{"version":1,"type":"story","url":"https://digestai.news/story/belt-reduces-llm-reasoning-tokens-by-41-on-qwen3-1-7b","json":"https://digestai.news/story/belt-reduces-llm-reasoning-tokens-by-41-on-qwen3-1-7b.json","markdown":"https://digestai.news/story/belt-reduces-llm-reasoning-tokens-by-41-on-qwen3-1-7b.md","slug":"belt-reduces-llm-reasoning-tokens-by-41-on-qwen3-1-7b","headline":"BELT reduces LLM reasoning tokens by 41% on Qwen3-1.7B","summary":"Researchers introduced BELT (BELief-based Termination), a method designed to stop Large Language Models from continuing to reason after they have already reached a correct answer. Current reasoning models often waste computation by generating extra tokens, which can sometimes lead to incorrect final outputs. BELT addresses this by separating the prediction of the answer from the decision of when to stop.\n\nThe system learns beliefs about current and future answers from a frozen LLM's hidden states. It then fits a linear stopping policy on these frozen features. Because the prediction targets do not depend on the specific cost of a token, a single probe can support different accuracy-cost trade-offs by simply adjusting thresholds or refitting the policy on CPU. This eliminates the need for additional controller training for new token prices.\n\nIn tests across five out-of-distribution math benchmarks, BELT reduced generated tokens by 41% on Qwen3-1.7B and 51% on OLMo-3-7B-Think without losing accuracy. The method also transferred to coding and agentic tasks without specific training data for those domains. On OLMo3, the BELT probe used 21.2x fewer parameters and 55x fewer training GPU-hours than the OS-Pruner baseline.","keyPoints":["BELT reduces generated tokens by 41% on Qwen3-1.7B and 51% on OLMo-3-7B-Think without accuracy loss.","The method uses a linear stopping policy on frozen features, requiring no new controller training for different token prices.","BELT transfers to coding and agentic tasks without domain-specific training, outperforming end-to-end controllers in efficiency."],"whyItMatters":"This research offers a practical way to cut inference costs for reasoning models. By stopping generation earlier, developers can reduce compute expenses and latency while maintaining accuracy, making complex reasoning tasks more viable for production use.","category":{"slug":"research","name":"Research","url":"https://digestai.news/category/research"},"entities":{"companies":[],"models":["Qwen3-1.7B","OLMo-3-7B-Think","OLMo3","OS-Pruner"],"people":[]},"firstPublishedAt":"2026-10-23T04:00:00Z","updatedAt":"2026-10-23T04:00:00Z","sourceCount":1,"hasPrimarySource":true,"sources":[{"outlet":"arXiv cs.AI","title":"Same Beliefs, Different Policies: Learning Reusable Representations for Early Stopping in LLM Reasoning","url":"https://arxiv.org/abs/2610.30450","publishedAt":"2026-10-23T04:00:00Z","type":"primary","primary":true,"lead":true}],"sourceNotes":null,"discussions":[],"thread":{"title":"Uncovering Hidden Flaws in AI Reasoning","url":"https://digestai.news/thread/researchers-find-qwen3s-reasoning-gains-may-stem-from-thinking-leakage","storyCount":3},"cite":{"text":"Digest AI, \"BELT reduces LLM reasoning tokens by 41% on Qwen3-1.7B\", 23 October 2026, https://digestai.news/story/belt-reduces-llm-reasoning-tokens-by-41-on-qwen3-1-7b","publisher":"Digest AI","title":"BELT reduces LLM reasoning tokens by 41% on Qwen3-1.7B","datePublished":"2026-10-23T04:00:00Z","url":"https://digestai.news/story/belt-reduces-llm-reasoning-tokens-by-41-on-qwen3-1-7b"},"generatedBy":"Written by Digest AI's editorial model from the linked sources; the sources are the record.","license":"Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse"}