Christopher D. Manning

One Frozen Simulator Is Not Enough: Simulator Collapse in Multi-Agent RL

单个冻结模拟器不够：多智能体强化学习中的模拟器崩溃问题

Simon Yu, Nicholas Tomlin, Marwa Abdulhai, Ximing Lu, Derek Chong, Abe Hou, Dilara Soylu, Sergey Levine, Christopher D. Manning, Weiyan Shi

机构 * Northeastern University（东北大学） ； New York University（纽约大学） ； UC Berkeley（加州大学伯克利分校） ； University of Washington（华盛顿大学） ； Stanford University（斯坦福大学） 纠错

AI总结 针对人机交互多智能体强化学习中单个LLM模拟器导致的策略泛化缺陷，提出Verbalized Sampling和Co-Training两种方案，在多轮基准测试和真实用户研究中显著提升了性能，发布了开源框架SCOPE。

Comments 41 pages, 28 figures

Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity

言语化采样：如何减轻模式坍塌并释放大语言模型的多样性

Jiayi Zhang, Simon Yu, Derek Chong, Anthony Sicilia, Michael R. Tomz, Christopher D. Manning, Weiyan Shi

AI总结 研究发现偏好数据中的典型性偏差致大语言模型模式坍塌，引入言语化采样策略。该策略能促使模型说出响应概率分布，经实验验证可显著提升创意写作等多方面性能，为模式坍塌提供新视角及实用补救方法。

Comments 83 pages, 31 figures, 44 tables. Code is available at https://github.com/CHATS-lab/verbalize-sampling

Shepherd: Enabling Programmable Meta-Agents via Reversible Agentic Execution Traces

Shepherd: 一个为元代理提供形式化执行迹的运行时基座

Simon Yu, Derek Chong, Ananjan Nandi, Dilara Soylu, Jiuding Sun, Christopher D Manning, Weiyan Shi

机构 * Northeastern University（东北大学） ； Stanford University（斯坦福大学） 纠错

AI总结 提出Shepherd，一个基于函数式编程的Python运行时基座，将代理执行作为一等对象，通过类似Git的执行迹支持元代理的检查、分叉和重放，在三个用例中显著提升性能。

Comments 50 pages, 22 figures, 14 tables

SecureForge: Finding and Preventing Vulnerabilities in LLM-Generated Code via Prompt Optimization

SecureForge：通过提示优化发现并防止LLM生成代码中的漏洞

Houjun Liu, Lisa Einstein, John Yang, Joachim Baumann, Duncan Eddy, Christopher D. Manning, Mykel Kochenderfer, Diyi Yang

机构 * Stanford University（斯坦福大学） 纠错

AI总结 本文提出SecureForge框架，通过自动化流程发现并减少LLM生成代码中的安全漏洞，提升代码安全性的同时保持单元测试性能，实验显示漏洞减少达48%。

Humans and transformer LMs: Abstraction drives language learning

人类与变换器语言模型：抽象驱动语言学习

Jasper Jian, Christopher D. Manning

机构 * Stanford University（斯坦福大学） 纠错

AI总结 研究比较变换器语言模型在训练过程中对语言类别的学习行为，发现抽象层面的行为早于词汇层面，揭示抽象对语言学习的重要性。

Comments EACL 2026

WebDS: An End-to-End Benchmark for Web-based Data Science

WebDS: 一种端到端的基于网络的数据科学基准

Ethan Hsu, Hong Meng Yam, Ines Bouissou, Aaron Murali John, Raj Thota, Josh Koe, Vivek Sarath Putta, G K Dharesan, Alexander Spangher, Shikhar Murty, Tenghao Huang, Christopher D. Manning

机构 * Stanford University（斯坦福大学） ； Pinetree Research（Pinetree研究公司） ； University of California, Berkeley（加州大学伯克利分校） ； Singapore University of Technology and Design（新加坡科技设计大学） ； University of Southern California（南加州大学） 纠错

AI总结 WebDS提出了一种端到端的基于网络的数据科学基准，旨在评估代理在复杂多步骤任务中的表现，揭示当前LLM在实际数据科学任务中的性能差距。

Comments 14 pages, ICLR 2026

DistillKac: Few-Step Image Generation via Damped Wave Equations

DistillKac：通过阻尼波动方程实现少步图像生成

Weiqiao Han, Chenlin Meng, Christopher D. Manning, Stefano Ermon

机构 * MIT（麻省理工学院） ； Stanford（斯坦福大学） 纠错

AI总结 DistillKac通过阻尼波动方程和随机Kac表示实现少步高质量图像生成，结合无分类指导和终点蒸馏方法，在保持数值稳定性的同时提升生成质量。

Comments Accepted to ICLR 2026

Thoughtbubbles: an Unsupervised Method for Parallel Thinking in Latent Space

Thoughtbubbles: 一种用于潜在空间中并行思维的无监督方法

Houjun Liu, Shikhar Murty, Christopher D. Manning, Róbert Csordás

机构 * Department of Computer Science, Stanford University, Stanford, CA, United States（计算机科学系，斯坦福大学，斯坦福，CA，美国） 纠错

AI总结 Thoughtbubbles是一种通过预训练学习并行自适应计算的转换器变体，能够在潜在空间中实现更高效的计算，提升模型的推理性能和零样本评估表现。

Stronger Baselines for Retrieval-Augmented Generation with Long-Context Language Models

更强的检索增强生成基线：长上下文语言模型

Alex Laitenberger, Christopher D. Manning, Nelson F. Liu

机构 * Stanford University（斯坦福大学） 纠错

AI总结 本文提出DOS RAG作为长上下文问答任务的强基线，通过保持文档结构和简单性，在多个基准上超越复杂方法。

Comments 11 pages, 6 figures, for associated source code, see https://github.com/alex-laitenberger/stronger-baselines-rag

Journal ref Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP 2025), pages 32559-32569

Mechanisms vs. Outcomes: Probing for Syntax Fails to Explain Performance on Targeted Syntactic Evaluations

Ananth Agarwal, Jasper Jian, Christopher D. Manning, Shikhar Murty

机构 * Stanford University（斯坦福大学） 纠错

Journal ref Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing

Do Language Models Use Their Depth Efficiently?

Róbert Csordás, Christopher D. Manning, Christopher Potts

机构 * Stanford University（斯坦福大学） 纠错

Comments Accepted to NeurIPS 2025

A Multimodal Benchmark for Framing of Oil & Gas Advertising and Potential Greenwashing Detection

Gaku Morio, Harri Rowlands, Dominik Stammbach, Christopher D. Manning, Peter Henderson

机构 * Hitachi, Ltd.（日立公司） ； Stanford University（斯坦福大学） ； Centre for the Acceleration of Social Technology（社会技术加速中心） ； Princeton University（普林斯顿大学） 纠错

Comments Forthcoming in NeurIPS 2025 Datasets and Benchmarks Track

Transcribe, Translate, or Transliterate: An Investigation of Intermediate Representations in Spoken Language Models

Tolúlopé Ògúnrèmí, Christopher D. Manning, Dan Jurafsky, Karen Livescu

机构 * Stanford University（斯坦福大学） ； Toyota Technological Institute at Chicago（芝加哥丰田技术研究所） 纠错

Comments ASRU 2025

Tversky Neural Networks: Psychologically Plausible Deep Learning with Differentiable Tversky Similarity

Moussa Koulako Bala Doumbouya, Dan Jurafsky, Christopher D. Manning

机构 * Department of Computer Science, 353 Jane Stanford Way（计算机科学系） 纠错

A New Pair of GloVes

Riley Carlson, John Bauer, Christopher D. Manning

机构 * Stanford NLP Group（斯坦福大学NLP组） ； Stanford University（斯坦福大学） 纠错

Drop Dropout on Single-Epoch Language Model Pretraining

Houjun Liu, John Bauer, Christopher D. Manning

机构 * Stanford University（斯坦福大学） 纠错

Comments Accepted to ACL Findings; 5 pages, 2 figures, 4 pages of appendix

Improved Representation Steering for Language Models

Zhengxuan Wu, Qinan Yu, Aryaman Arora, Christopher D. Manning, Christopher Potts

机构 * Stanford University（斯坦福大学） 纠错

Comments 46 pages, 23 figures, preprint

Osiris: A Lightweight Open-Source Hallucination Detection System

Alex Shan, John Bauer, Christopher D. Manning

机构 * Stanford University（斯坦福大学） ； Stanford HAI（斯坦福HAI） 纠错

Comments Stanford 191W

A Reasoning-Focused Legal Retrieval Benchmark

Lucia Zheng, Neel Guha, Javokhir Arifov, Sarah Zhang, Michal Skreta, Christopher D. Manning, Peter Henderson, Daniel E. Ho

机构 * Stanford University（斯坦福大学） ； Princeton University（普林斯顿大学） 纠错

Comments CS&Law 2025. For data, see https://reglab.github.io/legal-rag-benchmarks/

Synthetic Data Generation & Multi-Step RL for Reasoning & Tool Use

Anna Goldie, Azalia Mirhoseini, Hao Zhou, Irene Cai, Christopher D. Manning

机构 * Department of Computer Science Stanford University（计算机科学系 斯坦福大学） 纠错

AuroraCap: Efficient, Performant Video Detailed Captioning and a New Benchmark

Wenhao Chai, Enxin Song, Yilun Du, Chenlin Meng, Vashisht Madhavan, Omer Bar-Tal, Jenq-Neng Hwang, Saining Xie, Christopher D. Manning

Comments Accepted to ICLR 2025. Code, docs, weight, benchmark and training data are all avaliable at https://rese1f.github.io/aurora-web/

MrT5: Dynamic Token Merging for Efficient Byte-level Language Models

Julie Kallini, Shikhar Murty, Christopher D. Manning, Christopher Potts, Róbert Csordás

h4rm3l: A language for Composable Jailbreak Attack Synthesis

Moussa Koulako Bala Doumbouya, Ananjan Nandi, Gabriel Poesia, Davide Ghilardi, Anna Goldie, Federico Bianchi, Dan Jurafsky, Christopher D. Manning

Comments Accepted to the Thirteenth International Conference on Learning Representations (ICLR 2025)

Sneaking Syntax into Transformer Language Models with Tree Regularization

Ananjan Nandi, Christopher D. Manning, Shikhar Murty

Comments Accepted to NAACL 2025

AI for Scaling Legal Reform: Mapping and Redacting Racial Covenants in Santa Clara County

Faiz Surani, Mirac Suzgun, Vyoma Raman, Christopher D. Manning, Peter Henderson, Daniel E. Ho

Comments https://reglab.github.io/racialcovenants/

AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders

Zhengxuan Wu, Aryaman Arora, Atticus Geiger, Zheng Wang, Jing Huang, Dan Jurafsky, Christopher D. Manning, Christopher Potts

NNetNav: Unsupervised Learning of Browser Agents Through Environment Interaction in the Wild

Shikhar Murty, Hao Zhu, Dzmitry Bahdanau, Christopher D. Manning

Comments Code, Data and Models available at https://www.nnetnav.dev

LawInstruct: A Resource for Studying Language Model Adaptation to the Legal Domain

Joel Niklaus, Lucia Zheng, Arya D. McCarthy, Christopher Hahn, Brian M. Rosen, Peter Henderson, Daniel E. Ho, Garrett Honke, Percy Liang, Christopher Manning

Comments Accepted at Findings of NAACL 2025

MoEUT: Mixture-of-Experts Universal Transformers

Róbert Csordás, Kazuki Irie, Jürgen Schmidhuber, Christopher Potts, Christopher D. Manning

Comments Accepted to NeurIPS 2024

Instruction Following without Instruction Tuning

John Hewitt, Nelson F. Liu, Percy Liang, Christopher D. Manning