arXiv subjects

Caesar Wu

Publications and source records attributed to Caesar Wu.

17 recordsLinked to original sources

arXiv2026

Beyond Compression: Diagnosing How Post-Training Changes Mathematical Reasoning

Post-training is central to mathematical reasoning in modern large language models (LLMs), but endpoint pass@1 alone underidentifies what has changed. Gains may reflect newly reachable solutions, cheaper sampling of latent solutions, surface robustness, or memorisation. We compare three post-training paths under a common diagnostic readout: our sufficiently trained off-policy distillation trajectories, released Qwen3 off-policy-plus-on-policy distillation endpoints, and a released DeepSeek-Math endpoint trained with Group Relative Policy Optimisation (GRPO). Our probe uses cross-surface pass@K over verbatim prompts, paraphrases, numerical isomorphisms, and translations, plus consistency, distribution-shape, and verified supervised-fine-tuning (SFT) membership analyses. We find two regimes. On easier AMC problems, large-K ceilings are near saturation, so post-training mainly compresses sample cost. On harder AIME problems, post-training expands the large-K ceiling over the base model: sufficient off-policy distillation already raises this ceiling, Qwen3 released endpoints raise it further, and DeepSeek-Math GRPO does not dominate sufficient off-policy distillation at large K. English-dominant distillation improves non-English reasoning but preserves language-tier gaps. A controlled-overfit audit finds limited sensitivity in current SFT-membership probes. Compression is one regime of post-training, not a universal explanation.

cs.LG↗

arXiv2026

What Does Post-Training Change in Multilingual Reasoning?

Open-source reasoning models provide unequal access to reasoning capability across languages. When a model can solve a problem but cannot deliver a complete solution in the user's language, language becomes an access barrier rather than merely a source of performance variation. We audit Qwen3 checkpoints on competition-mathematics tasks in eleven languages. Across the ten non-English languages, only 15.4-17.9% of problems receive a correct, terminating solution with visible reasoning in the requested language in any of 16 samples, compared with 92.9% in English. To identify the source of this disparity, we evaluate thirteen endpoints from one model family, spanning released checkpoints, multilingual supervised fine-tuning (SFT) at two scales, controlled SFT ablations, and three reinforcement-learning (RL) reward formulations. We jointly track correctness, language adherence, termination, and delivery efficiency. The dominant bottleneck shifts across post-training stages. Released models often reason in English. Multilingual SFT restores target-language reasoning, but accuracy declines across multilingual, English-only, and single-language SFT runs, showing that this cost is not specific to multilingual mixing; non-English reasoning traces additionally become prone to non-terminating loops. RL restores termination in both arms at no cost in accuracy, but only the arm whose reward includes a language term delivers: rewarding correctness alone returns the model to English. Together, these stages establish a constructive post-training path from English-pivoted capability to multilingual reasoning that is reliably delivered.

cs.CL↗

arXiv2026

Unlocking the Critic: Reward-Free Policy Optimization for LLM Post-Training

Recent approaches to reinforcement learning (RL) post-training for large language models increasingly remove the critic to reduce training instability and memory overhead. Even where a critic is trained, it is discarded once training ends, although it has learned to predict outcomes. We revisit this trend and show that a pretrained critic's ability to predict future outcomes can make it a valuable asset for efficient long-horizon reasoning. First, we find that instability in critic-based RL for long chain-of-thought reasoning is largely an optimization artifact: keeping policy updates small and low in variance restores stable convergence. Second, a well-pretrained critic estimates the posterior probability of eventual success from later trajectory states and unfinished prefixes. Its predictions provide outcome-derived, dense, per-prefix learning signals that, during policy optimization, require neither completed rollouts, step-level annotations, nor external reward labels. Building on this insight, we introduce Reward-Free Policy Optimization (RFPO), which repurposes a single calibrated, frozen critic as a rollout-level reward, a value baseline for generalized advantage estimation, and a success forecaster for unfinished prefixes. We further show that binarizing the debiased score stops the policy from exploiting the critic's length bias. Binarized, RFPO matches supervised PPO without a single label in the training loop, while cutting compute and memory overhead. This makes RFPO well suited to long-horizon reasoning tasks, where outcomes arrive late and generation dominates cost: because rollouts can be rewarded before they finish, training no longer has to pay for waiting on every trajectory to complete. Our findings challenge the prevailing critic-free paradigm and establish critic-based, reward-free optimization as a scalable and computationally efficient path for LLM post-training.

cs.LG↗

arXiv2026

Solving Without Stopping: On-Policy Distillation at Small Scale

On-policy distillation, where a student learns from a stronger teacher's feedback on its own outputs, is a common way to pass reasoning to smaller models. We analyze what it transfers at small scale, distilling Qwen3-8B into Qwen3 4B, 1.7B and 0.6B students, in thinking mode (reason at length, then end the reasoning and answer) and, for comparison, in non-thinking mode (no separate reasoning phase). Long reasoning needs two abilities, solving a problem and knowing when it is solved, and we find that distillation transfers the first, but in thinking mode not the second. Solving improves at every size, up to two ceilings, which we measure comprehensively across both modes and all student sizes: a student's single attempt never exceeds what it could already reach in many attempts before training, and the smaller the student, the further it stays below the teacher. Stopping is where the modes part. In non-thinking mode every student keeps stopping; in thinking mode students stop ending their reasoning early in training, and the smaller the student, the less of this ability survives: the teacher signals a stop almost only where a student already ends its reasoning, so distillation teaches no new stops; it only keeps the student's existing stops that land on a right answer, and a weak student has few such stops. The smallest students often reach the right value but do not commit to it: they either rarely mark it or mark it and write past it. Together, these results describe how small students behave under on-policy distillation, and a diagnostic that separates answer marking, correctness and stopping.

cs.AI↗

arXiv2026

Reward-Guided Autoregressive Graph Generation for Efficient Multi-Agent Communication Topology Design

LLM-based Multi-Agent Systems (MAS) achieve strong performance on complex reasoning tasks by coordinating multiple agents, but at the cost of substantial token consumption. Recent work on automatic topology design, ARG-Designer, has reframed this problem as autoregressive graph generation. However, its training objective provides no explicit incentive for the model to generate sparse and efficient topologies. We address this limitation by introducing a Reward-Guided Autoregressive Graph Generation (RGA-Designer) inspired by Reinforcement Learning from Human Feedback (RLHF). We train a reward model that jointly captures task correctness and structural compactness, and then fine-tune the pretrained graph generator using the reward model as feedback. Our method preserves task accuracy at the level of ARG-Designer while reducing token consumption by an average of 20.5%.

cs.MA↗

arXiv2025

Trustworthy Efficient Communication for Distributed Learning using LQ-SGD Algorithm

We propose LQ-SGD (Low-Rank Quantized Stochastic Gradient Descent), an efficient communication gradient compression algorithm designed for distributed training. LQ-SGD further develops on the basis of PowerSGD by incorporating the low-rank approximation and log-quantization techniques, which drastically reduce the communication overhead, while still ensuring the convergence speed of training and model accuracy. In addition, LQ-SGD and other compression-based methods show stronger resistance to gradient inversion than traditional SGD, providing a more robust and efficient optimization path for distributed learning systems.

cs.LG↗

arXiv2025

Trustworthiness of Stochastic Gradient Descent in Distributed Learning

Distributed learning (DL) uses multiple nodes to accelerate training, enabling efficient optimization of large-scale models. Stochastic Gradient Descent (SGD), a key optimization algorithm, plays a central role in this process. However, communication bottlenecks often limit scalability and efficiency, leading to increasing adoption of compressed SGD techniques to alleviate these challenges. Despite addressing communication overheads, compressed SGD introduces trustworthiness concerns, as gradient exchanges among nodes are vulnerable to attacks like gradient inversion (GradInv) and membership inference attacks (MIA). The trustworthiness of compressed SGD remains unexplored, leaving important questions about its reliability unanswered. In this paper, we provide a trustworthiness evaluation of compressed versus uncompressed SGD. Specifically, we conducted empirical studies using GradInv attacks, revealing that compressed SGD demonstrates significantly higher resistance to privacy leakage compared to uncompressed SGD. In addition, our findings suggest that MIA may not be a reliable metric for assessing privacy risks in distributed learning.

cs.LG↗

arXiv2025

Lightweight Trustworthy Distributed Clustering

Ensuring data trustworthiness within individual edge nodes while facilitating collaborative data processing poses a critical challenge in edge computing systems (ECS), particularly in resource-constrained scenarios such as autonomous systems sensor networks, industrial IoT, and smart cities. This paper presents a lightweight, fully distributed k-means clustering algorithm specifically adapted for edge environments, leveraging a distributed averaging approach with additive secret sharing, a secure multiparty computation technique, during the cluster center update phase to ensure the accuracy and trustworthiness of data across nodes.

cs.DC↗

arXiv2025

A Unified Hyperparameter Optimization Pipeline for Transformer-Based Time Series Forecasting Models

Transformer-based models for time series forecasting (TSF) have attracted significant attention in recent years due to their effectiveness and versatility. However, these models often require extensive hyperparameter optimization (HPO) to achieve the best possible performance, and a unified pipeline for HPO in transformer-based TSF remains lacking. In this paper, we present one such pipeline and conduct extensive experiments on several state-of-the-art (SOTA) transformer-based TSF models. These experiments are conducted on standard benchmark datasets to evaluate and compare the performance of different models, generating practical insights and examples. Our pipeline is generalizable beyond transformer-based architectures and can be applied to other SOTA models, such as Mamba and TimeMixer, as demonstrated in our experiments. The goal of this work is to provide valuable guidance to both industry practitioners and academic researchers in efficiently identifying optimal hyperparameters suited to their specific domain applications. The code and complete experimental results are available on GitHub.

cs.LG↗

arXiv2024

Dataset | Mindset = Explainable AI | Interpretable AI

We often use "explainable" Artificial Intelligence (XAI)" and "interpretable AI (IAI)" interchangeably when we apply various XAI tools for a given dataset to explain the reasons that underpin machine learning (ML) outputs. However, these notions can sometimes be confusing because interpretation often has a subjective connotation, while explanations lean towards objective facts. We argue that XAI is a subset of IAI. The concept of IAI is beyond the sphere of a dataset. It includes the domain of a mindset. At the core of this ambiguity is the duality of reasons, in which we can reason either outwards or inwards. When directed outwards, we want the reasons to make sense through the laws of nature. When turned inwards, we want the reasons to be happy, guided by the laws of the heart. While XAI and IAI share reason as the common notion for the goal of transparency, clarity, fairness, reliability, and accountability in the context of ethical AI and trustworthy AI (TAI), their differences lie in that XAI emphasizes the post-hoc analysis of a dataset, and IAI requires a priori mindset of abstraction. This hypothesis can be proved by empirical experiments based on an open dataset and harnessed by High-Performance Computing (HPC). The demarcation of XAI and IAI is indispensable because it would be impossible to determine regulatory policies for many AI applications, especially in healthcare, human resources, banking, and finance. We aim to clarify these notions and lay the foundation of XAI, IAI, EAI, and TAI for many practitioners and policymakers in future AI applications and research.

cs.AI↗

arXiv2024

Survey and Taxonomy: The Role of Data-Centric AI in Transformer-Based Time Series Forecasting

Alongside the continuous process of improving AI performance through the development of more sophisticated models, researchers have also focused their attention to the emerging concept of data-centric AI, which emphasizes the important role of data in a systematic machine learning training process. Nonetheless, the development of models has also continued apace. One result of this progress is the development of the Transformer Architecture, which possesses a high level of capability in multiple domains such as Natural Language Processing (NLP), Computer Vision (CV) and Time Series Forecasting (TSF). Its performance is, however, heavily dependent on input data preprocessing and output data evaluation, justifying a data-centric approach to future research. We argue that data-centric AI is essential for training AI models, particularly for transformer-based TSF models efficiently. However, there is a gap regarding the integration of transformer-based TSF and data-centric AI. This survey aims to pin down this gap via the extensive literature review based on the proposed taxonomy. We review the previous research works from a data-centric AI perspective and we intend to lay the foundation work for the future development of transformer-based architecture and data-centric AI.

cs.LG↗

arXiv2024

Transformer Multivariate Forecasting: Less is More?

In the domain of multivariate forecasting, transformer models stand out as powerful apparatus, displaying exceptional capabilities in handling messy datasets from real-world contexts. However, the inherent complexity of these datasets, characterized by numerous variables and lengthy temporal sequences, poses challenges, including increased noise and extended model runtime. This paper focuses on reducing redundant information to elevate forecasting accuracy while optimizing runtime efficiency. We propose a novel transformer forecasting framework enhanced by Principal Component Analysis (PCA) to tackle this challenge. The framework is evaluated by five state-of-the-art (SOTA) models and four diverse real-world datasets. Our experimental results demonstrate the framework's ability to minimize prediction errors across all models and datasets while significantly reducing runtime. From the model perspective, one of the PCA-enhanced models: PCA+Crossformer, reduces mean square errors (MSE) by 33.3% and decreases runtime by 49.2% on average. From the dataset perspective, the framework delivers 14.3% MSE and 76.6% runtime reduction on Electricity datasets, as well as 4.8% MSE and 86.9% runtime reduction on Traffic datasets. This study aims to advance various SOTA models and enhance transformer-based time series forecasting for intricate data. Code is available at: https://github.com/jingjing-unilu/PCA_Transformer.

cs.LG↗

arXiv2023

Trustworthy AI: Deciding What to Decide

When engaging in strategic decision-making, we are frequently confronted with overwhelming information and data. The situation can be further complicated when certain pieces of evidence contradict each other or become paradoxical. The primary challenge is how to determine which information can be trusted when we adopt Artificial Intelligence (AI) systems for decision-making. This issue is known as deciding what to decide or Trustworthy AI. However, the AI system itself is often considered an opaque black box. We propose a new approach to address this issue by introducing a novel framework of Trustworthy AI (TAI) encompassing three crucial components of AI: representation space, loss function, and optimizer. Each component is loosely coupled with four TAI properties. Altogether, the framework consists of twelve TAI properties. We aim to use this framework to conduct the TAI experiments by quantitive and qualitative research methods to satisfy TAI properties for the decision-making context. The framework allows us to formulate an optimal prediction model trained by the given dataset for applying the strategic investment decision of credit default swaps (CDS) in the technology sector. Finally, we provide our view of the future direction of TAI research

cs.AI↗

arXiv2023

Survey of Trustworthy AI: A Meta Decision of AI

When making strategic decisions, we are often confronted with overwhelming information to process. The situation can be further complicated when some pieces of evidence are contradicted each other or paradoxical. The challenge then becomes how to determine which information is useful and which ones should be eliminated. This process is known as meta-decision. Likewise, when it comes to using Artificial Intelligence (AI) systems for strategic decision-making, placing trust in the AI itself becomes a meta-decision, given that many AI systems are viewed as opaque "black boxes" that process large amounts of data. Trusting an opaque system involves deciding on the level of Trustworthy AI (TAI). We propose a new approach to address this issue by introducing a novel taxonomy or framework of TAI, which encompasses three crucial domains: articulate, authentic, and basic for different levels of trust. To underpin these domains, we create ten dimensions to measure trust: explainability/transparency, fairness/diversity, generalizability, privacy, data governance, safety/robustness, accountability, reproducibility, reliability, and sustainability. We aim to use this taxonomy to conduct a comprehensive survey and explore different TAI approaches from a strategic decision-making perspective.

cs.AI↗

arXiv2023

The Emerging Artificial Intelligence Protocol for Hierarchical Information Network

The recent development of artificial intelligence enables a machine to achieve a human level of intelligence. Problem-solving and decision-making are two mental abilities to measure human intelligence. Many scholars have proposed different models. However, there is a gap in establishing an AI-oriented hierarchical model with a multilevel abstraction. This study proposes a novel model known as the emerged AI protocol that consists of seven distinct layers capable of providing an optimal and explainable solution for a given problem.

cs.AI↗

arXiv2022

Strategic Decisions Survey, Taxonomy, and Future Directions from Artificial Intelligence Perspective

Strategic Decision-Making is always challenging because it is inherently uncertain, ambiguous, risky, and complex. It is the art of possibility. We develop a systematic taxonomy of decision-making frames that consists of 6 bases, 18 categorical, and 54 frames. We aim to lay out the computational foundation that is possible to capture a comprehensive landscape view of a strategic problem. Compared with traditional models, it covers irrational, non-rational and rational frames c dealing with certainty, uncertainty, complexity, ambiguity, chaos, and ignorance.

cs.AI↗

arXiv2016

Big Data Analytics = Machine Learning + Cloud Computing

Big Data can mean different things to different people. The scale and challenges of Big Data are often described using three attributes, namely Volume, Velocity and Variety (3Vs), which only reflect some of the aspects of data. In this chapter we review historical aspects of the term "Big Data" and the associated analytics. We augment 3Vs with additional attributes of Big Data to make it more comprehensive and relevant. We show that Big Data is not just 3Vs, but 32 Vs, that is, 9 Vs covering the fundamental motivation behind Big Data, which is to incorporate Business Intelligence (BI) based on different hypothesis or statistical models so that Big Data Analytics (BDA) can enable decision makers to make useful predictions for some crucial decisions or researching results. History of Big Data has demonstrated that the most cost effective way of performing BDA is to employ Machine Learning (ML) on the Cloud Computing (CC) based infrastructure or simply, ML + CC -> BDA. This chapter is devoted to help decision makers by defining BDA as a solution and opportunity to address their business needs.

cs.DC↗