gpt-oss:20b

When the experiment added the voting network's intuition to the crew prompt, the intuition line stated that the model can pick the real impostor about 70% of the time, but that 70% was a rough estimate the author made when writing the prompt, not an actual measurement.

The author's earlier benchmark testing on ranking language models by their ability to identify liars in the game Mafia found that first-day accusations were entirely luck.

In tasks requiring precise alignment with the statistical features of a specific environment, a general large model may still fall short of a specialized small model trained extensively within the domain.

Developer Seth Wheeler replaced the meeting voting system originally driven by a small recurrent network with the open-weight model gpt-oss:20b in an Among Us simulator, and the result was that across six configurations, the large model never made the crew perform better.

gpt-oss:20b is an open-weight reasoning model released by OpenAI in 2025, using a MoE architecture, with about 21 billion total parameters, about 3.6 billion activated parameters, and support for adjustable reasoning effort.

In the configuration with eight players and two impostors using second-league bots, the trained network's crew win rate was 52.0%, with an average of 0.99 impostors ejected per round, whereas gpt-oss in the turn-based Q&A format won only 2 of 16 rounds (12.5%), ejected 0.06 impostors per round, and wrongly ejected 0.94 innocent crewmates.

In the eight-player configuration, the win-rate gap between gpt-oss and the trained network had a one-sided Fisher exact test p = 0.0017, far beyond random fluctuation.

After adding the trained network's intuition to the eight-player configuration and using the turn-based Q&A format, gpt-oss as crew won 8 of 16 rounds, almost indistinguishable from the network baseline (p = 0.54).

In the strictly comparable experiment with fifteen players and three impostors in the sixth league, the trained network's crew win rate over 300 rounds was 93.3%, while gpt-oss under the dialogue+intuition format won 16 of 20 rounds (80%), with a one-sided Fisher exact test p = 0.052.

In the sixth-league fifteen-player rounds, meetings using gpt-oss took an average of 524 seconds of wall-clock time, with almost all the time spent on turn-based speaking across about 4 meetings, calling the model once per player per round; the trained network's voting requires no model calls.

Across the six configurations, the language model was clearly worse than the network only in the eight-player no-intuition configuration, and among the five with intuition it tied once and lagged four times, but each gap on its own was too small or had confounders and could not be conclusive; in none of the configurations did it perform better.

The author believes the trained network was repeatedly calibrated over thousands of rounds in the same simulator, so its suspicion scores are highly aligned with who ultimately turns out to be the impostor, whereas the general knowledge brought by the language model offers no advantage in this environment, because the impostors' speech has no obvious anomaly and the only way to see through it is to check statements against who could see that room, but no results show the model can reliably perform that check.

The only real advantage of 120B over 20B is more reliable orchestration, but even this did not solve the deeper reliability problem it was supposed to solve.

After downgrading the coordinator to gpt-oss:20b (about 14GB), about 60GB of memory is left free, enough to perform coordination and image generation simultaneously, with no swapping and no daemon rejections.