Aixin Sun

Xiang Li, Yiqun Yao, Xin Jiang et al. · tencent-ai, tsinghua

Large language models (LLMs) are considered important approaches towards foundational machine intelligence, achieving remarkable success in Natural Language Processing and multimodal tasks, among others. However, the carbon footprints and financial costs originating from heavy pre-training computation is a non-negligible issue. Progressive training methods, inspired by the neurogenesis process that grows neural structures, have shown potential to accelerate LLM pre-training. However, the algorithms, implementation, and practices for progressively training LLMs beyond 100B parameters remain underexplored. In this paper, we show that our model, namely FLM-101B, trained with our growth strategy under a budget of \$100K, reaches 80\% of the baselines' performances with only 10\% of their floating-point operations. We believe that further studies on progressive training will benefit the community by cutting down the costs and promoting green AI. The checkpoint of FLM-101B is released at https://huggingface.co/CofeAI/FLM-101B.

Xinze Li, Yuhang Zang, Yixin Cao et al.

Markdown skill libraries for LLM agents ship as free-form prose, forcing the agent to re-derive both the input schema and the concrete invocation syntax on every retrieval. We observe that this often produces a "confused -> re-retrieve -> still confused" loop in which the agent issues a partially-correct action, receives uninformative environment feedback, and re-retrieves the same prose. We propose Skill-as-Pseudocode (SaP), an automatic conversion of markdown skill libraries into typed pseudocode with deterministic quality control. For each cluster of similar procedural passages drawn from one or more skills, SaP extracts a typed contract and filters it through a four-check deterministic verifier (coverage, binding, replacement, risk). Promoted contracts are inlined into a rewritten skill skeleton together with restored concrete action templates, giving the agent two complementary signals: a typed signature for what the skill does and a concrete template for how to invoke it. On the 134-game ALFWorld unseen split with gpt-4o-mini, pooled across three seeds, SaP wins 82/402 paired games versus 47/402 for the Graph-of-Skills (GoS) baseline (pooled McNemar p = 8.2e-5), at -22.8 +/- 6.4% input tokens and -14.5 +/- 4.1% LLM calls per game.

Yubo Ma, Yuhang Zang, Liangyu Chen et al. · pku, stanford

Understanding documents with rich layouts and multi-modal components is a long-standing and practical task. Recent Large Vision-Language Models (LVLMs) have made remarkable strides in various tasks, particularly in single-page document understanding (DU). However, their abilities on long-context DU remain an open problem. This work presents MMLongBench-Doc, a long-context, multi-modal benchmark comprising 1,062 expert-annotated questions. Distinct from previous datasets, it is constructed upon 130 lengthy PDF-formatted documents with an average of 49.4 pages and 20,971 textual tokens. Towards comprehensive evaluation, answers to these questions rely on pieces of evidence from (1) different sources (text, image, chart, table, and layout structure) and (2) various locations (i.e. page number). Moreover, 33.2% of the questions are cross-page questions requiring evidence across multiple pages. 22.8% of the questions are designed to be unanswerable for detecting potential hallucinations. Experiments on 14 LVLMs demonstrate that long-context DU greatly challenges current models. Notably, the best-performing model, GPT-4o, achieves an F1 score of only 42.7%, while the second-best, GPT-4V, scores 31.4%. Furthermore, 12 LVLMs (all except GPT-4o and GPT-4V) even present worse performance than their LLM counterparts which are fed with lossy-parsed OCR documents. These results validate the necessity of future research toward more capable long-context LVLMs. Project Page: https://mayubo2333.github.io/MMLongBench-Doc

Mathieu Ravaut, Aixin Sun, Nancy F. Chen et al.

Large language models (LLMs) excel in abstractive summarization tasks, delivering fluent and pertinent summaries. Recent advancements have extended their capabilities to handle long-input contexts, exceeding 100k tokens. However, in question answering, language models exhibit uneven utilization of their input context. They tend to favor the initial and final segments, resulting in a U-shaped performance pattern concerning where the answer is located within the input. This bias raises concerns, particularly in summarization where crucial content may be dispersed throughout the source document(s). Besides, in summarization, mapping facts from the source to the summary is not trivial as salient content is usually re-phrased. In this paper, we conduct the first comprehensive study on context utilization and position bias in summarization. Our analysis encompasses 6 LLMs, 10 datasets, and 5 evaluation metrics. We introduce a new evaluation benchmark called MiddleSum on the which we benchmark two alternative inference methods to alleviate position bias: hierarchical summarization and incremental summarization. Our code and data can be found here: https://github.com/ntunlp/MiddleSum.

Xiang Li, Yiqun Yao, Xin Jiang et al.

Large Language Models (LLMs) represent a significant stride toward Artificial General Intelligence. As scaling laws underscore the potential of increasing model sizes, the academic community has intensified its investigations into LLMs with capacities exceeding 50 billion parameters. This technical report builds on our prior work with Tele-FLM (also known as FLM-2), a publicly available 52-billion-parameter model. We delve into two primary areas: we first discuss our observation of Supervised Fine-tuning (SFT) on Tele-FLM-52B, which supports the "less is more" approach for SFT data construction; second, we demonstrate our experiments and analyses on the best practices for progressively growing a model from 52 billion to 102 billion, and subsequently to 1 trillion parameters. We will open-source a 1T model checkpoint, namely Tele-FLM-1T, to advance further training and research.

Xin Jiang, Xiang Li, Wenjia Ma et al.

Large language models (LLMs) represented by GPT family have achieved remarkable success. The characteristics of LLMs lie in their ability to accommodate a wide range of tasks through a generative approach. However, the flexibility of their output format poses challenges in controlling and harnessing the model's outputs, thereby constraining the application of LLMs in various domains. In this work, we present Sketch, an innovative toolkit designed to streamline LLM operations across diverse fields. Sketch comprises the following components: (1) a suite of task description schemas and prompt templates encompassing various NLP tasks; (2) a user-friendly, interactive process for building structured output LLM services tailored to various NLP tasks; (3) an open-source dataset for output format control, along with tools for dataset construction; and (4) an open-source model based on LLaMA3-8B-Instruct that adeptly comprehends and adheres to output formatting instructions. We anticipate this initiative to bring considerable convenience to LLM users, achieving the goal of ''plug-and-play'' for various applications. The components of Sketch will be progressively open-sourced at https://github.com/cofe-ai/Sketch.

Yiqun Yao, Wenjia Ma, Xuezhi Fang et al.

Controlling the format of outputs generated by large language models (LLMs) is a critical functionality in various applications. Current methods typically employ constrained decoding with rule-based automata or fine-tuning with manually crafted format instructions, both of which struggle with open-domain format requirements. To address this limitation, we introduce a novel framework for controlled generation in LLMs, leveraging user-provided, one-shot QA pairs. This study investigates LLMs' capabilities to follow open-domain, one-shot constraints and replicate the format of the example answers. We observe that this is a non-trivial problem for current LLMs. We also develop a dataset collection methodology for supervised fine-tuning that enhances the open-domain format control of LLMs without degrading output quality, as well as a benchmark on which we evaluate both the helpfulness and format correctness of LLM outputs. The resulting datasets, named OIFC-SFT, along with the related code, will be made publicly available at https://github.com/cofe-ai/OIFC.

Renjie Liang, Li Li, Chongzhi Zhang et al.

In this paper, we propose the task of \textit{Ranked Video Moment Retrieval} (RVMR) to locate a ranked list of matching moments from a collection of videos, through queries in natural language. Although a few related tasks have been proposed and studied by CV, NLP, and IR communities, RVMR is the task that best reflects the practical setting of moment search. To facilitate research in RVMR, we develop the TVR-Ranking dataset, based on the raw videos and existing moment annotations provided in the TVR dataset. Our key contribution is the manual annotation of relevance levels for 94,442 query-moment pairs. We then develop the $NDCG@K, IoU\geq μ$ evaluation metric for this new task and conduct experiments to evaluate three baseline models. Our experiments show that the new RVMR task brings new challenges to existing models and we believe this new dataset contributes to the research on multi-modality search. The dataset is available at \url{https://github.com/Ranking-VMR/TVR-Ranking}

Yubo Ma, Yixin Cao, YongChing Hong et al.

Large Language Models (LLMs) have made remarkable strides in various tasks. Whether LLMs are competitive few-shot solvers for information extraction (IE) tasks, however, remains an open problem. In this work, we aim to provide a thorough answer to this question. Through extensive experiments on nine datasets across four IE tasks, we demonstrate that current advanced LLMs consistently exhibit inferior performance, higher latency, and increased budget requirements compared to fine-tuned SLMs under most settings. Therefore, we conclude that LLMs are not effective few-shot information extractors in general. Nonetheless, we illustrate that with appropriate prompting strategies, LLMs can effectively complement SLMs and tackle challenging samples that SLMs struggle with. And moreover, we propose an adaptive filter-then-rerank paradigm to combine the strengths of LLMs and SLMs. In this paradigm, SLMs serve as filters and LLMs serve as rerankers. By prompting LLMs to rerank a small portion of difficult samples identified by SLMs, our preliminary system consistently achieves promising improvements (2.4% F1-gain on average) on various IE tasks, with an acceptable time and cost investment.

Wanying Ren, Xin Song, Futing Wang et al.

Parameter-based knowledge editing updates the internal knowledge of large language models (LLMs) via localized weight modifications and has attracted significant attention. However, most existing methods overlook fundamental theoretical limitations and are rarely evaluated under realistic, practice-oriented settings. In this paper, we first present a theoretical analysis based on the dimensional Collapse Hypothesis, explaining how localized parameter edits can propagate along fragile directions in the representation space, inducing global interference and ultimately causing reasoning collapse. Building on this insight, we conduct a comprehensive empirical evaluation by systematically varying knowledge complexity, number of edits, evaluation dimensions, and baseline methods. Our results show that parameter-based editing methods consistently damage core LLM capabilities. In contrast, a simple retrieval-based baseline achieves consistently stronger performance than all parameter-editing methods across all evaluated conditions. These findings highlight that preserving the fundamental capabilities of LLMs after knowledge editing should be a central concern for future research.

Yiqun Yao, Siqi fan, Xiusheng Huang et al. · tencent-ai, tsinghua

As language models scale up, it becomes increasingly expensive to verify research ideas because conclusions on small models do not trivially transfer to large ones. A possible solution is to establish a generic system that accurately predicts certain metrics for large models without training them. Existing scaling laws require hyperparameter search on the largest models, limiting their predicative capability. In this paper, we present an approach (namely μScaling) to predict the pre-training loss, based on our observations that Maximal Update Parametrization (μP) enables accurate fitting of scaling laws close to common loss basins in hyperparameter space. With μScaling, different model designs can be compared on large scales by training only their smaller counterparts. Further, we introduce nanoLM: an affordable LLM pre-training benchmark that facilitates this new research paradigm. With around 14% of the one-time pre-training cost, we can accurately forecast the loss for models up to 52B. Our goal with nanoLM is to empower researchers with limited resources to reach meaningful conclusions on large models. We also aspire for our benchmark to serve as a bridge between the academic community and the industry. Code for μScaling is available at https://github.com/cofe-ai/Mu-scaling. Code for nanoLLM will be available later.

Yequan Wang, Xuying Meng, Yiyi Liu et al. · tencent-ai, tsinghua

Many studies on dialog emotion analysis focus on utterance-level emotion only. These models hence are not optimized for dialog-level emotion detection, i.e. to predict the emotion category of a dialog as a whole. More importantly, these models cannot benefit from the context provided by the whole dialog. In real-world applications, annotations to dialog could fine-grained, including both utterance-level tags (e.g. speaker type, intent category, and emotion category), and dialog-level tags (e.g. user satisfaction, and emotion curve category). In this paper, we propose a Context-based Hierarchical Attention Capsule~(Chat-Capsule) model, which models both utterance-level and dialog-level emotions and their interrelations. On a dialog dataset collected from customer support of an e-commerce platform, our model is also able to predict user satisfaction and emotion curve category. Emotion curve refers to the change of emotions along the development of a conversation. Experiments show that the proposed Chat-Capsule outperform state-of-the-art baselines on both benchmark dataset and proprietary dataset. Source code will be released upon acceptance.

Chen Chen, Yufei Wang, Aixin Sun et al.

Knowledge Graph Completion (KGC) often requires both KG structural and textual information to be effective. Pre-trained Language Models (PLMs) have been used to learn the textual information, usually under the fine-tune paradigm for the KGC task. However, the fine-tuned PLMs often overwhelmingly focus on the textual information and overlook structural knowledge. To tackle this issue, this paper proposes CSProm-KG (Conditional Soft Prompts for KGC) which maintains a balance between structural information and textual knowledge. CSProm-KG only tunes the parameters of Conditional Soft Prompts that are generated by the entities and relations representations. We verify the effectiveness of CSProm-KG on three popular static KGC benchmarks WN18RR, FB15K-237 and Wikidata5M, and two temporal KGC benchmarks ICEWS14 and ICEWS05-15. CSProm-KG outperforms competitive baseline models and sets new state-of-the-art on these benchmarks. We conduct further analysis to show (i) the effectiveness of our proposed components, (ii) the efficiency of CSProm-KG, and (iii) the flexibility of CSProm-KG.

Xinyu Zou, Zhi Hu, Yiming Zhao et al.

Multi-scenario learning (MSL) enables a service provider to cater for users' fine-grained demands by separating services for different user sectors, e.g., by user's geographical region. Under each scenario there is a need to optimize multiple task-specific targets e.g., click through rate and conversion rate, known as multi-task learning (MTL). Recent solutions for MSL and MTL are mostly based on the multi-gate mixture-of-experts (MMoE) architecture. MMoE structure is typically static and its design requires domain-specific knowledge, making it less effective in handling both MSL and MTL. In this paper, we propose a novel Automatic Expert Selection framework for Multi-scenario and Multi-task search, named AESM^{2}. AESM^{2} integrates both MSL and MTL into a unified framework with an automatic structure learning. Specifically, AESM^{2} stacks multi-task layers over multi-scenario layers. This hierarchical design enables us to flexibly establish intrinsic connections between different scenarios, and at the same time also supports high-level feature extraction for different tasks. At each multi-scenario/multi-task layer, a novel expert selection algorithm is proposed to automatically identify scenario-/task-specific and shared experts for each input. Experiments over two real-world large-scale datasets demonstrate the effectiveness of AESM^{2} over a battery of strong baselines. Online A/B test also shows substantial performance gain on multiple metrics. Currently, AESM^{2} has been deployed online for serving major traffic.

Shaowen Zhou, Bowen Yu, Aixin Sun et al.

Open Information Extraction (OpenIE) facilitates domain-independent discovery of relational facts from large corpora. The technique well suits many open-world natural language understanding scenarios, such as automatic knowledge base construction, open-domain question answering, and explicit reasoning. Thanks to the rapid development in deep learning technologies, numerous neural OpenIE architectures have been proposed and achieve considerable performance improvement. In this survey, we provide an extensive overview of the-state-of-the-art neural OpenIE models, their key design decisions, strengths and weakness. Then, we discuss limitations of current solutions and the open issues in OpenIE problem itself. Finally we list recent trends that could help expand its scope and applicability, setting up promising directions for future research in OpenIE. To our best knowledge, this paper is the first review on this specific topic.

Xinze Li, Yixin Cao, Liangming Pan et al.

Although achieving great success, Large Language Models (LLMs) usually suffer from unreliable hallucinations. Although language attribution can be a potential solution, there are no suitable benchmarks and evaluation metrics to attribute LLMs to structured knowledge. In this paper, we define a new task of Knowledge-aware Language Model Attribution (KaLMA) that improves upon three core concerns with conventional attributed LMs. First, we extend attribution source from unstructured texts to Knowledge Graph (KG), whose rich structures benefit both the attribution performance and working scenarios. Second, we propose a new ``Conscious Incompetence" setting considering the incomplete knowledge repository, where the model identifies the need for supporting knowledge beyond the provided KG. Third, we propose a comprehensive automatic evaluation metric encompassing text quality, citation quality, and text citation alignment. To implement the above innovations, we build a dataset in biography domain BioKaLMA via evolutionary question generation strategy, to control the question complexity and necessary knowledge to the answer. For evaluation, we develop a baseline solution and demonstrate the room for improvement in LLMs' citation generation, emphasizing the importance of incorporating the "Conscious Incompetence" setting, and the critical role of retrieval accuracy.

Yu-chen Fan, Yitong Ji, Jie Zhang et al.

A typical benchmark dataset for recommender system (RecSys) evaluation consists of user-item interactions generated on a platform within a time period. The interaction generation mechanism partially explains why a user interacts with (e.g., like, purchase, rate) an item, and the context of when a particular interaction happened. In this study, we conduct a meticulous analysis of the MovieLens dataset and explain the potential impact of using the dataset for evaluating recommendation algorithms. We make a few main findings from our analysis. First, there are significant differences in user interactions at the different stages when a user interacts with the MovieLens platform. The early interactions largely define the user portrait which affects the subsequent interactions. Second, user interactions are highly affected by the candidate movies that are recommended by the platform's internal recommendation algorithm(s). Third, changing the order of user interactions makes it more difficult for sequential algorithms to capture the progressive interaction process. We further discuss the discrepancy between the interaction generation mechanism that is employed by the MovieLens system and that of typical real-world recommendation scenarios. In summary, the MovieLens platform demonstrates an efficient and effective way of collecting user preferences to address cold-starts. However, models that achieve excellent recommendation accuracy on the MovieLens dataset may not demonstrate superior performance in practice, for at least two kinds of differences: (i) the differences in the contexts of user-item interaction generation, and (ii) the differences in user knowledge about the item collections. While results on MovieLens can be useful as a reference, they should not be solely relied upon as the primary justification for the effectiveness of a recommendation system model.

Fuzhao Xue, Jianghai Chen, Aixin Sun et al.

Transformer-based models have delivered impressive results on many tasks, particularly vision and language tasks. In many model training situations, conventional configurations are typically adopted. For example, we often set the base model with hidden dimensions (i.e. model width) to be 768 and the number of transformer layers (i.e. model depth) to be 12. In this paper, we revisit these conventional configurations. Through theoretical analysis and experimental evaluation, we show that the masked autoencoder is effective in alleviating the over-smoothing issue in deep transformer training. Based on this finding, we propose Bamboo, an idea of using deeper and narrower transformer configurations, for masked autoencoder training. On ImageNet, with such a simple change in configuration, re-designed model achieves 87.1% top-1 accuracy and outperforms SoTA models like MAE and BEiT. On language tasks, re-designed model outperforms BERT with default setting by 1.1 points on average, on GLUE datasets.

Kuicai Dong, Aixin Sun, Jung-Jae Kim et al.

Open Information Extraction (OpenIE) aims to extract relational tuples from open-domain sentences. Traditional rule-based or statistical models have been developed based on syntactic structures of sentences, identified by syntactic parsers. However, previous neural OpenIE models under-explore the useful syntactic information. In this paper, we model both constituency and dependency trees into word-level graphs, and enable neural OpenIE to learn from the syntactic structures. To better fuse heterogeneous information from both graphs, we adopt multi-view learning to capture multiple relationships from them. Finally, the finetuned constituency and dependency representations are aggregated with sentential semantic representations for tuple generation. Experiments show that both constituency and dependency information, and the multi-view learning are effective.

Yequan Wang, Hengran Zhang, Aixin Sun et al.

Given comparative text, comparative relation extraction aims to extract two targets (\eg two cameras) in comparison and the aspect they are compared for (\eg image quality). The extracted comparative relations form the basis of further opinion analysis.Existing solutions formulate this task as a sequence labeling task, to extract targets and aspects. However, they cannot directly extract comparative relation(s) from text. In this paper, we show that comparative relations can be directly extracted with high accuracy, by generative model. Based on GPT-2, we propose a Generation-based Comparative Relation Extractor (GCRE-GPT). Experiment results show that \modelname achieves state-of-the-art accuracy on two datasets.

Yequan Wang, Xiang Li, Aixin Sun et al.

Quotation extraction aims to extract quotations from written text. There are three components in a quotation: source refers to the holder of the quotation, cue is the trigger word(s), and content is the main body. Existing solutions for quotation extraction mainly utilize rule-based approaches and sequence labeling models. While rule-based approaches often lead to low recalls, sequence labeling models cannot well handle quotations with complicated structures. In this paper, we propose the Context and Former-Label Enhanced Net (CofeNet) for quotation extraction. CofeNet is able to extract complicated quotations with components of variable lengths and complicated structures. On two public datasets (i.e., PolNeAR and Riqua) and one proprietary dataset (i.e., PoliticsZH), we show that our CofeNet achieves state-of-the-art performance on complicated quotation extraction.

Ziqi Jin, Bin Wang, Xiang Lin et al.

Diffusion models offer appealing properties for language generation, such as parallel decoding and iterative refinement, but the discrete and highly structured nature of text challenges the direct application of diffusion principles. In this paper, we revisit diffusion language modeling from the view of diffusion process and language modeling, and outline five properties that separate diffusion mechanics from language-specific requirements. We first categorize existing approaches into continuous diffusion in embedding space and discrete diffusion over tokens. We then show that each satisfies only part of the five essential properties and therefore reflects a structural trade-off. Through analyses of recent large diffusion language models, we identify two central issues: (i) uniform corruption does not respect how information is distributed across positions, and (ii) token-wise marginal training cannot capture multi-token dependencies during parallel decoding. These observations motivate diffusion processes that align more closely with the structure of text, and encourage future work toward more coherent diffusion language models.

Yuan Cheng, Fengzhuo Zhang, Yunlong Hou et al.

Large Language Models (LLMs) often exhibit slash attention patterns, where attention scores concentrate along the $Δ$-th sub-diagonal for some offset $Δ$. These patterns play a key role in passing information across tokens. But why do they emerge? In this paper, we demystify the emergence of these Slash-Dominant Heads (SDHs) from both empirical and theoretical perspectives. First, by analyzing open-source LLMs, we find that SDHs are intrinsic to models and generalize to out-of-distribution prompts. To explain the intrinsic emergence, we analyze the queries, keys, and Rotary Position Embedding (RoPE), which jointly determine attention scores. Our empirical analysis reveals two characteristic conditions of SDHs: (1) Queries and keys are almost rank-one, and (2) RoPE is dominated by medium- and high-frequency components. Under these conditions, queries and keys are nearly identical across tokens, and interactions between medium- and high-frequency components of RoPE give rise to SDHs. Beyond empirical evidence, we theoretically show that these conditions are sufficient to ensure the emergence of SDHs by formalizing them as our modeling assumptions. Particularly, we analyze the training dynamics of a shallow Transformer equipped with RoPE under these conditions, and prove that models trained via gradient descent exhibit SDHs. The SDHs generalize to out-of-distribution prompts.

Yequan Wang, Jiawen Deng, Aixin Sun et al.

Recently, amounts of works utilize perplexity~(PPL) to evaluate the quality of the generated text. They suppose that if the value of PPL is smaller, the quality(i.e. fluency) of the text to be evaluated is better. However, we find that the PPL referee is unqualified and it cannot evaluate the generated text fairly for the following reasons: (i) The PPL of short text is larger than long text, which goes against common sense, (ii) The repeated text span could damage the performance of PPL, and (iii) The punctuation marks could affect the performance of PPL heavily. Experiments show that the PPL is unreliable for evaluating the quality of given text. Last, we discuss the key problems with evaluating text quality using language models.

Xinze Li, Ziyue Zhu, Siyuan Liu et al.

We introduce EMemBench, a programmatic benchmark for evaluating long-term memory of agents through interactive games. Rather than using a fixed set of questions, EMemBench generates questions from each agent's own trajectory, covering both text and visual game environments. Each template computes verifiable ground truth from underlying game signals, with controlled answerability and balanced coverage over memory skills: single/multi-hop recall, induction, temporal, spatial, logical, and adversarial. We evaluate memory agents with strong LMs/VLMs as backbones, using in-context prompting as baselines. Across 15 text games and multiple visual seeds, results are far from saturated: induction and spatial reasoning are persistent bottlenecks, especially in visual setting. Persistent memory yields clear gains for open backbones on text games, but improvements are less consistent for VLM agents, suggesting that visually grounded episodic memory remains an open challenge. A human study further confirms the difficulty of EMemBench.

Tze-Kean Ng, Joshua Teng-Khing Khoo, Aixin Sun

A good number of toolkits have been developed in Recommender Systems (RecSys) research to promote fair evaluation and reproducibility. However, recent critical examinations of RecSys evaluation protocols have raised concerns regarding the validity of existing evaluation pipelines. In this demonstration, we present RecNextEval, a reference implementation of an evaluation framework specifically designed for next-batch recommendation. RecNextEval utilizes a time-window data split to ensure models are evaluated along a global timeline, effectively minimizing data leakage. Our implementation highlights the inherent complexities of RecSys evaluation and encourages a shift toward model development that more accurately simulates production environments. The RecNextEval library and its accompanying GUI interface are open-source and publicly accessible.

Bobo Li, Xudong Han, Jiang Liu et al.

Large language models (LLMs) and multimodal LLMs are changing event extraction (EE): prompting and generation can often produce structured outputs in zero shot or few shot settings. Yet LLM based pipelines face deployment gaps, including hallucinations under weak constraints, fragile temporal and causal linking over long contexts and across documents, and limited long horizon knowledge management within a bounded context window. We argue that EE should be viewed as a system component that provides a cognitive scaffold for LLM centered solutions. Event schemas and slot constraints create interfaces for grounding and verification; event centric structures act as controlled intermediate representations for stepwise reasoning; event links support relation aware retrieval with graph based RAG; and event stores offer updatable episodic and agent memory beyond the context window. This survey covers EE in text and multimodal settings, organizing tasks and taxonomy, tracing method evolution from rule based and neural models to instruction driven and generative frameworks, and summarizing formulations, decoding strategies, architectures, representations, datasets, and evaluation. We also review cross lingual, low resource, and domain specific settings, and highlight open challenges and future directions for reliable event centric systems. Finally, we outline open challenges and future directions that are central to the LLM era, aiming to evolve EE from static extraction into a structurally reliable, agent ready perception and memory layer for open world systems.

Xiang Li, Yiqun Yao, Xin Jiang et al.

Large language models (LLMs) have showcased profound capabilities in language understanding and generation, facilitating a wide array of applications. However, there is a notable paucity of detailed, open-sourced methodologies on efficiently scaling LLMs beyond 50 billion parameters with minimum trial-and-error cost and computational resources. In this report, we introduce Tele-FLM (aka FLM-2), a 52B open-sourced multilingual large language model that features a stable, efficient pre-training paradigm and enhanced factual judgment capabilities. Tele-FLM demonstrates superior multilingual language modeling abilities, measured by BPB on textual corpus. Besides, in both English and Chinese foundation model evaluation, it is comparable to strong open-sourced models that involve larger pre-training FLOPs, such as Llama2-70B and DeepSeek-67B. In addition to the model weights, we share the core designs, engineering practices, and training details, which we expect to benefit both the academic and industrial communities.

Sicong Leng, Jing Wang, Jiaxi Li et al.

Large multimodal reasoning models have achieved rapid progress, but their advancement is constrained by two major limitations: the absence of open, large-scale, high-quality long chain-of-thought (CoT) data, and the instability of reinforcement learning (RL) algorithms in post-training. Group Relative Policy Optimization (GRPO), the standard framework for RL fine-tuning, is prone to gradient vanishing when reward variance is low, which weakens optimization signals and impairs convergence. This work makes three contributions: (1) We propose Variance-Aware Sampling (VAS), a data selection strategy guided by Variance Promotion Score (VPS) that combines outcome variance and trajectory diversity to promote reward variance and stabilize policy optimization. (2) We release large-scale, carefully curated resources containing ~1.6M long CoT cold-start data and ~15k RL QA pairs, designed to ensure quality, difficulty, and diversity, along with a fully reproducible end-to-end training codebase. (3) We open-source a family of multimodal reasoning models in multiple scales, establishing standardized baselines for the community. Experiments across mathematical reasoning benchmarks demonstrate the effectiveness of both the curated data and the proposed VAS. Comprehensive ablation studies and analyses provide further insight into the contributions of each component. In addition, we theoretically establish that reward variance lower-bounds the expected policy gradient magnitude, with VAS serving as a practical mechanism to realize this guarantee. Our code, data, and checkpoints are available at https://github.com/LengSicong/MMR1.

Mathieu Ravaut, Hao Zhang, Lu Xu et al.

Conversational recommender systems (CRS) aim to recommend relevant items to users by eliciting user preference through natural language conversation. Prior work often utilizes external knowledge graphs for items' semantic information, a language model for dialogue generation, and a recommendation module for ranking relevant items. This combination of multiple components suffers from a cumbersome training process, and leads to semantic misalignment issues between dialogue generation and item recommendation. In this paper, we represent items in natural language and formulate CRS as a natural language processing task. Accordingly, we leverage the power of pre-trained language models to encode items, understand user intent via conversation, perform item recommendation through semantic matching, and generate dialogues. As a unified model, our PECRS (Parameter-Efficient CRS), can be optimized in a single stage, without relying on non-textual metadata such as a knowledge graph. Experiments on two benchmark CRS datasets, ReDial and INSPIRED, demonstrate the effectiveness of PECRS on recommendation and conversation. Our code is available at: https://github.com/Ravoxsg/efficient_unified_crs.

Yitong Ji, Aixin Sun, Jie Zhang et al.

Popularity is often included in experimental evaluation to provide a reference performance for a recommendation task. To understand how popularity baseline is defined and evaluated, we sample 12 papers from top-tier conferences including KDD, WWW, SIGIR, and RecSys, and 6 open source toolkits. We note that the widely adopted MostPop baseline simply ranks items based on the number of interactions in the training data. We argue that the current evaluation of popularity (i) does not reflect the popular items at the time when a user interacts with the system, and (ii) may recommend items released after a user's last interaction with the system. On the widely used MovieLens dataset, we show that the performance of popularity could be significantly improved by 70% or more, if we consider the popular items at the time point when a user interacts with the system. We further show that, on MovieLens dataset, the users having lower tendencies on movies tend to follow the crowd and rate more popular movies. Movie lovers who rate a large number of movies, rate movies based on their own preferences and interests. Through this study, we call for a re-visit of the popularity baseline in recommender system to better reflect its effectiveness.

Shuai Zhang, Yi Tay, Lina Yao et al.

Deep learning based recommender systems have been extensively explored in recent years. However, the large number of models proposed each year poses a big challenge for both researchers and practitioners in reproducing the results for further comparisons. Although a portion of papers provides source code, they adopted different programming languages or different deep learning packages, which also raises the bar in grasping the ideas. To alleviate this problem, we released the open source project: \textbf{DeepRec}. In this toolkit, we have implemented a number of deep learning based recommendation algorithms using Python and the widely used deep learning package - Tensorflow. Three major recommendation scenarios: rating prediction, top-N recommendation (item ranking) and sequential recommendation, were considered. Meanwhile, DeepRec maintains good modularity and extensibility to easily incorporate new models into the framework. It is distributed under the terms of the GNU General Public License. The source code is available at github: \url{https://github.com/cheungdaven/DeepRec}

Siqi Fan, Xin Jiang, Xiang Li et al.

Due to the large number of parameters, the inference phase of Large Language Models (LLMs) is resource-intensive. However, not all requests posed to LLMs are equally difficult to handle. Through analysis, we show that for some tasks, LLMs can achieve results comparable to the final output at some intermediate layers. That is, not all layers of LLMs are necessary during inference. If we can predict at which layer the inferred results match the final results (produced by evaluating all layers), we could significantly reduce the inference cost. To this end, we propose a simple yet effective algorithm named AdaInfer to adaptively terminate the inference process for an input instance. AdaInfer relies on easily obtainable statistical features and classic classifiers like SVM. Experiments on well-known LLMs like the Llama2 series and OPT, show that AdaInfer can achieve an average of 17.8% pruning ratio, and up to 43% on sentiment tasks, with nearly no performance drop (<1%). Because AdaInfer does not alter LLM parameters, the LLMs incorporated with AdaInfer maintain generalizability across tasks.

Yubo Ma, Zhibin Gou, Junheng Hao et al. · microsoft-research, tsinghua

Scientific reasoning poses an excessive challenge for even the most advanced Large Language Models (LLMs). To make this task more practical and solvable for LLMs, we introduce a new task setting named tool-augmented scientific reasoning. This setting supplements LLMs with scalable toolsets, and shifts the focus from pursuing an omniscient problem solver to a proficient tool-user. To facilitate the research of such setting, we construct a tool-augmented training corpus named MathFunc which encompasses over 30,000 samples and roughly 6,000 tools. Building on MathFunc, we develop SciAgent to retrieve, understand and, if necessary, use tools for scientific problem solving. Additionally, we craft a benchmark, SciToolBench, spanning five scientific domains to evaluate LLMs' abilities with tool assistance. Extensive experiments on SciToolBench confirm the effectiveness of SciAgent. Notably, SciAgent-Mistral-7B surpasses other LLMs with the same size by more than 13% in absolute accuracy. Furthermore, SciAgent-DeepMath-7B shows much superior performance than ChatGPT.

Zongmin Li, Jian Su, Farah Benamara et al.

Large language models (LLMs) are often assumed to contain ``safety regions'' -- parameter subsets whose modification directly influences safety behaviors. We conduct a systematic evaluation of four safety region identification methods spanning different parameter granularities, from individual weights to entire Transformer layers, across four families of backbone LLMs with varying sizes. Using ten safety identification datasets, we find that the identified safety regions exhibit only low to moderate overlap, as measured by IoU. The overlap drops significantly when the safety regions are further refined using utility datasets (\ie non-harmful queries). These results suggest that current techniques fail to reliably identify a stable, dataset-agnostic safety region.

Yixin Cao, Shibo Hong, Xinze Li et al.

Large Language Models (LLMs) are advancing at an amazing speed and have become indispensable across academia, industry, and daily applications. To keep pace with the status quo, this survey probes the core challenges that the rise of LLMs poses for evaluation. We identify and analyze two pivotal transitions: (i) from task-specific to capability-based evaluation, which reorganizes benchmarks around core competencies such as knowledge, reasoning, instruction following, multi-modal understanding, and safety; and (ii) from manual to automated evaluation, encompassing dynamic dataset curation and "LLM-as-a-judge" scoring. Yet, even with these transitions, a crucial obstacle persists: the evaluation generalization issue. Bounded test sets cannot scale alongside models whose abilities grow seemingly without limit. We will dissect this issue, along with the core challenges of the above two transitions, from the perspectives of methods, datasets, evaluators, and metrics. Due to the fast evolving of this field, we will maintain a living GitHub repository (links are in each section) to crowd-source updates and corrections, and warmly invite contributors and collaborators.

Xinze Li, Yixin Cao, Yubo Ma et al.

Extending context windows (i.e., Long Context, LC) and using retrievers to selectively access relevant information (i.e., Retrieval-Augmented Generation, RAG) are the two main strategies to enable LLMs to incorporate extremely long external contexts. This paper revisits recent studies on this topic, highlighting their key insights and discrepancies. We then provide a more comprehensive evaluation by filtering out questions answerable without external context, identifying the most effective retrieval methods, and expanding the datasets. We show that LC generally outperforms RAG in question-answering benchmarks, especially for Wikipedia-based questions. Summarization-based retrieval performs comparably to LC, while chunk-based retrieval lags behind. However, RAG has advantages in dialogue-based and general question queries. These insights underscore the trade-offs between RAG and LC strategies, offering guidance for future optimization of LLMs with external knowledge sources. We also provide an in-depth discussion on this topic, highlighting the overlooked importance of context relevance in existing studies.

Guoxiu He, Xin Song, Aixin Sun

As real-world knowledge evolves, the information embedded within large language models (LLMs) can become outdated, inadequate, or erroneous. Model editing has emerged as a prominent approach for updating LLMs' knowledge with minimal computational costs and parameter changes. This approach typically identifies and adjusts specific model parameters associated with newly acquired knowledge. However, existing methods often underestimate the adverse effects that parameter modifications can have on broadly distributed knowledge. More critically, post-edit LLMs frequently struggle with multi-hop reasoning and continuous knowledge updates. Although various studies have discussed these shortcomings, there is a lack of comprehensive evaluation. In this paper, we provide an evaluation of ten model editing methods along four dimensions: reliability, generalization, locality, and portability. Results confirm that all ten popular model editing methods show significant shortcomings across multiple dimensions, suggesting model editing is less promising. We then propose a straightforward method called Selective Contextual Reasoning (SCR), for knowledge updating. SCR does not modify model parameters but harnesses LLM's inherent contextual reasoning capabilities utilizing the updated knowledge pieces. Under SCR, an LLM first assesses whether an incoming query falls within the scope of an external knowledge base. If it does, the relevant external knowledge texts are contextualized to enhance reasoning; otherwise, the query is answered directly. We evaluate SCR against the ten model editing methods on two counterfactual datasets with three backbone LLMs. Empirical results confirm the effectiveness and efficiency of contextual reasoning for knowledge updating.

Jie Zou, Aixin Sun, Cheng Long et al.

In conversational recommender systems (CRSs), conversations usually involve a set of items and item-related entities or attributes, e.g., director is a related entity of a movie. These items and item-related entities are often mentioned along the development of a dialog, leading to potential sequential dependencies among them. However, most of existing CRSs neglect these potential sequential dependencies. In this article, we first propose a Transformer-based sequential conversational recommendation method, named TSCR, to model the sequential dependencies in the conversations to improve CRS. In TSCR, we represent conversations by items and the item-related entities, and construct user sequences to discover user preferences by considering both the mentioned items and item-related entities. Based on the constructed sequences, we deploy a Cloze task to predict the recommended items along a sequence. Meanwhile, in certain domains, knowledge graphs formed by the items and their related entities are readily available, which provide various different kinds of associations among them. Given that TSCR does not benefit from such knowledge graphs, we then propose a knowledge graph enhanced version of TSCR, called TSCRKG. In specific, we leverage the knowledge graph to offline initialize our model TSCRKG, and augment the user sequence of conversations (i.e., sequence of the mentioned items and item-related entities in the conversation) with multi-hop paths in the knowledge graph. Experimental results demonstrate that our TSCR model significantly outperforms state-of-the-art baselines, and the enhanced version TSCRKG further improves recommendation performance on top of TSCR.

Siqi Fan, Bowen Qin, Peng Han et al.

Recent thinking models trained with reinforcement learning and backward-checking CoT often suffer from overthinking: they produce excessively long outputs even on simple problems, wasting computation. Existing evaluations, based on token efficiency, give an incomplete view as they neglect problem difficulty and intermediate computation costs. We formalize reasoning efficiency as a relative measure between thinking and instruct models, treating instruct models as the minimal-effort baseline. A systematic study across four thinking models and multiple benchmarks reveals two consistent patterns: (i) instruct models achieve higher efficiency overall, and (ii) problem difficulty affects efficiency, with thinking models wasting computation on easy problems but providing value on harder ones. Building on this insight, we propose COTHINK, a simple two-stage pipeline: an instruct model drafts a brief outline, and a thinking model expands it. On GSM8K, MATH500, and AIME24, COTHINK cuts token usage by 21.1% while keeping accuracy on four thinking models, and remains competitive with strong efficiency baselines.

Yubo Ma, Jinsong Li, Yuhang Zang et al. · pku

Despite the strong performance of ColPali/ColQwen2 in Visualized Document Retrieval (VDR), it encodes each page into multiple patch-level embeddings and leads to excessive memory usage. This empirical study investigates methods to reduce patch embeddings per page at minimum performance degradation. We evaluate two token-reduction strategies: token pruning and token merging. Regarding token pruning, we surprisingly observe that a simple random strategy outperforms other sophisticated pruning methods, though still far from satisfactory. Further analysis reveals that pruning is inherently unsuitable for VDR as it requires removing certain page embeddings without query-specific information. Turning to token merging (more suitable for VDR), we search for the optimal combinations of merging strategy across three dimensions and develop Light-ColPali/ColQwen2. It maintains 98.2% of retrieval performance with only 11.8% of original memory usage, and preserves 94.6% effectiveness at 2.8% memory footprint. We expect our empirical findings and resulting Light-ColPali/ColQwen2 offer valuable insights and establish a competitive baseline for future research towards efficient VDR.

Jing Wang, Fengzhuo Zhang, Xiaoli Li et al.

A variety of Auto-Regressive Video Diffusion Models (ARVDM) have achieved remarkable successes in generating realistic long-form videos. However, theoretical analyses of these models remain scant. In this work, we develop theoretical underpinnings for these models and use our insights to improve the performance of existing models. We first develop Meta-ARVDM, a unified framework of ARVDMs that subsumes most existing methods. Using Meta-ARVDM, we analyze the KL-divergence between the videos generated by Meta-ARVDM and the true videos. Our analysis uncovers two important phenomena inherent to ARVDM -- error accumulation and memory bottleneck. By deriving an information-theoretic impossibility result, we show that the memory bottleneck phenomenon cannot be avoided. To mitigate the memory bottleneck, we design various network structures to explicitly use more past frames. We also achieve a significantly improved trade-off between the mitigation of the memory bottleneck and the inference efficiency by compressing the frames. Experimental results on DMLab and Minecraft validate the efficacy of our methods. Our experiments also demonstrate a Pareto-frontier between the error accumulation and memory bottleneck across different methods.

Yequan Wang, Aixin Sun · tencent-ai, tsinghua

Artificial General Intelligence (AGI) is often envisioned as inherently embodied. With recent advances in robotics and foundational AI models, we stand at the threshold of a new era-one marked by increasingly generalized embodied AI systems. This paper contributes to the discourse by introducing a systematic taxonomy of Embodied AGI spanning five levels (L1-L5). We review existing research and challenges at the foundational stages (L1-L2) and outline the key components required to achieve higher-level capabilities (L3-L5). Building on these insights and existing technologies, we propose a conceptual framework for an L3+ robotic brain, offering both a technical outlook and a foundation for future exploration.

Yiqun Yao, Xiang Li, Xin Jiang et al.

Humans naturally process real-world multimodal information in a full-duplex manner. In artificial intelligence, replicating this capability is essential for advancing model development and deployment, particularly in embodied contexts. The development of multimodal models faces two primary challenges: (1) effectively handling more than three modalities-such as vision, audio, and text; and (2) delivering full-duplex responses to rapidly evolving human instructions. To facilitate research on models that support both omnimodal processing and full duplexity, we present RoboEgo (alias: FLM-Ego), a unified model system designed to address both challenges. RoboEgo incorporates a backbone architecture and algorithms that natively support full duplexity, achieving a theoretical duplex latency of 80 ms. In streaming visually grounded conversations under real-world conditions, RoboEgo exhibits superior responsiveness and speech naturalness, while maintaining comparable content qualities to state-of-the-art semi-duplex omnimodal models-a feat previously considered unattainable by native full-duplex systems.

Siqi Fan, Xiusheng Huang, Yiqun Yao et al.

Large language models (LLMs) can carry out human-like dialogue, but unlike humans, they are stateless due to the superposition property. However, during multi-turn, multi-agent interactions, LLMs begin to exhibit consistent, character-like behaviors, hinting at a form of emergent lifelong learning. Despite this, existing benchmarks often fail to capture these dynamics, primarily focusing on static, open-ended evaluations. To address this gap, we introduce LIFESTATE-BENCH, a benchmark designed to assess lifelong learning in LLMs. It features two episodic datasets: Hamlet and a synthetic script collection, rich in narrative structure and character interactions. Our fact checking evaluation probes models' self-awareness, episodic memory retrieval, and relationship tracking, across both parametric and non-parametric approaches. Experiments on models like Llama3.1-8B, GPT-4-turbo, and DeepSeek R1, we demonstrate that nonparametric methods significantly outperform parametric ones in managing stateful learning. However, all models exhibit challenges with catastrophic forgetting as interactions extend, highlighting the need for further advancements in lifelong learning.

Ziqi Jin, Lei Wang, Ziwei Luo et al.

Autoregressive (AR) language models generate text one token at a time, even when consecutive tokens are highly predictable given earlier context. We introduce MARS (Mask AutoRegreSsion), a lightweight fine-tuning method that teaches an instruction-tuned AR model to predict multiple tokens per forward pass. MARS adds no architectural modifications, no extra parameters, and produces a single model that can still be called exactly like the original AR model with no performance degradation. Unlike speculative decoding, which maintains a separate draft model alongside the target, or multi-head approaches such as Medusa, which attach additional prediction heads, MARS requires only continued training on existing instruction data. When generating one token per forward pass, MARS matches or exceeds the AR baseline on six standard benchmarks. When allowed to accept multiple tokens per step, it maintains baseline-level accuracy while achieving 1.5-1.7x throughput. We further develop a block-level KV caching strategy for batch inference, achieving up to 1.71x wall-clock speedup over AR with KV cache on Qwen2.5-7B. Finally, MARS supports real-time speed adjustment via confidence thresholding: under high request load, the serving system can increase throughput on the fly without swapping models or restarting, providing a practical latency-quality knob for deployment.

Yiqun Yao, Naitong Yu, Xiang Li et al.

We introduce EgoMem, the first lifelong memory agent tailored for full-duplex models that process real-time omnimodal streams. EgoMem enables real-time models to recognize multiple users directly from raw audiovisual streams, to provide personalized response, and to maintain long-term knowledge of users' facts, preferences, and social relationships extracted from audiovisual history. EgoMem operates with three asynchronous processes: (i) a retrieval process that dynamically identifies user via face and voice, and gathers relevant context from a long-term memory; (ii) an omnimodal dialog process that generates personalized audio responses based on the retrieved context; and (iii) a memory management process that automatically detects dialog boundaries from omnimodal streams, and extracts necessary information to update the long-term memory. Unlike existing memory agents for LLMs, EgoMem relies entirely on raw audiovisual streams, making it especially suitable for lifelong, real-time, and embodied scenarios. Experimental results demonstrate that EgoMem's retrieval and memory management modules achieve over 95% accuracy on the test set. When integrated with a fine-tuned RoboEgo omnimodal chatbot, the system achieves fact-consistency scores above 87% in real-time personalized dialogs, establishing a strong baseline for future research.

Yiqun Yao, Xiang Li, Xin Jiang et al.

Full-duplex dialog models aim to listen and speak simultaneously, delivering rapid responses to dynamic user input. Among different solutions to full duplexity, a native solution merges multiple channels in each time step, achieving the lowest latency. However, prevailing designs break down the textual monologue sentences for word-level alignment with audio streams, which degrades language modeling abilities. To help address this issue, we introduce natural monologues, which are composed by continuous sentences and waiting intervals, mimicking humanoid cognitive behavior in dialogs. We find a proper training paradigm to be critical for semantically aligning natural monologues with audio. To this end, we develop a dual training paradigm that alternates the position of the monologues, either leading or trailing the audio, across different training stages. A combination of our natural monologue and dual training strategy is applied in developing FLM-Audio, our 7B spoken dialog chatbot with native full-duplexity. As confirmed by experimental results, FLM-Audio achieves superior response qualities and chatting experiences while requiring significantly less training data.

Qishuai Zhong, Zongmin Li, Siqi Fan et al.

Effective engagement by large language models (LLMs) requires adapting responses to users' sociodemographic characteristics, such as age, occupation, and education level. While many real-world applications leverage dialogue history for contextualization, existing evaluations of LLMs' behavioral adaptation often focus on single-turn prompts. In this paper, we propose a framework to evaluate LLM adaptation when attributes are introduced either (1) explicitly via user profiles in the prompt or (2) implicitly through multi-turn dialogue history. We assess the consistency of model behavior across these modalities. Using a multi-agent pipeline, we construct a synthetic dataset pairing dialogue histories with distinct user profiles and employ questions from the Value Survey Module (VSM 2013) (Hofstede and Hofstede, 2016) to probe value expression. Our findings indicate that most models adjust their expressed values in response to demographic changes, particularly in age and education level, but consistency varies. Models with stronger reasoning capabilities demonstrate greater alignment, indicating the importance of reasoning in robust sociodemographic adaptation.

KM Pooja, Cheng Long, Aixin Sun

The task of entity linking, which involves associating mentions with their respective entities in a knowledge graph, has received significant attention due to its numerous potential applications. Recently, various multimodal entity linking (MEL) techniques have been proposed, targeted to learn comprehensive embeddings by leveraging both text and vision modalities. The selection of high-quality negative samples can potentially play a crucial role in metric/representation learning. However, to the best of our knowledge, this possibility remains unexplored in existing literature within the framework of MEL. To fill this gap, we address the multimodal entity linking problem in a generative adversarial setting where the generator is responsible for generating high-quality negative samples, and the discriminator is assigned the responsibility for the metric learning tasks. Since the generator is involved in generating samples, which is a discrete process, we optimize it using policy gradient techniques and propose a policy gradient-based generative adversarial network for multimodal entity linking (PGMEL). Experimental results based on Wiki-MEL, Richpedia-MEL and WikiDiverse datasets demonstrate that PGMEL learns meaningful representation by selecting challenging negative samples and outperforms state-of-the-art methods.