Stop When Further Reasoning Won't Help: Attention-State Adaptive Generation in Reasoning Models

A training-free method called ASAG can reduce token waste in large reasoning models by inferring when a model should stop generating, according to a paper posted to arXiv on 13 June 2026 [1]. The approach adaptively adjusts generation strategy based on the model’s attention state, improving both efficiency and accuracy [1]. Large reasoning models (LRMs) use test-time compute scaling and explicit chain-of-thought processes to solve complex problems, but they frequently suffer from “overthinking,” producing redundant tokens that degrade accuracy [1]. Existing remedies are split between resource-intensive training-based methods and training-free approaches that depend on carefully engineered prompts or unreliable confidence signals [1]. The new framework, Attention-State Adaptive Generation (ASAG), takes a different path by examining attention distributions to infer the model’s reasoning state and decide when further reasoning is unlikely to help [1]. The method is plug-and-play, requiring no retraining, and can be integrated into existing LRMs without architectural changes [1]. In experiments across nine benchmarks, ASAG delivered consistent gains on mainstream LRMs at different parameter scales, including the DeepSeek-R1-Distill and Qwen3 series [1]. On Qwen3-8B, the method improved average accuracy by 3.2% while cutting the number of generated tokens by nearly 40% across all reasoning tasks [1]. The dual improvement in accuracy and efficiency addresses a known tension in the field: scaling test-time compute can boost performance but often at the cost of excessive computation and verbosity [1]. The paper arrives amid a broader push to make AI systems more capable and efficient. Since the introduction of the transformer architecture in 2017, large language models have driven an investment boom and rapid public releases of systems such as ChatGPT [3][5]. Researchers and companies including OpenAI, Google, and Meta have stated goals of building artificial general intelligence (AGI), a hypothetical AI that matches or surpasses human abilities across virtually all cognitive tasks [4]. A 2020 survey identified 72 active AGI research and development projects across 37 countries [4]. ASAG’s training-free design lowers the barrier to adoption compared with methods that require substantial computational resources for retraining [1]. The authors frame the work as an investigation of early stopping through attention distributions, a perspective that departs from prompt-based or confidence-based stopping rules [1]. The paper was submitted to arXiv’s Computation and Language section and is available with links to code and demos through Hugging Face’s paper-pages integration, which allows the community to discuss the work and link related models and datasets [9][10].