ChenHuajun@Zhejiang_University

Full professor of College of Computer Science of Zhejiang University (@ZJU_China), leading founder of in #AGI,#KG,#LLM,#AI4Science.

[ChenHuajun@Zhejiang_University @ChenHuajun]

TextGrad: Symbolic Gradient vs Neural Gradient — A symbolic feedback optimization mechanism and an improved approach for neural-symbolic system integration. It offers potential for optimizing multi-agent workflows involving various black-box modules. Combining it with agent protocols like MCP for enhanced multi-agent coordination could be promising.

Optimizing generative AI by backpropagating language model feedback | Nature
https://t.co/436gx5sBt7

[ChenHuajun@Zhejiang_University @ChenHuajun]

OntoTune: Our #WWW2025 paper leverages hierarchical knowledge to guide LLM tuning, enabling the generation of responses guided by the ontology such as SNOMED CT.

OntoTune: Ontology-Driven Self-training for Aligning Large Language Models.
https://t.co/ey3Oy1ScdM

[ChenHuajun@Zhejiang_University @ChenHuajun]

Human or LLM as Judger? Our #ICLR2025 paper, "SaMer: A Scenario-aware Multi-dimensional Evaluator for Large Language Models," introduce a fine-grained, scenario-adaptive evaluator that dynamically adjusts evaluation dimensions based on query context. As large models evolve, automated model-based evaluation, such as SaMer, will become increasingly vital, offering more scalable, consistent, and reliable assessments compared to traditional human evaluation.

[Physics In History @PhysInHistory]

Nikola Tesla once said, "It's the quality of thought, not quantity, that counts." ✍️

[Zhejiang University @ZJU_China]

#ZJU & #AntGroup Sign Strategic Cooperation Agreement 🤝

On Jan. 8, the two parties held a signing ceremony at Ant Group’s headquarters and officially launched the Zhejiang University-Ant Group Joint Research Center for Data and Intelligence.

This collaboration aims to drive innovation and talent development by combining ZJU’s cutting-edge research strengths with Ant Group’s industry expertise in real-world applications. Together, they will tackle challenges in AI, data science, and intelligent technologies, laying the groundwork for future breakthroughs in the digital economy. 💡

Photo: ZHE Ying | 浙大发布
#ZJUimpact

[ChenHuajun@Zhejiang_University @ChenHuajun]

Benchmarking Agentic Planning: Our #ICLR2025 paper "Benchmarking Agentic Workflow Generation" shows that LLMs struggle more with graph planning than sequential planning. We introduce WORFBENCH, an agentic workflow benchmark, and WORFEVAL, a corresponding evaluation protocol, to assess LLMs' graph planning abilities.
https://t.co/UcUL1wBXkw

[ChenHuajun@Zhejiang_University @ChenHuajun]

MoE for KG: our #ICLR2025 paper proposes Mixture of Modality Knowledge Experts (MoMoKE) to address the representation learning problem in multi-modal knowledge graphs.

https://t.co/Sd3eytn7SJ

[ChenHuajun@Zhejiang_University @ChenHuajun]

Our comprehensive survey on scientific large language models is now published in ACM Computing Surveys. Covering nearly 300 citations, it spans text, molecules, proteins, genomes, and single cells across biology, chemistry, materials, and medicine.
https://t.co/P42MIyTPjI

[ChenHuajun@Zhejiang_University @ChenHuajun]

Genes encode fundamental genetic information, forming the basis for RNA, proteins, and even complex biomolecular assemblies like CRISPR-Cas. With a gene foundation model such as Evo, a key question emerges: can genomic sequence signals alone enable the prediction and generative design of all these biological systems?

[Science Magazine @ScienceMagazine]

A new Science study presents “Evo”—a machine learning model capable of decoding and designing DNA, RNA, and protein sequences, from molecular to genome scale, with unparalleled accuracy.

Evo’s ability to predict, generate, and engineer entire genomic sequences could change the way synthetic biology is done. Learn more in this week's issue: https://t.co/rGWOLUsYZc

[Biology+AI Daily @BiologyAIDaily]

DePLM: Denoising Protein Language Models for Property Optimization

1. The study introduces DePLM, a cutting-edge framework that refines evolutionary information (EI) in protein language models (PLMs) to enhance property optimization by filtering out irrelevant noise.

2. DePLM employs a novel rank-based denoising diffusion process to isolate property-specific likelihoods, moving beyond traditional methods that often entangle multiple functional properties in their optimization.

3. By shifting the optimization objective from minimizing numerical errors to maximizing rank correlation, DePLM achieves robust generalization across novel proteins and diverse datasets.

4. Extensive benchmarks show DePLM outperforming state-of-the-art methods in mutation effect prediction, achieving superior results on datasets like ProteinGym and various protein fitness challenges.

5. The method integrates both sequence and structure-based representations using advanced attention mechanisms, allowing for comprehensive and expressive protein modeling.

6. DePLM’s modular architecture includes a feature encoder and denoising blocks, enabling fine-grained analysis of protein properties and adaptability to unseen sequences.

7. This approach has transformative potential in protein engineering, drug design, and understanding evolutionary dynamics, making it a versatile tool for biological research.

@ChenHuajun
📜Paper: https://t.co/g3yDY6EvXy

#ProteinEngineering #DeepLearning #DenoisingModels #Bioinformatics #AI #EvolutionaryBiology

[ChenHuajun@Zhejiang_University @ChenHuajun]

"Practice is the ultimate test of truth."

[Physics In History @PhysInHistory]

Richard Feynman at the Cornell University Lecture, 1964 says,

"It doesn't matter how beautiful your theory is, it doesn't matter how smart you are. If it doesn't agree with experiment, it's wrong. In that simple statement is the key to science."

[Physics In History @PhysInHistory]

"A good teacher does not teach facts, he or she teaches enthusiasm, open-mindedness and values."

- Gian-Carlo Rota

[ChenHuajun@Zhejiang_University @ChenHuajun]

How can we improve LLMs' step-wise reasoning and planning ability? Our #EMNLP2024 paper proposes a framework that, echoing O1's multi-step reasoning, enhances LLMs by leveraging knowledge graphs (KGs) to synthesize step-by-step instructions. Just as chain-of-thought reasoning breaks down tasks into logical steps, our framework extracts structured patterns from KGs to guide the decomposition of complex questions. This KG-driven approach boosts LLMs' capabilities to handle complex QA tasks with a more systematic, multi-step reasoning process.

https://t.co/YMHsIoVHo2

[Physics In History @PhysInHistory]

The world has changed far more in the past 100 years than in any other century in history. The reason is not political or economic but technological — technologies that flowed directly from advances in basic science.

- Stephen Hawking

[ChenHuajun@Zhejiang_University @ChenHuajun]

Start from Zero: Triple Set Prediction for Automatic Knowledge Graph Completion (KGC): In our #TKDE paper, we redefine the KGC task by introducing "Triple Set Level" completion. Unlike traditional methods that predict missing elements at the single-triple level, our approach works with a set of triples and autonomously identifies missing triples in one pass. This approach better aligns with real-world knowledge graph construction needs.
https://t.co/qwsVwMHFmI

[ChenHuajun@Zhejiang_University @ChenHuajun]

Our #NeurIPS2024 paper, 'DePLM: Denoising Protein Language Models for Property Optimization,' leverages the denoising process from diffusion models to filter noisy information from protein language models. This approach enables the model to focus more effectively on predicting specific protein properties, such as thermostability. #AI4Science
https://t.co/UtEFJMy0vr

[ChenHuajun@Zhejiang_University @ChenHuajun]

Even Hiton himself doesn't quite understand how someone working on neural networks ended up winning the Nobel Prize in Physics. 😂

[The Nobel Prize @NobelPrize]

“How could I be sure it wasn’t a spoof call?”

2024 physics laureate Geoffrey Hinton received a phone call from Stockholm in the early hours in a hotel room in California. Multiple Swedish accents helped reassure him that his #NobelPrize in Physics, awarded today, was real.

[ChenHuajun@Zhejiang_University @ChenHuajun]

@ZhenhengT It's just fine-tuning a small model, the computational cost won't be very high. I forward it to the first author for more details.@QSFByte

[ChenHuajun@Zhejiang_University @ChenHuajun]

Agent Planning with the World Knowledge Model (WKM). Our #NeurIPS2024 paper introduces WKM, an independent model that enhances large-model agents by synthesizing task and state knowledge from expert trajectories. Fine-tuned from a small base model, WKM applies a process supervision-like approach to guide agent planning. Experiments show that agents using world knowledge significantly outperform those relying solely on trial-and-error methods.
https://t.co/XRGrCeggOG