TBNAI
When Reasoning Traces Conflict with Evidence: How Answer Configuration Shapes Evidence Following
Publication: cs.AI updates on arXiv.org
Author: Shaoang Li, Yanhang Shi, Jian Li
Source published: 16/10/2026, 04:00:00 UTC
Briefing created: 16/10/2026, 04:46:01 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 16/10/2026, 04:03:25 UTC.
Briefings
Research on answer configuration shows that separate answer calls significantly improve evidence-following accuracy compared to continuing the model's own reasoning trace. Across three model families on HotpotQA and MuSiQue, separate calls raised evidence-following by 29.9 to 60.6 points while removing traces lowered accuracy by similar amounts. The final line of a trace carries most of its value, and gold evidence helps most before the trace is written. The study concludes that trace deletion alone cannot reveal which source an answer follows, so evaluations should report both properties together with the answer configuration.
Key points
- Separate answer calls raise evidence-following accuracy by 29.9 to 60.6 points compared to continuing the model's own reasoning trace across three model families.
- The final line of a reasoning trace carries most of its value, and gold evidence helps most before the trace is written.
Context
This study investigates how answer configuration affects evidence following in reasoning traces. The findings matter because they reveal that separate answer calls improve accuracy significantly more than continuing the model's own thinking, suggesting distinct properties of trace usage.
Limitations
Evidence is limited to HotpotQA and MuSiQue datasets. Results are from a preprint (arXiv:2610.20805v1) and remain subject to verification. The study does not report causal claims or opinions, only observed correlations between configurations and accuracy.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 16/10/2026, 04:46:01 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeiamtvhyhch3bdgkvy4albilos6exo4g5u7rce5234rczjmt7vpowu