arXiv subjects

Junwei Bao

Publications and source records attributed to Junwei Bao.

At least 19 recordsLinked to original sources

arXiv2026

PGPO: Potential-Guided Policy Optimization for Multi-Turn Agentic Tasks

Group-based reinforcement learning (RL) has become an effective paradigm for LLM post-training, but in multi-turn agentic tasks with sparse terminal rewards, it often provides coarse credit for intermediate actions. To obtain more fine-grained credit assignment, recent work such as GiGPO introduces step-level advantages for intermediate actions. However, these step-level signals still rely on the final outcome of each individual trajectory. As a result, actions within failed trajectories can remain poorly differentiated, so effective actions can receive the same unfavorable credit as erroneous ones. In this work, we propose Potential-Guided Policy Optimization (PGPO) for multi-turn agentic tasks. PGPO estimates empirical state potentials from anchor-state-group return statistics within each rollout group. It then derives action advantages from potential differences between adjacent states, enabling cross-trajectory credit propagation. This provides finer-grained step-level credit assignment, especially within failed trajectories. Experiments on ALFWorld and WebShop show strong overall performance relative to recent group-based RL methods. Further analysis provides evidence that PGPO yields more informative failure-side credit signals with negligible training overhead.

cs.AI↗

arXiv2026

FABSVer: Faster Training and Better Self-Verification for LLM Mathematical Reasoning

While large language models have made significant progress in mathematical reasoning, they remain unreliable at judging the correctness of their own solutions. Existing approaches that equip models with self-verification typically treat solution generation and verification as two separate tasks, leading to substantially increased training time. In this paper, we propose FABSVer, which fuses these two tasks into a single generation pass, dramatically reducing training overhead while jointly optimizing both capabilities. We further identify a convergence bottleneck both theoretically and empirically: as training progresses, the reward reaches a plateau because the policy is constrained by a fixed reference model. To overcome this, we introduce Dynamic Reference Model Update (DRMU), which raises the reward ceiling and enables sustained reward growth. Extensive experiments on math benchmarks demonstrate that FABSVer achieves superior self-verification and reasoning performance across three model scales, while requiring only 51%--71% of the training time of existing methods. Analysis further reveals distinct learning phases in how models acquire self-verification, and that the gap between verify and answer rewards shrinks noticeably as model size increases.

cs.CL↗

arXiv2026

Elo-Evolve: A Co-evolutionary Framework for Language Model Alignment

Current alignment methods for Large Language Models (LLMs) rely on compressing vast amounts of human preference data into static, absolute reward functions, leading to data scarcity, noise sensitivity, and training instability. We introduce Elo-Evolve, a co-evolutionary framework that redefines alignment as dynamic multi-agent competition within an adaptive opponent pool. Our approach makes two key innovations: (1) eliminating Bradley-Terry model dependencies by learning directly from binary win/loss outcomes in pairwise competitions, and (2) implementing Elo-orchestrated opponent selection that provides automatic curriculum learning through temperature-controlled sampling. We ground our approach in PAC learning theory, demonstrating that pairwise comparison achieves superior sample complexity and empirically validate a 4.5x noise reduction compared to absolute scoring approaches. Experimentally, we train a Qwen2.5-7B model using our framework with opponents including Qwen2.5-14B, Qwen2.5-32B, and Qwen3-8B models. Results demonstrate a clear performance hierarchy: point-based methods < static pairwise training < Elo-Evolve across Alpaca Eval 2.0 and MT-Bench, validating the progressive benefits of pairwise comparison and dynamic opponent selection for LLM alignment.

cs.CL↗

arXiv2026

Quality-constrained Entropy Maximization Policy Optimization for LLM Diversity

In many large language model (LLM) alignment applications, users expect not only high-quality outputs but also substantial diversity. However, existing methods often face a fundamental trade-off between these objectives: approaches that improve output quality tend to reduce diversity, while methods that increase diversity often do so at the expense of quality. In this work, we propose Quality-constrained Entropy Maximization Policy Optimization (QEMPO), a novel framework that enhances the diversity of LLM outputs while explicitly preserving output quality. QEMPO is grounded in a strong theoretical foundation: we derive a closed-form analytical solution that provably maximizes entropy-a principled measure of diversity-subject to a quality constraint, with guarantees on optimality under the defined objective. Leveraging this solution, QEMPO naturally supports both online and offline training settings. Empirical results demonstrate that QEMPO consistently improves output diversity without sacrificing quality, and in many cases yields gains in both dimensions compared to existing baselines, aligning with our theoretical guarantees.

cs.CL↗

arXiv2025

RSPO: Risk-Seeking Policy Optimization for Pass@k and Max@k Metrics in Large Language Models

Current large language model post-training optimizes a risk-neutral objective that maximizes expected reward, yet evaluation relies heavily on risk-seeking metrics like Pass@k (at least one success in k trials) and Max@k (maximum reward across k responses). This mismatch in risk preferences can inevitably lead to suboptimal performance. To bridge this gap, we propose Risk-Seeking Policy Optimization (RSPO), a novel method that directly targets Pass@k and Max@k during training. A key challenge in optimizing these metrics is the "hitchhiking" problem: low-reward responses are inadvertently reinforced if they co-occur with a high-reward response within a sample of k generations, resulting in inefficient optimization. RSPO addresses this problem by leveraging the closed-form probability that a given response is the maximum among k samplings. Despite the complexity of nested gradients over multiple responses, RSPO produces efficient, unbiased gradient estimators for both metrics. We validate our approach with both rigorous theoretical analysis and comprehensive experimental results.

cs.LG↗

arXiv2025

GVPO: Group Variance Policy Optimization for Large Language Model Post-Training

Post-training plays a crucial role in refining and aligning large language models to meet specific tasks and human preferences. While recent advancements in post-training techniques, such as Group Relative Policy Optimization (GRPO), leverage increased sampling with relative reward scoring to achieve superior performance, these methods often suffer from training instability that limits their practical adoption. As a next step, we present Group Variance Policy Optimization (GVPO). GVPO incorporates the analytical solution to KL-constrained reward maximization directly into its gradient weights, ensuring alignment with the optimal policy. The method provides intuitive physical interpretations: its gradient mirrors the mean squared error between the central distance of implicit rewards and that of actual rewards. GVPO offers two key advantages: (1) it guarantees a unique optimal solution, exactly the KL-constrained reward maximization objective, (2) it supports flexible sampling distributions that avoids on-policy and importance sampling limitations. By unifying theoretical guarantees with practical adaptability, GVPO establishes a new paradigm for reliable and versatile LLM post-training.

cs.AI↗

arXiv2024

Energy-Based Preference Model Offers Better Offline Alignment than the Bradley-Terry Preference Model

Since the debut of DPO, it has been shown that aligning a target LLM with human preferences via the KL-constrained RLHF loss is mathematically equivalent to a special kind of reward modeling task. Concretely, the task requires: 1) using the target LLM to parameterize the reward model, and 2) tuning the reward model so that it has a 1:1 linear relationship with the true reward. However, we identify a significant issue: the DPO loss might have multiple minimizers, of which only one satisfies the required linearity condition. The problem arises from a well-known issue of the underlying Bradley-Terry preference model: it does not always have a unique maximum likelihood estimator (MLE). Consequently,the minimizer of the RLHF loss might be unattainable because it is merely one among many minimizers of the DPO loss. As a better alternative, we propose an energy-based model (EBM) that always has a unique MLE, inherently satisfying the linearity requirement. To approximate the MLE in practice, we propose a contrastive loss named Energy Preference Alignment (EPA), wherein each positive sample is contrasted against one or more strong negatives as well as many free weak negatives. Theoretical properties of our EBM enable the approximation error of EPA to almost surely vanish when a sufficient number of negatives are used. Empirically, we demonstrate that EPA consistently delivers better performance on open benchmarks compared to DPO, thereby showing the superiority of our EBM.

cs.LG↗

arXiv2024

Preference-Oriented Supervised Fine-Tuning: Favoring Target Model Over Aligned Large Language Models

Alignment, endowing a pre-trained Large language model (LLM) with the ability to follow instructions, is crucial for its real-world applications. Conventional supervised fine-tuning (SFT) methods formalize it as causal language modeling typically with a cross-entropy objective, requiring a large amount of high-quality instruction-response pairs. However, the quality of widely used SFT datasets can not be guaranteed due to the high cost and intensive labor for the creation and maintenance in practice. To overcome the limitations associated with the quality of SFT datasets, we introduce a novel \textbf{p}reference-\textbf{o}riented supervised \textbf{f}ine-\textbf{t}uning approach, namely PoFT. The intuition is to boost SFT by imposing a particular preference: \textit{favoring the target model over aligned LLMs on the same SFT data.} This preference encourages the target model to predict a higher likelihood than that predicted by the aligned LLMs, incorporating assessment information on data quality (i.e., predicted likelihood by the aligned LLMs) into the training process. Extensive experiments are conducted, and the results validate the effectiveness of the proposed method. PoFT achieves stable and consistent improvements over the SFT baselines across different training datasets and base models. Moreover, we prove that PoFT can be integrated with existing SFT data filtering methods to achieve better performance, and further improved by following preference optimization procedures, such as DPO.

cs.CL↗

arXiv2024

BoRA: Bi-dimensional Weight-Decomposed Low-Rank Adaptation

In recent years, Parameter-Efficient Fine-Tuning (PEFT) methods like Low-Rank Adaptation (LoRA) have significantly enhanced the adaptability of large-scale pre-trained models. Weight-Decomposed Low-Rank Adaptation (DoRA) improves upon LoRA by separating the magnitude and direction components of the weight matrix, leading to superior performance. However, DoRA's improvements are limited to the vertical dimension, resulting in an asymmetrical pattern between horizontal and vertical dimensions. This paper introduces BoRA, an innovative extension of LoRA and DoRA, characterized by symmetrical properties across horizontal and vertical dimensions. Our approach optimizes the weight matrix symmetrically by adjusting both column-wise and row-wise magnitudes. Extensive experiments demonstrate that BoRA surpasses state-of-the-art PEFT methods, including LoRA and DoRA, achieving superior results across various benchmarks.

cs.CL↗

arXiv2024

Multi-Turn Interactions for Text-to-SQL with Large Language Models

This study explores text-to-SQL parsing by leveraging the powerful reasoning capabilities of large language models (LLMs). Despite recent advancements, existing LLM-based methods are still inefficient and struggle to handle cases with wide tables effectively. Furthermore, current interaction-based approaches either lack a step-by-step, interpretable SQL generation process or fail to provide a universally applicable interaction design. To address these challenges, we introduce Interactive-T2S, a framework that generates SQL queries through direct interactions with databases. This framework includes four general tools that facilitate proactive and efficient information retrieval by the LLM. Additionally, we have developed detailed exemplars to demonstrate the step-wise reasoning processes within our framework. Our approach achieves advanced performance on the Spider and BIRD datasets as well as their variants. Notably, we obtain state-of-the-art results on the BIRD leaderboard under the setting without oracle knowledge, demonstrating the effectiveness of our method.

cs.CL↗

arXiv2024

Interactive-KBQA: Multi-Turn Interactions for Knowledge Base Question Answering with Large Language Models

This study explores the realm of knowledge base question answering (KBQA). KBQA is considered a challenging task, particularly in parsing intricate questions into executable logical forms. Traditional semantic parsing (SP)-based methods require extensive data annotations, which result in significant costs. Recently, the advent of few-shot in-context learning, powered by large language models (LLMs), has showcased promising capabilities. However, fully leveraging LLMs to parse questions into logical forms in low-resource scenarios poses a substantial challenge. To tackle these hurdles, we introduce Interactive-KBQA, a framework designed to generate logical forms through direct interaction with knowledge bases (KBs). Within this framework, we have developed three generic APIs for KB interaction. For each category of complex question, we devised exemplars to guide LLMs through the reasoning processes. Our method achieves competitive results on the WebQuestionsSP, ComplexWebQuestions, KQA Pro, and MetaQA datasets with a minimal number of examples (shots). Importantly, our approach supports manual intervention, allowing for the iterative refinement of LLM outputs. By annotating a dataset with step-wise reasoning processes, we showcase our model's adaptability and highlight its potential for contributing significant enhancements to the field.

cs.CL↗

arXiv2023

HopPG: Self-Iterative Program Generation for Multi-Hop Question Answering over Heterogeneous Knowledge

The semantic parsing-based method is an important research branch for knowledge-based question answering. It usually generates executable programs lean upon the question and then conduct them to reason answers over a knowledge base. Benefit from this inherent mechanism, it has advantages in the performance and the interpretability. However, traditional semantic parsing methods usually generate a complete program before executing it, which struggles with multi-hop question answering over heterogeneous knowledge. On one hand, generating a complete multi-hop program relies on multiple heterogeneous supporting facts, and it is difficult for generators to understand these facts simultaneously. On the other hand, this way ignores the semantic information of the intermediate answers at each hop, which is beneficial for subsequent generation. To alleviate these challenges, we propose a self-iterative framework for multi-hop program generation (HopPG) over heterogeneous knowledge, which leverages the previous execution results to retrieve supporting facts and generate subsequent programs hop by hop. We evaluate our model on MMQA-T^2, and the experimental results show that HopPG outperforms existing semantic-parsing-based baselines, especially on the multi-hop questions.

cs.CL↗

arXiv2023

AUGUST: an Automatic Generation Understudy for Synthesizing Conversational Recommendation Datasets

High-quality data is essential for conversational recommendation systems and serves as the cornerstone of the network architecture development and training strategy design. Existing works contribute heavy human efforts to manually labeling or designing and extending recommender dialogue templates. However, they suffer from (i) the limited number of human annotators results in that datasets can hardly capture rich and large-scale cases in the real world, (ii) the limited experience and knowledge of annotators account for the uninformative corpus and inappropriate recommendations. In this paper, we propose a novel automatic dataset synthesis approach that can generate both large-scale and high-quality recommendation dialogues through a data2text generation process, where unstructured recommendation conversations are generated from structured graphs based on user-item information from the real world. In doing so, we comprehensively exploit: (i) rich personalized user profiles from traditional recommendation datasets, (ii) rich external knowledge from knowledge graphs, and (iii) the conversation ability contained in human-to-human conversational recommendation datasets. Extensive experiments validate the benefit brought by the automatically synthesized data under low-resource scenarios and demonstrate the promising potential to facilitate the development of a more effective conversational recommendation system.

cs.CL↗

arXiv2022

SegCLIP: Patch Aggregation with Learnable Centers for Open-Vocabulary Semantic Segmentation

Recently, the contrastive language-image pre-training, e.g., CLIP, has demonstrated promising results on various downstream tasks. The pre-trained model can capture enriched visual concepts for images by learning from a large scale of text-image data. However, transferring the learned visual knowledge to open-vocabulary semantic segmentation is still under-explored. In this paper, we propose a CLIP-based model named SegCLIP for the topic of open-vocabulary segmentation in an annotation-free manner. The SegCLIP achieves segmentation based on ViT and the main idea is to gather patches with learnable centers to semantic regions through training on text-image pairs. The gathering operation can dynamically capture the semantic groups, which can be used to generate the final segmentation results. We further propose a reconstruction loss on masked patches and a superpixel-based KL loss with pseudo-labels to enhance the visual representation. Experimental results show that our model achieves comparable or superior segmentation accuracy on the PASCAL VOC 2012 (+0.3% mIoU), PASCAL Context (+2.3% mIoU), and COCO (+2.2% mIoU) compared with baselines. We release the code at https://github.com/ArrowLuo/SegCLIP.

cs.CV↗

arXiv2022

MoNET: Tackle State Momentum via Noise-Enhanced Training for Dialogue State Tracking

Dialogue state tracking (DST) aims to convert the dialogue history into dialogue states which consist of slot-value pairs. As condensed structural information memorizing all history information, the dialogue state in the last turn is typically adopted as the input for predicting the current state by DST models. However, these models tend to keep the predicted slot values unchanged, which is defined as state momentum in this paper. Specifically, the models struggle to update slot values that need to be changed and correct wrongly predicted slot values in the last turn. To this end, we propose MoNET to tackle state momentum via noise-enhanced training. First, the previous state of each turn in the training data is noised via replacing some of its slot values. Then, the noised previous state is used as the input to learn to predict the current state, improving the model's ability to update and correct slot values. Furthermore, a contrastive context matching framework is designed to narrow the representation distance between a state and its corresponding noised variant, which reduces the impact of noised state and makes the model better understand the dialogue history. Experimental results on MultiWOZ datasets show that MoNET outperforms previous DST methods. Ablations and analysis verify the effectiveness of MoNET in alleviating state momentum and improving anti-noise ability.

cs.CL↗

arXiv2022

P$^3$LM: Probabilistically Permuted Prophet Language Modeling for Generative Pre-Training

Conventional autoregressive left-to-right (L2R) sequence generation faces two issues during decoding: limited to unidirectional target sequence modeling, and constrained on strong local dependencies. To address the aforementioned problem, we propose P$^3$LM, a probabilistically permuted prophet language model, which strengthens the modeling of bidirectional information and long token dependencies for sequence generation. Specifically, P$^3$LM learns to generate tokens in permuted order upon an order-aware transformer decoder, as well as to generate the corresponding future $N$ tokens with a multi-stream attention mechanism. Extensive experiments are conducted on the GLGE benchmark, which includes four datasets for summarization, two for question generation, one for conversational question answering, and one for dialog response generation, where P$^3$LM achieves state-of-the-art results compared with strong publicly available generative pre-training methods.

cs.CL↗

arXiv2022

MuGER$^2$: Multi-Granularity Evidence Retrieval and Reasoning for Hybrid Question Answering

Hybrid question answering (HQA) aims to answer questions over heterogeneous data, including tables and passages linked to table cells. The heterogeneous data can provide different granularity evidence to HQA models, e.t., column, row, cell, and link. Conventional HQA models usually retrieve coarse- or fine-grained evidence to reason the answer. Through comparison, we find that coarse-grained evidence is easier to retrieve but contributes less to the reasoner, while fine-grained evidence is the opposite. To preserve the advantage and eliminate the disadvantage of different granularity evidence, we propose MuGER$^2$, a Multi-Granularity Evidence Retrieval and Reasoning approach. In evidence retrieval, a unified retriever is designed to learn the multi-granularity evidence from the heterogeneous data. In answer reasoning, an evidence selector is proposed to navigate the fine-grained evidence for the answer reader based on the learned multi-granularity evidence. Experiment results on the HybridQA dataset show that MuGER$^2$ significantly boosts the HQA performance. Further ablation analysis verifies the effectiveness of both the retrieval and reasoning designs.

cs.CL↗

arXiv2022

Mars: Modeling Context & State Representations with Contrastive Learning for End-to-End Task-Oriented Dialog

Traditional end-to-end task-oriented dialog systems first convert dialog context into belief state and action state before generating the system response. The system response performance is significantly affected by the quality of the belief state and action state. We first explore what dialog context representation is beneficial to improving the quality of the belief state and action state, which further enhances the generated response quality. To tackle our exploration, we propose Mars, an end-to-end task-oriented dialog system with two contrastive learning strategies to model the relationship between dialog context and belief/action state representations. Empirical results show dialog context representations, which are more different from semantic state representations, are more conducive to multi-turn task-oriented dialog. Moreover, our proposed Mars achieves state-of-the-art performance on the MultiWOZ 2.0, CamRest676, and CrossWOZ.

cs.CL↗