According to God of Prompt on Twitter, a new paper from Beihang University and ByteDance finds that overthinking in reasoning models like DeepSeek R1 and Qwen3 stems from sampling, not training, and a stopping-aware decoding method reduces token usage by 44% while improving accuracy; as reported by the tweet, this implies businesses can lower inference costs and latency without retraining by adapting sampling to let models stop when confident.

According to God of Prompt on Twitter, a new paper by Beihang University and ByteDance finds that overthinking in reasoning models like DeepSeek R1 and Qwen3 stems from sampling, not training, and a revised stopping strategy reduces token usage by 44% while improving accuracy. As reported by the tweet, the method lets models stop when internal signals indicate solution completion, addressing inefficiencies in long-chain reasoning and enabling faster, cheaper inference. According to the authors cited by the tweet, the approach offers immediate business impact for LLM ops by lowering compute costs, stabilizing latency, and boosting win rates on reasoning benchmarks.