Hongyuan Lu

Hongyuan Lu, Haoyang Huang, Shuming Ma et al. · microsoft-research

Despite the success of multilingual sequence-to-sequence pre-training, most existing approaches rely on document-level monolingual corpora in many different languages, sentence-level bilingual corpora,\footnote{In this paper, we use `bilingual corpora' to denote parallel corpora with `bilingual translation pairs' in many different language pairs, each consisting of two sentences/documents with the same meaning written in different languages. We use `trilingual corpora' to denote parallel corpora with `trilingual translation pairs' in many different language combinations, each consisting of three sentences/documents.} and sometimes synthetic document-level bilingual corpora. This hampers the performance with cross-lingual document-level tasks such as document-level translation. Therefore, we propose to mine and leverage document-level trilingual parallel corpora to improve sequence-to-sequence multilingual pre-training. We present \textbf{Tri}angular Document-level \textbf{P}re-training (\textbf{TRIP}), which is the first in the field to accelerate the conventional monolingual and bilingual objectives into a trilingual objective with a novel method called Grafting. Experiments show that TRIP achieves several strong state-of-the-art (SOTA) scores on three multilingual document-level machine translation benchmarks and one cross-lingual abstractive summarization benchmark, including consistent improvements by up to 3.11 d-BLEU points and 8.9 ROUGE-L points.

Hongyuan Lu, Haoyang Huang, Shuming Ma et al. · microsoft-research

Despite the fact that multilingual agreement (MA) has shown its importance for multilingual neural machine translation (MNMT), current methodologies in the field have two shortages: (i) require parallel data between multiple language pairs, which is not always realistic and (ii) optimize the agreement in an ambiguous direction, which hampers the translation performance. We present \textbf{B}idirectional \textbf{M}ultilingual \textbf{A}greement via \textbf{S}witched \textbf{B}ack-\textbf{t}ranslation (\textbf{BMA-SBT}), a novel and universal multilingual agreement framework for fine-tuning pre-trained MNMT models, which (i) exempts the need for aforementioned parallel data by using a novel method called switched BT that creates synthetic text written in another source language using the translation target and (ii) optimizes the agreement bidirectionally with the Kullback-Leibler Divergence loss. Experiments indicate that BMA-SBT clearly improves the strong baselines on the task of MNMT with three benchmarks: TED Talks, News, and Europarl. In-depth analyzes indicate that BMA-SBT brings additive improvements to the conventional BT method.

Hongyuan Lu, Wai Lam

How to defend large language models (LLMs) from generating toxic content is an important research area. Yet, most research focused on various model training techniques to remediate LLMs by updating their weights. A typical related research area is safety alignment. This however is often costly and tedious and can expose the model to even more problems such as catastrophic forgetting if the trainings are not carefully handled by experienced NLP practitioners. We thus propose a simple yet effective and novel algorithm, namely \textbf{Tox}ic Subword \textbf{Prun}ing (ToxPrune) to prune the subword contained by the toxic words from BPE in trained LLMs. In contrast to the previous work that demonstrates pruning BPE tokens as harmful to the task of machine translation, we surprisingly found its usefulness in preventing toxic content from being generated on LLMs. Fortunately, our findings suggest that ToxPrune simultaneously improves the toxic language model NSFW-3B on the task of dialogue response generation obviously. We surprisingly found that ToxPrune can even obviously improve official Llama-3.1-6B in the metric of dialogue diversity. Extensive automatic results and human evaluation indicate that ToxPrune could be helpful for both remediating toxic LLMs and improving non-toxic LLMs on the task of dialogue response generation.\footnote{We plan to release the resources to facilitate future work.}

Hongyuan Lu, Wai Lam

Large language models (LLMs) have shown promising performance on various NLP tasks via task prompting. And their performance can be further improved by appending task demonstrations to the head of the prompt. And usually, a better performance can be achieved with more demonstrations. However, asking the users to write the demonstrations can be cumbersome. As a simple yet cost-effective workaround, this paper proposes a novel method called EPA (\textbf{E}asy \textbf{P}rompt \textbf{A}ugmentation)\footnote{While this paper considers augmenting prompts via demonstrations, we name it EPA as the name EDA is already taken by a well-known NLP method \citep{wei-zou-2019-eda}.} that effectively minimizes user efforts in writing demonstrations while improving the model performance at the same time. EPA achieves these goals by automatically augmenting the demonstrations with multiple sources/targets, where each of them paraphrases each other. This is well motivated as augmenting data via paraphrasing effectively improves neural language models. EPA thus employs paraphrasing as an augmentation method for in-context learning. Extensive experiments indicate that EPA effectively improves both NLU and NLG tasks, covering from natural language inference to machine translation in translating tens of languages.\footnote{Code and data will be released upon publication.}

Hongru Wang, Huimin Wang, Lingzhi Wang et al.

Large language models (LLMs) have demonstrated exceptional performance in planning the use of various functional tools, such as calculators and retrievers, particularly in question-answering tasks. In this paper, we expand the definition of these tools, centering on conceptual tools within the context of dialogue systems. A conceptual tool specifies a cognitive concept that aids systematic or investigative thought. These conceptual tools play important roles in practice, such as multiple psychological or tutoring strategies being dynamically applied in a single turn to compose helpful responses. To further enhance the reasoning and planning capability of LLMs with these conceptual tools, we introduce a multi-persona collaboration framework: Think-Plan-Execute (TPE). This framework decouples the response generation process into three distinct roles: Thinker, Planner, and Executor. Specifically, the Thinker analyzes the internal status exhibited in the dialogue context, such as user emotions and preferences, to formulate a global guideline. The Planner then generates executable plans to call different conceptual tools (e.g., sources or strategies), while the Executor compiles all intermediate results into a coherent response. This structured approach not only enhances the explainability and controllability of responses but also reduces token redundancy. We demonstrate the effectiveness of TPE across various dialogue response generation tasks, including multi-source (FoCus) and multi-strategy interactions (CIMA and PsyQA). This reveals its potential to handle real-world dialogue interactions that require more complicated tool learning beyond just functional tools. The full code and data will be released for reproduction.

Hongyuan Lu, Wai Lam

Curriculum Data Augmentation (CDA) improves neural models by presenting synthetic data with increasing difficulties from easy to hard. However, traditional CDA simply treats the ratio of word perturbation as the difficulty measure and goes through the curriculums only once. This paper presents \textbf{PCC}: \textbf{P}araphrasing with Bottom-k Sampling and \textbf{C}yclic Learning for \textbf{C}urriculum Data Augmentation, a novel CDA framework via paraphrasing, which exploits the textual paraphrase similarity as the curriculum difficulty measure. We propose a curriculum-aware paraphrase generation module composed of three units: a paraphrase candidate generator with bottom-k sampling, a filtering mechanism and a difficulty measure. We also propose a cyclic learning strategy that passes through the curriculums multiple times. The bottom-k sampling is proposed to generate super-hard instances for the later curriculums. Experimental results on few-shot text classification as well as dialogue generation indicate that PCC surpasses competitive baselines. Human evaluation and extensive case studies indicate that bottom-k sampling effectively generates super-hard instances, and PCC significantly improves the baseline dialogue agent.

Bowen Cao, Dongdong Zhang, Yixia Li et al.

Large language models now solve many benchmark math problems at near-expert levels, yet this progress has not fully translated into reliable performance in real-world applications. We study this gap through contextual mathematical reasoning, where the mathematical core must be formulated from descriptive scenarios. We introduce ContextMATH, a benchmark that repurposes AIME and MATH-500 problems into two contextual settings: Scenario Grounding (SG), which embeds abstract problems into realistic narratives without increasing reasoning complexity, and Complexity Scaling (CS), which transforms explicit conditions into sub-problems to capture how constraints often appear in practice. Evaluating 61 proprietary and open-source models, we observe sharp drops: on average, open-source models decline by 13 and 34 points on SG and CS, while proprietary models drop by 13 and 20. Error analysis shows that errors are dominated by incorrect problem formulation, with formulation accuracy declining as original problem difficulty increases. Correct formulation emerges as a prerequisite for success, and its sufficiency improves with model scale, indicating that larger models advance in both understanding and reasoning. Nevertheless, formulation and reasoning remain two complementary bottlenecks that limit contextual mathematical problem solving. Finally, we find that fine-tuning with scenario data improves performance, whereas formulation-only training is ineffective. However, performance gaps are only partially alleviated, highlighting contextual mathematical reasoning as a central unsolved challenge for LLMs.

Wenhong Zhu, Hongkun Hao, Zhiwei He et al.

We are currently in an era of fierce competition among various large language models (LLMs) continuously pushing the boundaries of benchmark performance. However, genuinely assessing the capabilities of these LLMs has become a challenging and critical issue due to potential data contamination, and it wastes dozens of time and effort for researchers and engineers to download and try those contaminated models. To save our precious time, we propose a novel and useful method, Clean-Eval, which mitigates the issue of data contamination and evaluates the LLMs in a cleaner manner. Clean-Eval employs an LLM to paraphrase and back-translate the contaminated data into a candidate set, generating expressions with the same meaning but in different surface forms. A semantic detector is then used to filter the generated low-quality samples to narrow down this candidate set. The best candidate is finally selected from this set based on the BLEURT score. According to human assessment, this best candidate is semantically similar to the original contamination data but expressed differently. All candidates can form a new benchmark to evaluate the model. Our experiments illustrate that Clean-Eval substantially restores the actual evaluation results on contaminated LLMs under both few-shot learning and fine-tuning scenarios.

Hao Yang, Hongyuan Lu, Xinhua Zeng et al.

In the rapidly evolving field of natural language processing, dialogue systems primarily employ a single-step dialogue paradigm. Although this paradigm is efficient, it lacks the depth and fluidity of human interactions and does not appear natural. We introduce a novel \textbf{Step}-by-Step Dialogue Paradigm (Stephanie), designed to mimic the ongoing dynamic nature of human conversations. By employing a dual learning strategy and a further-split post-editing method, we generated and utilized a high-quality step-by-step dialogue dataset to fine-tune existing large language models, enabling them to perform step-by-step dialogues. We thoroughly present Stephanie. Tailored automatic and human evaluations are conducted to assess its effectiveness compared to the traditional single-step dialogue paradigm. We will release code, Stephanie datasets, and Stephanie LLMs to facilitate the future of chatbot eras.

Hao Yang, Hongyuan Lu, Dingkang Yang et al.

Instant-messaging human social chat typically progresses through a sequence of short messages. Existing step-by-step AI chatting systems typically split a one-shot generation into multiple messages and send them sequentially, but they lack an active waiting mechanism and exhibit unnatural message pacing. In order to address these issues, we propose Stephanie2, a novel next-generation step-wise decision-making dialogue agent. With active waiting and message-pace adaptation, Stephanie2 explicitly decides at each step whether to send or wait, and models latency as the sum of thinking time and typing time to achieve more natural pacing. We further introduce a time-window-based dual-agent dialogue system to generate pseudo dialogue histories for human and automatic evaluations. Experiments show that Stephanie2 clearly outperforms Stephanie1 on metrics such as naturalness and engagement, and achieves a higher pass rate on human evaluation with the role identification Turing test.

Ruijie Hou, Yueyang Jiao, Hanxu Hu et al.

The problem of data contamination is now almost inevitable during the development of large language models (LLMs), with the training data commonly integrating those evaluation benchmarks even unintentionally. This problem subsequently makes it hard to benchmark LLMs fairly. Instead of constructing contamination-free datasets (quite hard), we propose a novel framework, \textbf{LNE-Blocking}, to restore model performance prior to contamination on potentially leaked datasets. Our framework consists of two components: contamination detection and disruption operation. For the prompt, the framework first uses the contamination detection method, \textbf{LNE}, to assess the extent of contamination in the model. Based on this, it adjusts the intensity of the disruption operation, \textbf{Blocking}, to elicit non-memorized responses from the model. Our framework is the first to efficiently restore the model's greedy decoding performance. This comes with a strong performance on multiple datasets with potential leakage risks, and it consistently achieves stable recovery results across different models and varying levels of data contamination. We release the code at https://github.com/RuijieH/LNE-Blocking to facilitate research.

Hongyuan Lu, Zixuan Li, Wai Lam

As current training data for Large Language Models (LLMs) are dominated by English corpus, they are English-centric and they present impressive performance on English reasoning tasks.\footnote{This paper primarily studies English-centric models, but our method could be universal by using the centric language in the dictionary for non-English-centric LLMs.} Yet, they usually suffer from lower performance in other languages. There are about 7,000 languages over the world, and many are low-resourced on English-centric LLMs. For the sake of people who primarily speak these languages, it is especially urgent to enable our LLMs in those languages. Model training is usually effective, but computationally expensive and requires experienced NLP practitioners. This paper presents a novel and simple yet effective method called \textbf{D}ictionary \textbf{I}nsertion \textbf{P}rompting (\textbf{DIP}). When providing a non-English prompt, DIP looks up a word dictionary and inserts words' English counterparts into the prompt for LLMs. It then enables better translation into English and better English model thinking steps which leads to obviously better results. We experiment with about 200 languages from FLORES-200. Since there are no adequate datasets, we use the NLLB translator to create synthetic multilingual benchmarks from the existing 4 English reasoning benchmarks such as GSM8K and AQuA. Despite the simplicity and computationally lightweight, we surprisingly found the effectiveness of DIP on math and commonsense reasoning tasks on multiple open-source and close-source LLMs.\footnote{Our dictionaries, code, and synthetic benchmarks will be open-sourced to facilitate future research.}

Tianyi Tang, Hongyuan Lu, Yuchen Eleanor Jiang et al.

Most research about natural language generation (NLG) relies on evaluation benchmarks with limited references for a sample, which may result in poor correlations with human judgements. The underlying reason is that one semantic meaning can actually be expressed in different forms, and the evaluation with a single or few references may not accurately reflect the quality of the model's hypotheses. To address this issue, this paper presents a simple and effective method, named Div-Ref, to enhance existing evaluation benchmarks by enriching the number of references. We leverage large language models (LLMs) to diversify the expression of a single reference into multiple high-quality ones to cover the semantic space of the reference sentence as much as possible. We conduct comprehensive experiments to empirically demonstrate that diversifying the expression of reference can significantly enhance the correlation between automatic evaluation and human evaluation. This idea is compatible with recent LLM-based evaluation which can similarly derive advantages from incorporating multiple references. We strongly encourage future generation benchmarks to include more references, even if they are generated by LLMs, which is once for all. We release all the code and data at https://github.com/RUCAIBox/Div-Ref to facilitate research.

Hanxu Hu, Hongyuan Lu, Huajian Zhang et al.

In this paper, we take the initiative to investigate the performance of LLMs on complex planning tasks that require LLMs to understand a virtual spatial environment simulated via natural language and act correspondingly in text. We propose a benchmark named Natural Language Planning and Action (Natala) composed of a set of novel tasks: Brick World, NLVR-based Manipulations, and Natural Language Navigation. We found that current popular LLMs such as ChatGPT still lack abilities in complex planning. This arises a question -- do the LLMs have a good understanding of the environments described in natural language, or maybe other alternatives such as symbolic representations are neater and hence better to be understood by LLMs? To this end, we propose a novel method called CoS (Chain-of-Symbol Prompting) that represents the complex environments with condensed symbolic spatial representations during the chained intermediate thinking steps. CoS is easy to use and does not need additional training on LLMs. Extensive experiments indicate that CoS clearly surpasses the performance of the Chain-of-Thought (CoT) Prompting in all three planning tasks with even fewer tokens used in the inputs compared with CoT on ChatGPT and InstructGPT. The performance gain is strong, by up to 60.8% accuracy (from 31.8% to 92.6%) on Brick World for ChatGPT. CoS also reduces the number of tokens in the prompt obviously, by up to 65.8% of the tokens (from 407 to 139) for the intermediate steps from demonstrations on Brick World. Code and data available at: https://github.com/hanxuhu/chain-of-symbol-planning

Shuaiyi Li, Yang Deng, Deng Cai et al.

As the typical retraining paradigm is unacceptably time- and resource-consuming, researchers are turning to model editing to find an effective way that supports both consecutive and batch scenarios to edit the model behavior directly. Despite all these practical expectations, existing model editing methods fail to realize all of them. Furthermore, the memory demands for such sequential model editing approaches tend to be prohibitive, frequently necessitating an external memory that grows incrementally over time. To cope with these challenges, we propose CoachHooK, a model editing method that simultaneously supports sequential and batch editing. CoachHooK is memory-friendly as it only needs a small amount of it to store several hook layers whose size remains unchanged over time. Experimental results demonstrate the superiority of our method over other batch-supportive model editing methods under both single-round and consecutive batch editing scenarios. Extensive analyses of CoachHooK have been conducted to verify the stability of our method over a number of consecutive steps.

Zefan Zhang, Kehua Zhu, Shijie Jiang et al.

Video Large Language Models (VideoLLMs) exhibit various types of hallucinations. Existing research has primarily focused on hallucinations involving the presence of events, objects, and scenes in videos, while largely neglecting event relation hallucination. In this paper, we introduce a novel benchmark for evaluating the Video Event Relation Hallucination, named VERHallu. This benchmark focuses on causal, temporal, and subevent relations between events, encompassing three types of tasks: relation classification, question answering, and counterfactual question answering, for a comprehensive evaluation of event relation hallucination. Additionally, it features counterintuitive video scenarios that deviate from typical pretraining distributions, with each sample accompanied by human-annotated candidates covering both vision-language and pure language biases. Our analysis reveals that current state-of-the-art VideoLLMs struggle with dense-event relation reasoning, often relying on prior knowledge due to insufficient use of frame-level cues. Although these models demonstrate strong grounding capabilities for key events, they often overlook the surrounding subevents, leading to an incomplete and inaccurate understanding of event relations. To tackle this, we propose a Key-Frame Propagating (KFP) strategy, which reallocates frame-level attention within intermediate layers to enhance multi-event understanding. Experiments show it effectively mitigates the event relation hallucination without affecting inference speed.

Hongyuan Lu, Zixuan Li, Zefan Zhang et al.

There are more than 7,000 languages around the world, and current Large Language Models (LLMs) only support hundreds of languages. Dictionary-based prompting methods can enhance translation on them, but most methods use all the available dictionaries, which could be expensive. Instead, it will be flexible to have a trade-off between token consumption and translation performance. This paper proposes a novel task called \textbf{A}utomatic \textbf{D}ictionary \textbf{S}election (\textbf{ADS}). The goal of the task is to automatically select which dictionary to use to enhance translation. We propose a novel and effective method which we call \textbf{S}elect \textbf{Lo}w-frequency \textbf{W}ords! (\textbf{SLoW}) which selects those dictionaries that have a lower frequency. Our methods have unique advantages. First, there is no need for access to the training data for frequency estimation (which is usually unavailable). Second, it inherits the advantage of dictionary-based methods, where no additional tuning is required on LLMs. Experimental results on 100 languages from FLORES indicate that SLoW surpasses strong baselines, and it can obviously save token usage, with many languages even surpassing the translation performance of the full dictionary baseline.\footnote{A shocking fact is that there is no need to use the actual training data (often unobtainable) for frequency estimation, and an estimation frequency obtained using public resources is still apparently effective in improving translation with ChatGPT and Llama, and DeepSeek.}\footnote{Code and data available upon publication.}

Haoran Yang, Yumeng Zhang, Jiaqi Xu et al.

While Large Language Models (LLMs) have demonstrated exceptional multitasking abilities, fine-tuning these models on downstream, domain-specific datasets is often necessary to yield superior performance on test sets compared to their counterparts without fine-tuning. However, the comprehensive effects of fine-tuning on the LLMs' generalization ability are not fully understood. This paper delves into the differences between original, unmodified LLMs and their fine-tuned variants. Our primary investigation centers on whether fine-tuning affects the generalization ability intrinsic to LLMs. To elaborate on this, we conduct extensive experiments across five distinct language tasks on various datasets. Our main findings reveal that models fine-tuned on generation and classification tasks exhibit dissimilar behaviors in generalizing to different domains and tasks. Intriguingly, we observe that integrating the in-context learning strategy during fine-tuning on generation tasks can enhance the model's generalization ability. Through this systematic investigation, we aim to contribute valuable insights into the evolving landscape of fine-tuning practices for LLMs.

Hongyuan Lu, Haoran Yang, Haoyang Huang et al.

Large language models (LLMs) have shown surprisingly good performance in multilingual neural machine translation (MNMT) even when trained without parallel data. Yet, despite the fact that the amount of training data is gigantic, they still struggle with translating rare words, particularly for low-resource languages. Even worse, it is usually unrealistic to retrieve relevant demonstrations for in-context learning with low-resource languages on LLMs, which restricts the practical use of LLMs for translation -- how should we mitigate this problem? To this end, we present a novel method, CoD, which augments LLMs with prior knowledge with the chains of multilingual dictionaries for a subset of input words to elicit translation abilities for LLMs. Extensive experiments indicate that augmenting ChatGPT with CoD elicits large gains by up to 13x chrF++ points for MNMT (3.08 to 42.63 for English to Serbian written in Cyrillic script) on FLORES-200 full devtest set. We further demonstrate the importance of chaining the multilingual dictionaries, as well as the superiority of CoD to few-shot demonstration for low-resource languages.

Hongyuan Lu, Wai Lam, Hong Cheng et al.

Incorporating personas information allows diverse and engaging responses in dialogue response generation. Unfortunately, prior works have primarily focused on self personas and have overlooked the value of partner personas. Moreover, in practical applications, the availability of ground truth partner personas is often not the case. This paper attempts to tackle these issues by offering a novel framework that leverages automatic partner personas generation to enhance the succeeding dialogue generation. We incorporate reinforcement learning with a dedicatedly designed critic network for reward judgement. Experimental results from both automatic and human evaluation demonstrate a) Our framework is capable of generating relevant, informative and coherent partner personas, even compared to the ground truth partner personas. b) Generated partner personas enhance the succeeding response generation, thus surpassing our baselines and comparison model when partner personas are missing during the inference stage. c) Our framework generates responses that are more informative and engaging than our baseline conditioned on the ground truth partner personas during inference. d) Our dedicatedly designed critic network reinforces our framework effectively. Finally, our framework gives better explainability and reduces the demands for external databases for partner personas.