When two neural networks learn to talk to each other from scratch, they tend to invent private codes: the sender exploits quirks of one particular receiver, and the "language" fails on new inputs or new partners. Human sociolinguistics has a well-tested account of when shared codes become simple and learnable. The Linguistic Niche Hypothesis holds that languages of large, mixed communities lose irregular complexity. Lab experiments show that this simplification needs more than a large group of speakers. It needs adult newcomers whose errors are made consistent by proficient speakers who adjust their own output for them. This proposal brings that account to machine agents. It asks whether emergent codes generalize because the population is large, or because some listeners in it are always newcomers, and whether the answer survives across datasets rather than resting on one image benchmark.
The two-sentence version is this. Existing machine work shows either that larger agent communities give better representations, or that weak, frequently reset listeners give better languages in a single pair. Neither separates population size from listener turnover, and each tests on one or two standard datasets. We cross population size with newcomer rate and evaluate on a ladder of datasets of increasing complexity. This tells us which ingredient removes the receiver-specific part of the error, and whether the conclusion is a property of the method or of the dataset it was tested on.
Core Claim
The human evidence predicts an interaction, not a main effect. In transmission chains, adults given a 50/50 mix of simplified and complex input rebuilt the complex system when they learned from 2 sources but abandoned it when they learned from 8 (Simplifications made early in learning can reshape language complexity, $p = .005$). Learners exposed only to other learners' uncoordinated simplifications did not converge on a simpler grammar. Convergence came when proficient speakers accommodated less proficient partners: they regularized irregular verbs 50% of the time in mixed pairs versus 3% in homogeneous pairs (Adult Learning and Language Simplification). A Bayesian model explains why. A speaker who tailors output to a partner with little exposure stops producing rare forms, and entropy drops from 1.96 toward 1.0 (A rational model of linguistic accommodation). Speaker variability alone is not the cause, since adults learn complex rules fine from many voices (Atkinson et al. 2015).
Carried over to neural agents, this gives a falsifiable claim. The receiver-specific part of emergent-language error is driven down mainly by the presence of low-exposure listeners that every sender must serve. Population size matters mostly by making that pressure unavoidable. Size alone, with only experienced listeners, should help less, and turnover in a single pair should help less than turnover spread across a large population. The claim is wrong if a single sender–listener pair with periodic listener resets matches the best population condition on every dataset. That outcome would itself be useful to report.
Relation to Prior Work
Two papers motivate this proposal and supply its pieces.
Emergent Communication: Generalization and Overfitting in Lewis Games splits the signaling-game loss into two parts:
- an information term, measuring how much the message says about the input;
- a co-adaptation term, measuring how far the current listener has drifted from an optimal one.
It isolates these terms by training a fresh probe listener on frozen messages, and shows that resetting or early-stopping the listener gives a 15-point generalization gain. Its experiments are confined to one sender and one listener, on CelebA, ImageNet and synthetic data. We reuse its probe decomposition as our main measurement and its early-stopping procedure as our strongest baseline.
Community size effect in artificial learning systems and its companion Shaping representations through communication randomly pair $K$ encoders with $K$ decoders. They find that larger communities give more abstract, more human-aligned and more seed-stable codes than a capacity-matched ensemble. However, the codes are continuous rather than discrete messages, and the effect weakened sharply on CIFAR-100 compared with MNIST. The authors themselves flag that dataset complexity modulates the effect. That weakening is the reason we treat dataset dependence as a first-class outcome rather than a footnote.
Two further papers constrain how results can be read. Compositionality and Generalization In Emergent Languages shows that agents can generalize with non-compositional codes, and that compositionality mainly predicts how easily naive receivers learn a language (correlations up to $\rho = 0.90$). Input-space size strongly predicts test accuracy ($\rho = 0.86$). It follows that held-out accuracy and compositionality scores alone cannot be the yardstick, and that dataset scale must be controlled. Compositionality and Capacity in Emergent Languages finds sharp capacity thresholds: total failure below about 72,400 parameters and roughly 22 bits of channel. We therefore keep every agent above those thresholds, so that differences between conditions cannot come from agents that simply cannot learn.
The second direction, reliance on existing datasets, enters as the evaluation lens rather than as a dataset to build. Both social-regularization results above rest on one or two standard image sets. DB-GPT-Hub shows in another field that tuning gains on a cleaner benchmark (Spider) shrink on a messier one (BIRD). We ask whether co-adaptation remedies behave the same way.
Research Program
We build discrete Lewis signaling games, in which a sender sees an input, emits a short symbol string, and a listener must pick the input out from distractors. Each game uses a pool of $N$ senders and $N$ listeners, paired at random on every step.
The first manipulation is population size, $N \in \{1, 2, 4, 8, 16\}$. The second is the newcomer rate $r$: the fraction of listeners re-initialized per training epoch, set to zero, low, or high. At any moment, a stationary share of the listener pool therefore has little exposure. With $N=1$, a nonzero $r$ reduces to the partial-reset regime of the Lewis-games paper, so prior work sits inside our grid as a special case.
A third factor tests accommodation directly. In the "aware" condition, each sender receives its partner's exposure count as an input; in the "blind" condition it does not. The rational accommodation model makes a pointed prediction here. Aware senders could keep a private, complex register for experienced listeners, which would weaken population-level regularization. Blind senders must produce one code that also serves newcomers.
We measure four things at the end of training:
- the probe-decomposed information and co-adaptation losses;
- held-out accuracy;
- learnability, defined as the number of steps a freshly initialized listener of a different architecture needs to reach criterion on frozen messages, following the transmission test of the compositionality–generalization paper;
- topographic similarity, the correlation between distances between inputs and distances between their messages.
We also record variance across seeds. The datasets form a complexity ladder:
- the scaled synthetic attribute-value spaces (from 16 objects to $10^8$);
- the $10^6$-object attribute space;
- dSprites;
- CelebA;
- ImageNet.
This ladder lets us test the CbAE observation, that social pressure fades on harder data, in the discrete setting.
Baselines
The naive reference point is a single pair trained continuously, with no resets and no regularization. Every gain is reported relative to it on the same splits.
The strongest existing method is the listener early-stopping and partial-reset procedure from Emergent Communication: Generalization and Overfitting in Lewis Games. It is the strongest because it targets exactly the co-adaptation term we measure, and it achieves the gain without any population. Its reported result is a 15-point gain in generalization accuracy on unseen CelebA data relative to continuous training. The file gives no absolute accuracy for this result. We found no reported result on Lacuna for its ImageNet number or for population variants. We run it with our own code, architectures, data splits, and an equal hyperparameter search budget, including a search over its reset interval.
Beating it means two things on CelebA and ImageNet. The best population condition must show a lower probe-measured co-adaptation loss and higher held-out accuracy, with non-overlapping seed intervals. It must also show better fresh-listener learnability, and do so on at least two other rungs of the dataset ladder.
The second baseline is a capacity-matched ensemble of $N$ independent pairs, following the control in Community size effect in artificial learning systems. It separates communication pressure from added parameters.
The third baseline is a population with $r = 0$. It is the discrete analogue of the community autoencoder, and it isolates size alone.
The key ablation compares turnover concentrated in one pair against the same total number of listener re-initializations spread across a large population. It also compares aware against blind senders.
To keep the comparison fair, all conditions are matched on four quantities:
- total sender and listener gradient steps;
- parameters per agent;
- message length and vocabulary;
- tuning trials.
For CbAE-style comparisons on continuous codes, we use the SimSiam Naming Game's dSprites topographic similarity of 0.22, against 0.19 for the Metropolis-Hastings naming game, as an external reference only. Those models are not population models (SimSiam Naming Game).
Why not just…?
Why not just reset or early-stop the listener in a single pair? This is the real threat, because it already delivers a 15-point CelebA gain. The $N=1, r>0$ cells of our grid are exactly this method. If it matches the best population condition on co-adaptation loss, accuracy and learnability across the dataset ladder, we will report that populations are unnecessary. The human evidence gives a reason to expect otherwise. Learners rebuilt complexity with 2 sources but not with 8, which suggests that newcomer pressure from many partners differs from newcomer pressure from one. The experiment decides it.
Why not just apply dropout or weight decay to the listener? In the Lewis-games study, listener-only dropout and weight decay helped but fell short of early stopping. We include a tuned listener-regularization pair so that this conclusion is re-checked on all five datasets rather than inherited from one.
Why not just use a bigger population, or a bigger ensemble, with no turnover? The community autoencoder already rules out a capacity-matched ensemble as the explanation in the continuous case. It does not test turnover, and its effect faded on CIFAR-100. Our $r=0$ populations measure how much size alone buys on each dataset. If size alone matches size plus turnover, the newcomer claim fails.
Proposed Experiments and Decisive Checks
The first decisive check is the interaction. On each dataset, we fit co-adaptation loss and learnability as a function of $N$, $r$ and their product. The claim requires a reliably negative product term whose size exceeds both main effects. It also requires that, at a fixed total number of resets, spreading them over a population beats concentrating them in one pair.
The second check is accommodation. If aware senders keep partner-specific registers, their messages to experienced listeners should show higher co-adaptation loss and lower learnability than blind senders' messages. Such a result would show that accommodation helps the population converge only when senders cannot tell their partners apart. That would be a sharper statement than the human literature currently supports.
The third check addresses the dataset question. We ask whether the ranking of methods is stable across the ladder. If population effects shrink from synthetic data and dSprites to ImageNet, as the CIFAR-100 attenuation suggests, we will report where on the ladder they vanish. Any remedy validated on a single standard set should then carry that boundary as a caveat.
Finally, because generalization can coexist with non-compositional codes, we never treat topographic similarity alone as success. A condition counts as producing a stronger language only if it lowers probe-measured co-adaptation loss and also speeds learning by fresh listeners of a different architecture.
References
- Simplifications made early in learning can reshape language complexity: an experimental test of the Linguistic Niche Hypothesis
- Adult Learning and Language Simplification
- A rational model of linguistic accommodation and its potential role in language simplification
- Speaker Input Variability Does Not Explain Why Larger Populations Have Simpler Languages
- Emergent Communication: Generalization and Overfitting in Lewis Games
- Community size effect in artificial learning systems.
- Shaping representations through communication: community size effect in artificial learning systems.
- Compositionality and Generalization In Emergent Languages
- Compositionality and Capacity in Emergent Languages
- DB-GPT-Hub: Towards Open Benchmarking Text-to-SQL Empowered by Large Language Models
- SimSiam Naming Game: A Unified Approach for Representation Learning and Emergent Communication