Han Shi

Jingyao Li, Senqiao Yang, Sitong Wu et al.

In recent years, developing compact and efficient large language models (LLMs) has emerged as a thriving area of research. Traditional Supervised Fine-Tuning (SFT), which relies on singular ground truth labels, often fails to capture token-level dependencies and linguistic diversity. To address these limitations, we propose a logits-based fine-tuning framework that integrates the strengths of supervised learning and knowledge distillation. Our approach constructs enriched training targets by combining teacher logits with ground truth labels, preserving both correctness and linguistic diversity. This ensures more reliable and effective training. We constructed a large-scale 1.2M logits dataset and trained a series of science-focused models. Experimental results demonstrate that our method achieves significant improvements, with accuracy gains of 18% on Mawps and 22.7% on TabMWP. Across nine widely used mathematical benchmarks, our method consistently outperforms prior SFT models, achieving an average improvement of 7.28%. Codes are available at https://github.com/dvlab-research/Logits-Based-Finetuning.

Haoyue Dai, Rui Ding, Yuanyuan Jiang et al.

Supervised Causal Learning (SCL) aims to learn causal relations from observational data by accessing previously seen datasets associated with ground truth causal relations. This paper presents a first attempt at addressing a fundamental question: What are the benefits from supervision and how does it benefit? Starting from seeing that SCL is not better than random guessing if the learning target is non-identifiable a priori, we propose a two-phase paradigm for SCL by explicitly considering structure identifiability. Following this paradigm, we tackle the problem of SCL on discrete data and propose ML4C. The core of ML4C is a binary classifier with a novel learning target: it classifies whether an Unshielded Triple (UT) is a v-structure or not. Specifically, starting from an input dataset with the corresponding skeleton provided, ML4C orients each UT once it is classified as a v-structure. These v-structures are together used to construct the final output. To address the fundamental question of SCL, we propose a principled method for ML4C featurization: we exploit the vicinity of a given UT (i.e., the neighbors of UT in skeleton), and derive features by considering the conditional dependencies and structural entanglement within the vicinity. We further prove that ML4C is asymptotically correct. Last but foremost, thorough experiments conducted on benchmark datasets demonstrate that ML4C remarkably outperforms other state-of-the-art algorithms in terms of accuracy, reliability, robustness and tolerance. In summary, ML4C shows promising results on validating the effectiveness of supervision for causal learning. Our codes are publicly available at https://github.com/microsoft/ML4C.

Zhihan Guo, Feiyang Xu, Yifan Li et al.

The explosive growth in academic literature necessitates automated deep research (DR) agents, yet their evaluation remains a significant challenge. First, existing benchmarks often focus narrowly on retrieval while neglecting high-level planning and reasoning. Second, existing benchmarks favor general domains over the scientific domains that are the core application for DR agents. To address these gaps, we introduce Dr.Mi-Bench, a Modular-integrated benchmark for scientific DR agents. Grounded in academic literature, our benchmark uses a human-annotated dataset of 200 instances across 10 scientific domains, including both research and review papers. Besides, we also propose a Modular-integrated Evaluation Paradigm for DR Agents (Dr.Mi-Eval), a novel modular-integrated evaluation paradigm, which leverages the rich structure of academic papers to assess the core competencies of planning, retrieval, and reasoning through two complementary modes: an end-to-end evaluation for DR agents and an isolated evaluation for foundational LLMs as potential backbones. Experimental results reveal a fragmented performance landscape: agents exhibit specialized strengths but share critical weaknesses, most notably in performing the multi-source retrieval required for review-style tasks and performing consistently across diverse scientific fields. Moreover, improving high-level planning capability is the crucial factor for unlocking the reasoning potential of foundational LLMs as backbones. By exposing these actionable failure modes, Dr.Mi-Bench provides a diagnostic tool to guide the development of more reliable academic research assistants.

Yihang Gao, Chuanyang Zheng, Enze Xie et al.

Besides natural language processing, transformers exhibit extraordinary performance in solving broader applications, including scientific computing and computer vision. Previous works try to explain this from the expressive power and capability perspectives that standard transformers are capable of performing some algorithms. To empower transformers with algorithmic capabilities and motivated by the recently proposed looped transformer, we design a novel transformer framework, dubbed Algorithm Transformer (abbreviated as AlgoFormer). We provide an insight that efficient transformer architectures can be designed by leveraging prior knowledge of tasks and the underlying structure of potential algorithms. Compared with the standard transformer and vanilla looped transformer, the proposed AlgoFormer can perform efficiently in algorithm representation in some specific tasks. In particular, inspired by the structure of human-designed learning algorithms, our transformer framework consists of a pre-transformer that is responsible for task preprocessing, a looped transformer for iterative optimization algorithms, and a post-transformer for producing the desired results after post-processing. We provide theoretical evidence of the expressive power of the AlgoFormer in solving some challenging problems, mirroring human-designed algorithms. Furthermore, some theoretical and empirical results are presented to show that the designed transformer has the potential to perform algorithm representation and learning. Experimental results demonstrate the empirical superiority of the proposed transformer in that it outperforms the standard transformer and vanilla looped transformer in some specific tasks. An extensive experiment on real language tasks (e.g., neural machine translation of German and English, and text classification) further validates the expressiveness and effectiveness of AlgoFormer.

Yihan Cao, Xu Chen, Lun Du et al.

Person-job fit is an essential part of online recruitment platforms in serving various downstream applications like Job Search and Candidate Recommendation. Recently, pretrained large language models have further enhanced the effectiveness by leveraging richer textual information in user profiles and job descriptions apart from user behavior features and job metadata. However, the general domain-oriented design struggles to capture the unique structural information within user profiles and job descriptions, leading to a loss of latent semantic correlations. We propose TAROT, a hierarchical multitask co-pretraining framework, to better utilize structural and semantic information for informative text embeddings. TAROT targets semi-structured text in profiles and jobs, and it is co-pretained with multi-grained pretraining tasks to constrain the acquired semantic information at each level. Experiments on a real-world LinkedIn dataset show significant performance improvements, proving its effectiveness in person-job fit tasks.

Yao Teng, Fuyun Wang, Xian Liu et al.

As a new paradigm of visual content generation, autoregressive text-to-image models suffer from slow inference due to their sequential token-by-token decoding process, often requiring thousands of model forward passes to generate a single image. To address this inefficiency, we propose Speculative Jacobi-Denoising Decoding (SJD2), a framework that incorporates the denoising process into Jacobi iterations to enable parallel token generation in autoregressive models. Our method introduces a next-clean-token prediction paradigm that enables the pre-trained autoregressive models to accept noise-perturbed token embeddings and predict the next clean tokens through low-cost fine-tuning. This denoising paradigm guides the model towards more stable Jacobi trajectories. During inference, our method initializes token sequences with Gaussian noise and performs iterative next-clean-token-prediction in the embedding space. We employ a probabilistic criterion to verify and accept multiple tokens in parallel, and refine the unaccepted tokens for the next iteration with the denoising trajectory. Experiments show that our method can accelerate generation by reducing model forward passes while maintaining the visual quality of generated images.

Yingfei Xiong, Jie Wang, Runfa Yan et al.

Due to the difficulty of repairing defect, many research efforts have been devoted into automatic defect repair. Given a buggy program that fails some test cases, a typical automatic repair technique tries to modify the program to make all tests pass. However, since the test suites in real world projects are usually insufficient, aiming at passing the test suites often leads to incorrect patches. In this paper we aim to produce precise patches, that is, any patch we produce has a relatively high probability to be correct. More concretely, we focus on condition synthesis, which was shown to be able to repair more than half of the defects in existing approaches. Our key insight is threefold. First, it is important to know what variables in a local context should be used in an "if" condition, and we propose a sorting method based on the dependency relations between variables. Second, we observe that the API document can be used to guide the repair process, and propose document analysis technique to further filter the variables. Third, it is important to know what predicates should be performed on the set of variables, and we propose to mine a set of frequently used predicates in similar contexts from existing projects. We develop a novel program repair system, ACS, that could generate precise conditions at faulty locations. Furthermore, given the generated conditions are very precise, we can perform a repair operation that is previously deemed to be too overfitting: directly returning the test oracle to repair the defect. Using our approach, we successfully repaired 17 defects on four projects of Defects4J, which is the largest number of fully automatically repaired defects reported on the dataset so far. More importantly, the precision of our approach in the evaluation is 73.9%, which is significantly higher than previous approaches, which are usually less than 40%.