Amplified Does Not Mean Predictive: Reasoning Behaviors in Thinking Models

• 类型：arxiv
• 标识：2608.13760
• 链接：https://arxiv.org/abs/2608.13760
• 主分类：engineering
• 形态：benchmark
• 被引：0
• 被引来源：Semantic Scholar
• S2被引：0
• 影响力被引：0
• TLDR：It is found that reasoning-oriented training does not preferentially amplify the highest-Lift behaviors, motivating process-level objectives that reward calibrated and grounded reasoning rather than surface form alone.
• 待LLM分类：否
• 标题中文：增强并不意味着可预测：思维模型中的推理行为
• TLDR中文：发现面向推理的训练并未优先放大具有最高 Lift 的行为，这促使研究者采用过程级目标，以奖励经过校准且有依据的推理，而不仅仅是表面形式。
• 来源文件：
• /inbox/tom/_candidates/2026-08-18-rag-retrieval-reranking-candidates.json
• /inbox/tom/_candidates/2026-08-18-agent-rag-longcontext-candidates.json
• [S2 enrich]