Pith. sign in

REVIEW 3 major objections 4 minor 16 references

Latent channels between LLM agents match text on cross-lingual concepts but never exceed it; text serialization destroys 88% of sparse features — surface form, not meaning.

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 ↗

Latent communication between LLM agents preserves far more SAE features than text, but those extra features encode surface form and provide no task-level advantage on text-expressible tasks.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection A solid framework for measuring latent vs. text communication, but the headline claim about surface-form loss rests on a mismatch between the feature-survival analysis and the actual text channel. the 3 major comments →

arxiv 2607.14103 v1 pith:RPU24YQM submitted 2026-05-06 cs.CL cs.MA

Latent Communication Between Language Model Agents: Channels, Alignment, and the Limits of Text

Markus Wenzel This is my paper
classification cs.CL cs.MA
keywords latent communicationmulti-agent systemssparse autoencoderstext serializationrepresentational alignmentProcrustes alignmentfeature survival analysisLLM agents
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.

The reading

The paper asks whether LLM agents can communicate more effectively by passing internal latent vectors directly instead of text, and it returns a deliberately negative answer for current tasks. It shows that text serialization destroys 88% of the Sparse Autoencoder features active in contextual encoding, replacing them with a different feature set rather than attenuating them. The decisive follow-up is that these lost features are surface form — tokenization artifacts, prompt structure, positional effects — not concept-discriminating semantics: re-injecting them degrades performance. On concept identification, attribute/sense disambiguation, and cross-lingual tasks, a latent channel matches text at best and trails by 3–10 percentage points otherwise; only bandwidth efficiency (28x compression at 99.4% probe fidelity) favors the sparse latent channel. The paper also provides clear evidence for geometric convergence between separately trained models, with a simple rotation aligning two 8B-parameter models' activation spaces to 92% top-1 concept retrieval.

Core claim

The central discovery is that text communication between LLM agents loses information at the feature level, but the lost information is not task-relevant. Using a text round-trip (a concept elicited by a contextual prompt versus the same concept encoded from its bare name), the paper finds that 88.3% of SAE features do not survive, with Jaccard overlap of only 5.8% and SAE-space cosine of 0.187. The loss is identity replacement: lost and surviving features have statistically indistinguishable activation magnitudes, so text moves the representation to a different neighborhood of feature space rather than weakening the original features. Augmentation experiments then show the lost features, wh

What carries the argument

The load-bearing object is the Sparse Autoencoder (SAE) feature dictionary — a 131,072-feature overcomplete decomposition of the 4,096-dimensional residual-stream hidden state, with about 70 active features per forward pass — used as a measurement instrument to compare two encoding paths for the same concept. The three communication channels are the dense latent channel (the full sender hidden state, 65,536 bits), the SAE-sparse channel (active feature indices and magnitudes, ~2,300 bits), and the text channel (the concept name or passage re-encoded by the receiver). Cross-architecture communication is implemented by Orthogonal Procrustes alignment — a pure rotation fitted on paired anchor c

Load-bearing premise

The negative conclusion rests on the assumption that the tested tasks — concept identification, disambiguation, and cross-lingual naming, all with text-expressible answers — are representative of real multi-agent communication needs; a task whose required information cannot be serialized to tokens would void it.

What would settle it

Run a communication task where the sender must convey information that cannot be faithfully serialized into text — e.g., a spatial layout, a novel visual shape, or a multi-hop reasoning state — and check whether a dense or sparse latent channel beats a text channel on a held-out receiver. A positive result would refute the paper's central negative claim. The paper itself proposes the cleanest version: align a pure vision model (no text training) to a language receiver via CCA and test whether a pathology image retrieves the correct text concept; above-chance retrieval would show latent geometr

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

If this is right

  • Current multi-agent systems that communicate in plain text are not leaving task-relevant information on the table: on all text-expressible tasks tested, text matches or beats direct latent transfer.
  • The SAE-sparse channel's 28x compression at 99.4% probe fidelity gives bandwidth- and interpretability-constrained deployments a practical reason to use sparse latent codes even when text remains the accuracy baseline.
  • Cross-architecture latent transfer is viable: a single orthogonal rotation maps one model's concept space into another's with 92% top-1 retrieval, so downstream agents could share latent states without joint training.
  • The 3–10pp latent-to-text gap on tasks is attributable to alignment quality, not channel fidelity or data size; better alignment (e.g., low-rank CCA, more anchors or prompts) would likely close the gap without granting latent channels an advantage on current task types.
  • Latent features add no value on top of a rich text passage, and the specifically 'lost' features actively hurt when re-injected — so text augmentation with hidden states is not a promising direction for concept-identification workloads.

Where Pith is reading between the lines

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

  • Editorial inference: the paper's negative conclusion is bounded by its task battery; a task requiring the sender to convey a non-textual relation (spatial layout, novel visual shape, unverbalized reasoning state) could break the ceiling, and the paper itself names cross-modal and deeper-than-language knowledge as the natural next probes.
  • Editorial inference: the paper's preliminary chain-of-thought results — pre-answer hidden states recover 33.1% accuracy on 'reverse-gap' reasoning tasks while text CoT reaches 96.2% — suggest latent states do carry reasoning signal; the decisive experiment would be a reasoning task where serializing the reasoning into tokens itself loses information.
  • Editorial inference: the finding that instruction-tuned Mistral aligns slightly worse than base Mistral hints that RLHF pushes representations away from the shared geometric core; comparing base versus chat models across more families would test whether the alignment tax is caused by instruction tuning specifically.
  • Editorial inference: the attention-sink injection failure in instruction-tuned models means practical latent communication cannot simply splice a vector into any layer; deployment will require template-aware injection positions, a constraint the paper's geometric-comparison evaluation sidesteps.
Share X Bluesky LinkedIn Reddit HN

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

3 major / 4 minor

Summary. The paper constructs three inter-agent communication channels—dense latent, SAE-sparse, and text—and compares them on concept-discrimination tasks involving Llama 3.1 8B-Instruct and Mistral 7B. The central positive results are that the SAE-sparse channel preserves high probe accuracy at substantial compression, and that Procrustes alignment enables cross-architecture concept retrieval well above chance. The central negative claim is that text serialization destroys 88% of SAE features, but that the lost features encode surface form rather than task-relevant semantics, so the latent channel never outperforms text on the tested tasks. The paper is structured as a gated experimental study with explicit prerequisites, claims C1–C3, task-level evaluation, and text-augmentation tests, and it is candid about the limitation that all tasks have text-expressible answers.

Significance. If the central claims hold, the paper makes a useful contribution by quantifying, at feature level, how much information is lost in text-based agent communication and by providing strong evidence for cross-architecture representational convergence (92% top-1 retrieval against a 0.87% chance baseline). The methodological apparatus—Procrustes with held-out generalization, SAE feature survival, text augmentation—is well conceived and goes beyond simple cosine-similarity comparisons. The honest negative result is valuable for the multi-agent systems community. However, the strongest negative conclusion depends on a feature-loss measurement that is currently performed on bare concept names rather than on the realistic text-passage channel; this needs to be fixed or the claims must be substantially weakened.

major comments (3)
  1. [§2.5, §3.3, Table 2] The 88% feature-loss rate is computed by comparing a contextual cloze-prompt encoding with a bare concept-name encoding (e.g., "Photosynthesis"). This measures the effect of removing context, not the effect of serializing the sender's latent representation into a natural-language passage. The task-level evaluation in §3.5 uses text-passage as the realistic text channel, and this channel outperforms the latent channels on all task types. Therefore the central claim that 'text serialization destroys 88% of SAE features' is not established for the text channel that actually matters. The authors should recompute the feature-survival analysis using the textual passages used in §2.7, or explicitly reframe the 88% result as 'context removal destroys features' and separate it from the serialization claim.
  2. [§3.6, Eq. (2)] The lost-feature augmentation experiment lacks a random-feature control. The lost features are defined as those active in contextual encoding but absent from bare-name encoding; when added to a text-passage representation, they degrade performance. Without a control of randomly selected SAE features of matched activation frequency and magnitude, the degradation cannot be attributed to the semantic content of the lost features. It may instead reflect that these features are out-of-distribution relative to the receiver's text-passage representation, or that the Procrustes mapping is poorly calibrated for them. A random-feature control is necessary to support the conclusion that the lost features 'act as noise' because they are surface-form features.
  3. [§3.2, Table 1] The headline C1 result—99.4% SAE-sparse vs 80.4% text accuracy—is reported with a text-channel standard deviation of 24.4%. No significance test is provided, and with this variance the 19.4pp gap may not be robust across concepts or prompt samples. The authors should report confidence intervals or a paired test across the 84 concept pairs, and clarify how probe accuracy varies across concepts.
minor comments (4)
  1. [Abstract / §4.4] The abstract states 'text serialization destroys 88% of SAE features' without the important qualifier that this is measured against bare-name re-encoding, not passage-level serialization. The limitations section does acknowledge the text-expressibility assumption, but the abstract and §3.3 overstate the specificity of the result.
  2. [§2.7 / §3.5] The cross-lingual concept result (+1.4pp over text-passage) is reported as the only regime where latent matches text. Given the small number of tasks (74), it would be useful to report a confidence interval or significance test; without it, the 'parity' claim is not statistically grounded.
  3. [§3.4] The 92% top-1 retrieval at 140 anchors is based on only 25 held-out concepts. The authors should acknowledge the small holdout size and report the binomial confidence interval; at n=25, the interval around 92% is roughly [74%, 99%].
  4. [Throughout] Several typographical issues: 'LLMS' in §1, 'an corresponding' in §4.4, 'T able' in Table captions, 'identityreplacement' in §3.3. Figure 1 caption is incomplete ('LLama... offers different positions'). The paper should be copy-edited before publication.

Circularity Check

1 steps flagged

Minor self-definitional step in the text round-trip; central claims otherwise empirically grounded.

specific steps
  1. self definitional [§2.5 (Step 2) and §3.3, Table 2; Abstract]
    "a concept is first represented in activation space via a contextual prompt, then, secondly, re-encoded by the same model, but this time reduced to the bare concept name, and the resulting representation is compared with the original. ... Using a text round-trip, we perform feature survival analysis to find that text serialization destroys 88% of SAE features, replacing them with a different feature set."

    The 'text round-trip' is defined as comparing a contextual-prompt encoding with a bare-concept-name encoding. The 88% feature-loss rate is therefore the complement of the overlap between two inputs that differ by construction in context, tokenization, and position. Calling this 'text serialization destroys 88% of SAE features' and attributing the loss to 'surface form' (tokenisation artifacts, prompt-structural features, positional encoding effects, §3.3) restates the operationalization: the lost set is definitionally the context-dependent difference between the two encodings. The task-level text-passage channel—the realistic text channel—is not what the 88% number measures, so the headline destruction rate is built into the round-trip definition rather than established for text serializat

full rationale

The Procrustes alignment is fitted on anchor concepts and evaluated on held-out concepts, so the 92% retrieval is genuine generalization, not a fitted value. The C1 channel-fidelity results and the C3 alignment results are empirical measurements with no self-citation chain. The main circularity concern is the text round-trip in §2.5: the paper defines 'text serialization' operationally as reducing a contextual prompt to the bare concept name, and then reports the resulting feature displacement as if it characterized text serialization broadly. This is a mild self-definitional step because the 88% loss is guaranteed by comparing two deliberately different inputs, and the paper's own 'surface form' components are exactly the dimensions along which those inputs differ. However, the paper is transparent about the bare-name vs. text-passage distinction and explicitly limits its negative conclusion to text-expressible tasks (§4.4). The augmentation experiment provides independent—though not fully controlled—evidence that the lost features do not help on the tested tasks. The central negative claim is therefore not merely a renaming of the input; it is an overreach in interpretation rather than a derivation equivalent to its own assumptions. No load-bearing self-citation or fitted-parameter-as-prediction circularity was found.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 0 invented entities

The central claims are empirical, not derivational. The main fitted object is the Procrustes rotation, which is properly evaluated on held-out concepts. The experimental conclusions depend on the domain assumption that SAE features capture semantics and that the task battery is representative; both are partly acknowledged by the authors.

free parameters (3)
  • Procrustes rotation matrix R = 4,096×4,096 orthogonal matrix fitted on 50–140 anchor concepts
    Fitted on anchor concept vectors from Llama and Mistral; tested on held-out concepts. The retrieval accuracy is the central evidence for cross-architecture communication.
  • Extraction layer (23 for Llama, 24 for Mistral) = approximately 72–75% of depth
    Chosen based on prior work; affects all representations and thus all results.
  • Mixing coefficient α in augmentation = sweep {0.01, 0.05, 0.1, 0.2, 0.5, 1.0}
    Hand-chosen sweep; the null augmentation result is conditional on this additive mixing protocol.
axioms (3)
  • domain assumption SAE features are a valid measurement of semantic content
    Assumed in §2.5 and used for feature survival analysis; authors verify with Δcos but SAEs may also capture surface features.
  • ad hoc to paper The task battery is representative of agent communication needs
    All tasks have text-expressible answers, so the negative conclusion is conditional on this assumption; stated in §4.4.
  • domain assumption Linear probes trained on sender representations are an appropriate measure of channel fidelity
    Probes are trained on ground-truth sender states and tested on channel outputs; if sender and receiver spaces are not aligned (as for text channel), probe accuracy may be misleading.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Latent Communication Between Language Model Agents: Channels, Alignment, and the Limits of Text." pith.science (2026). https://pith.science/paper/RPU24YQM

@misc{pith2026260714103,
  author       = {Pith},
  title        = {Pith review of: Latent Communication Between Language Model Agents: Channels, Alignment, and the Limits of Text},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RPU24YQM}},
  note         = {Machine review of arXiv:2607.14103}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Multi-agent systems (MAS) are utilized in many contexts and many professions. Those MAS rely on inter-agent communication, usually implemented by clear-text message passing. We hypothesize that Large Language Models may have a world model at their disposal that exceeds expressibility in text when complex concepts need to be communicated. Our aim is to approach a proof of this hypothesis with structured experiments. In this work, we show that LLM agents communicating via text lose information, which we quantify via Sparse Autoencoder (SAE) feature analysis. We construct three communication channels and measure concept-discriminating information in each. We first show that the SAE-sparse channel retains a 99.4% probe accuracy at 28-fold compression over the dense-latent channel vs 80.4% for the text channel. We then proceed to examine the same for cross-architecture communication by using sparse latent space alignment. We find for Procrustes alignment a 92% top-1 retrieval between Llama and Mistral. Using a text round-trip, we perform feature survival analysis to find that text serialization destroys 88% of SAE features, replacing them with a different feature set. We attribute the loss to identity replacement, not attenuation. By our analysis, we were able to attribute a 3-10pp performance penalty to the linear Procrustes alignment, improving with nonlinear alignment methods. In a task-level evaluation we find that the latent channel matches the text channel on cross-lingual concept tasks but never exceeds it. Text augmentation with latent features provides no benefit, leading us to negative conclusions for the initial hypothesis: lost features mostly or completely encode surface form, not task-relevant semantics. To pinpoint the practical advantage of latent communication over a text channel, deeper tasks eliciting complex concepts and an corresponding analysis framework are needed.

Figures

Figures reproduced from arXiv: 2607.14103 by Markus Wenzel.

Figure 1
Figure 1. Figure 1: LLama 3.1 8B offers different positions from which representations can be tapped. Of those, R is the most well researched which is used in this work. pairs constitute 76% of the dataset, supplemented by multilingual subsets in German, French, Spanish, and Mandarin Chinese. For task-level evaluation (§2.7), the dataset was expanded to 163 concepts with an average of 4.7 validated prompts per concept, plus 1… view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of communication channels used: Text channel transports concept names or descriptions, the dense latent channel transports the last token internal representation, and the SAE latent channel transports the sparse most activating neurons. re-encodes entirely: the concept string is processed by the receiver model from scratch, producing a new hidden state that is then used for comparison. 2.4. Li… view at source ↗
Figure 3
Figure 3. Figure 3: UMAP visualisation of SAE feature vectors; textual annotations indicate most/least displaced concepts [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: SAE feature cosine similarity distribution through text round-trip. 3.5. Task-Level Evaluation Concept identification (165 tasks). Dense la￾tent: 58.2%. Text-passage: 65.5% (chance: 25%). The 7.3 pp gap is consistent across concept categories. Using prompt-averaged sender vectors, the dense la￾tent channel rises to 62.4%, narrowing the gap to 3.1 pp. Within-concept representation variance provides a partia… 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

16 extracted references · 2 linked inside Pith

  1. [1]

    M. Huh, B. Cheung, T. Wang, and P. Isola. The Platonic Representation Hypothesis. InICML 2024, 2024. 12

    2024

  2. [2]

    Bansal, P

    Y. Bansal, P. Nakkiran, and B. Barak. Revisit- ing model similarity. InNeurIPS 2021, 2021

    2021

  3. [3]

    Lenc and A

    K. Lenc and A. Vedaldi. Understanding image representations by measuring their equivariance and equivalence. InCVPR 2015, 2015

    2015

  4. [4]

    Bricken, A

    T. Bricken, A. Templeton, J. Marcus, et al. To- wards monosemanticity: Decomposing language models with dictionary learning.Transformer Circuits Thread, 2023

    2023

  5. [5]

    Cunningham, A

    H. Cunningham, A. Ewart, L. Riggs, R. Huben, and L. Sharkey. Sparse autoencoders find highly interpretable features in language models. arXiv:2309.08600, 2023

    Pith/arXiv arXiv 2023

  6. [6]

    Conneau, S

    A. Conneau, S. Wu, H. Li, L. Zettlemoyer, and V. Stoyanov. Emerging cross-lingual structure in pretrained language models. InACL 2020, 2020

    2020

  7. [7]

    Schäfer, T

    R. Schäfer, T. Nicke, H. Höfener, A. Lange, D. Merhof, F. Feuerhake, V. Schulz, J. Lotz, and F. Kiessling. Overcoming data scarcity in biomedical imaging with a foundational multi- task model.Nature Computational Science, 2024

    2024

  8. [8]

    Cherti, R

    M. Cherti, R. Beaumont, R. Wightman, M. Wortsman, G. Ilharco, C. Gordon, C. Schuh- mann, L. Schmidt, and J. Jitsev. Repro- ducible scaling laws for contrastive language- image learning. InCVPR 2023, 2023

    2023

  9. [9]

    Gromov, K

    A. Gromov, K. Tirumala, H. Shapourian, P. Glo- rioso, and D. Roberts. The unreasonable ineffec- tiveness of the deeper layers.arXiv:2403.17887, 2024

    Pith/arXiv arXiv 2024

  10. [10]

    S. L. Smith, D. H. P. Turban, S. Hamblin, and N. Y. Hammerla. Offline bilingual word vectors, or- thogonal transformations and the inverted soft- max. InICLR 2017, 2017

    2017

  11. [11]

    Conneau, G

    A. Conneau, G. Lample, M. Ranzato, L. De- noyer, and H. Jégou. Word translation without parallel data. InICLR 2018, 2018

    2018

  12. [12]

    Elhage, T

    N. Elhage, T. Hume, C. Olsson, N. Schiefer, et al. Toy models of superposition.Transformer Circuits Thread, 2022

    2022

  13. [13]

    Templeton, T

    A. Templeton, T. Conerly, J. Marcus, J. Lind- sey, et al. Scaling monosemanticity: Extracting interpretable features from Claude 3 Sonnet. Transformer Circuits Thread, 2024

    2024

  14. [14]

    Pires, E

    T. Pires, E. Schlinger, and D. Garrette. How multilingual is multilingual BERT? InACL 2019, 2019

    2019

  15. [15]

    Kottur, J

    S. Kottur, J. M. F. Moura, S. Lee, and D. Batra. Natural language does not emerge ‘naturally’ in multi-agent dialog. InEMNLP 2017, 2017

    2017

  16. [16]

    Lester, R

    B. Lester, R. Al-Rfou, and N. Constant. The power of scale for parameter-efficient prompt tuning. InEMNLP 2021, 2021. 13

    2021

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.