arXiv subjects

Haitian Zhong

Publications and source records attributed to Haitian Zhong.

9 recordsLinked to original sources

arXiv2026

SCAN: Sparse Circuit Anchor Interpretable Neuron for Lifelong Knowledge Editing

Large Language Models (LLMs) often suffer from catastrophic forgetting and collapse during sequential knowledge editing. This vulnerability stems from the prevailing dense editing paradigm, which treats models as black boxes and relies on coarse-grained parameter interventions that inevitably disrupt preserved knowledge. To address this, we propose SCAN (a sparse editing framework based on Sparse Circuit Anchored Neuron) which transforms editing into a mechanism-aware manipulation by constructing a knowledge circuit via Sparse Transcoders. Experiments on Gemma2, Qwen3, and Llama3.1 across CounterFact, ZsRE and WikiFactDiff demonstrate that SCAN achieves a superior performance, maintaining model integrity on benchmarks like MMLU and GSM8K even after 3,000 sequential edits, whereas other existing methods deteriorate progressively as editing accumulates, eventually resulting in model collapse.

cs.AI↗

arXiv2026

RC-GRPO: Reward-Conditioned Group Relative Policy Optimization for Multi-Turn Tool Calling Agents

Multi-turn tool calling is challenging for Large Language Models (LLMs) because rewards are sparse and exploration is expensive. A common recipe, SFT followed by GRPO, can stall when within-group reward variation is low (e.g., more rollouts in a group receive the all 0 or all 1 reward), making the group-normalized advantage uninformative and yielding vanishing updates. To address this problem, we propose RC-GRPO (Reward-Conditioned Group Relative Policy Optimization), which treats exploration as a controllable steering problem via discrete reward tokens. We first fine-tune a Reward-Conditioned Trajectory Policy (RCTP) on mixed-quality trajectories with reward goal special tokens (e.g., <|high_reward|>, <|low_reward|>) injected into the prompts, enabling the model to learn how to generate distinct quality trajectories on demand. Then during RL, we sample diverse reward tokens within each GRPO group and condition rollouts on the sampled token to improve within-group diversity, improving advantage gains. On the Berkeley Function Calling Leaderboard v4 (BFCLv4) multi-turn benchmark, our method yields consistently improved performance than baselines, and the performance on Qwen-2.5-7B-Instruct even surpasses all closed-source API models.

cs.AI↗

arXiv2025

A general language model for peptide function identification

Accurate identification of bioactive peptides (BPs) and protein post-translational modifications (PTMs) is essential for understanding protein function and advancing therapeutic discovery. However, most computational methods remain limited in their generalizability across diverse peptide functions. Here, we present PDeepPP, a unified deep learning framework that integrates pretrained protein language models with a hybrid transformer-CNN architecture, enabling robust identification across diverse peptide classes and PTM sites. We curated comprehensive benchmark datasets and implemented strategies to address data imbalance, allowing PDeepPP to systematically extract both global and local sequence features. Through extensive analyses including dimensionality reduction and comparison studies, PDeepPP demonstrates strong, interpretable peptide representations and achieves state-of-the-art performance in 25 of the 33 biological identification tasks. Notably, PDeepPP attains high accuracy in antimicrobial (0.9726) and phosphorylation site (0.9984) identification, with 99.5% specificity in glycosylation site prediction and substantial reduction in false negatives in antimalarial tasks. By enabling large-scale, accurate peptide analysis, PDeepPP supports biomedical research and the discovery of novel therapeutic targets for disease treatment. All code, datasets, and pretrained models are publicly available via GitHub (https://github.com/fondress/PDeepPP) and Hugging Face (https://huggingface.co/fondress/PDeppPP)

cs.LG↗

arXiv2025

SciAgent: A Unified Multi-Agent System for Generalistic Scientific Reasoning

Recent advances in large language models have enabled AI systems to achieve expert-level performance on domain-specific scientific tasks, yet these systems remain narrow and handcrafted. We introduce SciAgent, a unified multi-agent system designed for generalistic scientific reasoning-the ability to adapt reasoning strategies across disciplines and difficulty levels. SciAgent organizes problem solving as a hierarchical process: a Coordinator Agent interprets each problem's domain and complexity, dynamically orchestrating specialized Worker Systems, each composed of interacting reasoning Sub-agents for symbolic deduction, conceptual modeling, numerical computation, and verification. These agents collaboratively assemble and refine reasoning pipelines tailored to each task. Across mathematics and physics Olympiads (IMO, IMC, IPhO, CPhO), SciAgent consistently attains or surpasses human gold-medalist performance, demonstrating both domain generality and reasoning adaptability. Additionally, SciAgent has been tested on the International Chemistry Olympiad (IChO) and selected problems from the Humanity's Last Exam (HLE) benchmark, further confirming the system's ability to generalize across diverse scientific domains. This work establishes SciAgent as a concrete step toward generalistic scientific intelligence-AI systems capable of coherent, cross-disciplinary reasoning at expert levels.

cs.AI↗

arXiv2025

MorphoBench: A Benchmark with Difficulty Adaptive to Model Reasoning

With the advancement of powerful large-scale reasoning models, effectively evaluating the reasoning capabilities of these models has become increasingly important. However, existing benchmarks designed to assess the reasoning abilities of large models tend to be limited in scope and lack the flexibility to adapt their difficulty according to the evolving reasoning capacities of the models. To address this, we propose MorphoBench, a benchmark that incorporates multidisciplinary questions to evaluate the reasoning capabilities of large models and can adjust and update question difficulty based on the reasoning abilities of advanced models. Specifically, we curate the benchmark by selecting and collecting complex reasoning questions from existing benchmarks and sources such as Olympiad-level competitions. Additionally, MorphoBench adaptively modifies the analytical challenge of questions by leveraging key statements generated during the model's reasoning process. Furthermore, it includes questions generated using simulation software, enabling dynamic adjustment of benchmark difficulty with minimal resource consumption. We have gathered over 1,300 test questions and iteratively adjusted the difficulty of MorphoBench based on the reasoning capabilities of models such as o3 and GPT-5. MorphoBench enhances the comprehensiveness and validity of model reasoning evaluation, providing reliable guidance for improving both the reasoning abilities and scientific robustness of large models. The code has been released in https://github.com/OpenDCAI/MorphoBench.

cs.AI↗

arXiv2025

REACT: Representation Extraction And Controllable Tuning to Overcome Overfitting in LLM Knowledge Editing

Large language model editing methods frequently suffer from overfitting, wherein factual updates can propagate beyond their intended scope, overemphasizing the edited target even when it's contextually inappropriate. To address this challenge, we introduce REACT (Representation Extraction And Controllable Tuning), a unified two-phase framework designed for precise and controllable knowledge editing. In the initial phase, we utilize tailored stimuli to extract latent factual representations and apply Principal Component Analysis with a simple learnbale linear transformation to compute a directional "belief shift" vector for each instance. In the second phase, we apply controllable perturbations to hidden states using the obtained vector with a magnitude scalar, gated by a pre-trained classifier that permits edits only when contextually necessary. Relevant experiments on EVOKE benchmarks demonstrate that REACT significantly reduces overfitting across nearly all evaluation metrics, and experiments on COUNTERFACT and MQuAKE shows that our method preserves balanced basic editing performance (reliability, locality, and generality) under diverse editing scenarios.

cs.AI↗

arXiv2024

VLKEB: A Large Vision-Language Model Knowledge Editing Benchmark

Recently, knowledge editing on large language models (LLMs) has received considerable attention. Compared to this, editing Large Vision-Language Models (LVLMs) faces extra challenges from diverse data modalities and complicated model components, and data for LVLMs editing are limited. The existing LVLM editing benchmark, which comprises three metrics (Reliability, Locality, and Generality), falls short in the quality of synthesized evaluation images and cannot assess whether models apply edited knowledge in relevant content. Therefore, we employ more reliable data collection methods to construct a new Large $\textbf{V}$ision-$\textbf{L}$anguage Model $\textbf{K}$nowledge $\textbf{E}$diting $\textbf{B}$enchmark, $\textbf{VLKEB}$, and extend the Portability metric for more comprehensive evaluation. Leveraging a multi-modal knowledge graph, our image data are bound with knowledge entities. This can be further used to extract entity-related knowledge, which constitutes the base of editing data. We conduct experiments of different editing methods on five LVLMs, and thoroughly analyze how do they impact the models. The results reveal strengths and deficiencies of these methods and hopefully provide insights for future research. The codes and dataset are available at: https://github.com/VLKEB/VLKEB.

cs.CL↗

arXiv2024

PTransIPs: Identification of phosphorylation sites enhanced by protein PLM embeddings

Phosphorylation is pivotal in numerous fundamental cellular processes and plays a significant role in the onset and progression of various diseases. The accurate identification of these phosphorylation sites is crucial for unraveling the molecular mechanisms within cells and during viral infections, potentially leading to the discovery of novel therapeutic targets. In this study, we develop PTransIPs, a new deep learning framework for the identification of phosphorylation sites. Independent testing results demonstrate that PTransIPs outperforms existing state-of-the-art (SOTA) methods, achieving AUCs of 0.9232 and 0.9660 for the identification of phosphorylated S/T and Y sites, respectively. PTransIPs contributes from three aspects. 1) PTransIPs is the first to apply protein pre-trained language model (PLM) embeddings to this task. It utilizes ProtTrans and EMBER2 to extract sequence and structure embeddings, respectively, as additional inputs into the model, effectively addressing issues of dataset size and overfitting, thus enhancing model performance; 2) PTransIPs is based on Transformer architecture, optimized through the integration of convolutional neural networks and TIM loss function, providing practical insights for model design and training; 3) The encoding of amino acids in PTransIPs enables it to serve as a universal framework for other peptide bioactivity tasks, with its excellent performance shown in extended experiments of this paper. Our code, data and models are publicly available at https://github.com/StatXzy7/PTransIPs.

q-bio.QM↗

arXiv2023

Capacity, Collision Avoidance and Shopping Rate under a Social Distancing Regime

Capacity restrictions in stores, maintained by mechanisms like spacing customer intake, became familiar features of retailing in the time of the pandemic. Shopping rates in a crowded store under a social distance regime is prone to considerable slowdown. Inspired by the random particle collision concepts of statistical mechanics, we introduce a dynamical model of the evolution of shopping rate as a function of a given customer intake rate. The slowdown of each individual customer is incorporated as an additive term to a baseline value shopping time, proportional to the number of other customers in the store. We determine analytically and by simulation the trajectory of the model as it approaches a Little's Law equilibrium, and identify the point beyond which equilibrium cannot be achieved. By relating customer shopping rate to the slowdown compared to the baseline, we can calculate the optimal intake rate leading to maximum equilibrium spending. This turns out to be the maximum rate compatible with equilibrium. The slowdown due to the largest possible number of shoppers is more than compensated for by the increased volume of shopping. This macroscopic model is validated by simulation experiments in which avoidance interactions between pairs of shoppers are responsible for shopping delays.

cs.CE↗