Yuanpeng Li

Yuanpeng Li

This survey includes systematic generalization and a history of how machine learning addresses it. We aim to summarize and organize the related information of both conventional and recent improvements. We first look at the definition of systematic generalization, then introduce Classicist and Connectionist. We then discuss different types of Connectionists and how they approach the generalization. Two crucial problems of variable binding and causality are discussed. We look into systematic generalization in language, vision, and VQA fields. Recent improvements from different aspects are discussed. Systematic generalization has a long history in artificial intelligence. We could cover only a small portion of many contributions. We hope this paper provides a background and is beneficial for discoveries in future work.

Yuanpeng Li

In this paper, we hypothesize that internal function sharing is one of the reasons to weaken o.o.d. or systematic generalization in deep learning for classification tasks. Under equivalent prediction, a model partitions an input space into multiple parts separated by boundaries. The function sharing prefers to reuse boundaries, leading to fewer parts for new outputs, which conflicts with systematic generalization. We show such phenomena in standard deep learning models, such as fully connected, convolutional, residual networks, LSTMs, and (Vision) Transformers. We hope this study provides novel insights into systematic generalization and forms a basis for new research directions.

Yuanpeng Li, Gefei Lin, Annie Qu et al.

Soft Actor-Critic (SAC) and its variants dominate Multi-Task Reinforcement Learning (MTRL) due to their off-policy sample efficiency, while on-policy methods such as Proximal Policy Optimization (PPO) remain underexplored. We diagnose that PPO in MTRL suffers from a previously overlooked issue: critic-side gradient ill-conditioning, which may cause tail tasks to stall while easy tasks dominate the value function's updates. To address this, we propose TOPPO (Tail-Optimized PPO), a reformulation of PPO via Critic Balancing -- a set of modules that improve gradient conditioning and balance learning dynamics across tasks. Unlike prior approaches that rely on modular architectures or large models, TOPPO targets the optimization bottleneck within PPO itself. Empirically, TOPPO achieves stronger mean and tail-task performance than published SAC-family and ARS-family baselines while using substantially fewer parameters and environment steps on Meta-World+ benchmark. Notably, TOPPO matches or surpasses strong SAC baselines early in training and maintains superior performance at full budget. Ablations confirm the effectiveness of each module in TOPPO and provide insights into their interactions. Our results demonstrate that, with proper optimization, on-policy methods can rival or exceed off-policy approaches in MTRL, challenging the prevailing reliance on SAC and highlighting critic-side gradient conditioning as the central bottleneck.

Yuanpeng Li, Joel Hestness, Mohamed Elhoseiny et al.

This paper proposes an efficient approach to learning disentangled representations with causal mechanisms based on the difference of conditional probabilities in original and new distributions. We approximate the difference with models' generalization abilities so that it fits in the standard machine learning framework and can be efficiently computed. In contrast to the state-of-the-art approach, which relies on the learner's adaptation speed to new distribution, the proposed approach only requires evaluating the model's generalization ability. We provide a theoretical explanation for the advantage of the proposed method, and our experiments show that the proposed technique is 1.9--11.0$\times$ more sample efficient and 9.4--32.4 times quicker than the previous method on various tasks. The source code is available at \url{https://github.com/yuanpeng16/EDCR}.

Xiaoyu Li, Yuanpeng Li, Yingyu Liang et al.

Modern Hopfield networks (MHNs) have emerged as powerful tools in deep learning, capable of replacing components such as pooling layers, LSTMs, and attention mechanisms. Recent advancements have enhanced their storage capacity, retrieval speed, and error rates. However, the fundamental limits of their computational expressiveness remain unexplored. Understanding the expressive power of MHNs is crucial for optimizing their integration into deep learning architectures. In this work, we establish rigorous theoretical bounds on the computational capabilities of MHNs using circuit complexity theory. Our key contribution is that we show that MHNs are $\mathsf{DLOGTIME}$-uniform $\mathsf{TC}^0$. Hence, unless $\mathsf{TC}^0 = \mathsf{NC}^1$, a $\mathrm{poly}(n)$-precision modern Hopfield networks with a constant number of layers and $O(n)$ hidden dimension cannot solve $\mathsf{NC}^1$-hard problems such as the undirected graph connectivity problem and the tree isomorphism problem. We also extended our results to Kernelized Hopfield Networks. These results demonstrate the limitation in the expressive power of the modern Hopfield networks. Moreover, Our theoretical analysis provides insights to guide the development of new Hopfield-based architectures.

Yuanpeng Li, Zhen Xu, Zongwei Lv et al.

Network stream mining is fundamental to many network operations. Sketches, as compact data structures that offer low memory overhead with bounded accuracy, have emerged as a promising solution for network stream mining. Recent studies attempt to optimize sketches using machine learning; however, these approaches face the challenges of lacking adaptivity to dynamic networks and incurring high training costs. In this paper, we propose LLM-Sketch, based on the insight that fields beyond the flow IDs in packet headers can also help infer flow sizes. By using a two-tier data structure and separately recording large and small flows, LLM-Sketch improves accuracy while minimizing memory usage. Furthermore, it leverages fine-tuned large language models (LLMs) to reliably estimate flow sizes. We evaluate LLM-Sketch on three representative tasks, and the results demonstrate that LLM-Sketch outperforms state-of-the-art methods by achieving a $7.5\times$ accuracy improvement.

Yuanpeng Li, Qi Long, Zhiyuan Yao et al.

As enterprise codebases continue to grow in scale and complexity, the volume of lint errors far exceeds engineers' manual remediation capacity, leading to continuous accumulation of technical debt and hindered development efficiency. This paper presents BitsAI-Fix, an automated lint error remediation workflow based on Large Language Models (LLMs), designed to address this critical challenge in industrial-scale environments. BitsAI-Fix employs tree-sitter for context expansion and generates search-and-replace format patches through specially trained LLMs, followed by lint scan re-verification to output final remediation results. Additionally, our approach introduces an innovative progressive reinforcement learning (RL) training strategy that can automatically acquire verifiable training data during the project cold-start phase and continuously iterate the model by collecting online samples through feedback after system deployment. Furthermore, we designed a targeted rule-based reward mechanism that combines format rewards and correctness rewards while penalizing redundant modifications. We also propose a "code diff matching" methodology to continuously track online effectiveness. In production deployment at ByteDance, our solution has supported over 5,000 engineers, resolved more than 12,000 static analysis issues, achieved approximately 85% remediation accuracy, with around 1,000 weekly active adopters. This work demonstrates the practical feasibility of LLM-based code remediation solutions in enterprise environments and serves as a reference for automated code fix in large-scale industrial scenarios.

Yuanpeng Li

Compositional generalization is a crucial property in artificial intelligence, enabling models to handle novel combinations of known components. While most deep learning models lack this capability, certain models succeed in specific tasks, suggesting the existence of governing conditions. This paper derives a necessary and sufficient condition for compositional generalization in neural networks. Conceptually, it requires that (i) the computational graph matches the true compositional structure, and (ii) components encode just enough information in training. The condition is supported by mathematical proofs. This criterion combines aspects of architecture design, regularization, and training data properties. A carefully designed minimal example illustrates an intuitive understanding of the condition. We also discuss the potential of the condition for assessing compositional generalization before training. This work is a fundamental theoretical study of compositional generalization in neural networks.

Yuanpeng Li

Compositional generalization is the capacity to recognize and imagine a large amount of novel combinations from known components. It is a key in human intelligence, but current neural networks generally lack such ability. This report connects a series of our work for compositional generalization, and summarizes an approach. The first part contains concepts and properties. The second part looks into a machine learning approach. The approach uses architecture design and regularization to regulate information of representations. This report focuses on basic ideas with intuitive and illustrative explanations. We hope this work would be helpful to clarify fundamentals of compositional generalization and lead to advance artificial intelligence.

Lei Wang, Wenbin Huang, Yuanpeng Li et al.

Predicting and modeling human behavior and finding trends within human decision-making processes is a major problem of social science. Rock Paper Scissors (RPS) is the fundamental strategic question in many game theory problems and real-world competitions. Finding the right approach to beat a particular human opponent is challenging. Here we use an AI (artificial intelligence) algorithm based on Markov Models of one fixed memory length (abbreviated as "single AI") to compete against humans in an iterated RPS game. We model and predict human competition behavior by combining many Markov Models with different fixed memory lengths (abbreviated as "multi-AI"), and develop an architecture of multi-AI with changeable parameters to adapt to different competition strategies. We introduce a parameter called "focus length" (a positive number such as 5 or 10) to control the speed and sensitivity for our multi-AI to adapt to the opponent's strategy change. The focus length is the number of previous rounds that the multi-AI should look at when determining which Single-AI has the best performance and should choose to play for the next game. We experimented with 52 different people, each playing 300 rounds continuously against one specific multi-AI model, and demonstrated that our strategy could win against more than 95% of human opponents.

Yuanpeng Li, Liang Zhao, Jianyu Wang et al.

Compositional generalization is a basic mechanism in human language learning, but current neural networks lack such ability. In this paper, we conduct fundamental research for encoding compositionality in neural networks. Conventional methods use a single representation for the input sentence, making it hard to apply prior knowledge of compositionality. In contrast, our approach leverages such knowledge with two representations, one generating attention maps, and the other mapping attended input words to output symbols. We reduce the entropy in each representation to improve generalization. Our experiments demonstrate significant improvements over the conventional methods in five NLP tasks including instruction learning and machine translation. In the SCAN domain, it boosts accuracies from 14.0% to 98.8% in Jump task, and from 92.0% to 99.7% in TurnLeft task. It also beats human performance on a few-shot learning task. We hope the proposed approach can help ease future research towards human-level compositional language learning.

Yuanpeng Li, Yi Yang, Jianyu Wang et al.

Acquiring a large vocabulary is an important aspect of human intelligence. Onecommon approach for human to populating vocabulary is to learn words duringreading or listening, and then use them in writing or speaking. This ability totransfer from input to output is natural for human, but it is difficult for machines.Human spontaneously performs this knowledge transfer in complicated multimodaltasks, such as Visual Question Answering (VQA). In order to approach human-levelArtificial Intelligence, we hope to equip machines with such ability. Therefore, toaccelerate this research, we propose a newzero-shot transfer VQA(ZST-VQA)dataset by reorganizing the existing VQA v1.0 dataset in the way that duringtraining, some words appear only in one module (i.e. questions) but not in theother (i.e. answers). In this setting, an intelligent model should understand andlearn the concepts from one module (i.e. questions), and at test time, transfer themto the other (i.e. predict the concepts as answers). We conduct evaluation on thisnew dataset using three existing state-of-the-art VQA neural models. Experimentalresults show a significant drop in performance on this dataset, indicating existingmethods do not address the zero-shot transfer problem. Besides, our analysis findsthat this may be caused by the implicit bias learned during training.