论文：验证器约束信用分配提升多轮Agent训练
This paper shows a better way to train multi-turn agents:
针对多轮Agent训练痛点给出具体可复现的算法改进，有明确基准对比，值得关注Agent训练范式的研究者阅读。
This paper shows a better way to train multi-turn agents:
本文展示了一种更好的多轮智能体训练方法：
score each turn separately, then use a self-teacher to focus learning without letting it override the reward.
分别对每一轮进行评分，然后使用自教师机制聚焦学习，同时防止其覆盖奖励信号。
Standard RL has a basic problem.
标准强化学习存在一个基本问题。
A long agent session can contain successful and failed turns, yet 1 overall reward can blur them together.
一次长周期的智能体会话可能包含成功和失败的轮次，但单一的总奖励会将它们混为一谈。
CREST fixes that by giving each turn its own verified credit, then using the same model as a teacher to put more learning weight on uncertain decisions inside that turn.
CREST 通过为每一轮分配独立的验证信用来解决这一问题，随后利用同一模型作为教师，在该轮内对不确定的决策赋予更高的学习权重。
The teacher can strengthen an update, but it cannot reverse the verifier's judgment.
教师可以增强更新效果，但不能推翻验证器的判断。
On Qwen3-4B-Instruct, it reaches 52.0% average BFCL V3 accuracy versus 49.25% for the strongest RL baseline.
在 Qwen3-4B-Instruct 上，其平均 BFCL V3 准确率达到 52.0%，而最强的 RL 基线仅为 49.25%。
– arxiv. org/abs/2608.13179
– arxiv.org/abs/2608.13179
Title: "Teach the Magnitude, Not the Direction: Verifier-Bounded Credit Assignment for Multi-Turn Multi-step LLM Agents"
标题：《教授幅度而非方向：面向多轮多步 LLM 智能体的验证器边界信用分配》
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力