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

Technical Summary: From Signals to Structure

Problem Statement
This paper investigates how Large Language Model (LLM) agents invent a shared language from scratch within a Lewis signaling game framework. While prior research on emergent communication has largely focused on gradient-trained neural agents and the role of channel capacity (bandwidth) in driving compositional structure, this study shifts the focus to memory architecture. The authors ask whether the ability of agents to retain and structure interaction history (via persistent memory) is a more critical determinant of language emergence than the raw capacity of the communication channel. Specifically, they challenge the information-theoretic hypothesis that a channel capacity exactly matching the number of referents (the "bottleneck") acts as an optimal pressure point for compositionality.

Methodology
The authors conducted three studies using gpt-5.4-mini as the base model for both sender and receiver agents. The agents played a referential signaling game for $N=200$ rounds, where the sender observed a target object (from a pool of 8 objects defined by 3 binary features) and transmitted a fixed-length symbolic message to the receiver, who then guessed the target.

• Memory Architectures: The study compared five distinct memory conditions, all sharing a baseline rolling window of the last 20 interactions:
  1. Memory Only: Agents rely solely on the rolling window.
  2. Scratchpad: Agents maintain a persistent, free-form notebook (overwritten each round) to track conventions.
  3. Codebook: Agents maintain a fixed-capacity (10 slots) slot list edited in-place.
  4. Codebook Meta: Codebook plus a persistent meta-note for language-level rules.
  5. Env Board: A shared public table compiled by the environment (not a private memory).
• Channel Configurations: Channel capacity ($C = |V|^L$) was swept from 4 to 125 by varying vocabulary size ($|V|$) and message length ($L$).
• Metrics: Performance was measured by late-game accuracy (rounds 151–200), Topographic Similarity (TopSim) to assess compositional structure, Mutual Information (MI) between signal positions and object features, and collision rates.

Key Results

1. Memory Architecture Dominates Channel Capacity:
   The study finds that memory architecture is a stronger predictor of coordination success than channel capacity. Agents with a persistent private notebook (Scratchpad) achieved the most reliable coordination (accuracy $0.867 \pm 0.023$ at capacity 25) and avoided the performance collapse seen in stateless agents. In contrast, stateless agents (Memory Only) peaked at moderate capacity (25) but degraded significantly as capacity increased to 64, where collision rates hit 1.0.

2. The Bottleneck is a Fragility Point, Not an Optimum:
   Contrary to information bottleneck theories suggesting that a channel capacity equal to the number of referents ($C=8$ for 8 objects) forces optimal compression, the authors found this point to be a fragility point. At $C=8$, outcomes were bimodal: agents either succeeded or failed completely, with no middle ground. The lack of surplus capacity prevented agents from repairing early collisions, leading to high variance and low average performance.

3. Surplus Capacity Stabilizes Conventions:
   Agents with persistent memory benefited from surplus capacity. The Scratchpad condition showed increasing accuracy as capacity grew, reaching 0.90 at $C=125$. Conversely, stateless agents suffered from "code proliferation" at high capacities; the rolling window could not accumulate sufficient evidence per code as the signal space expanded, leading to ambiguity and performance collapse.

4. Consolidation vs. History Length:
   Study 3 demonstrated that the failure of stateless agents at high capacity is a consolidation problem, not merely a history-length problem. Increasing the rolling window size for stateless agents from 20 to 40 rounds did not improve performance at $C=64$. However, Scratchpad agents achieved high accuracy with only 10 rounds of context, proving that the ability to externalize and persist learned conventions is the critical factor, not the volume of raw history.

5. Compositionality and Drift:
   While high accuracy was achieved, the resulting languages were often holistic lookup tables rather than fully compositional codes. The Scratchpad condition showed partial positional structure, but "convention drift" (where senders re-assign tokens to different objects mid-game) was identified as a cause for late-game accuracy dips in some runs.

Significance and Claims
The paper claims that channel capacity alone cannot predict coordination in LLM agents. Instead, the transition from signals to language depends on the joint interaction of channel capacity and memory architecture.

• Memory Architecture as a Determinant: The structure of intermediate representations (how agents store what they learn) determines whether interaction history is converted into stable conventions. Persistent notebooks allow agents to externalize conventions, freeing them from re-deriving codes each round.
• Re-evaluating the Bottleneck: The study argues that the "bottleneck" ($C=8$) is not a compositional optimum for LLM agents but a point of fragility where early errors become permanent due to a lack of redundancy. Surplus capacity is generally superior as it allows agents to repair collisions and stabilize conventions.
• Limitations: The authors modestly note that results are indicative rather than statistically conclusive due to the use of a single model and limited seeds for some conditions. They emphasize that the observed curves may reflect specific LLM properties rather than universal agent behaviors, suggesting future work should test open-weight models and larger compositional spaces.

In conclusion, the paper posits that for LLM agents, signals become language only when channel capacity and memory architecture jointly support the stabilization of conventions. Without persistent memory, agents cannot maintain stable codes in high-capacity environments, regardless of the theoretical optimality of the channel bandwidth.