The authors analyze a procedure that checks each intermediate result and retries from the last retained state upon rejection, stopping after a fixed local budget.
Conditions for reliable reuse of intermediate results in LLM reasoning
Theoretical bounds and empirical tests show that retrying locally from retained states can be efficient if checking costs and error rates are managed.
Chinese Tech
Deyuan Liu · Zecheng Wang · Xuanyi Liu · Jungang Li · Yihao Hu · Mingjie Tang · +2 more
Westlake University · Wechat AI · Peking University · HKUST · Ant Group
Research Digest··3 min read
The authors study when large language models can reliably reuse intermediate reasoning results across attempts.
§