ResearchBenchmarksNatural Language Grounding and Discrimination on B-hard with hard negatives v4 (held-out)Follow98.6B1 ScoreLBM94.33695.44396.5597.657Jul 8, 2026Evaluation ResultsMethodMethodLinksB1 ScoreB2 ScoreB3 ScoreB4 ScoreAverage ScoreLBMTraining=GRPO v5 + SFT...Training=GRPO v5 + SFT-Lazada2026.0798.698.981.65282.8GPT-5.52026.0794.551.470.951.667.1