Suyu Ge

Suyu Ge, Chenyan Xiong, Corby Rosset et al.

In this paper we improve the zero-shot generalization ability of language models via Mixture-Of-Memory Augmentation (MoMA), a mechanism that retrieves augmentation documents from multiple information corpora ("external memories"), with the option to "plug in" new memory at inference time. We develop a joint learning mechanism that trains the augmentation component with latent labels derived from the end retrieval task, paired with hard negatives from the memory mixture. We instantiate the model in a zero-shot dense retrieval setting by augmenting a strong T5-based retriever with MoMA. Our model, MoMA, obtains strong zero-shot retrieval accuracy on the eighteen tasks included in the standard BEIR benchmark. It outperforms systems that seek generalization from increased model parameters and computation steps. Our analysis further illustrates the necessity of augmenting with mixture-of-memory for robust generalization, the benefits of augmentation learning, and how MoMA utilizes the plug-in memory at inference time without changing its parameters. We plan to open source our code.

Suyu Ge, Chunting Zhou, Rui Hou et al.

Red-teaming is a common practice for mitigating unsafe behaviors in Large Language Models (LLMs), which involves thoroughly assessing LLMs to identify potential flaws and addressing them with responsible and accurate responses. While effective, manual red-teaming is costly, and existing automatic red-teaming typically discovers safety risks without addressing them. In this paper, we propose a Multi-round Automatic Red-Teaming (MART) method, which incorporates both automatic adversarial prompt writing and safe response generation, significantly increasing red-teaming scalability and the safety of the target LLM. Specifically, an adversarial LLM and a target LLM interplay with each other in an iterative manner, where the adversarial LLM aims to generate challenging prompts that elicit unsafe responses from the target LLM, while the target LLM is fine-tuned with safety aligned data on these adversarial prompts. In each round, the adversarial LLM crafts better attacks on the updated target LLM, while the target LLM also improves itself through safety fine-tuning. On adversarial prompt benchmarks, the violation rate of an LLM with limited safety alignment reduces up to 84.7% after 4 rounds of MART, achieving comparable performance to LLMs with extensive adversarial prompt writing. Notably, model helpfulness on non-adversarial prompts remains stable throughout iterations, indicating the target LLM maintains strong performance on instruction following.

Suyu Ge, Yunan Zhang, Liyuan Liu et al.

In this study, we introduce adaptive KV cache compression, a plug-and-play method that reduces the memory footprint of generative inference for Large Language Models (LLMs). Different from the conventional KV cache that retains key and value vectors for all context tokens, we conduct targeted profiling to discern the intrinsic structure of attention modules. Based on the recognized structure, we then construct the KV cache in an adaptive manner: evicting long-range contexts on attention heads emphasizing local contexts, discarding non-special tokens on attention heads centered on special tokens, and only employing the standard KV cache for attention heads that broadly attend to all tokens. Moreover, with the lightweight attention profiling used to guide the construction of the adaptive KV cache, FastGen can be deployed without resource-intensive fine-tuning or re-training. In our experiments across various asks, FastGen demonstrates substantial reduction on GPU memory consumption with negligible generation quality loss. We will release our code and the compatible CUDA kernel for reproducibility.

Lu Cheng, Suyu Ge, Huan Liu

Natural Language Processing (NLP) models have been found discriminative against groups of different social identities such as gender and race. With the negative consequences of these undesired biases, researchers have responded with unprecedented effort and proposed promising approaches for bias mitigation. In spite of considerable practical importance, current algorithmic fairness literature lacks an in-depth understanding of the relations between different forms of biases. Social bias is complex by nature. Numerous studies in social psychology identify the "generalized prejudice", i.e., generalized devaluing sentiments across different groups. For example, people who devalue ethnic minorities are also likely to devalue women and gays. Therefore, this work aims to provide a first systematic study toward understanding bias correlations in mitigation. In particular, we examine bias mitigation in two common NLP tasks -- toxicity detection and word embeddings -- on three social identities, i.e., race, gender, and religion. Our findings suggest that biases are correlated and present scenarios in which independent debiasing approaches dominant in current literature may be insufficient. We further investigate whether jointly mitigating correlated biases is more desired than independent and individual debiasing. Lastly, we shed light on the inherent issue of debiasing-accuracy trade-off in bias mitigation. This study serves to motivate future research on joint bias mitigation that accounts for correlated biases.

Xihui Lin, Yunan Zhang, Suyu Ge et al.

Sparse attention, which selectively attends to a subset of tokens in the context was supposed to be efficient. However, its theoretical reduction in FLOPs has rarely translated into wall-clock speed-up over its dense attention counterparts due to the lack of hardware-aware optimizations like FlashAttention. Meanwhile, it remains unclear whether sparse attention can maintain the model's quality at a scale of today's large language models (LLMs) and how. This paper presents Sparsely-Sharded(S2) Attention, a Triton library that provides kernel optimization for sparse attention customizable at both per-head and per-context-range levels. S2-Attention enables the exploration of novel and high-performance sparse attention techniques, which we demonstrate through extensive ablations across a wide range of sparse attention designs at various model scales. From these insights, we present several basic guidelines to design sparse attention that can achieve not only practical efficiency improvements, but also strong downstream performance. To achieve high parallelization and optimized memory IO, sparse attention should shard the context heterogeneously across attention heads, where each head attends to a different subset of tokens while collectively covering the full context. Meanwhile, we find hybrid architectures combining sparse and dense attention particularly beneficial in practice. S2-Attention achieves wall-clock speedup of 8.79X, 15.87X, 25.3X compared to the strong FlashAttention-2 baseline with strong downstream performance on-par with full attention and perfect retrieval performance at a 128k context length. At inference, for 7B models, our model, with the help of our S2-Attention kernel, achieves 4.5x speed-up compared to dense counterparts. S2-Attention is released with easy-to-customize APIs for direct usage in Megatron and vLLM.

Yida Chen, Yuning Mao, Xianjun Yang et al. · harvard

Current comparisons of large reasoning models (LRMs) focus on macro-level statistics such as task accuracy or reasoning length. Whether different LRMs reason differently remains an open question. To address this gap, we introduce the LLM-proposed Open Taxonomy (LOT), a classification method that uses a generative language model to compare reasoning traces from two LRMs and articulate their distinctive features in words. LOT then models how these features predict the source LRM of a reasoning trace based on their empirical distributions across LRM outputs. Iterating this process over a dataset of reasoning traces yields a human-readable taxonomy that characterizes how models think. We apply LOT to compare the reasoning of 12 open-source LRMs on tasks in math, science, and coding. LOT identifies systematic differences in their thoughts, achieving 80-100% accuracy in distinguishing reasoning traces from LRMs that differ in scale, base model family, or objective domain. Beyond classification, LOT's natural-language taxonomy provides qualitative explanations of how LRMs think differently. Finally, in a case study, we link the reasoning differences to performance: aligning the reasoning style of smaller Qwen3 models with that of the largest Qwen3 during test time improves their accuracy on GPQA by 3.3-5.7%.

Ming Zhong, Yang Liu, Suyu Ge et al.

Text summarization is a user-preference based task, i.e., for one document, users often have different priorities for summary. As a key aspect of customization in summarization, granularity is used to measure the semantic coverage between the summary and source document. However, developing systems that can generate summaries with customizable semantic coverage is still an under-explored topic. In this paper, we propose the first unsupervised multi-granularity summarization framework, GranuSum. We take events as the basic semantic units of the source documents and propose to rank these events by their salience. We also develop a model to summarize input documents with given events as anchors and hints. By inputting different numbers of events, GranuSum is capable of producing multi-granular summaries in an unsupervised manner. Meanwhile, we annotate a new benchmark GranuDUC that contains multiple summaries at different granularities for each document cluster. Experimental results confirm the substantial superiority of GranuSum on multi-granularity summarization over strong baselines. Further, by exploiting the event information, GranuSum also exhibits state-of-the-art performance under the conventional unsupervised abstractive setting. Dataset for this paper can be found at: https://github.com/maszhongming/GranuDUC

Zhenning Zhang, Yunan Zhang, Suyu Ge et al.

The advent of large language models (LLMs) brings an opportunity to minimize the effort in search engine result page (SERP) organization. In this paper, we propose GenSERP, a framework that leverages LLMs with vision in a few-shot setting to dynamically organize intermediate search results, including generated chat answers, website snippets, multimedia data, knowledge panels into a coherent SERP layout based on a user's query. Our approach has three main stages: (1) An information gathering phase where the LLM continuously orchestrates API tools to retrieve different types of items, and proposes candidate layouts based on the retrieved items, until it's confident enough to generate the final result. (2) An answer generation phase where the LLM populates the layouts with the retrieved content. In this phase, the LLM adaptively optimize the ranking of items and UX configurations of the SERP. Consequently, it assigns a location on the page to each item, along with the UX display details. (3) A scoring phase where an LLM with vision scores all the generated SERPs based on how likely it can satisfy the user. It then send the one with highest score to rendering. GenSERP features two generation paradigms. First, coarse-to-fine, which allow it to approach optimal layout in a more manageable way, (2) beam search, which give it a better chance to hit the optimal solution compared to greedy decoding. Offline experimental results on real-world data demonstrate how LLMs can contextually organize heterogeneous search results on-the-fly and provide a promising user experience.

Suyu Ge, Jiaxin Huang, Yu Meng et al.

Opinion summarization aims to profile a target by extracting opinions from multiple documents. Most existing work approaches the task in a semi-supervised manner due to the difficulty of obtaining high-quality annotation from thousands of documents. Among them, some use aspect and sentiment analysis as a proxy for identifying opinions. In this work, we propose a new framework, FineSum, which advances this frontier in three aspects: (1) minimal supervision, where only aspect names and a few aspect/sentiment keywords are available; (2) fine-grained opinion analysis, where sentiment analysis drills down to the sub-aspect level; and (3) phrase-based summarization, where opinion is summarized in the form of phrases. FineSum automatically identifies opinion phrases from the raw corpus, classifies them into different aspects and sentiments, and constructs multiple fine-grained opinion clusters under each aspect/sentiment. Each cluster consists of semantically coherent phrases, expressing uniform opinions towards certain sub-aspect or characteristics (e.g., positive feelings for ``burgers'' in the ``food'' aspect). An opinion-oriented spherical word embedding space is trained to provide weak supervision for the phrase classifier, and phrase clustering is performed using the aspect-aware contextualized embedding generated from the phrase classifier. Both automatic evaluation on the benchmark and quantitative human evaluation validate the effectiveness of our approach.

Suyu Ge, Lu Cheng, Huan Liu

Cyberbullying, identified as intended and repeated online bullying behavior, has become increasingly prevalent in the past few decades. Despite the significant progress made thus far, the focus of most existing work on cyberbullying detection lies in the independent content analysis of different comments within a social media session. We argue that such leading notions of analysis suffer from three key limitations: they overlook the temporal correlations among different comments; they only consider the content within a single comment rather than the topic coherence across comments; they remain generic and exploit limited interactions between social media users. In this work, we observe that user comments in the same session may be inherently related, e.g., discussing similar topics, and their interaction may evolve over time. We also show that modeling such topic coherence and temporal interaction are critical to capture the repetitive characteristics of bullying behavior, thus leading to better predicting performance. To achieve the goal, we first construct a unified temporal graph for each social media session. Drawing on recent advances in graph neural network, we then propose a principled graph-based approach for modeling the temporal dynamics and topic coherence throughout user interactions. We empirically evaluate the effectiveness of our approach with the tasks of session-level bullying detection and comment-level case study. Our code is released to public.

Suyu Ge, Chuhan Wu, Fangzhao Wu et al.

With the explosion of online news, personalized news recommendation becomes increasingly important for online news platforms to help their users find interesting information. Existing news recommendation methods achieve personalization by building accurate news representations from news content and user representations from their direct interactions with news (e.g., click), while ignoring the high-order relatedness between users and news. Here we propose a news recommendation method which can enhance the representation learning of users and news by modeling their relatedness in a graph setting. In our method, users and news are both viewed as nodes in a bipartite graph constructed from historical user click behaviors. For news representations, a transformer architecture is first exploited to build news semantic representations. Then we combine it with the information from neighbor news in the graph via a graph attention network. For user representations, we not only represent users from their historically clicked news, but also attentively incorporate the representations of their neighbor users in the graph. Improved performances on a large-scale real-world dataset validate the effectiveness of our proposed method.

Suyu Ge, Fangzhao Wu, Chuhan Wu et al.

Medical named entity recognition (NER) has wide applications in intelligent healthcare. Sufficient labeled data is critical for training accurate medical NER model. However, the labeled data in a single medical platform is usually limited. Although labeled datasets may exist in many different medical platforms, they cannot be directly shared since medical data is highly privacy-sensitive. In this paper, we propose a privacy-preserving medical NER method based on federated learning, which can leverage the labeled data in different platforms to boost the training of medical NER model and remove the need of exchanging raw data among different platforms. Since the labeled data in different platforms usually has some differences in entity type and annotation criteria, instead of constraining different platforms to share the same model, we decompose the medical NER model in each platform into a shared module and a private module. The private module is used to capture the characteristics of the local data in each platform, and is updated using local labeled data. The shared module is learned across different medical platform to capture the shared NER knowledge. Its local gradients from different platforms are aggregated to update the global shared module, which is further delivered to each platform to update their local shared modules. Experiments on three publicly available datasets validate the effectiveness of our method.