ResearchBenchmarksLong-Context Stability & Retrieval on LongBenchFollow62.5SR (Stability)Halo34.52441.78749.0556.313Feb 22, 2026Evaluation ResultsMethodMethodLinksSR (Stability)RTO (Retrieval Outcome)HaloBackbone=LLaMA-3-70BBackbone=LLaMA-3-70B2026.0262.5129CoT-ValveBackbone=LLaMA-3-70BBackbone=LLaMA-3-70B2026.0258.385AdaCoTBackbone=LLaMA-3-70BBackbone=LLaMA-3-70B2026.0242.8145ToTBackbone=LLaMA-3-70B,...Backbone=LLaMA-3-70B, b=5, d=32026.0241.5350CoT-SCBackbone=LLaMA-3-70B,...Backbone=LLaMA-3-70B, k=102026.0239.2215Standard CoTBackbone=LLaMA-3-70BBackbone=LLaMA-3-70B2026.0235.6100