'Donor pairs, the operative null, and transition classes'] The operative null as stated — 'under the byte-identity control with greedy decoding, the two donor runs would produce identical answers absent any contextual carryover' — is literally false, and the manuscript's own compatibility gate documents the failure mode: the retained root rows were computed with the source present, so the two donor runs never have bit-identical inputs, and greedy decoding on numerically different caches can diverge at bf16 near-ties (the gate's own example: top-1/top-2 margin 0.25 below the 0.53 path deviation; 3/24 wording-level divergences on Qwen3). Consequently the donor-dependence prevalence (131/256 in the replication; the 'donor-sensitive' conditioning behind Table 1) is a numerical-sensitivity measure that can absorb near-tie flips, not purely a semantic-carryover measure. The headline direction claim is nonetheless secure: noise-induced flips are donor-
  2. [§4, §7, Abstract/Conclusion — scope of the 'memory contract'] The positive contract (trigger/landing/access) is established entirely on synthetic register/mirror/flag constructions with prespecified event roles, and the manuscript's own evidence shows transfer to natural workloads is fragile: passive harvest on REALTALK/LoCoMo yields no detected advantage (Table 6; Gemma-4 LoCoMo equivalent to isolated encoding within ±.05), no construction is class W on all three recent models (Table 2), and the preregistered two-model criterion failed (App. C). The abstract does disclose the passive-harvest null, but 'a memory contract for sparse event-KV serving: what to write, where it lands, and what survives' reads as a serving-system result. Please state explicitly in the abstract and conclusion that the positive contract is calibrated per checkpoint on synthetic constructions and that deployment on natural agent trajectories is currently unestablished (exis
minor comments (7)
  1. [Table 1] Report per-cell denominators (donor-dependent pairs / total pairs), not only follow:anti among donor-sensitive items; the direction ratios are otherwise not interpretable as prevalence without footnote 2 and the replication paragraph. Relatedly, the Qwen3 mention-co-served cell (39:18, cited in §3) is far weaker directionally than every other cell,
  2. [Fig. 2A] Bars are rates over differing denominators (n=128/187/252/364); the caption notes comparability only as rates, but please also annotate raw follow/anti counts on the bars, since the Qwen3 co-served anti rate of 0.07 corresponds to the 18 anti pairs that constitute the weakest direction cell in the paper.
  3. [§2 'Where the query goes' vs App. D] Synthetic arms append the query after the last retained row, which 'compacts the gap left by the omitted source'; App. D shows compact vs. native geometry disagrees sharply on Gemma-4 beyond the 1024 window (argmax agreement 29/32 → 4/32). Please state in §2 that synthetic-arm effect sizes are optimistic relative to eviction-style native geometry, with a forward pointer to App. D and the Table 12 geometry comparison.
  4. [§4 write classes] The W/P/U thresholds (≥.75, <.625) and the TOST bands (±.05/±.075) are reader-visible free parameters; please state whether they were fixed before seeing the bank results, and one sentence on the rationale for the partial band would help.
  5. [§6] 'Three-digit payloads are absent entirely (0/192)' sits awkwardly next to the later 'Query form matters' paragraph (first-digit probe .20, quantile query .62). Consider reordering so the exact-recall failure and the partial-access-via-query-form result are presented together.
  6. [App. F / §8] Leyline (Ma et al., 2026) is cited only in App. F; given it proposes serving-side edit/remove/replace directives, it belongs in the §8 update/repair discussion as well.
  7. [General] The running head appears to have lost inter-word spacing ('COMPUTEGLOBALLY, MATERIALIZELOCALLY'); likely an extraction artifact, but please check the final PDF. In App. A, clarify which gate metric 'Gemma-4 23/24 exact-string agreement' refers to relative to the three reported levels (content, first-token, logit deviation).

Circularity Check

0 steps flagged

No significant circularity: empirical donor-pair measurement, not a derivation that folds inputs into claimed predictions.

full rationale

The paper’s load-bearing claims are experimental outcomes under a controlled donor-pair protocol (source omitted; served tokens and positions held fixed; greedy decoding), not first-principles derivations. Donor dependence is an operational definition—outputs differ across byte-identical served spans—and the headline 99:0 (and 130:0 replication) is a conditional sign test on direction among differing pairs, not a quantity fitted then re-predicted. Write-class thresholds (W/P/U), TOST equivalence bands, and menu/trace readout choices are analysis conventions that organize rates; they do not force the donor-following direction or the 6%→51% carrier lift. Related-work citations (Li 2026; Kontonis et al. 2026; SnapKV; EpiCache, etc.) situate the phenomenon and do not supply a uniqueness theorem or ansatz that the present results merely rename; author names do not overlap those citations in a load-bearing way. Skeptical concerns about bf16 near-tie flips under the operative null are identification/correctness issues, not circular reduction of a prediction to its inputs. The work is self-contained against its own benchmarks and synthetic/real-dialog arms.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 3 invented entities

The central claim is experimental, so the ledger is mostly protocol axioms and analysis thresholds rather than physical free parameters. Load-bearing commitments are: frozen native-interface recovery without trained readers; donor-pair byte-identity plus greedy decoding as the causal null; synthetic event roles as proxies for agent memory; and class/equivalence cutoffs that organize but do not create the headline directionality. Invented entities are conceptual labels for the measured channel, not new physical objects.

free parameters (3)
  • Write-class thresholds (W≥0.75, U<0.625, P in between) = W≥0.75; U<0.625
    Hand-chosen cutoffs that map continuous write-through rates into W/P/U classes in the 16-construction bank; they organize Table 2 but do not generate the donor-follow phenomenon.
  • Equivalence bands for TOST (±0.05 dialog; ±0.075 multi-way) = ±0.05 (dialogs); ±0.075 (cardinality)
    Pre-registered analysis bounds used to declare equivalence to chance or to isolated encoding; conventional but claim-shaping for the passive-harvest and multi-way access verdicts.
  • Per-checkpoint carrier construction choice = model-specific (e.g. consist/mirrors strong on Qwen3; none W on Gemma-4 trace)
    Which surface form writes is selected empirically per model; the programmable-primitive claim depends on this calibration rather than a universal syntax.
axioms (5)
  • domain assumption Under greedy decoding and byte-identical served tokens/positions, differing donor outputs witness contextual carryover (operative null of the donor pair).
    Stated in §2 and App. A; foundation of all causal claims.
  • domain assumption No auxiliary trained reader or fine-tuning; frozen model recovers only through its native generation/logit interface.
    Serving contract (ii)–(iii) in §2; scopes all accessibility verdicts as output-interface measurements.
  • ad hoc to paper Prespecified synthetic event roles (source/root/edge/mention/carrier) and lexical heuristics on real dialogs adequately stand in for sparse event-KV agent serving.
    §2 span selection; the passive-harvest null and construction sensitivity show this proxy is imperfect.
  • domain assumption Retained rows keep original position ids / RoPE; query placement is a protocol choice that can change effect size.
    §2 and App. D/G; motivates native vs compact serving split on Gemma-4.
  • standard math Standard causal attention and KV cache semantics of transformer LMs.
    Background computational model assumed throughout prefill and serve re-assembly.
invented entities (3)
  • Semantic materialization independent evidence
    purpose: Name the phenomenon that downstream event KV encodes upstream-derived state recoverable without the source or verbatim text.
    Core coined term; operationalized via donor-aligned behavior with source omitted. Independent evidence is the donor-pair experiments themselves, not an external observable outside this paradigm.
  • Carrier / trigger construction independent evidence
    purpose: Denote answer-free events whose surface form causes write-through of upstream state into their KV rows.
    Engineering handle in §4 and §7; existence supported by controlled rate lifts, but which strings work is checkpoint-specific.
  • Memory contract (trigger, landing, access) for sparse event-KV no independent evidence
    purpose: Package measured regularities into a systems-facing interface: what to write, where it lands, what survives.
    Synthesizing framework rather than a separately falsified object; risk of over-generalizing beyond compact binary synthetic state.

pith-pipeline@v1.2.0-grok45-kimik3 · 28762 in / 4105 out tokens · 91951 ms · 2026-07-30T15:43:40.352676+00:00 · methodology

0 comments
read the original abstract

Long-horizon agents increasingly reuse their KV cache as memory: a serving system keeps a subset of cached entries and drops the rest. Eviction and episodic-memory schemes therefore rest on a premise rarely tested directly, that a retained event is still informative once the observations that produced it are gone. We test it by omitting one earlier observation from what is served, across otherwise identical agent histories. Among items sensitive to that observation, the answer overwhelmingly follows the omitted value, though no served span says which value is correct. We call this semantic materialization: a downstream event's cached rows act as an independently servable view of computation whose inputs are gone. It can also be written on purpose. A deliberately phrased, answer-free event raises donor-aligned recovery from 6% to 51% on Qwen3-8B without ever naming the value, whereas passively harvesting natural mentions from long-term dialog yields no detected advantage. What such a row carries is specific and bounded. Compact state survives, larger payloads decay toward chance, and whether a construction writes at all turns on phrasing rather than on meaning alone, so two phrasings the model comprehends equally well can diverge sharply. The result is a memory contract for sparse event-KV serving: what to write, where it lands, and what survives once the source is gone. For anyone who evicts the corollary is that dropping a source event and observing no accuracy loss does not show the source was unnecessary.

Figures

Figures reproduced from arXiv: 2607.23693 by Zefeng Cai, Zerui Cai.

Figure 1
Figure 1. Figure 1: Semantic materialization, and the comparison that isolates it. A [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Roots carry the dominant donor-aligned signal; served reference edges mostly route it (sparse [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Native recovery is strong for binary state and falls toward chance as payload cardinality grows; [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Passive natural mentions are not a dependable write interface; deliberate carriers establish the [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Cross-generation construction bank (exploratory): Qwen3 as a recent anchor and four 2024 legacy models. Sixteen constructions × five models (Qwen3/Gemma-2/Qwen2.5/Llama/Mistral). A: KV write￾through (serve carrier rows only, source omitted), class W/P/U/X. B: full text visible — near-uniformly solved except Mistral. n per cell: 64/48/64/48/32. model. Target-addressed readout collapses on Gemma-2’s latent a… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

36 extracted references · 24 linked inside Pith

  1. [1]

    APEX-MEM: Agentic semi-structured memory with temporal reasoning for long-term conversa- tional AI.arXiv preprint arXiv:2604.14362,

    Pratyay Banerjee, Masud Moshtaghi, Shivashankar Subramanian, Amita Misra, and Ankit Chadha. APEX-MEM: Agentic semi-structured memory with temporal reasoning for long-term conversa- tional AI.arXiv preprint arXiv:2604.14362,

    Pith/arXiv arXiv

  2. [3]

    Human-inspired episodic memory for infinite con- text LLMs.arXiv preprint arXiv:2407.09450,

    Zafeirios Fountas, Martin A Benfeghoul, Adnan Oomerjee, Fenia Christopoulou, Gerasimos Lam- pouras, Haitham Bou-Ammar, and Jun Wang. Human-inspired episodic memory for infinite con- text LLMs.arXiv preprint arXiv:2407.09450,

    arXiv

  3. [5]

    Gemma 4 model card, 2026.https://ai.google.dev/gemma/docs/ core/model_card_4

    Gemma Team. Gemma 4 model card, 2026.https://ai.google.dev/gemma/docs/ core/model_card_4. Cheng-Ping Hsieh, Simeng Sun, Samuel Kriman, Shantanu Acharya, Dima Rekesh, Fei Jia, Yang Zhang, and Boris Ginsburg. RULER: What’s the real context size of your long-context language models?arXiv preprint arXiv:2404.06654,

    Pith/arXiv arXiv 2026

  4. [6]

    EPIC: Efficient position-independent caching for serving large language models.arXiv preprint arXiv:2410.15332,

    Junhao Hu, Wenrui Huang, Weidong Wang, et al. EPIC: Efficient position-independent caching for serving large language models.arXiv preprint arXiv:2410.15332,

    Pith/arXiv arXiv

  5. [7]

    EpiCache: Episodic KV cache management for long-term conversation on resource-constrained environments.arXiv preprint arXiv:2509.17396,

    Minsoo Kim, Arnav Kundu, Han-Byul Kim, et al. EpiCache: Episodic KV cache management for long-term conversation on resource-constrained environments.arXiv preprint arXiv:2509.17396,

    Pith/arXiv arXiv

  6. [8]

    Model Year Bank W Probe Edge (root) X9 latent X9 expl

    Legacy rows (2024) are exploratory. Model Year Bank W Probe Edge (root) X9 latent X9 expl. Mistral-7B 2024 excl. (13 X) — — — — Gemma-2-9B 2024 8/16 8/16≥.90174:0 .53 1.00 Llama-3.1-8B 2024 0/16 split — — — Qwen3-8B 2025 5/16≈gen 84:0 † (root 89:0†) .51 1.00 Ministral-3-8B 2025 7/16 —298:0abst. .95 Gemma-4-12B 2026 0/16≥.90on 8/16 41:0 † ≈1.0 1.00 Active ...

    2024

  7. [9]

    KV-PRM: Efficient process reward modeling via KV-cache transfer for multi-agent test- time scaling.arXiv preprint arXiv:2607.09153,

    Peng Kuang, Haibo Jin, Xiaoyu Han, Yanli Wang, Xiaopeng Yuan, Ye Yu, Kaidi Xu, and Haohan Wang. KV-PRM: Efficient process reward modeling via KV-cache transfer for multi-agent test- time scaling.arXiv preprint arXiv:2607.09153,

    Pith/arXiv arXiv

  8. [10]

    REALTALK: A 21-day real-world dataset for long-term conversation.arXiv preprint arXiv:2502.13270,

    Dong-Ho Lee, Adyasha Maharana, Jay Pujara, Xiang Ren, and Francesco Barbieri. REALTALK: A 21-day real-world dataset for long-term conversation.arXiv preprint arXiv:2502.13270,

    Pith/arXiv arXiv

  9. [11]

    Models take notes at prefill: KV cache can be editable and composable.arXiv preprint arXiv:2606.17107,

    Bojie Li. Models take notes at prefill: KV cache can be editable and composable.arXiv preprint arXiv:2606.17107,

    Pith/arXiv arXiv

  10. [12]

    KVEraser: Learning to steer KV cache for efficient localized context erasing.arXiv preprint arXiv:2606.17034,

    Mufei Li, Shikun Liu, Dongqi Fu, Haoyu Wang, Yinglong Xia, Hong Li, Hong Yan, and Pan Li. KVEraser: Learning to steer KV cache for efficient localized context erasing.arXiv preprint arXiv:2606.17034,

    Pith/arXiv arXiv

  11. [13]

    SnapKV: LLM knows what you are looking for before generation.arXiv preprint arXiv:2404.14469,

    Yuhong Li, Yingbing Huang, Bowen Yang, Bharat Venkitesh, Acyr Locatelli, Hanchen Ye, Tianle Cai, Patrick Lewis, and Deming Chen. SnapKV: LLM knows what you are looking for before generation.arXiv preprint arXiv:2404.14469,

    Pith/arXiv arXiv

  12. [14]

    HYPIC: Accelerating hybrid-attention LLM serving with position-independent caching.arXiv preprint arXiv:2607.01299,

    Yifei Liu, Juntong Wu, Yang Liu, Junhao Hu, Minghao Li, Xiaoxu Chen, and Weihang Chen. HYPIC: Accelerating hybrid-attention LLM serving with position-independent caching.arXiv preprint arXiv:2607.01299,

    Pith/arXiv arXiv

  13. [15]

    The llama 3 herd of models.arXiv preprint arXiv:2407.21783,

    Llama Team. The llama 3 herd of models.arXiv preprint arXiv:2407.21783,

    Pith/arXiv arXiv

  14. [16]

    Leyline: KV cache directives for agentic inference

    Bole Ma, Jan Eitzinger, and Harald Koestler. Leyline: KV cache directives for agentic inference. arXiv preprint arXiv:2606.01065,

    Pith/arXiv arXiv

  15. [17]

    Evaluating very long-term conversational memory of LLM agents.arXiv preprint arXiv:2402.17753,

    Adyasha Maharana, Dong-Ho Lee, Sergey Tulyakov, Mohit Bansal, Francesco Barbieri, and Yuwei Fang. Evaluating very long-term conversational memory of LLM agents.arXiv preprint arXiv:2402.17753,

    Pith/arXiv arXiv

  16. [18]

    11 Preprint — July 2026 Mistral AI

    Released 2024-05-22; architecture from Jiang et al., arXiv:2310.06825.https://docs.mistral.ai/models/model-cards/ mistral-7b-0-3. 11 Preprint — July 2026 Mistral AI. Mistral 3, 2025.https://mistral.ai/news/mistral-3/. Charles Packer, Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G. Patil, Ion Stoica, and Joseph E. Gonzalez. MemGPT: Towards LLMs as operat...

    Pith/arXiv arXiv 2024

  17. [20]

    Qwen2.5 technical report.arXiv preprint arXiv:2412.15115,

    Qwen Team. Qwen2.5 technical report.arXiv preprint arXiv:2412.15115,

    Pith/arXiv arXiv

  18. [21]

    Qwen3 technical report.arXiv preprint arXiv:2505.09388,

    Qwen Team. Qwen3 technical report.arXiv preprint arXiv:2505.09388,

    Pith/arXiv arXiv

  19. [22]

    Model check- pointQwen/Qwen3.6-27B, Apache-2.0.https://huggingface.co/Qwen/Qwen3. 6-27B. Preston Rasmussen, Pavlo Paliychuk, Travis Beauvais, Jack Ryan, and Daniel Chalef. Zep: A temporal knowledge graph architecture for agent memory.arXiv preprint arXiv:2501.13956,

    Pith/arXiv arXiv

  20. [23]

    AAFLOW+: Stateful operator ab- straction with zero-copy distributed KV cache orchestration for multi-agent workflows.arXiv preprint arXiv:2607.10987,

    Arup Kumar Sarker, Alexander James Halpern, Mills Staylor, Aymen Alsaadi, Gregor von Laszewski, Yue Cheng, Shantenu Jha, and Geoffrey Fox. AAFLOW+: Stateful operator ab- straction with zero-copy distributed KV cache orchestration for multi-agent workflows.arXiv preprint arXiv:2607.10987,

    Pith/arXiv arXiv

  21. [24]

    InfoFlow KV: Information-flow-aware KV recompu- tation for long context.arXiv preprint arXiv:2603.05353,

    Xin Teng, Canyu Zhang, Shaoyi Zheng, et al. InfoFlow KV: Information-flow-aware KV recompu- tation for long context.arXiv preprint arXiv:2603.05353,

    Pith/arXiv arXiv

  22. [25]

    KVLink: Accelerating large language models via efficient KV cache reuse.arXiv preprint arXiv:2502.16002,

    Jingbo Yang, Bairu Hou, Wei Wei, et al. KVLink: Accelerating large language models via efficient KV cache reuse.arXiv preprint arXiv:2502.16002,

    arXiv

  23. [26]

    KEEP: A KV-cache-centric memory management system for efficient embodied planning.arXiv preprint arXiv:2602.23592,

    Zebin Yang, Tong Xie, Baotong Lu, Shaoshan Liu, Bo Yu, and Meng Li. KEEP: A KV-cache-centric memory management system for efficient embodied planning.arXiv preprint arXiv:2602.23592,

    arXiv

  24. [27]

    CacheBlend: Fast large language model serving for RAG with cached knowledge fusion.arXiv preprint arXiv:2405.16444,

    Jiayi Yao, Hanchen Li, Yuhan Liu, et al. CacheBlend: Fast large language model serving for RAG with cached knowledge fusion.arXiv preprint arXiv:2405.16444,

    Pith/arXiv arXiv

  25. [28]

    H 2O: Heavy-hitter oracle for efficient generative inference of large language models.arXiv preprint arXiv:2306.14048,

    Zhenyu Zhang, Ying Sheng, Tianyi Zhou, Tianlong Chen, Lianmin Zheng, Ruisi Cai, Zhao Song, Yuandong Tian, Christopher R´e, Clark Barrett, and Zhangyang Wang. H 2O: Heavy-hitter oracle for efficient generative inference of large language models.arXiv preprint arXiv:2306.14048,

    Pith/arXiv arXiv

  26. [29]

    what is recorded in registerT

    A PROTOCOL DETAILS Serving mechanics.Events are encoded once per trajectory under the checkpoint’s native causal at- tention pattern; serving re-assembles selected rows at their originalposition ids(position-budgeted across donors), so retained keys keep the rotation they were written with, and appends the query fresh. The rows are concatenated into conti...

    2026

  27. [30]

    ifNORTHorSOUTHanswerRED, elseBLUE

    Paired difference with McNemar exactp and discordants (injection-helps:injection-hurts), plus a conversation-clustered bootstrap95%CI (seed-pinned B=4000, matching Table 6); the McNemarpis item-level, and clustering changes neither conclusion. Qwen3’s forced choice is option-prior-saturated (a counterfactual carrier still selects gold∼80%), so its control...

    2024

  28. [31]

    To answer this question, we need to analyze the trajectory

    is in Appendix C; the tested Qwen3.6-27B is out of scope by substrate (§9). X9 transition classes.Table 7 gives the full counts on the three recent models. Unrelated-query flips0/192with accuracy1.00in every arm — the materialization channel does not echo into un- related readouts;anti≤4/192throughout. On Qwen3 and Gemma-4 the non-follow remainder is cons...

    2026

  29. [32]

    — not inert. Ministral-3’s passive-carrier drop under co-serving (pas.09→.00) isnot identified: the query-controlledr→bcontrast shows no detected change (1→0,p=1.0), so the cause (query offset vs. review membership) is unresolved. Areview-dominantshift, the reviewleading, appears only in the exploratory Gemma-2 run (App. C) and is not reproduced on any re...

    2024

  30. [34]

    directive>passive on both models

    — but none is a review-dominant landing, so the shift did not reproduce on any recent 17 Preprint — July 2026 Qwen3-8B Gemma-2-9B Qwen2.5-7B Llama-3.1-8B Mistral-7B W W W W P P U U W P P U U U U U W W U X W U X U W U W W W U W P P P P W P U U U P U U U P U U U U U U U U P U P U U U U U U U U U X X X P X X X X X X U X X X X A KV write-through (serve carrie...

    2026

  31. [35]

    the value at stept?

    tracks the uncompressed sequence length explicitly so the query stays on the original timeline. Serving geometry matters at this scale. Under native serving harvested and isolated land on the same rate (.562vs..562; LoCoMo−.011,95%CI[−.026,+.004]), whereas routing the same items through a compact cache — retained rows re-packed adjacent to the query — rep...

    2026

  32. [36]

    than co-served (+.083). Those two protocols differ in serve-set membershipandin the appended query’s absolute position: the query is placed atmax(served position)+1, the root sits at event slot13and the mention at slot18, so dropping the mention also moves the query five slots closer to the root. X11 separates the two. Design.Same 24-event log, Qwen3-8B,n...

    2026

  33. [2023]

    Language models use lookbacks to track beliefs.arXiv preprint arXiv:2505.14685,

    Nikhil Prakash, Natalie Shapira, Arnab Sen Sharma, Christoph Riedl, Yonatan Belinkov, Tamar Rott Shaham, David Bau, and Atticus Geiger. Language models use lookbacks to track beliefs.arXiv preprint arXiv:2505.14685,

    arXiv

  34. [2024]

    Gemma 2: Improving open language models at a practical size.arXiv preprint arXiv:2408.00118,

    Gemma Team. Gemma 2: Improving open language models at a practical size.arXiv preprint arXiv:2408.00118,

    Pith/arXiv arXiv

  35. [2025]

    MEMENTO: Teaching LLMs to manage their own context.arXiv preprint arXiv:2604.09852,

    Vasilis Kontonis, Yuchen Zeng, Shivam Garg, Lingjiao Chen, Hao Tang, Ziyan Wang, Ahmed Awadallah, Eric Horvitz, John Langford, and Dimitris Papailiopoulos. MEMENTO: Teaching LLMs to manage their own context.arXiv preprint arXiv:2604.09852,

    Pith/arXiv arXiv

  36. [2026]

    C 2KV: Compressed and composable KV cache reuse for efficient LLM inference.arXiv preprint arXiv:2607.17715,

    10 Preprint — July 2026 Chuheng Du, Junyi Chen, Hanlin Tang, Kan Liu, Tao Lan, Lin Qu, Chaoyue Niu, Shengzhong Liu, Guihai Chen, and Fan Wu. C 2KV: Compressed and composable KV cache reuse for efficient LLM inference.arXiv preprint arXiv:2607.17715,

    Pith/arXiv arXiv 2026