AI

Same Beliefs, Different Policies: Learning Reusable Representations for Early Stopping in LLM Reasoning

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 23/10/2026, 04:03:25 UTC.

Briefings

BELT (BELief-based Termination) is a new method that separates outcome prediction from stopping control in LLMs. It learns beliefs about future answers from a frozen LLM's hidden states and fits a linear stopping policy on these features. This approach reduces generated tokens by 41% on Qwen3-1.7B and 51% on OLMo-3-7B-Think without losing accuracy, and transfers to coding and agentic tasks without additional training. Performance on OLMo3 is competitive with or better than end-to-end controllers.

Key points

• BELT learns beliefs about current and future answers from a frozen LLM's hidden states to determine when to stop reasoning.
• BELT reduces generated tokens by 41% on Qwen3-1.7B and 51% on OLMo-3-7B-Think without loss of accuracy.

Context

Reasoning LLMs often continue reasoning after reaching a correct answer, wasting computation. BELT addresses this by learning beliefs about future answers from a frozen LLM's hidden states to determine when to stop reasoning.

Limitations

Evidence is limited to five out-of-distribution math benchmarks; performance on coding and agentic tasks is supported but no specific metrics are provided. The method requires CPU policy refitting or threshold adjustment for different token prices, and probe parameters are significantly smaller than baselines.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 23/10/2026, 04:11:28 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeicf3ocawrb35vh7u5vxmaktvdyfo43okivdmho5h7dgn4rvoizlwy