Tianle Wang

Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao et al. · allen-ai, amazon-science

Language models demonstrate both quantitative improvement and new qualitative capabilities with increasing scale. Despite their potentially transformative impact, these new capabilities are as yet poorly characterized. In order to inform future research, prepare for disruptive new model capabilities, and ameliorate socially harmful effects, it is vital that we understand the present and near-future capabilities and limitations of language models. To address this challenge, we introduce the Beyond the Imitation Game benchmark (BIG-bench). BIG-bench currently consists of 204 tasks, contributed by 450 authors across 132 institutions. Task topics are diverse, drawing problems from linguistics, childhood development, math, common-sense reasoning, biology, physics, social bias, software development, and beyond. BIG-bench focuses on tasks that are believed to be beyond the capabilities of current language models. We evaluate the behavior of OpenAI's GPT models, Google-internal dense transformer architectures, and Switch-style sparse transformers on BIG-bench, across model sizes spanning millions to hundreds of billions of parameters. In addition, a team of human expert raters performed all tasks in order to provide a strong baseline. Findings include: model performance and calibration both improve with scale, but are poor in absolute terms (and when compared with rater performance); performance is remarkably similar across model classes, though with benefits from sparsity; tasks that improve gradually and predictably commonly involve a large knowledge or memorization component, whereas tasks that exhibit "breakthrough" behavior at a critical scale often involve multiple steps or components, or brittle metrics; social bias typically increases with scale in settings with ambiguous context, but this can be improved with prompting.

Xubin Ren, Lianghao Xia, Yuhao Yang et al. · microsoft-research

Self-supervised learning (SSL) has gained significant interest in recent years as a solution to address the challenges posed by sparse and noisy data in recommender systems. Despite the growing number of SSL algorithms designed to provide state-of-the-art performance in various recommendation scenarios (e.g., graph collaborative filtering, sequential recommendation, social recommendation, KG-enhanced recommendation), there is still a lack of unified frameworks that integrate recommendation algorithms across different domains. Such a framework could serve as the cornerstone for self-supervised recommendation algorithms, unifying the validation of existing methods and driving the design of new ones. To address this gap, we introduce SSLRec, a novel benchmark platform that provides a standardized, flexible, and comprehensive framework for evaluating various SSL-enhanced recommenders. The SSLRec framework features a modular architecture that allows users to easily evaluate state-of-the-art models and a complete set of data augmentation and self-supervised toolkits to help create SSL recommendation models with specific needs. Furthermore, SSLRec simplifies the process of training and evaluating different recommendation models with consistent and fair settings. Our SSLRec platform covers a comprehensive set of state-of-the-art SSL-enhanced recommendation models across different scenarios, enabling researchers to evaluate these cutting-edge models and drive further innovation in the field. Our implemented SSLRec framework is available at the source code repository https://github.com/HKUDS/SSLRec.

Tianxing He, Jingyu Zhang, Tianle Wang et al.

In this work, we explore a useful but often neglected methodology for robustness analysis of text generation evaluation metrics: stress tests with synthetic data. Basically, we design and synthesize a wide range of potential errors and check whether they result in a commensurate drop in the metric scores. We examine a range of recently proposed evaluation metrics based on pretrained language models, for the tasks of open-ended generation, translation, and summarization. Our experiments reveal interesting insensitivities, biases, or even loopholes in existing metrics. For example, we find that BERTScore is confused by truncation errors in summarization, and MAUVE (built on top of GPT-2) is insensitive to errors at the beginning or middle of generations. Further, we investigate the reasons behind these blind spots and suggest practical workarounds for a more reliable evaluation of text generation. We have released our code and data at https://github.com/cloudygoose/blindspot_nlg.

Tianle Wang, Zhongyuan Wu, Shenghao Jin et al.

Reinforcement learning with verifiable rewards (RLVR) has become a central component of large language model (LLM) post-training. Unlike supervised fine-tuning (SFT), RLVR lets an LLM generate multiple candidate solutions and reinforces those that lead to a verifiably correct final answer. However, in practice, RLVR often requires thousands of training steps to reach strong performance, incurring substantial computation largely attributed to prolonged exploration. In this work, we make a surprising observation: during RLVR, LLMs evolve in a strongly linear manner. Specifically, both model weights and model output log-probabilities exhibit strong linear correlations with RL training steps. This suggests that RLVR predominantly amplifies trends that emerge early in training, rather than continuously discovering new behaviors throughout the entire optimization trajectory. Motivated by this linearity, we investigate whether future model states can be predicted from intermediate checkpoints via extrapolation, avoiding continued expensive training. We show that Weight Extrapolation produces models with performance comparable to standard RL training while requiring significantly less computation. Moreover, Logits Extrapolation consistently outperforms continued RL training on mathematics and code benchmarks by extrapolating beyond the step range where RL training remains stable. Our code is available at https://github.com/Miaow-Lab/RLVR-Linearity

Mengke Zhang, Tianxing He, Tianle Wang et al.

In the current user-server interaction paradigm of prompted generation with large language models (LLM) on cloud, the server fully controls the generation process, which leaves zero options for users who want to keep the generated text to themselves. We propose LatticeGen, a cooperative framework in which the server still handles most of the computation while the user controls the sampling operation. The key idea is that the true generated sequence is mixed with noise tokens by the user and hidden in a noised lattice. Considering potential attacks from a hypothetically malicious server and how the user can defend against it, we propose the repeated beam-search attack and the mixing noise scheme. In our experiments we apply LatticeGen to protect both prompt and generation. It is shown that while the noised lattice degrades generation quality, LatticeGen successfully protects the true generation to a remarkable degree under strong attacks (more than 50% of the semantic remains hidden as measured by BERTScore).

Xiaojun Chen, Tianle Wang, Tianhao Qiu et al.

Despite the success of large language models (LLMs) in Text-to-SQL tasks, open-source LLMs encounter challenges in contextual understanding and response coherence. To tackle these issues, we present \ours, a systematic methodology tailored for Text-to-SQL with open-source LLMs. Our contributions include a comprehensive evaluation of open-source LLMs in Text-to-SQL tasks, the \openprompt strategy for effective question representation, and novel strategies for supervised fine-tuning. We explore the benefits of Chain-of-Thought in step-by-step inference and propose the \openexample method for enhanced few-shot learning. Additionally, we introduce token-efficient techniques, such as \textbf{Variable-length Open DB Schema}, \textbf{Target Column Truncation}, and \textbf{Example Column Truncation}, addressing challenges in large-scale databases. Our findings emphasize the need for further investigation into the impact of supervised fine-tuning on contextual learning capabilities. Remarkably, our method significantly improved Llama2-7B from 2.54\% to 41.04\% and Code Llama-7B from 14.54\% to 48.24\% on the BIRD-Dev dataset. Notably, the performance of Code Llama-7B surpassed GPT-4 (46.35\%) on the BIRD-Dev dataset.

Zihan Wang, Tianle Wang, Dheeraj Mekala et al.

Etremely Weakly Supervised Text Classification (XWS-TC) refers to text classification based on minimal high-level human guidance, such as a few label-indicative seed words or classification instructions. There are two mainstream approaches for XWS-TC, however, never being rigorously compared: (1) training classifiers based on pseudo-labels generated by (softly) matching seed words (SEED) and (2) prompting (and calibrating) language models using classification instruction (and raw texts) to decode label words (PROMPT). This paper presents the first XWS-TC benchmark to compare the two approaches on fair grounds, where the datasets, supervisions, and hyperparameter choices are standardized across methods. Our benchmarking results suggest that (1) Both SEED and PROMPT approaches are competitive and there is no clear winner; (2) SEED is empirically more tolerant than PROMPT to human guidance (e.g., seed words, classification instructions, and label words) changes; (3) SEED is empirically more selective than PROMPT to the pre-trained language models; (4) Recent SEED and PROMPT methods have close connections and a clustering post-processing step based on raw in-domain texts is a strong performance booster to both. We hope this benchmark serves as a guideline in selecting XWS-TC methods in different scenarios and stimulate interest in developing guidance- and model-robust XWS-TC methods. We release the repo at https://github.com/ZihanWangKi/x-TC.

Tianle Wang, Zhaoyang Wang, Guangchen Lan et al.

Reinforcement learning (RL) has been applied to improve large language model (LLM) reasoning, yet the systematic study of how training scales with task difficulty has been hampered by the lack of controlled, scalable environments. We introduce ScaleLogic, a synthetic logical reasoning framework that offers independent control over two axes of difficulty: the depth of the required proof planning (i.e., the horizon) and the expressiveness of the underlying logic. Our proposed framework supports a wide range of logics: from simple implication-only logic ("if-then") towards more expressive first-order reasoning with conjunction ("and"), disjunction ("or"), negation ("not"), and universal quantification ("for all"). Using this framework, we show that the RL training compute $T$ follows a power law with respect to reasoning depth $D$ ($T \propto D^γ$, $R^{2} > 0.99$), and that the scaling exponent $γ$ increases monotonically with logical expressiveness, from $1.04$ to $2.60$. On downstream mathematics and general reasoning benchmarks, more expressive training settings yield both larger performance gains (up to $+10.66$ points) and more compute-efficient transfer compared to less expressive settings, demonstrating that what a model is trained on, not just how much it is trained, shapes downstream transfer. We further show that the power-law relationship holds across multiple RL methods, and curriculum-based training substantially improves scaling efficiency.

Yiyang Zhou, Zhaoyang Wang, Tianle Wang et al.

High-quality preference data is essential for aligning foundation models with human values through preference learning. However, manual annotation of such data is often time-consuming and costly. Recent methods often adopt a self-rewarding approach, where the target model generates and annotates its own preference data, but this can lead to inaccuracies since the reward model shares weights with the target model, thereby amplifying inherent biases. To address these issues, we propose Anyprefer, a framework designed to synthesize high-quality preference data for aligning the target model. Anyprefer frames the data synthesis process as a cooperative two-player Markov Game, where the target model and the judge model collaborate together. Here, a series of external tools are introduced to assist the judge model in accurately rewarding the target model's responses, mitigating biases in the rewarding process. In addition, a feedback mechanism is introduced to optimize prompts for both models, enhancing collaboration and improving data quality. The synthesized data is compiled into a new preference dataset, Anyprefer-V1, consisting of 58K high-quality preference pairs. Extensive experiments show that Anyprefer significantly improves model alignment performance across four main applications, covering 21 datasets, achieving average improvements of 18.55% in five natural language generation datasets, 3.66% in nine vision-language understanding datasets, 30.05% in three medical image analysis datasets, and 16.00% in four visuo-motor control tasks.

Guangchen Lan, Sipeng Zhang, Tianle Wang et al.

As the era of large language models (LLMs) on behalf of users unfolds, Preference Optimization (PO) methods have become a central approach to aligning LLMs with human preferences and improving performance. We propose Maximum a Posteriori Preference Optimization (MaPPO), a framework for learning from preferences that explicitly incorporates prior reward knowledge into the optimization objective. While existing methods such as Direct Preference Optimization (DPO) and its variants treat preference learning as a Maximum Likelihood Estimation (MLE) problem, MaPPO extends this paradigm by integrating prior reward estimates into a principled Maximum a Posteriori (MaP) objective. This not only generalizes DPO and its variants, but also enhances alignment by mitigating the oversimplified binary classification of responses. More importantly, MaPPO introduces no additional hyperparameter, and supports preference optimization in both offline and online settings. In addition, MaPPO can be used as a plugin with consistent improvement on DPO variants, including widely used SimPO, IPO, and CPO. Extensive empirical evaluations of different model sizes and model series on three standard benchmarks, including MT-Bench, AlpacaEval 2.0, and Arena-Hard, demonstrate consistent improvements in alignment performance without sacrificing computational efficiency.

Sanket Jantre, Tianle Wang, Gilchan Park et al.

Identification of protein-protein interactions (PPIs) helps derive cellular mechanistic understanding, particularly in the context of complex conditions such as neurodegenerative disorders, metabolic syndromes, and cancer. Large Language Models (LLMs) have demonstrated remarkable potential in predicting protein structures and interactions via automated mining of vast biomedical literature; yet their inherent uncertainty remains a key challenge for deriving reproducible findings, critical for biomedical applications. In this study, we present an uncertainty-aware adaptation of LLMs for PPI analysis, leveraging fine-tuned LLaMA-3 and BioMedGPT models. To enhance prediction reliability, we integrate LoRA ensembles and Bayesian LoRA models for uncertainty quantification (UQ), ensuring confidence-calibrated insights into protein behavior. Our approach achieves competitive performance in PPI identification across diverse disease contexts while addressing model uncertainty, thereby enhancing trustworthiness and reproducibility in computational biology. These findings underscore the potential of uncertainty-aware LLM adaptation for advancing precision medicine and biomedical research.

Derek Anderson, Amit Bashyal, Markus Diefenthaler et al.

The Production and Distributed Analysis (PanDA) system, originally developed for the ATLAS experiment at the CERN Large Hadron Collider (LHC), has evolved into a robust platform for orchestrating large-scale workflows across distributed computing resources. Coupled with its intelligent Distributed Dispatch and Scheduling (iDDS) component, PanDA supports AI/ML-driven workflows through a scalable and flexible workflow engine. We present an AI-assisted framework for detector design optimization that integrates multi-objective Bayesian optimization with the PanDA--iDDS workflow engine to coordinate iterative simulations across heterogeneous resources. The framework addresses the challenge of exploring high-dimensional parameter spaces inherent in modern detector design. We demonstrate the framework using benchmark problems and realistic studies of the ePIC and dRICH detectors for the Electron-Ion Collider (EIC). Results show improved automation, scalability, and efficiency in multi-objective optimization. This work establishes a flexible and extensible paradigm for AI-driven detector design and other computationally intensive scientific applications.

Tianle Wang, Sirui Zhang, Xinyi Tong et al.

This paper presents an unsupervised machine learning algorithm that identifies recurring patterns -- referred to as ``music-words'' -- from symbolic music data. These patterns are fundamental to musical structure and reflect the cognitive processes involved in composition. However, extracting these patterns remains challenging because of the inherent semantic ambiguity in musical interpretation. We formulate the task of music-word discovery as a statistical optimization problem and propose a two-stage Expectation-Maximization (EM)-based learning framework: 1. Developing a music-word dictionary; 2. Reconstructing the music data. When evaluated against human expert annotations, the algorithm achieved an Intersection over Union (IoU) score of 0.61. Our findings indicate that minimizing code length effectively addresses semantic ambiguity, suggesting that human optimization of encoding systems shapes musical semantics. This approach enables computers to extract ``basic building blocks'' from music data, facilitating structural analysis and sparse encoding. The method has two primary applications. First, in AI music, it supports downstream tasks such as music generation, classification, style transfer, and improvisation. Second, in musicology, it provides a tool for analyzing compositional patterns and offers insights into the principle of minimal encoding across diverse musical styles and composers.

Mohammad Atif, Kriti Chopra, Ozgur Kilic et al.

Next-generation High Energy Physics (HEP) experiments will generate unprecedented data volumes, necessitating High Performance Computing (HPC) integration alongside traditional high-throughput computing. However, HPC adoption in HEP is hindered by the challenge of porting legacy software to heterogeneous architectures and the sparse documentation of these complex scientific codebases. We present CelloAI, a locally hosted coding assistant that leverages Large Language Models (LLMs) with retrieval-augmented generation (RAG) to support HEP code documentation and generation. This local deployment ensures data privacy, eliminates recurring costs and provides access to large context windows without external dependencies. CelloAI addresses two primary use cases, code documentation and code generation, through specialized components. For code documentation, the assistant provides: (a) Doxygen style comment generation for all functions and classes by retrieving relevant information from RAG sources (papers, posters, presentations), (b) file-level summary generation, and (c) an interactive chatbot for code comprehension queries. For code generation, CelloAI employs syntax-aware chunking strategies that preserve syntactic boundaries during embedding, improving retrieval accuracy in large codebases. The system integrates callgraph knowledge to maintain dependency awareness during code modifications and provides AI-generated suggestions for performance optimization and accurate refactoring. We evaluate CelloAI using real-world HEP applications from ATLAS, CMS, and DUNE experiments, comparing different embedding models for code retrieval effectiveness. Our results demonstrate the AI assistant's capability to enhance code understanding and support reliable code generation while maintaining the transparency and safety requirements essential for scientific computing environments.

Andre Merzky, Mikhail Titov, Matteo Turilli et al.

Hybrid workflows combining traditional HPC and novel ML methodologies are transforming scientific computing. This paper presents the architecture and implementation of a scalable runtime system that extends RADICAL-Pilot with service-based execution to support AI-out-HPC workflows. Our runtime system enables distributed ML capabilities, efficient resource management, and seamless HPC/ML coupling across local and remote platforms. Preliminary experimental results show that our approach manages concurrent execution of ML models across local and remote HPC/cloud resources with minimal architectural overheads. This lays the foundation for prototyping three representative data-driven workflow applications and executing them at scale on leadership-class HPC platforms.

Tianle Wang, Zihan Wang, Weitang Liu et al.

State-of-the-art weakly supervised text classification methods, while significantly reduced the required human supervision, still requires the supervision to cover all the classes of interest. This is never easy to meet in practice when human explore new, large corpora without complete pictures. In this paper, we work on a novel yet important problem of weakly supervised open-world text classification, where supervision is only needed for a few examples from a few known classes and the machine should handle both known and unknown classes in test time. General open-world classification has been studied mostly using image classification; however, existing methods typically assume the availability of sufficient known-class supervision and strong unknown-class prior knowledge (e.g., the number and/or data distribution). We propose a novel framework WOT-Class that lifts those strong assumptions. Specifically, it follows an iterative process of (a) clustering text to new classes, (b) mining and ranking indicative words for each class, and (c) merging redundant classes by using the overlapped indicative words as a bridge. Extensive experiments on 7 popular text classification datasets demonstrate that WOT-Class outperforms strong baselines consistently with a large margin, attaining 23.33% greater average absolute macro-F1 over existing approaches across all datasets. Such competent accuracy illuminates the practical potential of further reducing human effort for text classification.

Wenhui Cui, Yanlin Liu, Yuxing Li et al.

Automated brain lesion segmentation provides valuable information for the analysis and intervention of patients. In particular, methods based on convolutional neural networks (CNNs) have achieved state-of-the-art segmentation performance. However, CNNs usually require a decent amount of annotated data, which may be costly and time-consuming to obtain. Since unannotated data is generally abundant, it is desirable to use unannotated data to improve the segmentation performance for CNNs when limited annotated data is available. In this work, we propose a semi-supervised learning (SSL) approach to brain lesion segmentation, where unannotated data is incorporated into the training of CNNs. We adapt the mean teacher model, which is originally developed for SSL-based image classification, for brain lesion segmentation. Assuming that the network should produce consistent outputs for similar inputs, a loss of segmentation consistency is designed and integrated into a self-ensembling framework. Specifically, we build a student model and a teacher model, which share the same CNN architecture for segmentation. The student and teacher models are updated alternately. At each step, the student model learns from the teacher model by minimizing the weighted sum of the segmentation loss computed from annotated data and the segmentation consistency loss between the teacher and student models computed from unannotated data. Then, the teacher model is updated by combining the updated student model with the historical information of teacher models using an exponential moving average strategy. For demonstration, the proposed approach was evaluated on ischemic stroke lesion segmentation, where it improves stroke lesion segmentation with the incorporation of unannotated data.