pith. sign in

arxiv: 2607.01308 · v1 · pith:6TZANIPQnew · submitted 2026-07-01 · 💻 cs.MA

Cache Merging as a Convergent Replicated State for Multi-Agent Latent Reasoning

Carlos Baquero , Lu\'is Brito This is my paper

Pith reviewed 2026-07-03 01:32 UTC · model grok-4.3

classification 💻 cs.MA
keywords KV cache mergingmulti-agent latent reasoningconvergent replicated stateCanonicalMergeCvRDTpermutation invarianceBagMerge
0
0 comments X

The pith

Ordering caches by mean K-norm at a middle layer produces byte-identical merges under any input permutation.

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

The paper shows how to turn non-commutative KV-cache concatenation into a convergent replicated state for multi-agent latent reasoning. CanonicalMerge orders caches by their mean K-norm value at a chosen middle layer, which makes the final merged cache identical no matter the order of the input agents. This ordering is content-based, so the merge satisfies the properties of a state-based CvRDT whose state is a set of content-addressed fragments and whose operation is set union. The deterministic render of that state preserves the accuracy of the best possible ordering on partitioned-reasoning and HotpotQA tasks while absorbing duplicate deliveries instead of re-concatenating them.

Core claim

CanonicalMerge fixes the layout by content: ordering caches by mean K-norm at a middle layer renders the merged cache byte-identical under any input permutation. The replicated state is a set of content-addressed latent fragments whose merge is set union, and CanonicalMerge is its deterministic render.

What carries the argument

CanonicalMerge, the deterministic ordering of caches by mean K-norm at a middle layer that produces permutation-invariant byte-identical merges.

If this is right

  • Every N=2 accuracy number carries over unchanged to the merged state.
  • Re-delivered duplicate caches are absorbed by set union rather than re-concatenated.
  • The method matches the accuracy of the best BagMerge ordering in every regime-by-budget cell without knowing which order is best.
  • The behaviour transfers to real multi-document QA while remaining distinct from output-level fusion methods.

Where Pith is reading between the lines

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

  • The separation of the abstract set state from its byte render could support distributed collection of latent fragments across independent agents.
  • At larger agent counts the ordering stability would need verification beyond the N<=5 algorithmic checks performed.
  • The current method colocates but does not compose latent traces, which leaves open the design of operators that would actually combine the fragments.

Load-bearing premise

The mean K-norm computed at a chosen middle layer supplies a stable, permutation-invariant total order that produces byte-identical merges for the tested models and regimes.

What would settle it

Two different permutations of the same set of caches produce byte-different merged outputs when both are ordered by mean K-norm at the chosen middle layer on Qwen3-1.7B or 4B.

Figures

Figures reproduced from arXiv: 2607.01308 by Carlos Baquero, Lu\'is Brito.

Figure 1
Figure 1. Figure 1: The two evaluation regimes. The KV-cache flow (solid arrows), thinkers [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Four ways to combine three latent traces at [PITH_FULL_IMAGE:figures/full_fig_p020_2.png] view at source ↗
read the original abstract

Multi-agent latent reasoning composes agents' KV-caches into one context for a final agent. Prior work (Agent Primitives) does this by concatenating caches along the sequence axis with RoPE re-encoding, which we call BagMerge. BagMerge is non-commutative, and the best input ordering is unpredictable, shifting with the regime, the latent-step budget, and the model scale. We make this exchange a convergent replicated state. First, CanonicalMerge fixes the layout by content: ordering caches by mean K-norm at a middle layer renders the merged cache byte-identical under any input permutation, verified algorithmically (arity N<=5) and bit-for-bit on real Qwen3-1.7B and 4B state. Second, we separate the replicated state from decode-time layout: the state is a set of content-addressed latent fragments whose merge is set union, a state-based CvRDT (commutative, associative, idempotent, absorbing), and CanonicalMerge is its deterministic render. Because the render is byte-equivalent, every N=2 accuracy number carries over unchanged and re-delivered duplicates are absorbed rather than re-concatenated. On a partitioned-reasoning benchmark, CanonicalMerge matches the best BagMerge ordering in every regime-by-budget-by-ordering cell without knowing which order is best, trading a small, statistically insignificant accuracy margin for an unconditional structural guarantee. The behaviour transfers to real multi-document QA (HotpotQA), while the closest training-free output-fusion baseline (PackLLM) loses by 45 points at matched budget, placing cache-level merging in a regime distinct from output-level fusion. Finally, at k>2 the approach transports and colocates latent traces but does not by itself compose them, which we characterize to motivate future work.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes CanonicalMerge for multi-agent latent reasoning: KV-caches are ordered by mean K-norm at a chosen middle layer to produce a permutation-invariant total order, yielding byte-identical merged caches under any input permutation. This is verified algorithmically for arity N≤5 and bit-for-bit on Qwen3-1.7B/4B models. The approach separates the replicated state (a set of content-addressed fragments whose merge is set union, forming a CvRDT) from its deterministic render (CanonicalMerge). On partitioned-reasoning and HotpotQA benchmarks, it matches the accuracy of the best BagMerge ordering in every regime-by-budget cell while absorbing duplicates, and outperforms output-fusion baselines like PackLLM.

Significance. If the central claims hold, the work supplies a structural guarantee of convergence and reproducibility for cache-level merging in multi-agent systems, converting a non-commutative operation into a CvRDT with explicit algorithmic and empirical verification of byte-identity. The fact that every N=2 accuracy number carries over unchanged and that the method colocates but does not yet compose traces at k>2 are useful characterizations. The distinction from output-level fusion is clearly drawn.

major comments (2)
  1. [Experimental validation] Experimental validation section: the claim of matching best BagMerge accuracy 'in every regime-by-budget-by-ordering cell' and the bit-for-bit equivalence on Qwen3 models rest on results whose support is described as moderate because error bars, full dataset details, and statistical tests are not provided in the reported text; these details are load-bearing for the empirical equivalence claim.
  2. [CanonicalMerge definition] Definition of CanonicalMerge (middle-layer ordering rule): mean K-norm at a single chosen middle layer is treated as supplying a stable, permutation-invariant total order, yet the manuscript identifies this layer index as a free parameter with no sensitivity analysis or justification for its selection across model scales or regimes.
minor comments (2)
  1. Notation for the set-union semantics and content-addressed fragments could be introduced earlier with an explicit small example to clarify the distinction between the CvRDT state and its render.
  2. The abstract states 'statistically insignificant accuracy margin' but the main text should cite the precise test and p-value used.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the positive assessment and recommendation of minor revision. We address each major comment below, committing to revisions where the manuscript requires strengthening.

read point-by-point responses

  1. Referee: [Experimental validation] Experimental validation section: the claim of matching best BagMerge accuracy 'in every regime-by-budget-by-ordering cell' and the bit-for-bit equivalence on Qwen3 models rest on results whose support is described as moderate because error bars, full dataset details, and statistical tests are not provided in the reported text; these details are load-bearing for the empirical equivalence claim.

    Authors: We agree that the reported equivalence claims would be more robust with additional statistical detail. In the revised manuscript we will add error bars (standard deviation over repeated runs where applicable), full dataset descriptions including splits and sizes, and results of statistical tests (paired t-tests for accuracy comparisons and exact-match verification for bit-identity) to support the claims of matching best BagMerge performance and byte-identical outputs. revision: yes

  2. Referee: [CanonicalMerge definition] Definition of CanonicalMerge (middle-layer ordering rule): mean K-norm at a single chosen middle layer is treated as supplying a stable, permutation-invariant total order, yet the manuscript identifies this layer index as a free parameter with no sensitivity analysis or justification for its selection across model scales or regimes.

    Authors: The middle-layer index is indeed presented as a fixed but arbitrary hyperparameter. We will revise the manuscript to include a short justification (middle layers encode higher-level semantics in the models studied) together with a sensitivity table showing that accuracy and bit-identity remain stable across a range of layer choices for both Qwen3-1.7B and 4B. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation self-contained

full rationale

The central construction defines CanonicalMerge via a content-derived total order (mean K-norm at a fixed middle layer) whose scalar is computed independently per cache and is therefore permutation-invariant by definition of sorting. Byte-identity of the merged cache is then a direct algorithmic consequence of that deterministic render plus set-union semantics, and the paper supplies explicit verification (algorithmic for N<=5, bit-for-bit on Qwen3 models) rather than deriving the property from any fitted accuracy metric or external theorem. No load-bearing step reduces to a self-citation, fitted input renamed as prediction, or ansatz smuggled from prior work; the accuracy equivalence to best-case BagMerge is presented as an observed consequence, not the justification for the structural claim. The derivation therefore stands on its own definitions and direct checks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that mean K-norm ordering at a middle layer yields a stable total order; no free parameters are fitted to accuracy data and no new entities are postulated.

free parameters (1)
  • middle_layer
    The specific layer chosen for mean K-norm computation is stated as 'middle' but its exact index is a modeling choice not derived from first principles.
axioms (1)
  • domain assumption Mean K-norm at a fixed middle layer produces a unique total order that is invariant under input permutation and sufficient for byte-identical merges
    Invoked directly in the definition of CanonicalMerge to fix layout by content.

pith-pipeline@v0.9.1-grok · 5862 in / 1496 out tokens · 42978 ms · 2026-07-03T01:32:45.225608+00:00 · methodology

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

40 extracted references

  1. [1]

    2024 , note =

    Su, Jianlin and Lu, Yu and Pan, Shengfeng and Murtadha, Ahmed and Wen, Bo and Liu, Yunfeng , journal =. 2024 , note =

    2024

  2. [2]

    arXiv preprint , year =

    Training Large Language Models to Reason in a Continuous Latent Space , author =. arXiv preprint , year =

  3. [3]

    Symposium on Self-Stabilizing Systems (SSS) , year =

    Conflict-Free Replicated Data Types , author =. Symposium on Self-Stabilizing Systems (SSS) , year =

  4. [4]

    Conflict-free Replicated Data Types

    Pregui. Conflict-free Replicated Data Types. 2018 , note =

    2018

  5. [5]

    arXiv preprint , year =

    Latent Collaboration in Multi-Agent Systems , author =. arXiv preprint , year =

  6. [6]

    arXiv preprint , year =

    Agent Primitives: Reusable Latent Building Blocks for Multi-Agent Systems , author =. arXiv preprint , year =

  7. [7]

    arXiv preprint , year =

    Enabling Agents to Communicate Entirely in Latent Space , author =. arXiv preprint , year =

  8. [8]

    International Conference on Learning Representations (ICLR) , year =

    Self-Consistency Improves Chain of Thought Reasoning in Language Models , author =. International Conference on Learning Representations (ICLR) , year =

  9. [9]

    arXiv preprint , year =

    Improving Factuality and Reasoning in Language Models through Multiagent Debate , author =. arXiv preprint , year =

  10. [10]

    NeurIPS , year =

    Tree of Thoughts: Deliberate Problem Solving with Large Language Models , author =. NeurIPS , year =

  11. [11]

    arXiv preprint , year =

    Mixture-of-Agents Enhances Large Language Model Capabilities , author =. arXiv preprint , year =

  12. [12]

    arXiv preprint , year =

    Universal Self-Consistency for Large Language Model Generation , author =. arXiv preprint , year =

  13. [13]

    International Conference on Learning Representations (ICLR) , year =

    Efficient Streaming Language Models with Attention Sinks , author =. International Conference on Learning Representations (ICLR) , year =

  14. [14]

    2023 , note =

    Zhang, Zhenyu and others , booktitle =. 2023 , note =

    2023

  15. [15]

    2024 , note =

    Li, Yuhong and others , journal =. 2024 , note =

    2024

  16. [16]

    2024 , note =

    Peng, Bowen and Quesnelle, Jeffrey and Fan, Honglu and Shippole, Enrico , booktitle =. 2024 , note =

    2024

  17. [17]

    Retrieval-Augmented Generation for Knowledge-Intensive

    Lewis, Patrick and others , booktitle =. Retrieval-Augmented Generation for Knowledge-Intensive. 2020 , note =

    2020

  18. [18]

    2023 , note =

    Douillard, Arthur and others , journal =. 2023 , note =

    2023

  19. [19]

    and Zhang, Hao and Stoica, Ion , booktitle =

    Kwon, Woosuk and Li, Zhuohan and Zhuang, Siyuan and Sheng, Ying and Zheng, Lianmin and Yu, Cody and Gonzalez, Joseph E. and Zhang, Hao and Stoica, Ion , booktitle =. Efficient Memory Management for Large Language Model Serving with. 2023 , note =

    2023

  20. [20]

    Not what you've signed up for: Compromising Real-World

    Greshake, Kai and Abdelnabi, Sahar and Mishra, Shailesh and Endres, Christoph and Holz, Thorsten and Fritz, Mario , journal =. Not what you've signed up for: Compromising Real-World. 2023 , note =

    2023

  21. [21]

    Tenney, Ian and Das, Dipanjan and Pavlick, Ellie , booktitle =

  22. [22]

    EMNLP , year =

    Transformer Feed-Forward Layers Are Key-Value Memories , author =. EMNLP , year =

  23. [23]

    arXiv preprint , year =

    Eliciting Latent Predictions from Transformers with the Tuned Lens , author =. arXiv preprint , year =

  24. [24]

    Representational Collapse in Multi-Agent

    Patel, Dipkumar , journal =. Representational Collapse in Multi-Agent. 2026 , note =

    2026

  25. [25]

    Mavromatis, Costas and Karypis, Petros and Karypis, George , journal =. Pack of. 2024 , note =

    2024

  26. [26]

    and Salakhutdinov, Ruslan and Manning, Christopher D

    Yang, Zhilin and Qi, Peng and Zhang, Saizheng and Bengio, Yoshua and Cohen, William W. and Salakhutdinov, Ruslan and Manning, Christopher D. , booktitle =. Hotpot

  27. [27]

    , journal =

    Chen, Zhijun and Lu, Xiaodong and Li, Jingzheng and Chen, Pengpeng and Li, Zhuoran and Sun, Kai and Luo, Yuankai and Mao, Qianren and Li, Ming and Xiao, Likang and Yang, Dingqi and Huang, Xiao and Ban, Yikun and Sun, Hailong and Yu, Philip S. , journal =. Harnessing Multiple Large Language Models: A Survey on. 2025 , note =

    2025

  28. [28]

    arXiv preprint , year =

    Merge, Ensemble, and Cooperate! A Survey on Collaborative Strategies in the Era of Large Language Models , author =. arXiv preprint , year =

  29. [29]

    arXiv preprint , year =

    Deep Model Fusion: A Survey , author =. arXiv preprint , year =

  30. [30]

    Conflict-Free Replicated Data Types for Neural Network Model Merging: A Two-Layer Architecture Enabling

    Ryan Gillespie , year =. Conflict-Free Replicated Data Types for Neural Network Model Merging: A Two-Layer Architecture Enabling

  31. [31]

    Towards Direct Latent-Space Synthesis for Parallel Branches in

    Shikun Liu and Mufei Li and Dongqi Fu and Haoyu Wang and Yinglong Xia and Hong Li and Hong Yan and Pan Li , year =. Towards Direct Latent-Space Synthesis for Parallel Branches in

  32. [32]

    Beyond Tokens: A Unified Framework for Latent Communication in

    Yingzhuo Liu , year =. Beyond Tokens: A Unified Framework for Latent Communication in

  33. [33]

    2025 , note =

    Do Latent Tokens Think? A Causal and Adversarial Analysis of Chain-of-Continuous-Thought , author =. 2025 , note =

    2025

  34. [34]

    Bottlenecked Transformers: Periodic

    Adnan Oomerjee and Zafeirios Fountas and Haitham Bou-Ammar and Jun Wang , year =. Bottlenecked Transformers: Periodic

  35. [35]

    2025 , note =

    Reasoning by Superposition: A Theoretical Perspective on Chain of Continuous Thought , author =. 2025 , note =

    2025

  36. [36]

    Model Tells You Where to Merge: Adaptive

    Zheng Wang and Boxiao Jin and Zhongzhi Yu and Minjia Zhang , year =. Model Tells You Where to Merge: Adaptive

  37. [37]

    Chengming Cui and Tianxin Wei and Ziyi Chen and Ruizhong Qiu and Zhichen Zeng and Zhining Liu and Xuying Ning and Duo Zhou and Jingrui He , year =

  38. [38]

    2024 , note =

    Determine-Then-Ensemble: Necessity of Top-k Union for Large Language Model Ensembling , author =. 2024 , note =

    2024

  39. [39]

    2026 , note =

    Recursive Multi-Agent Systems , author =. 2026 , note =

    2026

  40. [40]

    Xing and Hao Zhang and Joseph E

    Lianmin Zheng and Wei-Lin Chiang and Ying Sheng and Siyuan Zhuang and Zhanghao Wu and Yonghao Zhuang and Zi Lin and Zhuohan Li and Dacheng Li and Eric P. Xing and Hao Zhang and Joseph E. Gonzalez and Ion Stoica , booktitle =. Judging. 2023 , note =

    2023