pith. sign in

arxiv: 2607.00233 · v1 · pith:FM5AN7LVnew · submitted 2026-06-30 · 💻 cs.AI · cs.CL· cs.IT· cs.MA· math.IT

From Signals to Structure: How Memory Architecture Drives Language Emergence in LLM Agents

Yashar Talebirad , Eden Redman , Ali Parsaee , Osmar R. Zaiane This is my paper

Pith reviewed 2026-07-02 18:58 UTC · model grok-4.3

classification 💻 cs.AI cs.CLcs.ITcs.MAmath.IT
keywords LLM agentslanguage emergencememory architecturesignaling gamescoordinationchannel capacityLewis game
0
0 comments X

The pith

Memory architecture decides whether LLM agents turn signals into stable shared languages, more than channel capacity does.

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

In a Lewis signaling game, two LLM agents must coordinate on a code using only interaction history. Tests across five memory architectures show that a persistent private notebook lets agents use surplus channel capacity to reach reliable coordination, while stateless agents degrade once vocabulary exceeds what a rolling context window can track. The notebook externalizes learned conventions so agents avoid re-deriving codes each round. An information-bottleneck prediction of an optimum at capacity equal to the number of objects fails; the bottleneck instead marks a fragility point and surplus capacity performs better overall. Channel capacity by itself cannot predict success; memory architecture must be considered together with it.

Core claim

Agents equipped with a persistent private notebook achieve the highest coordination reliability (0.867 ± 0.023 at capacity = 25) by externalizing conventions, whereas stateless agents peak at moderate capacity and then decline as the vocabulary grows. The predicted bottleneck capacity of 8 acts as a fragility point rather than an optimum, and surplus capacity is generally advantageous. Memory architecture therefore determines whether interaction history becomes stable conventions.

What carries the argument

Persistent private notebook memory architecture, which stores learned conventions outside the agents' immediate context and frees them from re-deriving codes each round.

If this is right

  • Agents with notebook memory benefit from surplus channel capacity instead of suffering high-capacity collapse.
  • Stateless agents reach a performance peak at moderate capacity and then degrade as vocabulary size exceeds context tracking limits.
  • An information bottleneck at capacity equal to the number of objects is a fragility point rather than an optimum.
  • Channel capacity alone cannot predict coordination success; both memory architecture and capacity must be examined together.

Where Pith is reading between the lines

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

  • External memory may become a necessary design choice when scaling multi-agent language emergence beyond small vocabularies.
  • The same memory-capacity interaction could be tested in non-LLM agents to check whether the pattern is architecture-specific or general.
  • Human language evolution might similarly depend on external memory aids once signal sets grow large.

Load-bearing premise

Observed differences in coordination success are caused by the choice of memory architecture rather than by other unmeasured differences in how the LLMs process or retain information.

What would settle it

Re-running the same signaling-game trials while holding memory architecture fixed but varying other LLM processing details, and finding that coordination success no longer tracks the original memory-type differences, would falsify the claim.

Figures

Figures reproduced from arXiv: 2607.00233 by Ali Parsaee, Eden Redman, Osmar R. Zaiane, Yashar Talebirad.

Figure 1
Figure 1. Figure 1: The referential signaling game. Each round, [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Prompt template and output schema for both agents. The common preamble (top) is identical in both system prompts. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Learning dynamics across memory architectures (cap [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Capacity curves for scratchpad and memory [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Effect of memory window size on late-game accu [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
read the original abstract

How do two agents invent a shared language from scratch? In a Lewis signaling game, a sender and receiver must coordinate on a code using only their interaction history. We study five memory architectures across varying channel configurations with LLM agents and find that memory architecture matters more than channel capacity. Agents with a persistent private notebook benefit from surplus channel capacity and avoid the high-capacity collapse seen in stateless agents, achieving the most reliable coordination ($0.867 \pm 0.023$ at capacity = 25). Stateless agents peak at moderate capacity and then degrade as the vocabulary grows beyond what a rolling context window can track The notebook externalizes learned conventions, freeing agents from having to re-derive codes each round. An information bottleneck-inspired argument predicts an optimal capacity equal to the number of objects. Instead, the bottleneck (capacity = 8) proves to be a fragility point, and surplus capacity is generally better. We show that channel capacity alone cannot predict coordination; memory architecture determines whether agents turn interaction history into stable conventions, and both dimensions are needed to understand how signals become language.

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 / 1 minor

Summary. The manuscript presents an empirical study of language emergence in LLM agents playing a Lewis signaling game. It compares five memory architectures across channel capacities and reports that a persistent private notebook achieves the highest coordination success (0.867 ± 0.023 at capacity=25), benefits from surplus capacity, and avoids the high-capacity collapse seen in stateless agents. The authors conclude that memory architecture determines whether agents convert interaction history into stable conventions and that channel capacity alone cannot predict coordination success, contrary to an information-bottleneck prediction that capacity=8 should be optimal.

Significance. If the results hold after addressing potential confounds, the work would provide concrete evidence that externalized memory stabilizes conventions in LLM agents more effectively than capacity scaling, with implications for multi-agent system design. The inclusion of quantitative metrics with error bars is a positive feature of the empirical approach.

major comments (2)
  1. [Abstract] Abstract: the reported coordination success of 0.867 ± 0.023 lacks any details on the number of trials, the computation of error bars, statistical tests performed, or controls for LLM stochasticity, leaving the central empirical claim only weakly supported.
  2. [Methods] Methods/Experimental Design: all five architectures are realized inside the same LLM via different prompt structures, yet the manuscript supplies no evidence that prompt length, token overhead, or auxiliary instructions were matched across conditions; this leaves open the possibility that observed gaps (e.g., notebook vs. stateless) arise from effective context-length differences rather than the memory mechanism itself.
minor comments (1)
  1. [Abstract] The abstract lists five architectures but does not name them explicitly; adding the names would improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our empirical study of memory architectures in LLM agents. The comments highlight opportunities to strengthen the reporting of results and experimental controls. We address each major comment below and commit to revisions that directly respond to the concerns raised.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the reported coordination success of 0.867 ± 0.023 lacks any details on the number of trials, the computation of error bars, statistical tests performed, or controls for LLM stochasticity, leaving the central empirical claim only weakly supported.

    Authors: We agree that the abstract would be strengthened by including these details. The reported value is the mean across 10 independent trials per condition, each using distinct random seeds to mitigate LLM stochasticity; error bars are standard error of the mean. Statistical tests (paired t-tests with Bonferroni correction) appear in the Methods section. We will revise the abstract to incorporate a concise statement of these elements so the central claim is more fully supported on first reading. revision: yes

  2. Referee: [Methods] Methods/Experimental Design: all five architectures are realized inside the same LLM via different prompt structures, yet the manuscript supplies no evidence that prompt length, token overhead, or auxiliary instructions were matched across conditions; this leaves open the possibility that observed gaps (e.g., notebook vs. stateless) arise from effective context-length differences rather than the memory mechanism itself.

    Authors: The referee correctly notes that the manuscript provides no explicit verification that prompt lengths and token overhead were matched. This is a legitimate potential confound. In the revision we will add to the Methods section both (i) measured average token counts for the prompt templates of each architecture and (ii) a description of how auxiliary instructions were minimized and standardized. Where lengths differ materially we will either re-run the affected conditions with length-matched prompts or report the measured differences and argue why they do not explain the observed performance gaps. This will allow readers to evaluate whether the memory mechanism itself drives the results. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical comparison of LLM memory architectures with no derivations or fitted predictions

full rationale

The paper reports experimental results from Lewis signaling games run with LLM agents under five memory architectures and varying channel capacities. The central claim—that memory architecture determines coordination success more than channel capacity—is grounded in observed performance differences (e.g., notebook at 0.867 vs. stateless degradation). No equations, parameter fitting, or first-principles derivations are present. The reference to an 'information bottleneck-inspired argument' is external and contrasted with results rather than used as a load-bearing derivation. No self-citations, ansatzes, or renamings reduce any claim to its inputs by construction. The work is self-contained as an empirical study.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The paper is an empirical simulation study; no free parameters, axioms, or invented entities are stated in the abstract.

pith-pipeline@v0.9.1-grok · 5736 in / 1052 out tokens · 16989 ms · 2026-07-02T18:58:08.630571+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

25 extracted references · 3 canonical work pages · 1 internal anchor

  1. [1]

    Schulz, E. (2025). Playing repeated games with large lan- guage models.Nature Human Behaviour, 9(7):1380–1390

    2025

  2. [2]

    F., Aiello, L

    Ashery, A. F., Aiello, L. M., and Baronchelli, A. (2025). Emergent social conventions and collective bias in LLM populations. Science Advances, 11(20):eadu9368

    2025

  3. [3]

    and Kirby, S

    Brighton, H. and Kirby, S. (2006). Understanding linguistic evolu- tion by visualizing the emergence of topographic mappings. Artificial Life, 12:229–242

    2006

  4. [4]

    Brown, T. B. et al. (2020). Language models are few-shot learners. InAdvances in Neural Information Processing Systems

    2020

  5. [5]

    Kirby, S. (2001). Spontaneous evolution of linguistic structure: An iterated learning model of the emergence of regularity and ir- regularity.IEEE Transactions on Evolutionary Computation, 5:102–110

    2001

  6. [6]

    Kirby, S., Cornish, H., and Smith, K. (2008). Cumulative cultural evolution in the laboratory: An experimental approach to the origins of structure in human language.Proceedings of the National Academy of Sciences, 105:10681–10686

    2008

  7. [7]

    Kirby, S., Tamariz, M., Cornish, H., and Smith, K. (2015). Com- pression and communication in the cultural evolution of lin- guistic structure.Cognition, 141:87–102

    2015

  8. [8]

    Kouwenhoven, T., Peeperkorn, M., and Verhoef, T. (2025). Search- ing for structure: Investigating emergent communication with large language models. InInternational Conference on Com- putational Linguistics

    2025

  9. [9]

    M., Tuyls, K., and Clark, S

    Lazaridou, A., Hermann, K. M., Tuyls, K., and Clark, S. (2018). Emergence of linguistic communication from referential games with symbolic and pixel input. InInternational Con- ference on Learning Representations

    2018

  10. [10]

    Lazaridou, A., Peysakhovich, A., and Baroni, M. (2017). Multi- agent cooperation and the emergence of (natural) language. InInternational Conference on Learning Representations

    2017

  11. [11]

    (1969).Convention: A Philosophical Study

    Lewis, D. (1969).Convention: A Philosophical Study. Harvard University Press

    1969

  12. [12]

    and Bowling, M

    Li, F. and Bowling, M. (2019). Ease-of-teaching and language structure from emergent communication. InAdvances in Neu- ral Information Processing Systems

    2019

  13. [13]

    Lowe, R., Foerster, J., Boureau, Y ., Pineau, J., and Dauphin, Y . (2019). On the pitfalls of measuring emergent communica- tion. InInternational Conference on Autonomous Agents and Multi-Agent Systems

    2019

  14. [14]

    Miller, G. A. (1956). The magical number seven, plus or minus two: Some limits on our capacity for processing information. Psychological Review, 63:81–97

    1956

  15. [15]

    Luan, D., Sutton, C., and Odena, A. (2021). Show your work: Scratchpads for intermediate computation with lan- guage models.arXiv preprint arXiv:2112.00114

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  16. [16]

    Parsaee, A., Talebirad, Y ., Szepesv ´ari, C., Ohal, V ., and Red- man, E. (2025). LoopBench: Discovering emergent sym- metry breaking strategies with LLM swarms.arXiv preprint arXiv:2512.13713

    work page arXiv 2025

  17. [17]

    Regier, T., Kemp, C., and Kay, P. (2015). Word meanings across languages support efficient communication. In MacWhin- ney, B. and O’Grady, W., editors,The Handbook of Language Emergence, pages 237–263. Wiley

    2015

  18. [18]

    B., and Kirby, S

    Ren, Y ., Guo, S., Labeau, M., Cohen, S. B., and Kirby, S. (2020). Compositional languages emerge in a neural iterated learning model. InInternational Conference on Learning Representa- tions

    2020

  19. [19]

    M., and Cho, K

    Resnick, C., Gupta, A., Foerster, J., Dai, A. M., and Cho, K. (2020). Capacity, bandwidth, and compositionality in emer- gent language learning. InInternational Conference on Au- tonomous Agents and Multi-Agent Systems

    2020

  20. [20]

    Shannon, C. E. (1948). A mathematical theory of communication. Bell System Technical Journal, 27:379–423

    1948

  21. [21]

    (2010).Signals: Evolution, Learning, and Information

    Skyrms, B. (2010).Signals: Evolution, Learning, and Information. Oxford University Press

    2010

  22. [22]

    Sutton, R

    Talebirad, Y ., Parsaee, A., Szepesv ´ari, C. Y ., Nadiri, A., and Za¨ıane, O. R. (2026). Toward a theory of hierarchical memory for language agents. InICLR 2026 Workshop on Memory for LLM-Based Agentic Systems. arXiv preprint arXiv:2603.21564

    work page arXiv 2026

  23. [23]

    C., and Bialek, W

    Tishby, N., Pereira, F. C., and Bialek, W. (1999). The information bottleneck method. In37th Annual Allerton Conference on

    1999

  24. [24]

    Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q., and Zhou, D. (2022). Chain-of-thought prompting elicits reasoning in large language models. InAd- vances in Neural Information Processing Systems

    2022

  25. [25]

    Zaslavsky, N., Kemp, C., Regier, T., and Tishby, N. (2018). Effi- cient compression in color naming and its evolution.Pro- ceedings of the National Academy of Sciences, 115:7937– 7942

    2018