Reasoning gaps between base LLMs and LRMs concentrate on ~8% of early planning tokens; intervening with the reasoning model only at high-disagreement positions recovers performance.
All evaluations are conducted on a single compute node equipped with an NVIDIA A100 GPU
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.AI 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Reasoning Can Be Restored by Correcting a Few Decision Tokens
Reasoning gaps between base LLMs and LRMs concentrate on ~8% of early planning tokens; intervening with the reasoning model only at high-disagreement positions recovers performance.