Gemini 3 Flash wins 72.9% of 98 canonical games

G3-F achieved a win rate of 72.9% in the combined-comp1 98-game slice.

Source paper

extracted_from

(2026) · Robert Müller · Clemens Müller

claim

Two heuristic code agents outperform most tested LLMs
supports
author assertion that deterministic heuristics surpass many LLMs

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Gemini 3 Flash median final score 5,250 in 98 canonical gamesfinding0.899
G3-F's median score across games was 5,250 points.
Gemini 3 Flash wins 67.9% of its 28 mixed-format games against code agentsfinding0.895
In the 172-game exp2 slice, G3-F has the highest LLM win rate against deterministic baselines.
Gemini 3 Flash TrueSkill µ = 30.1 ± 3.3 in 98 canonical gamesfinding0.851
The Bayesian skill posterior mean for G3-F is 30.1 with 3σ confidence interval ±3.3.
Gemini 3 Flash averages 3.96 quartets per gamefinding0.845
G3-F completed on average 3.96 quartets per game.
G3-F win rate=72.9% in 98 canonical gamesfinding0.797
Gemini 3 Flash won nearly 3/4 of its games
Gemini 2.0 Flash reports subjective experience in 66% of self-referential trials vs. 0% in all control conditionsfinding0.783
Specific result for Gemini 2.0 Flash in Experiment 1; lowest rate among tested models
Gemini 2.5 Flash Lite overbid rate 1.20%finding0.778
G2.5-FL has the highest overbid frequency among all agents.
Gemini 3 Flash bid aggressiveness ramp from 0.26 early to 2.49 latefinding0.777
G3-F escalates bidding by nearly 10× from early to late game, adapting to phase.