ResearchBenchmarksLong-horizon evaluation on LMSYS-Chat-1M (144 trajectories > 50 turns)Follow3.67Overall MALEvoSpec3.39963.46983.543.6102Apr 17, 2026Evaluation ResultsMethodMethodLinksOverall MALThroughput (tok/s)EvoSpec2026.043.67168.6FR-SpecVocabulary adaptation...Vocabulary adaptation setting=Static 32k2026.043.41158.7