Shuo Wu

Xinxin Lin, Guangxin Dai, Yi Zhong et al.

Large language models (LLMs) hold transformative potential for medical decision support yet their application in psychiatry remains constrained by hallucinations and superficial reasoning. This limitation is particularly acute in light-parameter LLMs which are essential for privacy-preserving and efficient clinical deployment. Existing training paradigms prioritize linguistic fluency over structured clinical logic and result in a fundamental misalignment with professional diagnostic cognition. Here we introduce ClinMPO, a reinforcement learning framework designed to align the internal reasoning of LLMs with professional psychiatric practice. The framework employs a specialized reward model trained independently on a dataset derived from 4,474 psychiatry journal articles and structured according to evidence-based medicine principles. We evaluated ClinMPO on a unseen subset of the benchmark designed to isolate reasoning capabilities from rote memorization. This test set comprises items where leading large-parameter LLMs consistently fail. We compared the ClinMPO-aligned light LLM performance against a cohort of 300 medical students. The ClinMPO-tuned Qwen3-8B model achieved a diagnostic accuracy of 31.4% and surpassed the human benchmark of 30.8% on these complex cases. These results demonstrate that medical evidence-guided optimization enables light-parameter LLMs to master complex reasoning tasks. Our findings suggest that explicit cognitive alignment offers a scalable pathway to reliable and safe psychiatric decision support.

Yongqiang Yao, Jingru Tan, Kaihuan Liang et al.

Training Long-Context Large Language Models (LLMs) is challenging, as hybrid training with long-context and short-context data often leads to workload imbalances. Existing works mainly use data packing to alleviate this issue, but fail to consider imbalanced attention computation and wasted communication overhead. This paper proposes Hierarchical Balance Packing (HBP), which designs a novel batch-construction method and training recipe to address those inefficiencies. In particular, the HBP constructs multi-level data packing groups, each optimized with a distinct packing length. It assigns training samples to their optimal groups and configures each group with the most effective settings, including sequential parallelism degree and gradient checkpointing configuration. To effectively utilize multi-level groups of data, we design a dynamic training pipeline specifically tailored to HBP, including curriculum learning, adaptive sequential parallelism, and stable loss. Our extensive experiments demonstrate that our method significantly reduces training time over multiple datasets and open-source models while maintaining strong performance. For the largest DeepSeek-V2 (236B) MoE model, our method speeds up the training by 2.4$\times$ with competitive performance. Codes will be released at https://github.com/ModelTC/HBP.

Yu Wu, Shuo Wu, Ye Tao et al.

Large Language Models (LLMs) vary widely in their capabilities, with larger models often having better performance but higher cost: choosing an LLM model often involves trading off performance and cost. The LLM Cascade is a paradigm that defers difficult queries from weak/cheap to strong/expensive models. This approach is nonadaptive: the deferral decision is trained offline. When confronted with similar or repeated queries, the LLM Cascade may then repeatedly consult the expensive model and incur higher cost. To improve the cascading efficiency, we propose Inter-Cascade, an online and interactive LLM Cascade that extends the role of strong model from a backup helper to a long-term teacher. In our system, when a strong model resolves a difficult query, it also distills its solution into a generalized, reusable problem-solving strategy that boosts the weak model on subsequent queries. Adding strategies to queries enables the weak model to dynamically improve its performance over time, avoiding computationally and time-intensive fine-tuning. Empirically, compared with standard LLM Cascade baselines across multiple benchmarks, the Inter-Cascade significantly improves the accuracy of the weak model (by up to 33.06 absolute percentage points) and the overall system (by up to 5.53 absolute percentage points), while reducing the calls to strong models (by up to 48.05% relative reduction) and saving the corresponding fees (by up to 49.63% relative reduction). Inter-Cascade demonstrates the effective in-context knowledge transfer between LLMs, and provides a general, scalable framework applicable to both open-source and API-based LLMs.

Yeji Wang, Shuo Wu, Yanwen Duan et al.

Motivation: Protein-ligand affinity prediction is an important part of structure-based drug design. It includes molecular docking and affinity prediction. Although molecular dynamics can predict affinity with high accuracy at present, it is not suitable for large-scale virtual screening. The existing affinity prediction and evaluation functions based on deep learning mostly rely on experimentally-determined conformations. Results: We build a predictive model of protein-ligand affinity through the ResNet neural network with added attention mechanism. The resulting ResAtom-Score model achieves Pearson's correlation coefficient R = 0.833 on the CASF-2016 benchmark test set. At the same time, we evaluated the performance of a variety of existing scoring functions in combination with ResAtom-Score in the absence of experimentally-determined conformations. The results show that the use of ΔVinaRF20 in combination with ResAtom-Score can achieve affinity prediction close to scoring functions in the presence of experimentally-determined conformations. These results suggest that ResAtom system may be used for in silico screening of small molecule ligands with target proteins in the future. Availability: https://github.com/wyji001/ResAtom

Yansong Li, Zeyu Dong, Ertai Luo et al.

Reinforcement learning (RL) algorithms can be divided into two classes: model-free algorithms, which are sample-inefficient, and model-based algorithms, which suffer from model bias. Dyna-style algorithms combine these two approaches by using simulated data from an estimated environmental model to accelerate model-free training. However, their efficiency is compromised when the estimated model is inaccurate. Previous works address this issue by using model ensembles or pretraining the estimated model with data collected from the real environment, increasing computational and sample complexity. To tackle this issue, we introduce an out-of-distribution (OOD) data filter that removes simulated data from the estimated model that significantly diverges from data collected in the real environment. We show theoretically that this technique enhances the quality of simulated data. With the help of the OOD data filter, the data simulated from the estimated model better mimics the data collected by interacting with the real model. This improvement is evident in the critic updates compared to using the simulated data without the OOD data filter. Our experiment integrates the data filter into the model-based policy optimization (MBPO) algorithm. The results demonstrate that our method requires fewer interactions with the real environment to achieve a higher level of optimality than MBPO, even without a model ensemble.

Shuo Wu, Pawan Poojary, Randall Berry

We consider a model of Bayesian observational learning in which a sequence of agents receives a private signal about an underlying binary state of the world. Each agent makes a decision based on its own signal and its observations of previous agents. A central planner seeks to improve the accuracy of these signals by allocating a limited budget to enhance signal quality across agents. We formulate and analyze the budget allocation problem and propose two optimal allocation strategies. At least one of these strategies is shown to maximize the probability of achieving a correct information cascade.

Shuo Wu, Haoxiang Ma, Jie Fu et al.

The problem of reward design examines the interaction between a leader and a follower, where the leader aims to shape the follower's behavior to maximize the leader's payoff by modifying the follower's reward function. Current approaches to reward design rely on an accurate model of how the follower responds to reward modifications, which can be sensitive to modeling inaccuracies. To address this issue of sensitivity, we present a solution that offers robustness against uncertainties in modeling the follower, including 1) how the follower breaks ties in the presence of nonunique best responses, 2) inexact knowledge of how the follower perceives reward modifications, and 3) bounded rationality of the follower. Our robust solution is guaranteed to exist under mild conditions and can be obtained numerically by solving a mixed-integer linear program. Numerical experiments on multiple test cases demonstrate that our solution improves robustness compared to the standard approach without incurring significant additional computing costs.