OPEN LINK ↗
// 1h agoRESEARCH PAPER
Spurious CoT Termination Exposes Early-Exit Flaw
KAIST researchers show that injecting an end-of-think token can fail to switch reasoning models cleanly into answer generation, causing them to continue reasoning before emitting another termination token. Their Exit-Token Attention Biasing method reduces this behavior across four models, five benchmarks, and two early-exit methods. [Paper](https://arxiv.org/abs/2609.03633)
// ANALYSIS
This research challenges the assumption that matching a model’s visible think-block format is enough to control its internal generation state.
- –Early exit can create misleading token savings when reasoning resurfaces in the supposed answer phase.
- –The issue appears across multiple models and evaluation settings, making it relevant to inference optimization broadly.
- –EAB offers a lightweight inference-time intervention that increases attention to the injected termination token without retraining.
- –Developers should measure answer-phase reasoning and regenerated EoT tokens, not just total output length.
- –The released implementation makes the findings practical to reproduce and test. [Code](https://github.com/Seunghee-Koh/Spurious-CoT-Termination)
// TAGS
spurious-cot-terminationreasoningllminferenceevaluationresearch
DISCOVERED
1h ago
2026-09-07
PUBLISHED
1h ago
2026-09-07
RELEVANCE
9/ 10
AUTHOR
Discover AI