ScholarPulse 日报 2026-08-15

2026-08-15 学术简报：2 篇。AI科学家代理Faraday在科学论文复制任务中性能超越现有大模型。

今日速览

序号	标题	来源	日期	主题	推荐等级
1	Training AI Scientists to Replicate Research	arXiv	2026-08-13	RAG	高
2	Teach the Magnitude, Not the Direction: Verifier-Bounded Credit Assignment for Multi-Turn Multi-step LLM Agents	arXiv	2026-08-13	AI-Agent	高

重点论文与技术动态

1. Training AI Scientists to Replicate Research

来源：arXiv
日期：2026-08-13
作者/机构：Damon Falck, Samer Sabri, Anja Surina, Thom Foster, Anya Sims, Sam Devlin
主题标签：RAG, arXiv
推荐等级：高
分类：cs.LG, cs.AI

一句话结论

AI科学家代理Faraday在科学论文复制任务中性能超越现有大模型。

一段话：本研究开发了Replica可扩展任务空间和auto-generated rubric-based judge，用于系统化论文复制。通过微调27B参数的Faraday代理（作为“AI科学家”），其利用编码代理作为工具，在复制任务上显著优于Claude Opus 4.8和GPT-5.5。定性分析显示Faraday采用更科学的探索方法，为AI实现长期科学创新提供基础，无需复杂框架。

核心内容

构建Replica任务空间，支持高效、系统化的科学论文复制。
引入自动评分裁判，确保低噪声且与人类评估高度一致。
Faraday代理在复制任务上性能超越Claude Opus 4.8和GPT-5.5。

方法与数据

方法：基于Replica任务空间和自动评分裁判进行训练。
数据：摘要未明确。

价值判断

值得关注：AI代理在提升科学可复现性中的关键作用，增强研究可靠性。
可复用点：Replica框架和自动评分系统可推广至其他科学复制场景。
局限/待核查：摘要未明确。

摘要

The replicability of papers is a cornerstone of scientific knowledge, ensuring the reliability of existing results and providing a base for further experiments. The act of replication typically illuminates details that were previously underspecified, and thus requires similar hypothesis-driven exploration to open-ended research. In this work, we develop Replica, a scalable task space for paper replication. To provide reward signal, we introduce an auto-generated rubric-based judge that has low noise and agrees with human assessment of replication quality. We post-train Faraday, a 27B-parameter "AI Scientist" agent that leverages coding agents as tools, surpassing the performance of Claude Opus 4.8 and GPT-5.5 on held-out replication tasks. Qualitative analysis of individual rollouts reveals that Faraday adopts a more scientifically-principled approach. We believe that our results provide a stepping stone towards AI agents capable of long-horizon scientific innovation without requiring complex harnesses.

2. Teach the Magnitude, Not the Direction: Verifier-Bounded Credit Assignment for Multi-Turn Multi-step LLM Agents

来源：arXiv
日期：2026-08-13
作者/机构：Zechuan Wang, Siyuan Lu, Hongxuan Zhang, Linjian Mo, Chenyi Zhuang, Leilei Gan
主题标签：AI-Agent, arXiv
推荐等级：高
分类：cs.AI

一句话结论

CrEST框架通过将教师角色从决定更新方向转变为调节更新幅度，实现了密集信用分配而不牺牲验证器约束的上限。

核心内容

RLVR的轨迹级信用分配将异质性每轮结果混为单一奖励信号，导致性能瓶颈。
CrEST采用分层信用分配框架，保留验证器约束上限，整合自教师的密集令牌级信号。
实验在BFCL V3和WildToolBench上，CrEST跨两个模型规模一致优于RL和蒸馏基线，长轨迹和严格会话级指标提升显著。

方法与数据

CrEST通过轮段验证优势解决跨轮稀释，熵门控自教师调制优化轮内令牌贡献。
实验数据集为BFCL V3和WildToolBench。

价值判断

值得关注：有效解决多轮工具使用代理训练中的信用分配异质性问题。
可复用点：框架可迁移至其他强化学习任务以实现密集监督。
局限/待核查：摘要未明确。

摘要

Reinforcement learning with verifiable rewards (RLVR) offers a verifier-bounded performance ceiling for training multi-turn tool-use agents, yet its trajectory-level credit assignment conflates heterogeneous per-turn outcomes into a single reward signal. On-policy distillation provides dense per-token supervision but is either teacher-bounded or prone to gradient concentration collapse. We introduce $\textbf{CrEST}$, a hierarchical credit assignment framework that retains RL's verifier-bounded ceiling while incorporating dense token-level signals from a privileged self-teacher. $\textbf{CrEST}$ resolves credit at two levels: turn-segmented verified advantages address inter-turn dilution, while entropy-gated self-teacher modulation refines intra-turn token contributions. Experiments on BFCL V3 and WildToolBench show that $\textbf{CrEST}$ consistently outperforms both RL and distillation baselines across two model scales, with the largest gains on long-trajectory and strict session-level metrics. Our work demonstrates that the teacher's role in policy optimization can be reduced from determining update directions to modulating update magnitudes, unlocking dense credit assignment without sacrificing the verifier-bounded ceiling.