Yaxuan Wang

Author directory

2026

Multimodal large language models show significant potential in mathematical reasoning, yet they remain constrained by a scarcity of high-quality training data for visually grounded problems. Existing approaches commonly suffer from narrow coverage, image-text inconsistency, and heavy imitation of seed problems. To address these limitations, this paper proposes AgSynth, a multi-agent collaborative reverse synthesis framework. It first constructs solution logic programmatically and renders the corresponding images, then generates problem statements backward from the visual facts, thereby embedding answer correctness and image-text consistency directly into the generation pipeline. Meanwhile, AgSynth distills the core competencies tested in seed problems and performs divergent reconstruction, enabling synthesized problems to be substantially novel in both structure and visual presentation. Based on this framework, we construct the AgSynth-10K dataset, comprising 10,477 multimodal mathematical reasoning problems that have undergone two-tier quality control, spanning four major domains: geometry, algebra, combinatorics, and number theory. Experiments demonstrate that reinforcement learning training on AgSynth-10K significantly improves model performance on reasoning-intensive benchmarks (WeMath +7.42, LogicVista+9.15), validating the effectiveness of this traceable synthesis paradigm. The framework is publicly available at https://github.com/Yaxuan7690/AgSynth.
The rapid advancement of large language models (LLMs) has led to growing interest in using synthetic data to train future models. However, this creates a self-consuming retraining loop, where models are trained on their own outputs and may cause performance drops and induce emerging biases. In real-world applications, previously deployed LLMs may influence the data they generate, leading to a dynamic system driven by user feedback. For example, if a model continues to underserve users from a group, less query data will be collected from this particular demographic of users. In this study, we introduce the concept of Self-Consuming Performative Loop (SCPL) and investigate the role of synthetic data in shaping bias during these dynamic iterative training processes under controlled performative feedback. This controlled setting is motivated by the inaccessibility of real-world user preference data from dynamic production systems, and enables us to isolate and analyze feedback-driven bias evolution in a principled manner. We focus on two types of loops, including the typical retraining setting and the incremental fine-tuning setting, which is largely underexplored. Through experiments on three real-world tasks, we find that the performative loop increases preference bias and decreases disparate bias. We design a reward-based rejection sampling strategy to mitigate the bias, moving towards more trustworthy self-improving systems. The code is available at https://github.com/UCSC-REAL/SCPL.git.