Pith. sign in

REVIEW 1 major objections 6 minor 44 references

When Does Latent Communication Pay? A Causal Audit of Relayed KV Caches in Multi-Agent LLMs

T0 review · 1 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Relayed KV caches carry example-specific 'latent thoughts' only where the receiver needs the sender's private information; benchmark deltas alone cannot establish such transfer.

desk verdict A careful, reusable audit that settles what benchmark deltas cannot prove, but overreaches on the 'when' answer because receiver necessity is still confounded with task surface. read the letter →

arxiv 2608.04893 v1 pith:B2HSRLFL submitted 2026-08-05 cs.CR cs.AIcs.LG

Jiaming Cheng , Subhransu Das , Rajiv Ramnath This is my paper
classification cs.CRcs.AIcs.LG
keywords latentcommunicationKVcacherelaymulti-agentLLMsystemscausalauditpairingeffectderanged-cachecontrolequivalencetestingreceivernecessity
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Multi-agent LLM systems that relay key–value (KV) caches — the per-layer attention tensors a transformer accumulates while processing a sequence — between agents credit their benchmark gains to 'latent thoughts,' example-specific content the receiver uses. This paper argues that credit is a causal claim about which example's cache is relayed, and that benchmark deltas cannot settle it. Replacing the relayed cache with deranged (mismatched-example), zeroed, and moment-matched random counterparts, the audit finds the pairing effect — true cache minus deranged cache — at or near ceiling wherever the receiver genuinely needs the sender's private information: roughly 100% against 23–25% for answer-irrelevant relays, replicated across three model families, five checkpoints, and a prose surface. The same battery, run unchanged on standard benchmarks where the receiver can solve alone, bounds the pairing effect within ±2.8 points on all seven Qwen3 natural cells, with one small detected advantage (+1.52 points) inside the margin, and finds no detected advantage on a second family. The dissociation shows a large cache effect need not be a pairing effect — on MedQA, zeroing the relay costs 14.7 points while a mismatched cache moves nothing — and the paper concludes that establishing latent-thought transmission requires a mismatched-cache audit, which it releases.

What carries the argument

The carrying object is the pairing-effect estimand $\tau = \mathbb{E}[Y \mid do(C:=C_i)] - \mathbb{E}[Y \mid do(C:=C_{\pi(i)})]$ — the accuracy difference between receiving the matched example's cache and receiving a different example's cache, where the deranged cache $C_{\pi(i)}$ is delivered by a fixed-point-free within-batch permutation that preserves the cache multiset exactly, isolating pairing from malformedness. Three supporting instruments do the work: zeroed and moment-matched random caches, whose accuracy collapse in every cell proves the intervention reaches the receiver's computation; a procedurally generated sender-private calibration instrument (entity–value registries sampled fresh at evaluation time) that makes receiver necessity a designed property rather than an assumption; and the algebraic identity $\tau_{TZ} = \tau_{TD} + \tau_{DZ}$, which splits the total cache effect into a content (pairing) term and a generic (computation) term. Natural-regime equivalence is established with two one-sided tests (TOST) at a ±2.8 percentage-point margin derived from the audited system's own reported aggregate sequential gain, Holm-corrected within pre-registered hypothesis families and run at the seed level.

What would settle it

Run the audit's four-arm battery on a natural surface where the receiver's task genuinely depends on sender-held context (for example, document QA where only the sender sees the document): a seed-stable true-minus-deranged contrast beyond ±2.8 points would break the natural-regime equivalence claim. Separately, estimate the necessity interaction on items hard enough that the informed row falls below ceiling; if cache pairing moves accuracy even when the receiver holds the content as visible text, the 'receiver need' account collapses.

Watch

Extended reading notes

Core claim

The paper's claim is that the value of latent communication lives in the pairing of cache to query, not in the mere presence of a relayed cache, and that this pairing value is governed by receiver necessity. On the audited native relay, the true-minus-deranged contrast reads above +60 points in every calibrated cell — across Qwen3-4B/8B/14B, Mistral-Nemo-12B, and phi-4, and again on a procedurally generated prose document-QA surface — while the same intervention yields seed-level means between −1.12 and +1.52 points on GSM8K, ARC-Challenge, and MedQA, with Holm-corrected equivalence within the ±2.8-point margin (anchored to the audited system's own reported gain) established on all seven Qwen3 cells. MedQA makes the dissociation concrete: the total cache contrast is +14.7 points and the pairing contrast +0.4 points, so the relay's measurable value there is interface-level, not content-level. Porting the audit to two further released systems shows receiver need is necessary but not sufficient: LatentMAS's full-cache prepend transfers at ceiling, KVComm's layer-selected relay transfers partially, and C2C's released projector shows no detected example-specific transfer under the same information-asymmetry construction. The conclusion is stated as an evidence standard: benchmark deltas do not by themselves establish latent-thought transmission; establishing it takes a mismatched-cache audit, which the paper releases.

Load-bearing premise

The cross-regime dissociation assumes the flip between calibrated and natural readings is driven by receiver necessity rather than by task surface — the calibrated regime uses a synthetic registry QA task while the natural regime uses GSM8K, ARC-Challenge, and MedQA — and the paper's own within-task factorial supports moderation without establishing it, because the informed row sits at ceiling.

Editorial extensions

If this is right

  • Benchmark deltas on standard tasks cannot by themselves be credited to latent-thought transmission; a mismatched-cache control is the evidentiary bar for any relayed-state channel.
  • Where the receiver genuinely lacks the sender's private information, the relayed cache carries the answer: the audited native relay reads roughly 100% against 23–25% for answer-irrelevant relays, replicated across three families, five checkpoints, and a prose surface.
  • In the natural regime, the example-pairing contribution of the relay is bounded below the audited system's own reported gain (±2.8 points) on all seven Qwen3 cells, with one cell showing a small detected advantage (+1.52 points) inside the margin.
  • A large total cache effect need not imply content transfer: on MedQA, zeroing the relay costs 14.7 points while a mismatched cache costs 0.4, so interface-level effects can dominate the relay's measurable value.
  • Receiver need is necessary but not sufficient: under the same information-asymmetry test, three delivered channels read ceiling (LatentMAS), partial (KVComm), and no detected example-specific transfer (C2C).

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the audit intervenes only on the relayed object and leaves the rest of the system fixed, the same derangement logic transfers to other latent interfaces — activation relays, compressed or quantized cache handoffs, and end-to-end trained communication protocols — and would be the natural next check of whether those channels carry content rather than interface effects.
  • A testable prediction follows for the field: latent-communication systems that report gains on tasks without real information asymmetry should generally fail a mismatched-cache audit, since the pairing term appears to sit near zero wherever the receiver can solve alone.
  • The moderator hypothesis — that receiver necessity, not task surface, drives the flip — can be settled prospectively with matched surfaces that cross necessity and difficulty within one distribution so the informed row falls below ceiling; the paper's factorial could not do this because its informed row sat at ceiling.
  • The MedQA decomposition locates most of the relay's natural-regime value in generic, format-level effects, which suggests that some published latent-communication gains may be interface-compliance effects rather than content transfer; re-running the audit on compressed or quantized relays, where generic and content contributions may trade off differently, would test that reading.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

1 major / 6 minor

Summary. This paper audits released multi-agent LLM systems that relay KV caches in place of text, asking whether benchmark gains can be credited to example-specific latent content. The audit intervenes on the relayed cache with deranged, zeroed, and moment-matched random counterparts, under a calibrated regime in which the receiver needs sender-private registry content and a natural regime on GSM8K/ARC/MedQA in which the receiver can solve alone. The paper reports near-ceiling true-minus-deranged contrasts in the calibrated regime across five checkpoints, equivalence within ±2.8 points under Holm-corrected TOST in seven Qwen3 natural cells, no established bound in a Mistral-Nemo second family, and a three-level pattern across the delivered channels of LatentMAS, KVComm, and C2C. Its central methodological claim is that benchmark deltas do not by themselves establish latent-thought transmission; a mismatched-cache control is required.

Significance. If the measurements hold, the paper makes a valuable methodological contribution: it supplies a reusable causal audit for relayed KV caches, with a marginal-preserving derangement, fixed cache geometry across arms, seed-level inference, a pre-registered equivalence margin, raw per-example ledgers, and a disclosed run-deviation record. These design choices are careful and the within-regime results are credible, especially the calibrated ceiling and the natural-regime null with liveness controls. However, the answer to the title question—when latent communication pays—is not established, because receiver necessity is confounded with task surface and the within-task factorial degenerates at ceiling; the manuscript's own appendix statements acknowledge this. The core evidence standard, that benchmark deltas require a mismatched-cache audit, is independent of that gap and is well supported.

major comments (1)
  1. [§4 and Appendix C.1] The answer to the title question—that latent communication pays 'when the receiver needs the sender's private information'—is not established by the reported dissociation. The calibrated and natural regimes differ in task surface (synthetic registry/document QA versus GSM8K/ARC/MedQA) as well as in receiver necessity, and the within-task factorial in Appendix C.1 does not remove this confound: the informed row injects the true registry as visible text into the receiver's prompt, changing the prompt surface, and both informed-row cells sit at 500/500 on all seeds, so the interaction estimate degenerates to the uninformed-row contrast. The manuscript itself states that 'across regimes necessity is observational and remains a moderator hypothesis' and that the factorial 'supports moderation without establishing it.' Because the abstract and conclusion present receiver necessity as the 'when' answer, this is a load-bearing gap. Please either reword the claim (for example, 'pairing value is detected only under constructed information asymmetry in the audited configurations') or add a manipulation that varies necessity while holding the task surface fixed and avoids ceiling (for example, graded information asymmetry or per-example receiver-capability measurement).
minor comments (6)
  1. [Abstract and Conclusion] The paper switches between 'latent thoughts' and 'example pairing value of the relayed cache'; since the audit measures the latter on the accuracy scale, the former should be explicitly defined as the attribution under test and kept in quotes throughout.
  2. [§4.1] The sentence 'The Qwen3 scale span rules out a size artifact; the two further families, a family artifact' is too strong; it rules out artifacts specific to these two families, not family artifacts in general.
  3. [Appendix G.2] The term 'pre-registered' is defined as a dated design document rather than a third-party registry; this caveat should appear at the first use of the term in the main text.
  4. [§4.2 and Appendix D.1] The premise that 'the receiver can solve the task without the sender' is supported only by the receiver-only baseline in Appendix D.1; adding a pointer to that baseline in Section 4.2 would make the argument easier to follow.
  5. [Appendix F.4] The KVComm cell uses a Qwen3-8B port with a replacement calibration split and a new attention tracer; the main text should state more prominently that this is a port rather than the released configuration, to avoid overstating the 'three delivered channels' framing.
  6. [Reproducibility] The release is described in detail, but no repository URL or persistent identifier is provided; adding one would help readers access the harness and ledgers.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the causal audit is an independent intervention; the receiver-need moderator limitation is a validity caveat, not a circular step.

full rationale

The paper's derivation chain is an experimental audit, not a derivation: the pairing estimand τ = E[Y|do(C:=C_i)] − E[Y|do(C:=C_{π(i)})] is defined by an intervention that permutes which example's cache is delivered, holding the receiver, prompts, and cache geometry fixed, with a fixed-point-free permutation preserving the marginal cache multiset. The calibrated/natural dissociation is an empirical comparison across constructed regimes, not an identity. The natural-regime equivalence margin (±2.8 points) is anchored to the audited system's reported aggregate gain, but this is a pre-fixed decision threshold, not a parameter fitted to the audit's own outcomes; the measured true-minus-deranged contrast is computed from raw per-seed ledgers and could have fallen outside the margin (as two Mistral-Nemo cells show, where equivalence was not established). No load-bearing argument reduces to a self-citation: the audited systems are cited as external artifacts (Zou et al., Fu et al., Shi et al.), and the paper's authors have no prior-work self-citations in the chain. The acknowledged limitation in Appendix C.1—that the informed-row ceiling makes the within-task factorial degenerate and that 'across regimes necessity is observational and remains a moderator hypothesis'—is an external-validity/confounding concern about the causal interpretation of the regime difference, not a circular reduction: the measured pairing effects themselves are independently obtained and would stand even if the moderator claim were withdrawn. The one small caveat is that the ±2.8 margin equals the audited system's claimed gain, so 'equivalence within margin' does not by itself rule out a pairing effect as large as the claimed gain; the paper's stronger 'large cache effect need not be a pairing effect' point rests on the direct MedQA contrast (+14.7 total vs +0.4 pairing), which is not circular.

Assumptions & free parameters 1 free parameters · 6 assumptions · 0 invented entities

The central claim rests on the audit's design assumptions (marginal-preserving derangement, fresh private bindings, identical cache geometry), the statistical margin, and the seed-level inference unit. No new physical or theoretical entities are introduced; the 'latent thoughts' attribution is the external claim under audit, not an entity the paper postulates.

free parameters (1)
  • Equivalence margin delta = ±2.8 percentage points
    Chosen in advance as the TOST margin, anchored to LatentMAS's reported aggregate 2.8% sequential gain. It is an analytic threshold, not fitted to outcome data, but the natural-regime 'bounded' verdicts are defined relative to it.
assumptions (6)
  • domain assumption The deranged arm's within-batch fixed-point-free permutation preserves the marginal cache multiset and isolates example pairing from distribution shift.
    Section 3, Identification. The entire pairing estimand tau_TD relies on this.
  • domain assumption Fresh-sampled registry bindings are absent from training corpora, so the receiver cannot recover sender-private information except through the relay.
    Section 3, The calibration instrument. Needed for the calibrated-regime ceiling to be a valid reading of content transfer.
  • domain assumption All arms deliver identical cache geometry, so accuracy differences are attributable to cache content.
    Section 3, Audited mechanism. The intervention holds layer count, sequence length, dtype, and masks fixed.
  • domain assumption The ±2.8 point margin, anchored to the audited system's own claimed aggregate gain, is the correct smallest effect of interest for the natural-regime bound.
    Section 3, Statistical conventions. The equivalence verdicts are defined relative to this margin.
  • standard math The seed is the valid inference unit because examples within a seed share one stochastic receiver realization.
    Appendix A, Statistical unit. Supports the use of five seed-level contrasts for TOST.
  • domain assumption The random-arm collapse establishes causal coupling at the relayed-cache intervention point in natural cells.
    Section 4.2. Needed to rule out a dead instrument in the natural regime.

how reviews work

0 comments
Cite this review

Pith. "Pith review of When Does Latent Communication Pay? A Causal Audit of Relayed KV Caches in Multi-Agent LLMs." pith.science (2026). https://pith.science/paper/B2HSRLFL

@misc{pith2026260804893,
  author       = {Pith},
  title        = {Pith review of: When Does Latent Communication Pay? A Causal Audit of Relayed KV Caches in Multi-Agent LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B2HSRLFL}},
  note         = {Machine review of arXiv:2608.04893}
}
read the original abstract

Multi-agent LLM systems relay key--value caches instead of text and credit their gains to exchanged ``latent thoughts''. That credit is a claim about \emph{which} example's cache is relayed, not merely that one is. We audit it causally in released systems. The cache is replaced with deranged (mismatched-example), zeroed, and moment-matched random counterparts, under two regimes defined by whether the receiver needs the sender's private information. Where it does, the battery reads ceiling: 100\% against 23--25\% for answer-irrelevant relays on the primary backbone, a contrast replicated across three families, five checkpoints, and a prose document-QA surface. Where it does not, a pre-registered five-seed protocol establishes equivalence within 2.8 points, a margin anchored to the audited system's reported gain, under Holm-corrected TOST on GSM8K and ARC-Challenge across three Qwen3 scales and on MedQA at 8B (one cell shows a small detected advantage inside the margin); a second family shows no detected advantage. A large cache effect need not be a pairing effect. In one natural cell, zeroing the relay costs 14.7 points; a mismatched cache, 0.4. Nor is need sufficient: under the same test, delivered channels span ceiling (LatentMAS's native relay), partial (KVComm's layer subset), and no detected example-specific transfer (C2C's released projector). Benchmark deltas do not by themselves establish latent-thought transmission; establishing it takes a mismatched-cache audit, which we release.

Figures

Figures reproduced from arXiv: 2608.04893 by the authors.

Figure 1
Figure 1. Causal audit design. The primary contrast holds the receiver and the batch of caches fixed, permuting only which cache reaches each query. Both arms relay an equally well-formed cache, so the true-minus-deranged difference isolates example pairing from any generic benefit of relaying one. et al., 2026a). Methodologically closest to us is activation patching / causal tracing, which swaps internal states between runs … view at source ↗
Figure 2
Figure 2. One estimand, two regimes. True-minus-deranged seed means (diamonds; per-seed contrasts as circles) span +61 to +94 points where the receiver needs sender-private content and cluster near zero where it solves alone. The zoomed panel magnifies the natural block, showing unadjusted 90% t CIs against the ±2.8-point margin operationalized from the audited system’s reported gain; hollow diamonds mark cells where Holm-cor… view at source ↗
Figure 3
Figure 3. One asymmetry test, three delivered channels. Per-seed true-minus-deranged (diamonds; the pairing estimand) and true-minus-zero (circles; total cache contrast) under the same receiver-necessary construction, with direct-text validity arms at ceiling throughout. C2C’s sign-mixed pairing contrasts beside its large total cache contrast read as no detected example-specific transfer. pre-registered rule scores it rather … view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Margin sensitivity ladder. All nine unadjusted intervals lie inside the operationalized ±2.8 band, two already inside ±1; the seven Qwen3 cells establish Holm-corrected equivalence there, and the two Mistral-Nemo-12B cells do not. Unadjusted seed-level 90% t CIs (thick…
Figure 5
Figure 5. Figure 5: Receiver-only placement. Receiver-only (no-relay) accuracy tracks the zero arm on the four Qwen3 GSM8K/ARC-Challenge cells, and at 4B the delivered arms sit below the receiver-only level; the MedQA row splits its total cache contrast into interface damage (zero to no-r…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 24 canonical work pages

  1. [1]

    Hewett, Mojan Javaheripi, Piero Kauffmann, James R

    Marah Abdin, Jyoti Aneja, Harkirat Behl, Sébastien Bubeck, Ronen Eldan, Suriya Gunasekar, Michael Harrison, Russell J. Hewett, Mojan Javaheripi, Piero Kauffmann, James R. Lee, Yin Tat Lee, Yuanzhi Li, Weishung Liu, Caio C. T. Mendes, Anh Nguyen, Eric Price, Gustavo de Rosa, Olli Saarikivi, Adil Salim, Shital Shah, Xin Wang, Rachel Ward, Yue Wu, Dingli Yu,...

    arXiv

  2. [4]

    When Latent Agents Lie: KV-Cache Integrity in Multi-Agent LLM Collaboration

    Luís Brito and Carlos Baquero. When latent agents lie: KV-cache integrity in multi-agent LLM collaboration.ArXiv preprint, abs/2606.28958,

    work page Pith review arXiv

  3. [5]

    Make each token count: Towards improving long-context performance with KV cache eviction.ArXiv preprint, abs/2605.09649,

    Ngoc Bui, Hieu Trung Nguyen, Arman Cohan, and Rex Ying. Make each token count: Towards improving long-context performance with KV cache eviction.ArXiv preprint, abs/2605.09649,

    arXiv

  4. [6]

    Transferring linear features across language models with model stitching

    Alan Chen, Jack Merullo, Alessandro Stolfo, and Ellie Pavlick. Transferring linear features across language models with model stitching. InAdvances in Neural Information Processing Systems 38, NeurIPS 2025, pages 48531–48563,

    work page 2025

  5. [8]

    See What I See, Know What I Think: Dense Latent Communication Across Heterogeneous Agents

    Siyi Chen, Xiaoyan Zhang, Meng Wu, Jonathan Tremblay, Valts Blukis, Stan Birchfield, Rene Vidal, Alvaro Velasquez, Sijia Liu, and Qing Qu. See what I see, know what I think: Dense latent communication across heterogeneous agents. ArXiv preprint, abs/2606.13594, 2026b. Hyeong Kyu Choi, Jiatong Li, Wendi Li, Xin Eric Wang, and Sharon Li. Multi-agent LLMs fa...

    work page Pith review arXiv

  6. [9]

    Think you have solved question answering? try ARC, the AI2 reasoning challenge.ArXiv preprint, abs/1803.05457,

    Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try ARC, the AI2 reasoning challenge.ArXiv preprint, abs/1803.05457,

    arXiv

  7. [12]

    Enabling agents to communicate entirely in latent space

    Zhuoyun Du, Runze Wang, Huiyu Bai, Zouying Cao, Xiaoyong Zhu, Yu Cheng, Bo Zheng, Wei Chen, and Haochao Ying. Enabling agents to communicate entirely in latent space. InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2026, San Diego, California, United States, July 2-7, 2026, pages 27106–27129,

    work page 2026

  8. [13]

    doi: 10.18653/v1/2026.acl-long.1248. Jakob N. Foerster, Yannis M. Assael, Nando de Freitas, and Shimon Whiteson. Learning to communicate with deep multi-agent reinforcement learning. InAdvances in Neural Information Processing Systems 29: Annual Conference on Neural Information Processing Systems 2016, December 5-10, 2016, Barcelona, Spain, pages 2137–2145,

    doi:10.18653/v1/2026.acl-long.1248 2026

Show all 44 references
  1. [15]

    Training large language models to reason in a continuous latent space.ArXiv preprint, abs/2412.06769,

    Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, Xian Li, Zhiting Hu, Jason Weston, and Yuandong Tian. Training large language models to reason in a continuous latent space.ArXiv preprint, abs/2412.06769,

    arXiv

  2. [20]

    When less latent leads to better relay: Information-preserving compression for latent multi-agent LLM collaboration.ArXiv preprint, abs/2604.13349,

    Yiping Li, Zhiyu An, and Wan Du. When less latent leads to better relay: Information-preserving compression for latent multi-agent LLM collaboration.ArXiv preprint, abs/2604.13349,

    arXiv

  3. [21]

    The vision wormhole: Latent-space communication in heterogeneous multi-agent systems

    Xiaoze Liu, Ruowang Zhang, Weichen Yu, Siheng Xiong, Liu He, Feijie Wu, Hoin Jung, Matt Fredrikson, Xiaoqian Wang, and Jing Gao. The vision wormhole: Latent-space communication in heterogeneous multi-agent systems. ArXiv preprint, abs/2602.15382,

    arXiv

  4. [22]

    Beyond tokens: a unified framework for latent communication in LLM-based multi-agent systems

    Yingzhuo Liu. Beyond tokens: a unified framework for latent communication in LLM-based multi-agent systems. ArXiv preprint, abs/2606.05711,

    arXiv

  5. [23]

    Locating and editing factual associations in GPT

    Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Locating and editing factual associations in GPT. InAdvances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 2...

    2022

  6. [25]

    Igor Mordatch and Pieter Abbeel

    Model release. Igor Mordatch and Pieter Abbeel. Emergence of grounded compositional language in multi-agent populations. InProceedings of the Thirty-Second AAAI Conference on Artificial Intelligence, (AAAI-18), the 30th innovative Applications of Artificial Intelligence (IAAI-...

    2018

  7. [26]

    HyLaT: Efficient multi-agent communication via hybrid latent-text protocol.ArXiv preprint, abs/2605.25421,

    Xinyi Mou, Siyuan Wang, Zejun Li, Yulan He, and Zhongyu Wei. HyLaT: Efficient multi-agent communication via hybrid latent-text protocol.ArXiv preprint, abs/2605.25421,

    arXiv

  8. [28]

    Gonzalez, M Waleed Kadous, and Ion Stoica

    Isaac Ong, Amjad Almahairi, Vincent Wu, Wei-Lin Chiang, Tianhao Wu, Joseph E. Gonzalez, M Waleed Kadous, and Ion Stoica. RouteLLM: Learning to route LLMs with preference data.ArXiv preprint, abs/2406.18665,

    arXiv

  9. [29]

    Activation space interventions can be transferred between large language models

    11 When Does Latent Communication Pay?A PREPRINT Narmeen Fatimah Oozeer, Dhruv Nathawani, Nirmalendu Prakash, Michael Lan, Abir Harrasse, and Amir Abdullah. Activation space interventions can be transferred between large language models. InForty-second International Conference...

    2025

  10. [30]

    Representational collapse in multi-agent LLM committees: Measurement and diversity-aware consensus.ArXiv preprint, abs/2604.03809,

    Dipkumar Patel. Representational collapse in multi-agent LLM committees: Measurement and diversity-aware consensus.ArXiv preprint, abs/2604.03809,

    arXiv

  11. [31]

    Communicating activations between language model agents

    Vignav Ramesh and Kenneth Li. Communicating activations between language model agents. InForty-second International Conference on Machine Learning, ICML 2025, Vancouver, BC, Canada, July 13-19, 2025,

    2025

  12. [32]

    Xiangyu Shi, Marco Chiesa, Gerald Q

    doi: 10.1007/BF01068419. Xiangyu Shi, Marco Chiesa, Gerald Q. Maguire Jr., and Dejan Kosti´c. KVComm: Enabling efficient LLM communica- tion through selective KV sharing. InThe Fourteenth International Conference on Learning Representations, ICLR 2026,

    doi

  13. [33]

    Learning multiagent communication with backpropagation

    Sainbayar Sukhbaatar, Arthur Szlam, and Rob Fergus. Learning multiagent communication with backpropagation. InAdvances in Neural Information Processing Systems 29: Annual Conference on Neural Information Processing Systems 2016, December 5-10, 2016, Barcelona, Spain, pages 2244–2252,

    2016

  14. [34]

    Out of sight, not out of mind: Unveiling latent attack in latent-based multi-agent systems.ArXiv preprint, abs/2605.28214, 2026a

    Chenxi Wang, Ruiyang Huang, Jiayan Sun, Lei Wei, and Yifan Wu. Out of sight, not out of mind: Unveiling latent attack in latent-based multi-agent systems.ArXiv preprint, abs/2605.28214, 2026a. Junlin Wang, Jue Wang, Ben Athiwaratkun, Ce Zhang, and James Zou. Mixture-of-agents ...

    2025 arXiv

  15. [35]

    Interpretability in the wild: a circuit for indirect object identification in GPT-2 small

    Kevin Ro Wang, Alexandre Variengien, Arthur Conmy, Buck Shlegeris, and Jacob Steinhardt. Interpretability in the wild: a circuit for indirect object identification in GPT-2 small. InThe Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, M...

    2023

  16. [36]

    Learning to communicate: Toward end-to-end optimization of multi-agent language systems.ArXiv preprint, abs/2604.21794,

    Ye Yu, Heming Liu, Haibo Jin, Xiaopeng Yuan, Peng Kuang, and Haohan Wang. Learning to communicate: Toward end-to-end optimization of multi-agent language systems.ArXiv preprint, abs/2604.21794,

    arXiv

  17. [37]

    Towards best practices of activation patching in language models: Metrics and methods

    Fred Zhang and Neel Nanda. Towards best practices of activation patching in language models: Metrics and methods. InThe Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024,

    2024

  18. [38]

    Do latent channels actually communicate? a causal audit of latent multi-agent LLM communication.ArXiv preprint, abs/2607.26773,

    Huixiang Zhang and Mahzabeen Emu. Do latent channels actually communicate? a causal audit of latent multi-agent LLM communication.ArXiv preprint, abs/2607.26773,

    arXiv

  19. [39]

    A negative result on cross-model activation transfer in a Pythia multi-hop setting.ArXiv preprint, abs/2606.03280,

    Peiyan Zhang and Jason Xin. A negative result on cross-model activation transfer in a Pythia multi-hop setting.ArXiv preprint, abs/2606.03280,

    arXiv

  20. [40]

    Thought communication in multiagent collaboration

    Yujia Zheng, Zhuokai Zhao, Zijian Li, Yaqi Xie, Mingze Gao, Lizhu Zhang, and Kun Zhang. Thought communication in multiagent collaboration. InAdvances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2025, NeurIPS 2025, San...

    2025

  21. [41]

    Latent collaboration in multi-agent systems

    Jiaru Zou, Ruizhong Qiu, Gaotang Li, Xiyuan Yang, Katherine Tieu, Pan Lu, Ke Shen, Hanghang Tong, Yejin Choi, Jingrui He, James Zou, Mengdi Wang, and Ling Yang. Latent collaboration in multi-agent systems. InProceedings of the 43rd International Conference on Machine Learning,...

    2026

  22. [42]

    sign-mixed

    These arms carry the quoted true-minus-zero and true-minus-random contrasts; the random arm’s collapse in every cell shows the intervention reaches the receiver’s computation. pre-intervention caches matching the production true-private cell batch by batch. The relay-cache eff...

    2026

  23. [44]

    See Liu (2026); Chen et al

    produces learned channels rather than delivered fixed mechanisms, and such systems are the natural next audit target — a protocol trained for communication is where example-specific pairing value would most plausibly appear, and our instrument applies unchanged. See Liu (2026)...

    2026

  24. [47]

    recorded in advance

    per the replacement rule recorded in advance, before any unblinding; the replaced seed’s partial outputs are pre- served unread MedQA cell hit the same failure class in its smoke None: failure preceded production Cell re-venued to A100-SXM4-80GB before production phi-4 calibra...

    2025

  25. [1979]

    QKVShare: Quantized KV-cache handoff for multi-agent on-device LLMs.ArXiv preprint, abs/2605.03884,

    Pratik Honavar and Tejpratap GVSL. QKVShare: Quantized KV-cache handoff for multi-agent on-device LLMs.ArXiv preprint, abs/2605.03884,

    arXiv

  26. [1987]

    Ascend Cluster

    Ohio Supercomputer Center. Ascend Cluster. https://doi.org/10.82404/6JBT-FA57,

    doi

  27. [2016]

    LatentMem: Customizing latent memory for multi-agent systems.ArXiv preprint, abs/2602.03036, 2026a

    10 When Does Latent Communication Pay?A PREPRINT Muxin Fu, Xiangyuan Xue, Yafu Li, Zefeng He, Siyuan Huang, Xiaoye Qu, Yu Cheng, and Yang Yang. LatentMem: Customizing latent memory for multi-agent systems.ArXiv preprint, abs/2602.03036, 2026a. Tianyu Fu, Zihan Min, Hanling Zha...

    2026

  28. [2017]

    Angeliki Lazaridou, Alexander Peysakhovich, and Marco Baroni

    doi: 10.1177/1948550617697177. Angeliki Lazaridou, Alexander Peysakhovich, and Marco Baroni. Multi-agent cooperation and the emergence of (natural) language. In5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Tr...

    doi

  29. [2018]

    Training verifiers to solve math word problems.ArXiv preprint, abs/2110.14168,

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. Training verifiers to solve math word problems.ArXiv preprint, abs/2110.14168,

    arXiv

  30. [2020]

    Agent primitives: Reusable latent building blocks for multi-agent systems.ArXiv preprint, abs/2602.03695,

    Haibo Jin, Peng Kuang, Ye Yu, Xiaopeng Yuan, and Haohan Wang. Agent primitives: Reusable latent building blocks for multi-agent systems.ArXiv preprint, abs/2602.03695,

    arXiv

  31. [2021]

    Tenenbaum, and Igor Mordatch

    Yilun Du, Shuang Li, Antonio Torralba, Joshua B. Tenenbaum, and Igor Mordatch. Improving factuality and reasoning in language models through multiagent debate. InForty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024,

    2024

  32. [2022]

    Llama 3.2: Revolutionizing edge AI and vision with open, customizable models

    Meta AI. Llama 3.2: Revolutionizing edge AI and vision with open, customizable models. https://ai.meta.com/ blog/llama-3-2-connect-2024-vision-edge-mobile-devices/,

    2024

  33. [2023]

    Di Jin, Eileen Pan, Nassim Oufattole, Wei-Hung Weng, Hanyi Fang, and Peter Szolovits

    doi: 10.18653/v1/2023.acl-long.792. Di Jin, Eileen Pan, Nassim Oufattole, Wei-Hung Weng, Hanyi Fang, and Peter Szolovits. What disease does this patient have? a large-scale open domain question answering dataset from medical exams.ArXiv preprint, abs/2009.13081,

    2023 arXiv

  34. [2024]

    ReST-KV: Robust KV cache eviction with layer-wise output reconstruction and spatial-temporal smoothing.ArXiv preprint, abs/2605.08840,

    Yongqi An, Chang Lu, Kuan Zhu, Tao Yu, Chaoyang Zhao, Hong Wu, Ming Tang, and Jinqiao Wang. ReST-KV: Robust KV cache eviction with layer-wise output reconstruction and spatial-temporal smoothing.ArXiv preprint, abs/2605.08840,

    arXiv

  35. [2025]

    Jingdi Chen, Hanqing Yang, Zongjun Liu, and Carlee Joe-Wong. The five Ws of multi-agent communication: Who talks to whom, when, what, and why – a survey from MARL to emergent language and LLMs.Transactions on Machine Learning Research, 2026, 2026a. Lingjiao Chen, Matei Zaharia...

    2026

  36. [2026]

    LCGuard: Latent communication guard for safe KV sharing in multi-agent systems.ArXiv preprint, abs/2605.22786,

    Sadia Asif, Mohammad Mohammadi Amiri, Momin Abbas, Prasanna Sattigeri, and Karthikeyan Natesan Ramamurthy. LCGuard: Latent communication guard for safe KV sharing in multi-agent systems.ArXiv preprint, abs/2605.22786,

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.