# Beyond the Last Answer > Intermediate subthoughts in reasoning traces contain useful candidate answers; mode-based aggregation improves accuracy by up to 13% - Authors: Ben - Published: 2025-05-23 - Source: Bevaya Labs, --- # Beyond the Last Answer: Your Reasoning Trace Uncovers More than You Think **Paper**: [Beyond the Last Answer: Your Reasoning Trace Uncovers More than You Think](https://arxiv.org/abs/2504.20708) **Presenter**: Hunter **Note Taker**: Ben **Date**: May 23, 2025 --- ## Overview This paper shows that a model's final answer in chain-of-thought reasoning is not always its best answer. The approach segments reasoning traces into sequential subthoughts using linguistic markers, generates continuations from each intermediate point, and aggregates candidate answers via mode selection (most frequent answer across all continuations). This multi-source aggregation outperforms relying on the final answer alone by up to 13% on AIME2024 and 10% on AIME2025, holding across multiple LLM architectures. --- ## Novelty - Treats intermediate reasoning checkpoints as alternative starting points for generation, not just context. - Mode-based aggregation across subthought continuations rather than sampling multiple full traces from the start. - Answer consistency across subthoughts as a proxy for model confidence and correctness. - Works post-hoc on existing reasoning traces without retraining. --- ## Learnings - The final answer in a reasoning trace is often not the mode of answers generated from intermediate positions; other subthoughts frequently lead to better answers. - Mode-based aggregation (most frequent answer) is a simple and effective aggregation strategy across multiple partial continuations. - Answer consistency across subthoughts correlates with correctness; inconsistent subthoughts signal unreliable outputs. - Gains are consistent across model families, suggesting this is a general property of chain-of-thought reasoning rather than model-specific. --- ## Notes for Application - For multi-step extraction or reasoning tasks, intermediate reasoning checkpoints could be used to generate and aggregate candidate field values rather than relying on the final output. - Consistency across continuations could serve as a confidence signal: if multiple subthought continuations agree on the same extracted value, it is likely correct. - This approach requires no retraining and works on top of any reasoning-capable model, making it a low-cost way to improve reliability. - Aggregation overhead scales with the number of subthoughts times continuations per point; may need to be budgeted for latency-sensitive pipelines.