arXiv subjects

Zhengzhou Zhu

Publications and source records attributed to Zhengzhou Zhu.

6 recordsLinked to original sources

arXiv2026

MathDebugger: Detecting and Diagnosing Errors in Synthetic Mathematical Data

Synthetic mathematical data has become an important resource for scaling the reasoning capabilities of large language models, yet errors in generated questions and solutions can substantially undermine its value. We introduce MathDebugger, a type-aware benchmark for evaluating whether models can detect and diagnose errors in synthetic mathematical data. MathDebugger contains 2,000 correct questions, 2,000 erroneous questions balanced across four error types, and 2,000 annotated solutions, including 610 erroneous solutions spanning three error types. Each instance is manually verified and labeled for correctness, with erroneous instances further assigned a fine-grained error category. Human annotation achieves substantial to near-perfect agreement, with per-type Fleiss kappa ranging from 0.69 to 0.91. We evaluate 14 representative large language models and three process reward models. The results show that even strong reasoning models remain far from saturating MathDebugger, particularly when identifying fine-grained error types. We also uncover a consistent solving-verification gap: models specialized for mathematical reasoning or long-form reasoning do not necessarily outperform their general-purpose counterparts when auditing mathematical data. Finally, we show that explicit error-type information provides actionable guidance for correcting erroneous questions and solutions, yielding consistent and statistically significant improvements. MathDebugger provides a practical benchmark for developing more reliable mathematical data synthesis and quality-control pipelines.

cs.CL↗

arXiv2026

PATS: Policy-Aware Training Scaffolding for Agentic Reinforcement Learning

In long-horizon LLM agent reinforcement learning, weak policies often repeat similar failures, producing uninformative rollout trajectories and limiting effective policy optimization. Existing skill-centric methods improve exploration by optimizing, filtering, or internalizing reusable skills. However, they remain centered on the skills themselves rather than being designed as adaptive training-time support for the evolving policy. To address this, we propose a policy-centric training paradigm that reframes skills as a dynamic training scaffold. Our framework, PATS, converts rollout groups from the latest policy into evidence cards and uses task-specific evaluation to adjust the context used in subsequent rollouts. Concrete guidance helps weak policies to complete challenging tasks. As policy improves, redundant context is revised or removed to reduce reliance on explicit guidance while preserving useful rollout variation. The policy is optimized with environmental rewards using standard RLVR, and the training scaffold is discarded at deployment. Across ALFWorld, WebShop, and seven search-augmented QA benchmarks, PATS achieves performance competitive with SOTA baselines while using 25%-50% fewer tokens.

cs.AI↗

arXiv2026

FoE: Forest of Errors Makes the First Solution the Best in Large Reasoning Models

Recent Large Reasoning Models (LRMs) like DeepSeek-R1 have demonstrated remarkable success in complex reasoning tasks, exhibiting human-like patterns in exploring multiple alternative solutions. Upon closer inspection, however, we uncover a surprising phenomenon: The First is The Best, where alternative solutions are not merely suboptimal but potentially detrimental. This observation challenges widely accepted test-time scaling laws, leading us to hypothesize that errors within the reasoning path scale concurrently with test time. Through comprehensive empirical analysis, we characterize errors as a forest-structured Forest of Errors (FoE) and conclude that FoE makes the First the Best, which is underpinned by rigorous theoretical analysis. Leveraging these insights, we propose RED, a self-guided efficient reasoning framework comprising two components: I) Refining First, which suppresses FoE growth in the first solution; and II) Discarding Subs, which prunes subsequent FoE via dual-consistency. Extensive experiments across five benchmarks and six backbone models demonstrate that RED outperforms eight competitive baselines, achieving performance gains of up to 19.0% while reducing token consumption by 37.7% ~ 70.4%. Moreover, comparative experiments on FoE metrics shed light on how RED achieves effectiveness.

cs.AI↗

arXiv2026

Let's Verify Math Questions Step by Step

Large Language Models (LLMs) have recently achieved remarkable progress in mathematical reasoning. To enable such capabilities, many existing works distill strong reasoning models into long chains of thought or design algorithms to construct high-quality math question-answer (QA) data for training. However, these efforts primarily focus on generating correct reasoning paths and answers, while largely overlooking the correctness of the questions themselves. In this work, we present ValiMath, a benchmark consisting of 2147 human-verified mathematical questions covering a wide range of domains such as arithmetic, algebra, and geometry, which are synthesized and curated from the NuminaMath dataset. Each question is annotated with its logical structure, domain coverage, and question correctness, enabling fine-grained evaluation of question quality. ValiMath serves as a high-quality gold-standard test set for validating mathematical questions in LLM training corpora. Building upon this benchmark, we further propose MathQ-Verify, a pipeline that performs fine-grained parsing of mathematical questions into atomic assumptions and conclusions, and evaluates their semantic soundness through consistency checks. This pipeline achieves high precision in detecting flawed questions and provides a reliable foundation for cleaning noisy mathematical datasets. Experiments show that MathQ-Verify achieves state-of-the-art performance across multiple benchmarks, improving the F1 score by up to 25 percentage points over the direct verification baseline. MathQ-Verify offers a scalable and accurate solution for curating reliable mathematical datasets, reducing label noise and avoiding unnecessary computation on invalid questions. Our code and data are available at the repository https://github.com/OpenDCAI/MathQ-Verify.

cs.CL↗

arXiv2025

AIComposer: Any Style and Content Image Composition via Feature Integration

Image composition has advanced significantly with large-scale pre-trained T2I diffusion models. Despite progress in same-domain composition, cross-domain composition remains under-explored. The main challenges are the stochastic nature of diffusion models and the style gap between input images, leading to failures and artifacts. Additionally, heavy reliance on text prompts limits practical applications. This paper presents the first cross-domain image composition method that does not require text prompts, allowing natural stylization and seamless compositions. Our method is efficient and robust, preserving the diffusion prior, as it involves minor steps for backward inversion and forward denoising without training the diffuser. Our method also uses a simple multilayer perceptron network to integrate CLIP features from foreground and background, manipulating diffusion with a local cross-attention strategy. It effectively preserves foreground content while enabling stable stylization without a pre-stylization network. Finally, we create a benchmark dataset with diverse contents and styles for fair evaluation, addressing the lack of testing datasets for cross-domain image composition. Our method outperforms state-of-the-art techniques in both qualitative and quantitative evaluations, significantly improving the LPIPS score by 30.5% and the CSD metric by 18.1%. We believe our method will advance future research and applications. Code and benchmark at https://github.com/sherlhw/AIComposer.

cs.CV↗

arXiv2020

A Pairwise Probe for Understanding BERT Fine-Tuning on Machine Reading Comprehension

Pre-trained models have brought significant improvements to many NLP tasks and have been extensively analyzed. But little is known about the effect of fine-tuning on specific tasks. Intuitively, people may agree that a pre-trained model already learns semantic representations of words (e.g. synonyms are closer to each other) and fine-tuning further improves its capabilities which require more complicated reasoning (e.g. coreference resolution, entity boundary detection, etc). However, how to verify these arguments analytically and quantitatively is a challenging task and there are few works focus on this topic. In this paper, inspired by the observation that most probing tasks involve identifying matched pairs of phrases (e.g. coreference requires matching an entity and a pronoun), we propose a pairwise probe to understand BERT fine-tuning on the machine reading comprehension (MRC) task. Specifically, we identify five phenomena in MRC. According to pairwise probing tasks, we compare the performance of each layer's hidden representation of pre-trained and fine-tuned BERT. The proposed pairwise probe alleviates the problem of distraction from inaccurate model training and makes a robust and quantitative comparison. Our experimental analysis leads to highly confident conclusions: (1) Fine-tuning has little effect on the fundamental and low-level information and general semantic tasks. (2) For specific abilities required for downstream tasks, fine-tuned BERT is better than pre-trained BERT and such gaps are obvious after the fifth layer.

cs.CL↗