New papers propose methods to curb overthinking in LLM reasoning

Arxiv CS.CL (NLP/language models)

SuCo, CoRA, and SuperThoughts are among several new works addressing excessive chain-of-thought computation in large reasoning models. Approaches include adaptive generation, confidence-rationale alignment, and continuous latent reasoning to reduce token overhead.

ReasoningChain-of-thoughtLLM

See 4 more sources

CoRA: Confidence-Rationale Alignment for Reliable Chain-of-Thought Reasoning2 days agoJuming Xiong, Weixin Liu, Kevin Guo, Congning Ni, Junchao Zhu, Chongyu Qu, Chao Yan, Katherine Brown, Avinash Baidya, Xiang Gao, Bradley Malin, Zhijun Yin

SuperThoughts: Reasoning Tokens in Superposition3 days agoZheyang Xiong, Shivam Garg, Max Yu, Vaishnavi Shrivastava, Haoyu Zhao, Anastasios Kyrillidis, Dimitris Papailiopoulos

SuCo: Sufficiency-guided Continuous Adaptive Reasoning4 hours agoJiahao Wang, Bingyu Liang, Chenhao Hu, Longhui Zhang, Xuebo Liu, Min zhang, Jing Li, Xuelong Li

Stop When Further Reasoning Won't Help: Attention-State Adaptive Generation in Reasoning Models2 days agoJiakai Li, Ke Qin, Rongzheng Wang, Yizhuo Ma, Qizhi Chen, Muquan Li, Shuang Liang

Arxiv CS.CL (NLP/language models)·Junlong TongWenqi XuYingqi FanAnhao ZhaoXuan LuYang TanXiaoyu Shen

3 days ago