Do language models converge on one voice?

We used one completed ElectionBench run to ask a narrow question: when language models share a simulated town for a week, does their speech converge on one voice? The run contains 2,872 conversation turns in 618 conversations over 84 ticks, or seven in-world days. Ten language models supplied the residents’ speech; the scripted poll interviewer was excluded from the convergence calculations.

The answer is no, with an important qualification. The models discuss the same election throughout the run, and their residual styles become modestly more aligned than shuffled model labels would predict. That movement does not produce a common voice. Most of the structure is already present on day one, and the later changes are concentrated in a few models and pairs.

Raw embedding similarity mixes subject matter with expression. Two models discussing the same campaign can be close in embedding space even when their phrasing and register remain different. We therefore compute two measurements for each centered five-tick window.

Topic similarity is the mean pairwise cosine between each model’s speech centroid. Residual style similarity uses the same centroids after subtracting the pooled centroid for that window. Removing this shared direction reduces the contribution from whatever the town is discussing at that time.

Centered residuals are negatively correlated by construction, so zero is not the right baseline. We permute model labels within each window 30 times while preserving the group sizes, then compare the observed residual-style cosine with the resulting null distribution. The figure reports the null mean and its 5th–95th percentile band.¹²

Topic cosine ranges from 0.69 to 0.85, with a mean of 0.77. The models remain semantically close because they share the election context. Raw embeddings alone would therefore overstate convergence in voice.

The residual-style result is smaller. The first sequence of at least five consecutive ticks above the shuffle band starts at tick 20, late on day two. From day three onward, 55 of 60 ticks are above the band. The observed-minus-null gap reaches its maximum at tick 79, at +0.056 cosine.

This is evidence of structure beyond the label-shuffle baseline, not evidence that the models become interchangeable. The residual-style values remain heterogeneous across pairs, as the next sections show.

The day-level pair matrix shows that the largest group did not form during the week. On day one, Claude Opus 5, Kimi K3, Qwen 3.8 Max, and Muse Spark 1.1 have a mean within-group residual-style cosine of 0.29. Their mean cosine to MiniMax M2.5 and DeepSeek V4 Flash is -0.31. The difference is 0.60 on day one and 0.77 on day six.

Assignment role follows a similar partition. Across the full week, candidate–candidate pairs have a residual-style cosine of 0.24, villager–villager pairs -0.05, and mixed pairs -0.26. The villager daily mean falls from 0.01 on day one to -0.17 on day seven, while candidate pairs remain positive on every day.³

Claude Sonnet 5 shows the clearest change in relative position. Its residual-style cosine to MiniMax is 0.28 on day one and -0.10 on day seven. Its mean cosine to the four-model group rises from -0.36 to -0.13. Sonnet moves away from MiniMax and toward that group, but its day-seven mean remains negative, so the run does not show it joining the group.

A second view compares each model’s mean residual-style cosine to the rest of the field over the first and last twelve ticks. MiniMax is the only model with a negative change. In the pooled corpus it also has a standardized distinct-2 score of 0.88 and a normalized duplicate-turn rate of 0.16. Those measures show high repetition, while its negative field change shows that repetition did not make its speech more similar to the other models.

For each of the six day boundaries, the observed-minus-null style gap at the first tick after the boundary is at least as large as it was at the last tick before it. The largest change is +0.020 between days three and four; the day-six to day-seven boundary is close behind.

This is a continuity check, not evidence that an unobserved overnight process caused the increase. The two measurements use centered five-tick windows, so adjacent dusk and dawn estimates share most of their turns. The result only shows that the measured alignment does not reset at the recorded day boundary.

We also computed one residual speech centroid per model over a pooled corpus of 25,401 turns from 40 runs and 18 models. We repeated the calculation with Qwen3 Embedding 4B and Gemini Embedding 2, then used the consensus distances only to lay out the map.

Both embedders place Claude Opus 5, Kimi K3, Qwen 3.8 Max, and Muse Spark 1.1 in the same neighborhood. The two GPT-5.6 Sol serving tiers also pair closely. This reduces the chance that the neighborhood is specific to one embedding model, but it is not an independent replication: the pooled corpus includes this run and other runs from the same simulation system.

Direct pairwise mimicry makes a simple prediction: model pairs that share more conversation turns should show more positive residual-style drift. We tested that prediction for the 35 pairs with measurements at 40 or more ticks. Drift is the pair’s fitted residual-style slope across the run, scaled to the run length; exposure is the number of turns the pair shared in conversations.

The prediction fails in this sample. Exposure has a negative rank association with drift (Spearman ρ = -0.34). Villager–villager pairs average 147 shared turns and a drift of -0.21. Candidate–candidate pairs average 30 shared turns and a drift of +0.03.

The result does not support more direct contact as the explanation for the observed convergence. It does not establish the cause of the remaining movement. Shared broadcasts and campaign context, model-family priors, role, and tier all vary in this run and could produce the same pattern.⁴

This analysis covers one seven-day run in one scenario. The per-tick analysis uses one embedder, sparse models are absent from some windows, and the centered windows make adjacent measurements dependent. Model, role, tier, exposure, and character assignment were not randomized independently, so the run supports associations rather than causal estimates.

Within those limits, the evidence answers the original question. The run has high topic similarity and a modest increase in residual-style alignment, but the ten models do not converge on one voice. The clearest structure is a group visible on day one, plus later movement by a small number of models. Testing why that movement occurs requires repeated runs that vary role, model, and exposure separately.