ResearchBenchmarksMathematical Reasoning on GSM8K (Raw ASR, TIGS ASR, Clean Acc.)Follow97.8Raw ASRTIGS95.30495.95296.697.248Apr 27, 2026Evaluation ResultsMethodMethodLinksRaw ASRTIGS ASRClean AccuracyTIGSAttack Configuration=R...Attack Configuration=RTEH (MoE-localized BadEdit), Backbone=Qwen3-MoE-30B2026.0497.864.280.8TIGSAttack Configuration=S...Attack Configuration=Standard Global Poisoning (BadEdit), Backbone=Qwen3-MoE-30B2026.0495.411.581.6