General Bachelor Projects on itu

Training-free Verifiable Process Reward for LLM Reinforcement Finetuning

Mon, 15 Jun 2026 00:00:00 +0900

This thesis project, started in October 2025, develops a training-free verifiable process reward for reinforcement finetuning of large language models. The implementation is available in the verl repository.

Approach

Forcing LLMs generate natural-language reasoning steps in Horn clause form, such as \(p \land q \to r\), so the reasoning process can be segmented into semantically meaningful steps.
A theorem prover or solver acts as a verifier and checks each reasoning step.
A teacher LLM formalizes natural-language reasoning into verifier-compatible expressions.
A GRPO-like advantage normalization method estimates critic-free values for process rewards.

Evaluation

Experiments use logical and mathematical question-answering datasets. The system is evaluated by comparing reasoning-model accuracy before and after reinforcement finetuning with the proposed process reward.

Reconstruction and Re-evaluation of SFCNN for PLA Scoring

Tue, 01 Apr 2025 00:00:00 +0900

This project reconstructs and re-evaluates the SFCNN (Scoring Function 3D Convolutional Neural Network) for protein-ligand binding affinity prediction. The implementation is available in the SFCNN_PLA_torch repository.

Report: Re-analyzing Scoring Function Convolutional Neural Networks (PDF), by Chuyan Zhou, Hanze Li and Xinyang Xu.

Work completed

Constructed a PyTorch implementation of the model architecture, training procedure, and evaluation pipeline.
Developed a benchmark that allows related models to infer directly from the 3D structure of protein-ligand complexes instead of using decoupled protein and ligand structures.

LLM-powered Lecture Generation

Thu, 01 Aug 2024 00:00:00 +0900

This project developed an LLM-powered lecture generation system for the UC Berkeley COMPSCI 194-196 research project. The implementation is available in the pdf2lec_llm repository.

Report: PDF2Lec: Streamlining LLM-driven Lecture Generation (PDF), by Rachel Lin, Mutian Hong, Chuyan Zhou, Jan Chen and Lucy Struefing.

Work completed

Independently developed the backend framework as a deployable FastAPI web service.
Added asynchronous task execution with multithreading, API-based task management backed by Redis, and a metadata system for generated data.
Integrated the different model components into the backend framework as the main developer.
Developed an LLM-powered question-answering agent that uses the generated lectures as long context and dynamically indexes grounding sources, such as textbooks, with a RAG system.