BELT reduces LLM reasoning tokens by 41% on Qwen3-1.7B
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…
Key points
- 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.
The 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.
In 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.
The story so far
3 episodes →- BELT reduces LLM reasoning tokens by 41% on Qwen3-1.7Bthis story
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- MIT historian says early AI engineers added randomness to mimic human creativity · 1 src
- Researchers introduce MemPlan for text-based planning with PDDL · 1 src
- Researchers present CiteAnchor framework for detecting hallucinated citations · 1 src
- Researchers find self‑rewarded moral argumentation makes language models decisive but pliable · 1 src
- AQL-Lite improves Arabic LLMs with curated data and partial-parameter training · 1 src
Comments
via GitHub Discussions