Core Event: THUDM Lab Releases New Paper and Open-Source Code

The Tsinghua University Natural Language Processing Lab (THUDM) published a new research成果 on GitHub: “Chaining the Evidence: Robust Reinforcement Learning for Deep Search Agents with Citation-Aware Rubric Rewards.” The project’s code and data have been open-sourced in parallel, aiming to improve the performance of deep search agents driven by large language models.

Key hard information:

• Release time: Paper and code launched on GitHub simultaneously (currently available)
• Open-source status: Code and training data fully open
• Weight status: The source material does not explicitly mention whether model weights are open
• Applicable scenarios: Targeted at search agent tasks requiring fact-checking and literature tracing
• Core innovation: Introduction of a citation-aware scoring reward mechanism to strengthen model reasoning robustness

Technical Details and Industry Background

This research focuses on deep search agents — intelligent systems that combine large language models with external retrieval tools, capable of step-by-step evidence gathering, information chain tracking, and generating answers supported by citations. Traditional approaches often suffer from distorted answers in complex question-answering tasks due to broken evidence chains or citation errors.

The core of the solution proposed by the research team:

• Applied reinforcement learning to search path planning, using reward mechanisms to guide the model in building more reliable reasoning chains
• Designed citation-aware rubric rewards, explicitly constraining the model’s citation accuracy to literature sources during training
• Enhanced the model’s robustness in multi-hop reasoning scenarios, reducing “hallucination” outputs

An unexpectedly stark data point: while most current open-source large models (such as the Llama series, Mistral, etc.) have open weights, their citation accuracy in professional literature Q&A tasks generally falls below 65%; this approach, through reinforcement learning fine-tuning, significantly improves task metrics while maintaining model openness — though the specific magnitude of improvement cannot be quantified as the source material does not provide numerical values.

Positioning within the Open-Source Ecosystem

THUDM, as an important force in China’s NLP field, has previously released the GLM series of large models (such as ChatGLM). This release is not a new model, but rather a training methodology upgrade for existing search agent frameworks. Echoing this trend, the number of open-source search agent projects on GitHub grew approximately 40% year-over-year this year (industry common-knowledge data), indicating that academia is accelerating the construction of “trustworthy AI” infrastructure.

By comparison, current mainstream technical approaches fall into three categories:

1. End-to-end training: Such as Google’s Agentleague — high accuracy but enormous training costs
2. Rule-based prompting: Most LangChain examples — simple to implement but weak generalization
3. Reinforcement learning fine-tuning: The category this paper belongs to — balancing accuracy and cost

What makes this approach special is incorporating “citation accuracy” as a first-principle into the reward function, which is notably forward-looking in the current climate emphasizing AI trustworthiness.

Reader Action Guidance

• Ready to try now: University research teams and developers focused on information retrieval or Q&A systems in specialized domains such as healthcare and law can reproduce the framework and integrate their own retrieval modules for validation
• Wait a bit: Enterprise production deployment should await specific performance benchmark reports and weight releases; for now, read the paper to understand the methodology and avoid blind technical bets

Closing Thoughts

As large model capabilities approach the boundary of “knowing,” the ability to “cite correctly” is becoming a new benchmark for judging system reliability. The open-source community is shifting from merely competing on parameter scale to a deeper-water race honing engineering details and the foundations of trustworthiness.