Shining Liang

Ning Wu, Ming Gong, Linjun Shou et al.

Text summarization has a wide range of applications in many scenarios. The evaluation of the quality of the generated text is a complex problem. A big challenge to language evaluation is that there is a clear divergence between existing metrics and human evaluation. A document summary's quality can be assessed by human annotators on various criteria, both objective ones like grammar and correctness, and subjective ones like informativeness, succinctness, and appeal. Most of the automatic evaluation methods like BLUE/ROUGE may be not able to adequately capture the above dimensions. In this paper, we propose a new evaluation framework based on LLMs, which provides a comprehensive evaluation framework by comparing generated text and reference text from both objective and subjective aspects. First, we propose to model objective and subjective dimensions of generated text based on roleplayers prompting mechanism. Furthermore, we introduce a context-based prompting mechanism that is able to generate dynamic roleplayer profiles based on input context. Finally, we design a multi-roleplayer prompting technology based on batch prompting and integrate multiple outputs into the final evaluation results. Experimental results on three real datasets for summarization show that our model is highly competitive and has a very high consistency with human annotators.

Shining Liang, Linjun Shou, Jian Pei et al.

Despite the great success of spoken language understanding (SLU) in high-resource languages, it remains challenging in low-resource languages mainly due to the lack of labeled training data. The recent multilingual code-switching approach achieves better alignments of model representations across languages by constructing a mixed-language context in zero-shot cross-lingual SLU. However, current code-switching methods are limited to implicit alignment and disregard the inherent semantic structure in SLU, i.e., the hierarchical inclusion of utterances, slots, and words. In this paper, we propose to model the utterance-slot-word structure by a multi-level contrastive learning framework at the utterance, slot, and word levels to facilitate explicit alignment. Novel code-switching schemes are introduced to generate hard negative examples for our contrastive learning framework. Furthermore, we develop a label-aware joint model leveraging label semantics to enhance the implicit alignment and feed to contrastive learning. Our experimental results show that our proposed methods significantly improve the performance compared with the strong baselines on two zero-shot cross-lingual SLU benchmark datasets.

Nuo Chen, Ning Wu, Shining Liang et al.

This paper presents an in-depth analysis of Large Language Models (LLMs), focusing on LLaMA, a prominent open-source foundational model in natural language processing. Instead of assessing LLaMA through its generative output, we design multiple-choice tasks to probe its intrinsic understanding in high-order tasks such as reasoning and computation. We examine the model horizontally, comparing different sizes, and vertically, assessing different layers. We unveil several key and uncommon findings based on the designed probing tasks: (1) Horizontally, enlarging model sizes almost could not automatically impart additional knowledge or computational prowess. Instead, it can enhance reasoning abilities, especially in math problem solving, and helps reduce hallucinations, but only beyond certain size thresholds; (2) In vertical analysis, the lower layers of LLaMA lack substantial arithmetic and factual knowledge, showcasing logical thinking, multilingual and recognitive abilities, with top layers housing most computational power and real-world knowledge.

Shiping Yang, Jie Wu, Wenbiao Ding et al.

Robustness has become a critical attribute for the deployment of RAG systems in real-world applications. Existing research focuses on robustness to explicit noise (e.g., document semantics) but overlooks spurious features (a.k.a. implicit noise). While previous works have explored spurious features in LLMs, they are limited to specific features (e.g., formats) and narrow scenarios (e.g., ICL). In this work, we statistically confirm the presence of spurious features in the RAG paradigm, a robustness problem caused by the sensitivity of LLMs to semantic-agnostic features. Moreover, we provide a comprehensive taxonomy of spurious features and empirically quantify their impact through controlled experiments. Further analysis reveals that not all spurious features are harmful and they can even be beneficial sometimes. Extensive evaluation results across multiple LLMs suggest that spurious features are a widespread and challenging problem in the field of RAG. The code and dataset will be released to facilitate future research. We release all codes and data at: $\\\href{https://github.com/maybenotime/RAG-SpuriousFeatures}{https://github.com/maybenotime/RAG-SpuriousFeatures}$.

Shangjian Yin, Shining Liang, Wenbiao Ding et al.

Reinforcement Learning from Human Feedback (RLHF) has become a cornerstone for aligning large language models (LLMs). However, its effectiveness depends on high-quality instruction data. Most existing alignment datasets are either private or require costly human annotation, which limits reproducibility and scalability. Even with Reinforcement Learning from AI Feedback (RLAIF), concerns about data quality remain. Moreover, it is unclear how much data is actually required to fine-tune a base model into a strong instruction-following model. Current approaches often rely on over 300k examples even at the supervised fine-tuning (SFT) stage, yet they still underperform compared to proprietary models, creating barriers for academic and resource-limited communities. To address this gap, we introduce PiKa, a data-efficient family of expert-level alignment datasets. In particular, the PiKa-SFT dataset uses only 30k SFT examples, far fewer than state-of-the-art datasets like Magpie. Through evaluations by fine-tuning Llama-3-8B-Base on PiKa and other public datasets, we show that PiKa-SFT outperforms models trained on much larger data. On AlpacaEval 2.0 and Arena-Hard benchmarks, PiKa-SFT fine-tuning even surpasses the official Llama-3-8B-Instruct model trained on over 10 million proprietary examples. We further extend our study by training the Qwen2.5 series (0.5B to 7B) on PiKa-SFT, achieving consistent gains. These findings demonstrate that high-quality alignment can be achieved with significantly less data, offering a scalable path for open-source LLM alignment. Code and data: https://github.com/SJY8460/PiKa.

Shining Liang, Ming Gong, Jian Pei et al.

Named entity recognition (NER) is a fundamental component in many applications, such as Web Search and Voice Assistants. Although deep neural networks greatly improve the performance of NER, due to the requirement of large amounts of training data, deep neural networks can hardly scale out to many languages in an industry setting. To tackle this challenge, cross-lingual NER transfers knowledge from a rich-resource language to languages with low resources through pre-trained multilingual language models. Instead of using training data in target languages, cross-lingual NER has to rely on only training data in source languages, and optionally adds the translated training data derived from source languages. However, the existing cross-lingual NER methods do not make good use of rich unlabeled data in target languages, which is relatively easy to collect in industry applications. To address the opportunities and challenges, in this paper we describe our novel practice in Microsoft to leverage such large amounts of unlabeled data in target languages in real production settings. To effectively extract weak supervision signals from the unlabeled data, we develop a novel approach based on the ideas of semi-supervised learning and reinforcement learning. The empirical study on three benchmark data sets verifies that our approach establishes the new state-of-the-art performance with clear edges. Now, the NER techniques reported in this paper are on their way to become a fundamental component for Web ranking, Entity Pane, Answers Triggering, and Question Answering in the Microsoft Bing search engine. Moreover, our techniques will also serve as part of the Spoken Language Understanding module for a commercial voice assistant. We plan to open source the code of the prototype framework after deployment.

Jie Deng, Hanshuang Tong, Jun Li et al.

Large language models (LLMs) have made impressive strides in mathematical reasoning, often fine-tuned using rejection sampling that retains only correct reasoning trajectories. While effective, this paradigm treats supervision as a binary filter that systematically excludes teacher-generated errors, leaving a gap in how reasoning failures are modeled during training. In this paper, we propose TrajFusion, a fine-tuning strategy that reframes rejection sampling as a structured supervision construction process. Specifically, TrajFusion forms fused trajectories that explicitly model trial-and-error reasoning by interleaving selected incorrect trajectories with reflection prompts and correct trajectories. The length of each fused sample is adaptively controlled based on the frequency and diversity of teacher errors, providing richer supervision for challenging problems while safely reducing to vanilla rejection sampling fine-tuning (RFT) when error signals are uninformative. TrajFusion requires no changes to the architecture or training objective. Extensive experiments across multiple math benchmarks demonstrate that TrajFusion consistently outperforms RFT, particularly on challenging and long-form reasoning problems.

Weihao Liu, Ning Wu, Shiping Yang et al.

Large Language Models (LLMs) frequently show distracted attention due to irrelevant information in the input, which severely impairs their long-context capabilities. Inspired by recent studies on the effectiveness of retrieval heads in long-context factutality, we aim at addressing this distraction issue through improving such retrieval heads directly. We propose Multi-Document Attention Focusing (MuDAF), a novel method that explicitly optimizes the attention distribution at the head level through contrastive learning. According to the experimental results, MuDAF can significantly improve the long-context question answering performance of LLMs, especially in multi-document question answering. Extensive evaluations on retrieval scores and attention visualizations show that MuDAF possesses great potential in making attention heads more focused on relevant information and reducing attention distractions.

Jie Deng, Shining Liang, Jun Li et al.

Large reasoning models (LRMs) typically solve reasoning-intensive tasks by generating long chain-of-thought (CoT) traces, leading to substantial inference overhead. We identify a reproducible inference-time phenomenon, termed Self-Compression: when multiple independent and answerable questions are presented within a single prompt, the model spontaneously produces shorter reasoning traces for each question. This phenomenon arises from multi-question contextual pressure during generation and consistently manifests across models and benchmarks. Building on this observation, we propose ConPress (Learning from Contextual Pressure), a lightweight self-supervised fine-tuning approach. ConPress constructs multi-question prompts to induce self-compression, samples the resulting model outputs, and parses and filters per-question traces to obtain concise yet correct reasoning trajectories. These trajectories are directly used for supervised fine-tuning, internalizing compressed reasoning behavior in single-question settings without external teachers, manual pruning, or reinforcement learning. With only 8k fine-tuning examples, ConPress reduces reasoning token usage by 59% on MATH500 and 33% on AIME25, while maintaining competitive accuracy.

Weihao Liu, Ning Wu, Wenbiao Ding et al.

Truthfulness stands out as an essential challenge for Large Language Models (LLMs). Although many works have developed various ways for truthfulness enhancement, they seldom focus on truthfulness in multilingual scenarios. Meanwhile, contemporary multilingual aligning technologies struggle to balance numerous languages and often exhibit serious truthfulness gaps across different languages, especially those that differ greatly from English. In our work, we extend truthfulness evaluation to multilingual contexts and propose a practical method for cross-lingual truthfulness transfer called Fact-aware Multilingual Selective Synergy (FaMSS). FaMSS is able to select an optimal subset of all tested languages by language bias and transfer contributions, and then employ translation instruction tuning for cross-lingual truthfulness transfer. Experimental results demonstrate that our approach can effectively reduce the multilingual representation disparity and boost cross-lingual truthfulness transfer of LLMs.

Shining Liang, Linjun Shou, Jian Pei et al.

Lack of training data in low-resource languages presents huge challenges to sequence labeling tasks such as named entity recognition (NER) and machine reading comprehension (MRC). One major obstacle is the errors on the boundary of predicted answers. To tackle this problem, we propose CalibreNet, which predicts answers in two steps. In the first step, any existing sequence labeling method can be adopted as a base model to generate an initial answer. In the second step, CalibreNet refines the boundary of the initial answer. To tackle the challenge of lack of training data in low-resource languages, we dedicatedly develop a novel unsupervised phrase boundary recovery pre-training task to enhance the multilingual boundary detection capability of CalibreNet. Experiments on two cross-lingual benchmark datasets show that the proposed approach achieves SOTA results on zero-shot cross-lingual NER and MRC tasks.

Shining Liang, Wanli Zuo, Zhenkun Shi et al.

Mining causality from text is a complex and crucial natural language understanding task corresponding to the human cognition. Existing studies at its solution can be grouped into two primary categories: feature engineering based and neural model based methods. In this paper, we find that the former has incomplete coverage and inherent errors but provide prior knowledge; while the latter leverages context information but causal inference of which is insufficiency. To handle the limitations, we propose a novel causality detection model named MCDN to explicitly model causal reasoning process, and furthermore, to exploit the advantages of both methods. Specifically, we adopt multi-head self-attention to acquire semantic feature at word level and develop the SCRN to infer causality at segment level. To the best of our knowledge, with regards to the causality tasks, this is the first time that the Relation Network is applied. The experimental results show that: 1) the proposed approach performs prominent performance on causality detection; 2) further analysis manifests the effectiveness and robustness of MCDN.