Yulan Hu

Lulu Zheng, Wenjin Yang, Xiangwen Zhang et al.

Multi-stakeholder tasks require one output to satisfy users with conflicting preferences. Holistic LLM judges conflate utility estimation and utility aggregation, yielding unstable implicit weights. We show empirically and theoretically that this aggregation-specific \emph{weighting noise} can create large score shifts when stakeholder satisfaction is dispersed; in our experiments, these weight-induced shifts also increase with stakeholder count. We propose \textsc{DecompR}: counterfactual-calibrated weights are fixed from query structure before candidate scoring, while per-role utilities are estimated independently, removing candidate-dependent weight drift and reducing estimation noise.

Xiang Cheng, Yulan Hu, Lulu Zheng et al.

Travel planning is a realistic task for evaluating the planning and tool-use abilities of LLM agents. However, existing benchmarks typically assume only a single user, thereby avoiding one of the most challenging aspects of real-world scenarios: an agent's ability to identify and resolve conflicts among multiple users. To address this gap, we introduce \textbf{GroupTravelBench}, the first benchmark for \textbf{multi-user, multi-turn} travel planning. Based on real user profiles, POI data, and ticket price data, we synthesize 650 tasks and divide them into three difficulty levels. Beyond standard abilities in single-user itinerary planning, such as multi-step reasoning and tool use, our benchmark further evaluates three key capabilities required for travel agents: \emph{(i) elicitation} -- proactively engaging in multi-turn dialogue to gather preferences from each user; \emph{(ii) coordination} -- resolving conflicts among users through compromise or subgrouping strategies; and \emph{(iii) planning} -- searching for travel plans that maximize overall group utility while maintaining fairness and feasibility. To simulate real-world conversational itinerary planning while enabling reliable tool use and offline evaluation, we build an interactive sandbox environment with cached real-world tool data. We evaluate a wide range of LLMs and find that even frontier models still show substantial weaknesses in preference coverage and group fairness. \textit{GroupTravelBench} provides a practical and reproducible benchmark for advancing research on LLM agents for real-world travel planning.

Zechang Xiong, Da Li, Kexin Tang et al.

Multimodal models often converge to a dominant-modality solution, in which a stronger, faster-converging modality overshadows weaker ones. This modality imbalance causes suboptimal performance. Existing methods attempt to balance different modalities by reweighting gradients or losses. However, they overlook the fact that each modality has finite information capacity. In this work, we propose IIBalance, a multimodal learning framework that aligns the modality contributions with Intrinsic Information Budgets (IIB). We propose a task-grounded estimator of each modality's IIB, transforming its capacity into a global prior over modality contributions. Anchored by the highest-budget modality, we design a prototype-based relative alignment mechanism that corrects semantic drift only when weaker modalities deviate from their budgeted potential, rather than forcing imitation. During inference, we propose a probabilistic gating module that integrates the global budgets with sample-level uncertainty to generate calibrated fusion weights. Experiments on three representative benchmarks demonstrate that IIBalance consistently outperforms state-of-the-art balancing methods and achieves better utilization of complementary modality cues. Our code is available at: https://github.com/XiongZechang/IIBalance.

Jinman Zhao, Zifan Qian, Linbo Cao et al.

Role-play in large language models (LLMs) enhances their ability to generate contextually relevant and high-quality responses by simulating diverse cognitive perspectives. However, our study identifies significant risks associated with this technique. First, we demonstrate that autotuning, a method used to auto-select models' roles based on the question, can lead to the generation of harmful outputs, even when the model is tasked with adopting neutral roles. Second, we investigate how different roles affect the likelihood of generating biased or harmful content. Through testing on benchmarks containing stereotypical and harmful questions, we find that role-play consistently amplifies the risk of biased outputs. Our results underscore the need for careful consideration of both role simulation and tuning processes when deploying LLMs in sensitive or high-stakes contexts.

Zhicong Li, Lingjie Jiang, Yulan Hu et al.

Critique-guided reinforcement learning (RL) has emerged as a powerful paradigm for training LLM agents by augmenting sparse outcome rewards with natural-language feedback. However, current methods often rely on static or offline critic models, which fail to adapt as the policy evolves. In on-policy RL, the agent's error patterns shift over time, causing stationary critics to become stale and providing feedback of diminishing utility. To address this, we introduce ECHO (Evolving Critic for Hindsight-Guided Optimization)}, a framework that jointly optimizes the policy and critic through a synchronized co-evolutionary loop. ECHO utilizes a cascaded rollout mechanism where the critic generates multiple diagnoses for an initial trajectory, followed by policy refinement to enable group-structured advantage estimation. We address the challenge of learning plateaus via a saturation-aware gain shaping objective, which rewards the critic for inducing incremental improvements in high-performing trajectories. By employing dual-track GRPO updates, ECHO ensures the critic's feedback stays synchronized with the evolving policy. Experimental results show that ECHO yields more stable training and higher long-horizon task success across open-world environments.

Yulan Hu, Sheng Ouyang, Jingyu Liu et al.

Graph contrastive learning (GCL) has emerged as a representative graph self-supervised method, achieving significant success. The currently prevalent optimization objective for GCL is InfoNCE. Typically, it employs augmentation techniques to obtain two views, where a node in one view acts as the anchor, the corresponding node in the other view serves as the positive sample, and all other nodes are regarded as negative samples. The goal is to minimize the distance between the anchor node and positive samples and maximize the distance to negative samples. However, due to the lack of label information during training, InfoNCE inevitably treats samples from the same class as negative samples, leading to the issue of false negative samples. This can impair the learned node representations and subsequently hinder performance in downstream tasks. While numerous methods have been proposed to mitigate the impact of false negatives, they still face various challenges. For instance, while increasing the number of negative samples can dilute the impact of false negatives, it concurrently increases computational burden. Thus, we propose GraphRank, a simple yet efficient graph contrastive learning method that addresses the problem of false negative samples by redefining the concept of negative samples to a certain extent, thereby avoiding the issue of false negative samples. The effectiveness of GraphRank is empirically validated through experiments on the node, edge, and graph level tasks.

AMAP AI Agent Team, Yulan Hu, Xiangwen Zhang et al.

We present STAgent, an agentic large language model tailored for spatio-temporal understanding, designed to solve complex tasks such as constrained point-of-interest discovery and itinerary planning. STAgent is a specialized model capable of interacting with ten distinct tools within spatio-temporal scenarios, enabling it to explore, verify, and refine intermediate steps during complex reasoning. Notably, STAgent effectively preserves its general capabilities. We empower STAgent with these capabilities through three key contributions: (1) a stable tool environment that supports over ten domain-specific tools, enabling asynchronous rollout and training; (2) a hierarchical data curation framework that identifies high-quality data like a needle in a haystack, curating high-quality queries by retaining less than 1\% of the raw data, emphasizing both diversity and difficulty; and (3) a cascaded training recipe that starts with a seed SFT stage acting as a guardian to measure query difficulty, followed by a second SFT stage fine-tuned on queries with high certainty, and an ultimate RL stage that leverages data of low certainty. Initialized with Qwen3-30B-A3B to establish a strong SFT foundation and leverage insights into sample difficulty, STAgent yields promising performance on TravelBench while maintaining its general capabilities across a wide range of general benchmarks, thereby demonstrating the effectiveness of our proposed agentic model.

Sheng Ouyang, Yulan Hu, Ge Chen et al.

Large Language Models (LLMs) have achieved impressive results in processing text data, which has sparked interest in applying these models beyond textual data, such as graphs. In the field of graph learning, there is a growing interest in harnessing LLMs to comprehend and manipulate graph-structured data. Existing research predominantly focuses on graphs with rich textual features, such as knowledge graphs or text attribute graphs, leveraging LLMs' ability to process text but inadequately addressing graph structure. This work specifically aims to assess and enhance LLMs' abilities to comprehend and utilize the structural knowledge inherent in graph data itself, rather than focusing solely on graphs rich in textual content. To achieve this, we introduce the \textbf{G}raph \textbf{U}nderstanding for \textbf{N}atural Language \textbf{D}riven \textbf{A}nalytical \textbf{M}odel (\model). This model adapts LLMs to better understand and engage with the structure of graph data, enabling them to perform complex reasoning tasks by leveraging the graph's structure itself. Our experimental evaluations on graph reasoning benchmarks not only substantiate that \model~ outperforms the SOTA baselines for comparisons. But also reveals key factors affecting the graph reasoning capabilities of LLMs. Moreover, we provide a theoretical analysis illustrating how reasoning paths can enhance LLMs' reasoning capabilities.

Hao Yi, Qingyang Li, Yulan Hu et al.

Recently, enhancing the numerical and logical reasoning capability of Large Language Models (LLMs) has emerged as a research hotspot. Existing methods face several limitations: inference-phase techniques (e.g., Chain of Thoughts) rely on prompt selection and the pretrained knowledge; sentence-level Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO) struggle with step-wise mathematical correctness and depend on stronger models distillation or human annotations; while Reinforcement Learning (RL) approaches incur high GPU memory costs and unstable training. To address these, we propose \textbf{S}elf-training framework integrating \textbf{P}rocess \textbf{P}reference learning using \textbf{D}ynamic value margin (SPPD). SPPD leverages a process-based Markov Decision Process (MDP) and Bellman optimality equation to derive \textbf{dynamic value margin} on step-level preference optimization, which employs tree-based self-sampling on model responses \textbf{without any distillation} from other models. Furthermore, we theoretically prove that SPPD is \textbf{equivalent to on-policy policy gradient methods} under reward constraints. Experiments on 7B-scale models demonstrate superior performance across in-domain and out-domain mathematical benchmarks. We open-source our code at \href{https://anonymous.4open.science/r/SSDPO-D-DCDD}{https://anonymous.4open.science/r/SPPD-DCDD}.

Kaihui Chen, Hao Yi, Qingyang Li et al.

Enhancing the conformity of large language models (LLMs) to human preferences remains an ongoing research challenge. Recently, offline approaches such as Direct Preference Optimization (DPO) have gained prominence as attractive options due to offering effective improvement in simple, efficient, and stable without interactions with reward models. However, these offline preference optimization methods highly rely on the quality of pairwise preference samples. Meanwhile, numerous iterative methods require additional training of reward models to select positive and negative samples from the model's own generated responses for preference learning. Furthermore, as LLMs' capabilities advance, it is quite challenging to continuously construct high-quality positive and negative preference instances from the model's outputs due to the lack of diversity. To tackle these challenges, we propose TSO, or Self-Training with Scaled Preference Optimization, a framework for preference optimization that conducts self-training preference learning without training an additional reward model. TSO enhances the diversity of responses by constructing a model matrix and incorporating human preference responses. Furthermore, TSO introduces corrections for model preference errors through human and AI feedback. Finally, TSO adopts iterative and dual clip reward strategies to update the reference model and its responses, adaptively adjusting preference data and balancing the optimization process. Experimental results demonstrate that TSO outperforms existing mainstream methods on various alignment evaluation benchmarks, providing practical insight into preference data construction and model training strategies in the alignment domain.

Jiaxuan Wang, Xuan Ouyang, Zhiyu Chen et al.

On-policy self-distillation (self-OPD) densifies reinforcement learning with verifiable rewards (RLVR) by letting a policy teach itself under privileged context. We find that when this guidance spans the full response, all-token KL spends gradients on mostly redundant positions and amplifies privileged-information leakage, causing entropy rise, shortened reasoning, and out-of-distribution degradation in long-horizon math training. We propose Token-Routed Alignment for Critical rEasoning (TRACE), which distills only on annotator-marked critical spans: forward KL on key spans of correct rollouts, optional reverse KL on localized error spans, and GRPO on all remaining tokens, with the KL channel annealed away after a short warm-up. Our analysis explains TRACE through two effects: forward KL provides non-vanishing lift to teacher-supported tokens that the student under-allocates, while span masking and decay keep cumulative privileged-gradient exposure finite. On four held-out math benchmarks plus GPQA-Diamond, TRACE improves over GRPO by 2.76 percentage points on average and preserves the Qwen3-8B base OOD score on GPQA-Diamond, where GRPO and all-token self-OPD baselines degrade. Gains persist under online self-annotation (+1.90 percentage points, about 69% of the strong-API gain), reducing the concern that TRACE merely imports external annotator capability. Across scales, the best routed action is base-dependent: on Qwen3-8B it is forward KL on key spans, while on Qwen3-1.7B it shifts to reverse KL on error spans.

Yulan Hu, Sheng Ouyang, Zhirui Yang et al.

Class imbalance in graph data presents significant challenges for node classification. While existing methods, such as SMOTE-based approaches, partially mitigate this issue, they still exhibit limitations in constructing imbalanced graphs. Generative self-supervised learning (SSL) methods, exemplified by graph autoencoders (GAEs), offer a promising solution by directly generating minority nodes from the data itself, yet their potential remains underexplored. In this paper, we delve into the shortcomings of SMOTE-based approaches in the construction of imbalanced graphs. Furthermore, we introduce VIGraph, a simple yet effective generative SSL approach that relies on the Variational GAE as the fundamental model. VIGraph strictly adheres to the concept of imbalance when constructing imbalanced graphs and innovatively leverages the variational inference (VI) ability of Variational GAE to generate nodes for minority classes. VIGraph introduces comprehensive training strategies, including cross-view contrastive learning at the decoding phase to capture semantic knowledge, adjacency matrix reconstruction to preserve graph structure, and alignment strategy to ensure stable training. VIGraph can generate high-quality nodes directly usable for classification, eliminating the need to integrate the generated nodes back to the graph as well as additional retraining found in SMOTE-based methods. We conduct extensive experiments, results from which demonstrate the superiority and generality of our approach.

Yulan Hu, Zhirui Yang, Sheng Ouyang et al.

Self-Supervised Learning (SSL) has shown significant potential and has garnered increasing interest in graph learning. However, particularly for generative SSL methods, its potential in Heterogeneous Graph Learning (HGL) remains relatively underexplored. Generative SSL utilizes an encoder to map the input graph into a latent representation and a decoder to recover the input graph from the latent representation. Previous HGL SSL methods generally design complex strategies to capture graph heterogeneity, which heavily rely on contrastive view construction strategies that are often non-trivial. Yet, refining the latent representation in generative SSL can effectively improve graph learning results. In this study, we propose HGVAE, a generative SSL method specially designed for HGL. Instead of focusing on designing complex strategies to capture heterogeneity, HGVAE centers on refining the latent representation. Specifically, HGVAE innovatively develops a contrastive task based on the latent representation. To ensure the hardness of negative samples, we develop a progressive negative sample generation (PNSG) mechanism that leverages the ability of Variational Inference (VI) to generate high-quality negative samples. As a pioneer in applying generative SSL for HGL, HGVAE refines the latent representation, thereby compelling the model to learn high-quality representations. Compared with various state-of-the-art (SOTA) baselines, HGVAE achieves impressive results, thus validating its superiority.

Hao Yi, Yulan Hu, Xin Li et al.

Large Language Models (LLMs) have recently improved mathematical reasoning through Reinforcement Learning with Verifiable Reward (RLVR). However, existing RLVR algorithms require large query budgets, making annotation costly. We investigate whether fewer but more informative queries can yield similar or superior performance, introducing active learning (AL) into RLVR. We identify that classic AL sampling strategies fail to outperform random selection in this setting, due to ignoring objective uncertainty when only selecting by subjective uncertainty. This work proposes an uncertainty consistency metric to evaluate how well subjective uncertainty aligns with objective uncertainty. In the offline setting, this alignment is measured using the Point-Biserial Correlation Coefficient (PBC). For online training, because of limited sampling and dynamically shifting output distributions, PBC estimation is difficult. Therefore, we introduce a new online variant, computed from normalized advantage and subjective uncertainty. Theoretically, we prove that the online variant is strictly negatively correlated with offline PBC and supports better sample selection. Experiments show our method consistently outperforms random and classic AL baselines, achieving full-dataset performance while training on only 30% of the data, effectively reducing the cost of RLVR for reasoning tasks.

Xiang Cheng, Yulan Hu, Xiangwen Zhang et al.

Travel planning is a natural real-world task to test large language models (LLMs) planning and tool-use abilities. Although prior work has studied LLM performance on travel planning, existing settings still differ from real-world needs, mainly due to limited domain coverage, insufficient modeling of users' implicit preferences in multi-turn conversations, and a lack of clear evaluation of agents' capability boundaries. To mitigate these gaps, we propose \textbf{TravelBench}, a benchmark for fully real-world travel planning. We collect user queries, user profile and tools from real scenarios, and construct three subtasks-Single-Turn, Multi-Turn, and Unsolvable-to evaluate agent's three core capabilities in real settings: (1) solving problems autonomously, (2) interacting with users over multiple turns to refine requirements, and (3) recognizing the limits of own abilities. To enable stable tool invocation and reproducible evaluation, we cache real tool-call results and build a sandbox environment that integrates ten travel-related tools. Agents can combine these tools to solve most practical travel planning problems, and our systematic verification demonstrates the stability of the proposed benchmark. We further evaluate multiple LLMs on TravelBench and conduct an in-depth analysis of their behaviors and performance. TravelBench provides a practical and reproducible evaluation benchmark to advance research on LLM agents for travel planning.\footnote{Our code and data will be available after internal review.

Yulan Hu, Sheng Ouyang, Jinman Zhao et al.

The Process Reward Model (PRM) plays a crucial role in mathematical reasoning tasks, requiring high-quality supervised process data. However, we observe that reasoning steps generated by Large Language Models (LLMs) often fail to exhibit strictly incremental information, leading to redundancy that can hinder effective reasoning. To address this issue, we propose CFPRM, a simple yet effective coarse-to-fine strategy. Instead of focusing on the detection of redundant steps, our approach first establishes a coarse-grained window to merge adjacent reasoning steps into unified, holistic steps. The window size is then progressively reduced to extract fine-grained reasoning steps, enabling data collection at multiple granularities for training. By leveraging this hierarchical refinement process, CFPRM mitigates redundancy while preserving essential fine-grained knowledge. Extensive experiments on two reasoning datasets across three loss criteria validate the CFPRM's effectiveness and versatility.

Jiaxuan Wang, Yulan Hu, Wenjin Yang et al.

In classical Reinforcement Learning from Human Feedback (RLHF), Reward Models (RMs) serve as the fundamental signal provider for model alignment. As Large Language Models evolve into agentic systems capable of autonomous tool invocation and complex reasoning, the paradigm of reward modeling faces unprecedented challenges--most notably, the lack of benchmarks specifically designed to assess RM capabilities within tool-integrated environments. To address this gap, we present Plan-RewardBench, a trajectory-level preference benchmark designed to evaluate how well judges distinguish preferred versus distractor agent trajectories in complex tool-using scenarios. Plan-RewardBench covers four representative task families -- (i) Safety Refusal, (ii) Tool-Irrelevance / Unavailability, (iii) Complex Planning, and (iv) Robust Error Recovery -- comprising validated positive trajectories and confusable hard negatives constructed via multi-model natural rollouts, rule-based perturbations, and minimal-edit LLM perturbations. We benchmark representative RMs (generative, discriminative, and LLM-as-Judge) under a unified pairwise protocol, reporting accuracy trends across varying trajectory lengths and task categories. Furthermore, we provide diagnostic analyses of prevalent failure modes. Our results reveal that all three evaluator families face substantial challenges, with performance degrading sharply on long-horizon trajectories, underscoring the necessity for specialized training in agentic, trajectory-level reward modeling. Ultimately, Plan-RewardBench aims to serve as both a practical evaluation suite and a reusable blueprint for constructing agentic planning preference data.

Yulan Hu, Sheng Ouyang, Zhirui Yang et al.

Graphs are ubiquitous in real-world scenarios and encompass a diverse range of tasks, from node-, edge-, and graph-level tasks to transfer learning. However, designing specific tasks for each type of graph data is often costly and lacks generalizability. Recent endeavors under the "Pre-training + Fine-tuning" or "Pre-training + Prompt" paradigms aim to design a unified framework capable of generalizing across multiple graph tasks. Among these, graph autoencoders (GAEs), generative self-supervised models, have demonstrated their potential in effectively addressing various graph tasks. Nevertheless, these methods typically employ multi-stage training and require adaptive designs, which on one hand make it difficult to be seamlessly applied to diverse graph tasks and on the other hand overlook the negative impact caused by discrepancies in task objectives between the different stages. To address these challenges, we propose GA^2E, a unified adversarially masked autoencoder capable of addressing the above challenges seamlessly. Specifically, GA^2E proposes to use the subgraph as the meta-structure, which remains consistent across all graph tasks (ranging from node-, edge-, and graph-level to transfer learning) and all stages (both during training and inference). Further, GA^2E operates in a \textbf{"Generate then Discriminate"} manner. It leverages the masked GAE to reconstruct the input subgraph whilst treating it as a generator to compel the reconstructed graphs resemble the input subgraph. Furthermore, GA^2E introduces an auxiliary discriminator to discern the authenticity between the reconstructed (generated) subgraph and the input subgraph, thus ensuring the robustness of the graph representation through adversarial training mechanisms. We validate GA^2E's capabilities through extensive experiments on 21 datasets across four types of graph tasks.

Sheng Ouyang, Yulan Hu, Ge Chen et al.

Rewards serve as proxies for human preferences and play a crucial role in Reinforcement Learning from Human Feedback (RLHF). However, if these rewards are inherently imperfect, exhibiting various biases, they can adversely affect the alignment of large language models (LLMs). In this paper, we collectively define the various biases present in rewards as the problem of reward unfairness. We propose a bias-agnostic method to address the issue of reward fairness from a resource allocation perspective, without specifically designing for each type of bias, yet effectively mitigating them. Specifically, we model preference learning as a resource allocation problem, treating rewards as resources to be allocated while considering the trade-off between utility and fairness in their distribution. We propose two methods, Fairness Regularization and Fairness Coefficient, to achieve fairness in rewards. We apply our methods in both verification and reinforcement learning scenarios to obtain a fairness reward model and a policy model, respectively. Experiments conducted in these scenarios demonstrate that our approach aligns LLMs with human preferences in a more fair manner.

Hao Yi, Qingyang Li, Yulan Hu et al.

High-quality video-text preference data is crucial for Multimodal Large Language Models (MLLMs) alignment. However, existing preference data is very scarce. Obtaining VQA preference data for preference training is costly, and manually annotating responses is highly unreliable, which could result in low-quality pairs. Meanwhile, AI-generated responses controlled by temperature adjustment lack diversity. To address these issues, we propose a high-quality VQA preference dataset, called \textit{\textbf{M}ultiple \textbf{M}ultimodal \textbf{A}rtificial \textbf{I}ntelligence \textbf{P}reference Datasets in \textbf{V}QA} (\textbf{MMAIP-V}), which is constructed by sampling from the response distribution set and using an external scoring function for response evaluation. Furthermore, to fully leverage the preference knowledge in MMAIP-V and ensure sufficient optimization, we propose \textit{\textbf{Iter}ative \textbf{W}eak-to-\textbf{S}trong \textbf{R}einforcement \textbf{L}earning from \textbf{AI} \textbf{F}eedback for video MLLMs} (\textbf{Iter-W2S-RLAIF}), a framework that gradually enhances MLLMs' alignment capabilities by iteratively updating the reference model and performing parameter extrapolation. Finally, we propose an unbiased and information-complete evaluation scheme in VQA evaluation. Experiments demonstrate that MMAIP-V is beneficial for MLLMs in preference learning and Iter-W2S-RLAIF fully exploits the alignment information in MMAIP-V. We believe that the proposed automatic VQA preference data generation pipeline based on AI feedback can greatly promote future work in the MLLMs alignment. \textbf{Code and dataset are available} \href{https://anonymous.4open.science/r/MMAIP-V_Iter-W2S-RLAIF-702F}{MMAIP-V\_Iter-W2S-RLAIF-702F}.

Ge Chen, Yulan Hu, Sheng Ouyang et al.

Graph autoencoders (GAEs), as a kind of generative self-supervised learning approach, have shown great potential in recent years. GAEs typically rely on distance-based criteria, such as mean-square-error (MSE), to reconstruct the input graph. However, relying solely on a single reconstruction criterion may lead to a loss of distinctiveness in the reconstructed graph, causing nodes to collapse into similar representations and resulting in sub-optimal performance. To address this issue, we have developed a simple yet effective strategy to preserve the necessary distinctness in the reconstructed graph. Inspired by the knowledge distillation technique, we found that the dual encoder-decoder architecture of GAEs can be viewed as a teacher-student relationship. Therefore, we propose transferring the knowledge of distinctness from the raw graph to the reconstructed graph, achieved through a simple KL constraint. Specifically, we compute pairwise node similarity scores in the raw graph and reconstructed graph. During the training process, the KL constraint is optimized alongside the reconstruction criterion. We conducted extensive experiments across three types of graph tasks, demonstrating the effectiveness and generality of our strategy. This indicates that the proposed approach can be employed as a plug-and-play method to avoid vague reconstructions and enhance overall performance.

Yulan Hu, Qingyang Li, Sheng Ouyang et al.

Reinforcement Learning from Human Feedback (RLHF) facilitates the alignment of large language models (LLMs) with human preferences, thereby enhancing the quality of responses generated. A critical component of RLHF is the reward model, which is trained on preference data and outputs a scalar reward during the inference stage. However, the collection of preference data still lacks thorough investigation. Recent studies indicate that preference data is collected either by AI or humans, where chosen and rejected instances are identified among pairwise responses. We question whether this process effectively filters out noise and ensures sufficient diversity in collected data. To address these concerns, for the first time, we propose a comprehensive framework for preference data collection, decomposing the process into four incremental steps: Prompt Generation, Response Generation, Response Filtering, and Human Labeling. This structured approach ensures the collection of high-quality preferences while reducing reliance on human labor. We conducted comprehensive experiments based on the data collected at different stages, demonstrating the effectiveness of the proposed data collection method.

Yulan Hu, Yong Liu

Benefiting from the strong ability of the pre-trained model, the research on Chinese Word Segmentation (CWS) has made great progress in recent years. However, due to massive computation, large and complex models are incapable of empowering their ability for industrial use. On the other hand, for low-resource scenarios, the prevalent decode method, such as Conditional Random Field (CRF), fails to exploit the full information of the training data. This work proposes a fast and accurate CWS framework that incorporates a light-weighted model and an upgraded decode method (PCRF) towards industrially low-resource CWS scenarios. First, we distill a Transformer-based student model as an encoder, which not only accelerates the inference speed but also combines open knowledge and domain-specific knowledge. Second, the perplexity score to evaluate the language model is fused into the CRF module to better identify the word boundaries. Experiments show that our work obtains relatively high performance on multiple datasets with as low as 14\% of time consumption compared with the original BERT-based model. Moreover, under the low-resource setting, we get superior results in comparison with the traditional decoding methods.