finding
active
finding:gemini-prefers-rlfr-completion-in-50-9-of-1000-pairs-508-vs-491-of-999Gemini prefers RLFR completion in 50.9% of 1000 pairs (508 vs 491 of 999)
Even stylistic distribution.
Source paper
extracted_from(2026) · Aaditya Vikram Prasad · Connor Watts · Jack Merullo · Dhruvil Gala +3
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Interpretation of KL asymmetry.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Specific result for Gemini 2.0 Flash in Experiment 1; lowest rate among tested models
- G2.5-FL has the highest overbid frequency among all agents.
- Best VS result in synthetic data generation for math, demonstrating downstream improvement through diversity
- RLFR policy matches base on standard benchmarks (e.g., MMLU 67.2 vs 69.2, GSM8K 73.0 vs 72.3)finding0.760No capability degradation.
- Ablation removing inline interventions.
- Shows RL reduces but does not eliminate unmonitored non-compliance
- G2.5-FL shows no phase adaptation in bidding intensity.
- G3-F achieved a win rate of 72.9% in the combined-comp1 98-game slice.