Xiaoying Zhang

Team Seedance, De Chen, Liyang Chen et al. · gatech

Seedance 2.0 is a new native multi-modal audio-video generation model, officially released in China in early February 2026. Compared with its predecessors, Seedance 1.0 and 1.5 Pro, Seedance 2.0 adopts a unified, highly efficient, and large-scale architecture for multi-modal audio-video joint generation. This allows it to support four input modalities: text, image, audio, and video, by integrating one of the most comprehensive suites of multi-modal content reference and editing capabilities available in the industry to date. It delivers substantial, well-rounded improvements across all key sub-dimensions of video and audio generation. In both expert evaluations and public user tests, the model has demonstrated performance on par with the leading levels in the field. Seedance 2.0 supports direct generation of audio-video content with durations ranging from 4 to 15 seconds, with native output resolutions of 480p and 720p. For multi-modal inputs as reference, its current open platform supports up to 3 video clips, 9 images, and 3 audio clips. In addition, we provide Seedance 2.0 Fast version, an accelerated variant of Seedance 2.0 designed to boost generation speed for low-latency scenarios. Seedance 2.0 has delivered significant improvements to its foundational generation capabilities and multi-modal generation performance, bringing an enhanced creative experience for end users.

Ziyue Wang, Aomufei Yuan, Yongfu Zhu et al.

Reinforcement Learning from Verifiable Rewards (RLVR) has become the dominant approach for improving mathematical reasoning in large language models, yet current methods reduce each correct rollout to a single reward bit, ignoring the geometric structure shared among their hidden states. Investigating this structure, we find that at the anchor token (the position immediately before the answer marker), correct rollouts converge naturally because they must produce the same answer (cosine similarity ~0.84), yet each retains residual variance from its unique reasoning path. Encouraging full alignment at this point pushes the model to extract a unified "correct decision" representation, reducing sensitivity to which reasoning path was taken. Based on this observation, we propose Hidden-Align, an auxiliary loss function that aligns the last-layer hidden states of correct rollouts at the anchor token during RL training, with zero overhead in both training and inference. On eight mathematical reasoning benchmarks, Hidden-Align improves average pass@1 over the DAPO baseline by 3.8, 6.2, and 5.4 percentage points on Qwen3-1.7B, 4B, and 14B respectively, with consistent pass@k gains across all three scales, supported by ablations on loss type, anchor position, layer depth, and loss weight.

Yang Liu, Yuanshun Yao, Jean-Francois Ton et al.

Ensuring alignment, which refers to making models behave in accordance with human intentions [1,2], has become a critical task before deploying large language models (LLMs) in real-world applications. For instance, OpenAI devoted six months to iteratively aligning GPT-4 before its release [3]. However, a major challenge faced by practitioners is the lack of clear guidance on evaluating whether LLM outputs align with social norms, values, and regulations. This obstacle hinders systematic iteration and deployment of LLMs. To address this issue, this paper presents a comprehensive survey of key dimensions that are crucial to consider when assessing LLM trustworthiness. The survey covers seven major categories of LLM trustworthiness: reliability, safety, fairness, resistance to misuse, explainability and reasoning, adherence to social norms, and robustness. Each major category is further divided into several sub-categories, resulting in a total of 29 sub-categories. Additionally, a subset of 8 sub-categories is selected for further investigation, where corresponding measurement studies are designed and conducted on several widely-used LLMs. The measurement results indicate that, in general, more aligned models tend to perform better in terms of overall trustworthiness. However, the effectiveness of alignment varies across the different trustworthiness categories considered. This highlights the importance of conducting more fine-grained analyses, testing, and making continuous improvements on LLM alignment. By shedding light on these key dimensions of LLM trustworthiness, this paper aims to provide valuable insights and guidance to practitioners in the field. Understanding and addressing these concerns will be crucial in achieving reliable and ethically sound deployment of LLMs in various applications.

Jingyan Zhou, Minda Hu, Junan Li et al.

Making moral judgments is an essential step toward developing ethical AI systems. Prevalent approaches are mostly implemented in a bottom-up manner, which uses a large set of annotated data to train models based on crowd-sourced opinions about morality. These approaches have been criticized for overgeneralizing the moral stances of a limited group of annotators and lacking explainability. This work proposes a flexible top-down framework to steer (Large) Language Models (LMs) to perform moral reasoning with well-established moral theories from interdisciplinary research. The theory-guided top-down framework can incorporate various moral theories. Our experiments demonstrate the effectiveness of the proposed framework on datasets derived from moral theories. Furthermore, we show the alignment between different moral theories and existing morality datasets. Our analysis exhibits the potential and flaws in existing resources (models and datasets) in developing explainable moral judgment-making systems.

Xiaoying Zhang, Junpu Chen, Hongning Wang et al.

Off-policy learning, referring to the procedure of policy optimization with access only to logged feedback data, has shown importance in various real-world applications, such as search engines, recommender systems, and etc. While the ground-truth logging policy, which generates the logged data, is usually unknown, previous work simply takes its estimated value in off-policy learning, ignoring both high bias and high variance resulted from such an estimator, especially on samples with small and inaccurately estimated logging probabilities. In this work, we explicitly model the uncertainty in the estimated logging policy and propose a Uncertainty-aware Inverse Propensity Score estimator (UIPS) for improved off-policy learning, with a theoretical convergence guarantee. Experiment results on synthetic and three real-world recommendation datasets demonstrate the advantageous sample efficiency of the proposed UIPS estimator against an extensive list of state-of-the-art baselines.

Qing Zhang, Xiaoying Zhang, Yang Liu et al.

Recommendation systems aim to predict users' feedback on items not exposed to them. Confounding bias arises due to the presence of unmeasured variables (e.g., the socio-economic status of a user) that can affect both a user's exposure and feedback. Existing methods either (1) make untenable assumptions about these unmeasured variables or (2) directly infer latent confounders from users' exposure. However, they cannot guarantee the identification of counterfactual feedback, which can lead to biased predictions. In this work, we propose a novel method, i.e., identifiable deconfounder (iDCF), which leverages a set of proxy variables (e.g., observed user features) to resolve the aforementioned non-identification issue. The proposed iDCF is a general deconfounded recommendation framework that applies proximal causal inference to infer the unmeasured confounders and identify the counterfactual feedback with theoretical guarantees. Extensive experiments on various real-world and synthetic datasets verify the proposed method's effectiveness and robustness.

Xiaoying Zhang, Hongning Wang, Hang Li

Accuracy and diversity have long been considered to be two conflicting goals for recommendations. We point out, however, that as the diversity is typically measured by certain pre-selected item attributes, e.g., category as the most popularly employed one, improved diversity can be achieved without sacrificing recommendation accuracy, as long as the diversification respects the user's preference about the pre-selected attributes. This calls for a fine-grained understanding of a user's preferences over items, where one needs to recognize the user's choice is driven by the quality of the item itself, or the pre-selected attributes of the item. In this work, we focus on diversity defined on item categories. We propose a general diversification framework agnostic to the choice of recommendation algorithms. Our solution disentangles the learnt user representation in the recommendation module into category-independent and category-dependent components to differentiate a user's preference over items from two orthogonal perspectives. Experimental results on three benchmark datasets and online A/B test demonstrate the effectiveness of our solution in improving both recommendation accuracy and diversity. In-depth analysis suggests that the improvement is due to our improved modeling of users' categorical preferences and refined ranking within item categories.

Kaituo Feng, Kaixiong Gong, Bohao Li et al.

Inspired by DeepSeek-R1's success in eliciting reasoning abilities through rule-based reinforcement learning (RL), we introduce Video-R1 as the first attempt to systematically explore the R1 paradigm for incentivizing video reasoning within multimodal large language models (MLLMs). However, directly applying RL training with the GRPO algorithm to video reasoning presents two primary challenges: (i) a lack of temporal modeling for video reasoning, and (ii) the scarcity of high-quality video-reasoning data. To address these issues, we first propose the T-GRPO algorithm, which encourages models to utilize temporal information in videos for reasoning. Additionally, instead of relying solely on video data, we incorporate high-quality image-reasoning data into the training process. We have constructed two datasets: Video-R1-CoT-165k for SFT cold start and Video-R1-260k for RL training, both comprising image and video data. Experimental results demonstrate that Video-R1 achieves significant improvements on video reasoning benchmarks such as VideoMMMU and VSI-Bench, as well as on general video benchmarks including MVBench and TempCompass, etc. Notably, Video-R1-7B attains a 37.1% accuracy on video spatial reasoning benchmark VSI-bench, surpassing the commercial proprietary model GPT-4o. All code, models, and data are released in: https://github.com/tulerfeng/Video-R1.

Xiaoying Zhang, Zichen Liu, Yipeng Zhang et al.

Standard reinforcement learning (RL) for large language model (LLM)-based agents typically optimizes extrinsic task-success rewards, prioritizing one-off task solving over continual adaptation. As a result, agents may converge to suboptimal policies due to limited exploration, and accumulated experience remains implicitly stored in model parameters, hindering efficient experiential learning. Inspired by humans' capacity for retrospective self-improvement, we introduce RetroAgent, an online RL framework that enables agents to master complex interactive environments not only by solving, but also by evolving under the joint guidance of extrinsic task-success rewards and retrospective dual intrinsic feedback. Concretely, RetroAgent features a hindsight self-reflection mechanism that produces: (1) intrinsic numerical feedback, which tracks incremental subtask completion relative to prior attempts to reward promising exploration; and (2) intrinsic language feedback, which distills reusable lessons into a memory buffer retrieved via our proposed Similarity & Utility-Aware Upper Confidence Bound (SimUtil-UCB) strategy, jointly balancing relevance, utility, and exploration. Extensive experiments across four challenging agentic tasks show that RetroAgent achieves state-of-the-art (SOTA) performance, substantially outperforming RL fine-tuning, memory-augmented RL, exploration-guided RL, and meta-RL methods -- e.g., exceeding Group Relative Policy Optimization (GRPO)-trained agents by +18.3% on ALFWorld, +15.4% on WebShop, +27.1% on Sokoban, and +8.9% on MineSweeper -- while maintaining strong test-time adaptation and out-of-distribution generalization.

Shuhua Yang, Hui Yuan, Xiaoying Zhang et al.

Conversational recommendation systems elicit user preferences by interacting with users to obtain their feedback on recommended commodities. Such systems utilize a multi-armed bandit framework to learn user preferences in an online manner and have received great success in recent years. However, existing conversational bandit methods have several limitations. First, they only enable users to provide explicit binary feedback on the recommended items or categories, leading to ambiguity in interpretation. In practice, users are usually faced with more than one choice. Relative feedback, known for its informativeness, has gained increasing popularity in recommendation system design. Moreover, current contextual bandit methods mainly work under linear reward assumptions, ignoring practical non-linear reward structures in generalized linear models. Therefore, in this paper, we introduce relative feedback-based conversations into conversational recommendation systems through the integration of dueling bandits in generalized linear models (GLM) and propose a novel conversational dueling bandit algorithm called ConDuel. Theoretical analyses of regret upper bounds and empirical validations on synthetic and real-world data underscore ConDuel's efficacy. We also demonstrate the potential to extend our algorithm to multinomial logit bandits with theoretical and experimental guarantees, which further proves the applicability of the proposed framework.

Wenpu Liu, Yuqi Xu, Weichu Xie et al.

Reinforcement Learning from Verifiable Rewards (RLVR) typically samples multiple responses per prompt and assigns binary rewards based on individual correctness, yet the collective structure of the group output, specifically the distribution of errors, is largely discarded. We identify this as a missed opportunity: empirical analysis reveals that error diversity within a group is a strong predictor of training success, with problems eliciting diverse wrong answers benefiting substantially more from RLVR than those producing homogeneous failures. Motivated by this observation, we propose Error Diversity Advantage Shaping (EDAS), a lightweight, algorithm-agnostic technique that modulates the advantage signal for incorrect rollouts based on intra-group error diversity. EDAS amplifies penalties for dominant, repeated errors and attenuates penalties for rare, exploratory ones, thereby encouraging the model to maintain diverse reasoning paths and discouraging error perseveration. Crucially, EDAS operates as a simple post-hoc adjustment that can be seamlessly integrated into any RLVR algorithm. We validate EDAS on top of several mainstream RLVR methods across a series of models and seven challenging math benchmarks, demonstrating consistent improvements. Notably, EDAS yields an average improvement of 6.29 points over DAPO on Qwen3-8B across seven benchmarks, confirming that exploiting the latent information in group rollouts is a broadly effective strategy for strengthening RLVR.

Weichu Xie, Haozhe Zhao, Wenpu Liu et al.

Reinforcement Learning with Verifiable Rewards (RLVR) is widely used to improve reasoning in large language models, but rewards only final-answer correctness with no supervision over intermediate steps. Rubric-based methods such as Rubrics as Rewards (RaR) introduce finer-grained supervision by scoring rollouts against structured criteria, yet the rubric scores are still aggregated into a single scalar applied to the entire response, causing three weaknesses: loss of multi-criterion structure, uniform supervision of correct and incorrect steps, and reward hacking through unbounded self-correction. On 1,000 problems, we find 18.2% of steps in correct-answer responses are wrong yet positively rewarded, while 49.9% of steps in incorrect-answer responses are correct yet penalized. We introduce Step-wise Rubrics as Rewards (SRaR), an RLVR framework that (i) uses an LLM judge to attribute each rubric item to a specific reasoning step, (ii) normalizes per-step rubric scores across rollouts so only steps whose quality varies produce a learning signal, and (iii) combines the per-step reward with the outcome reward through a decoupled advantage estimator that keeps the outcome baseline stable. We further build a 16K-problem rubric dataset by contrastively distilling rubric items from correct and flawed reasoning paths sampled from a strong model. Across six mathematical reasoning benchmarks, SRaR improves average accuracy over RaR by 3.57 points on Qwen3-8B and 2.75 points on Qwen3-32B, raises the Faithful Reasoning Rate on AIME 2025 from 34.5% to 46.7%, and reduces self-correction looping from 48.1% to 26.5%.

Genqiang Wu, Xiaoying Zhang, Yu Qi et al.

The exponential growth of data collection necessitates robust privacy protections that preserve data utility. We address information disclosure against adversaries with bounded prior knowledge, modeled by an entropy constraint $H(X) \geq b$. Within this information privacy framework -- which replaces differential privacy's independence assumption with a bounded-knowledge model -- we study three core problems: maximal per-record leakage, the primal leakage-distortion tradeoff (minimizing worst-case leakage under distortion $D$), and the dual distortion minimization (minimizing distortion under leakage constraint $L$). These problems resemble classical information-theoretic ones (channel capacity, rate-distortion) but are more complex due to high dimensionality and the entropy constraint. We develop efficient alternating optimization algorithms that exploit convexity-concavity duality, with theoretical guarantees including local convergence for the primal problem and convergence to a stationary point for the dual. Experiments on binary symmetric channels and modular sum queries validate the algorithms, showing improved privacy-utility tradeoffs over classical differential privacy mechanisms. This work provides a computational framework for auditing privacy risks and designing certified mechanisms under realistic adversary assumptions.

Tao Lin, Kun Jin, Andrew Estornell et al.

Recommender systems serve the dual purpose of presenting relevant content to users and helping content creators reach their target audience. The dual nature of these systems naturally influences both users and creators: users' preferences are affected by the items they are recommended, while creators may be incentivized to alter their content to attract more users. We define a model, called user-creator feature dynamics, to capture the dual influence of recommender systems. We prove that a recommender system with dual influence is guaranteed to polarize, causing diversity loss in the system. We then investigate, both theoretically and empirically, approaches for mitigating polarization and promoting diversity in recommender systems. Unexpectedly, we find that common diversity-promoting approaches do not work in the presence of dual influence, while relevancy-optimizing methods like top-$k$ truncation can prevent polarization and improve diversity of the system.

Jialong Wu, Xiaoying Zhang, Hongyi Yuan et al.

Humans construct internal world models and reason by manipulating the concepts within these models. Recent advances in AI, particularly chain-of-thought (CoT) reasoning, approximate such human cognitive abilities, where world models are believed to be embedded within large language models. Expert-level performance in formal and abstract domains such as mathematics and programming has been achieved in current systems by relying predominantly on verbal reasoning. However, they still lag far behind humans in domains like physical and spatial intelligence, which require richer representations and prior knowledge. The emergence of unified multimodal models (UMMs) capable of both verbal and visual generation has therefore sparked interest in more human-like reasoning grounded in complementary multimodal pathways, though their benefits remain unclear. From a world-model perspective, this paper presents the first principled study of when and how visual generation benefits reasoning. Our key position is the visual superiority hypothesis: for certain tasks--particularly those grounded in the physical world--visual generation more naturally serves as world models, whereas purely verbal world models encounter bottlenecks arising from representational limitations or insufficient prior knowledge. Theoretically, we formalize internal world modeling as a core component of CoT reasoning and analyze distinctions among different forms of world models. Empirically, we identify tasks that necessitate interleaved visual-verbal CoT reasoning, constructing a new evaluation suite, VisWorld-Eval. Controlled experiments on a state-of-the-art UMM show that interleaved CoT significantly outperforms purely verbal CoT on tasks that favor visual world modeling, but offers no clear advantage otherwise. Together, this work clarifies the potential of multimodal world modeling for more powerful, human-like multimodal AI.

Dong Guo, Faming Wu, Feida Zhu et al. · pku

We present Seed1.5-VL, a vision-language foundation model designed to advance general-purpose multimodal understanding and reasoning. Seed1.5-VL is composed with a 532M-parameter vision encoder and a Mixture-of-Experts (MoE) LLM of 20B active parameters. Despite its relatively compact architecture, it delivers strong performance across a wide spectrum of public VLM benchmarks and internal evaluation suites, achieving the state-of-the-art performance on 38 out of 60 public benchmarks. Moreover, in agent-centric tasks such as GUI control and gameplay, Seed1.5-VL outperforms leading multimodal systems, including OpenAI CUA and Claude 3.7. Beyond visual and video understanding, it also demonstrates strong reasoning abilities, making it particularly effective for multimodal reasoning challenges such as visual puzzles. We believe these capabilities will empower broader applications across diverse tasks. In this report, we mainly provide a comprehensive review of our experiences in building Seed1.5-VL across model design, data construction, and training at various stages, hoping that this report can inspire further research. Seed1.5-VL is now accessible at https://www.volcengine.com/ (Volcano Engine Model ID: doubao-1-5-thinking-vision-pro-250428)

Xiaoying Zhang, Baolin Peng, Ye Tian et al.

Despite showing increasingly human-like abilities, large language models (LLMs) often struggle with factual inaccuracies, i.e. "hallucinations", even when they hold relevant knowledge. To address these hallucinations, current approaches typically necessitate high-quality human factuality annotations. In this work, we explore Self-Alignment for Factuality, where we leverage the self-evaluation capability of an LLM to provide training signals that steer the model towards factuality. Specifically, we incorporate Self-Eval, a self-evaluation component, to prompt an LLM to validate the factuality of its own generated responses solely based on its internal knowledge. Additionally, we design Self-Knowledge Tuning (SK-Tuning) to augment the LLM's self-evaluation ability by improving the model's confidence estimation and calibration. We then utilize these self-annotated responses to fine-tune the model via Direct Preference Optimization algorithm. We show that the proposed self-alignment approach substantially enhances factual accuracy over Llama family models across three key knowledge-intensive tasks on TruthfulQA and BioGEN.

Hongyi Guo, Yuanshun Yao, Wei Shen et al.

Aligning large language models (LLMs) with human values is a vital task for LLM practitioners. Current alignment techniques have several limitations: (1) requiring a large amount of annotated data; (2) demanding heavy human involvement; (3) lacking a systematic mechanism to continuously improve. In this work, we study aligning LLMs to a new domain with limited samples (e.g. < 100). We propose an algorithm that can self-align LLMs iteratively without active human involvement. Unlike existing works, our algorithm relies on neither human-crafted instructions nor labeled rewards, significantly reducing human involvement. In addition, our algorithm can self-improve the alignment continuously. The key idea is to first retrieve high-quality samples related to the target domain and use them as In-context Learning examples to generate more samples. Then we use the self-generated samples to finetune the LLM iteratively. We show that our method can unlock the LLMs' self-generalization ability to perform alignment with near-zero human supervision. We test our algorithm on three benchmarks in safety, truthfulness, and instruction-following, and show good performance in alignment, domain adaptability, and scalability.

Xiaoying Zhang, Hao Sun, Yipeng Zhang et al.

Recent advances in reinforcement learning (RL) with numerical feedback, such as scalar rewards, have significantly enhanced the complex reasoning capabilities of large language models (LLMs). Despite this success, we identify three key challenges encountered by RL with solely numerical feedback: performance plateaus, limited effectiveness of spontaneous self-reflection, and persistent failures. We then demonstrate that RL-finetuned models, even after exhibiting performance plateaus, can generate correct refinements on persistently failed problems by leveraging natural language feedback in the form of critiques. Building on this insight, we propose Critique-GRPO, an online RL framework that integrates both natural language and numerical feedback for effective policy optimization. Critique-GRPO enables LLMs to learn from initial responses and critique-guided self-refinements simultaneously while maintaining exploration. Additionally, we employ a shaping function to amplify learning from correct, especially unfamiliar, refinements and penalize incorrect ones. Extensive experiments with Qwen2.5-7B-Base, Qwen2.5-Math-7B-Base, and Qwen3-8B demonstrate that Critique-GRPO consistently outperforms supervised learning and RL-based fine-tuning methods across eight challenging mathematical, STEM, and general reasoning tasks. Specifically, Critique-GRPO improves average pass@1 scores across all compared methods by approximately +4.4% on Qwen2.5-7B-Base and +3.8% on Qwen3-8B. Notably, Critique-GRPO enables effective self-improvement through self-critiquing, achieving significant gains over GRPO, e.g., +16.7% pass@1 improvement on AIME 2024.

Wei Shen, Xiaoying Zhang, Yuanshun Yao et al.

Reinforcement learning from human feedback (RLHF) is the mainstream paradigm used to align large language models (LLMs) with human preferences. Yet existing RLHF heavily relies on accurate and informative reward models, which are vulnerable and sensitive to noise from various sources, e.g. human labeling errors, making the pipeline fragile. In this work, we improve the effectiveness of the reward model by introducing a penalty term on the reward, named as \textit{contrastive rewards}. %Contrastive rewards Our approach involves two steps: (1) an offline sampling step to obtain responses to prompts that serve as baseline calculation and (2) a contrastive reward calculated using the baseline responses and used in the Proximal Policy Optimization (PPO) step. We show that contrastive rewards enable the LLM to penalize reward uncertainty, improve robustness, encourage improvement over baselines, calibrate according to task difficulty, and reduce variance in PPO. We show empirically contrastive rewards can improve RLHF substantially, evaluated by both GPTs and humans, and our method consistently outperforms strong baselines.

Xiaoying Zhang, Jean-Francois Ton, Wei Shen et al.

We introduce Adversarial Policy Optimization (AdvPO), a novel solution to the pervasive issue of reward over-optimization in Reinforcement Learning from Human Feedback (RLHF) for Large Language Models (LLMs). Over-optimization occurs when a reward model serves as an imperfect proxy for human preference, and RL-driven policy optimization erroneously exploits reward inaccuracies. In this paper, we begin by introducing a lightweight way to quantify uncertainties in rewards, relying solely on the last layer embeddings of the reward model, without the need for computationally expensive reward ensembles. AdvPO then addresses a distributionally robust optimization problem centred around the confidence interval of the reward model's predictions for policy improvement. Through comprehensive experiments on the Anthropic HH and TL;DR summarization datasets, we illustrate the efficacy of AdvPO in mitigating the overoptimization issue, consequently resulting in enhanced performance as evaluated through human-assisted evaluation.

Xiaoying Zhang, Da Peng, Yipeng Zhang et al.

Recent progress in (multimodal) large language models ((M)LLMs) has shifted focus from pre-training to inference-time computation and post-training optimization, largely due to concerns over the availability of high-quality human data. However, these strategies alone are insufficient to drive substantial model improvements. We argue that effective model advancement requires strong synergy among pre-training, inference-time computation, and post-training optimization. In this paper, we introduce Self-Improving cognition (SIcog), a self-learning framework for constructing next-generation foundation MLLMs by imparting multimodal knowledge and enhancing systematic cognitive capabilities through multimodal pre-training with self-generated data. Specifically, we propose Chain-of-Description for step-by-step visual understanding and integrate structured Chain-of-Thought (CoT) reasoning to support in-depth multimodal reasoning. SIcog first equips a base model with systematic perception and reasoning using minimal external supervision. The enhanced models then generate candidate image captions and CoT reasoning responses for unlabeled images and image-question pairs across diverse tasks, which are filtered through a semantic-similarity-guided self-consistency mechanism. These high-quality, self-generated samples enable large-scale multimodal pre-training, creating a self-improvement loop. Experiments demonstrate SIcog's effectiveness in developing MLLMs with enhanced multimodal cognition. Using only 213K self-generated pre-training samples, SIcog achieves significant improvements, including +3.6% on MMStar and +3.5% on AI2D, outperforming previous pre-training approaches. When combined with post-training techniques for CoT reasoning, SIcog yields +9% gains on MMVet and +8.5% on ScienceQA.

Xinchen Zhang, Xiaoying Zhang, Youbin Wu et al.

We introduce Generative Universal Verifier, a novel concept and plugin designed for next-generation multimodal reasoning in vision-language models and unified multimodal models, providing the fundamental capability of reflection and refinement on visual outcomes during the reasoning and generation process. This work makes three main contributions: (1) We build ViVerBench, a comprehensive benchmark spanning 16 categories of critical tasks for evaluating visual outcomes in multimodal reasoning. Results show that existing VLMs consistently underperform across these tasks, underscoring a substantial gap from human-level capability in reliable visual verification. (2) We design two automated pipelines to construct large-scale visual verification data and train OmniVerifier-7B, the first omni-capable generative verifier trained for universal visual verification and achieves notable gains on ViVerBench(+8.3). Through training, we identify three atomic capabilities in visual verification and demonstrate how they generalize and interact synergistically. (3) We propose OmniVerifier-TTS, a sequential test-time scaling paradigm that leverages the universal verifier to bridge image generation and editing within unified models, enhancing the upper bound of generative ability through iterative fine-grained optimization. Beyond generation, we extend universal verifier to broader world-modeling interleaved reasoning scenarios. Empirically, OmniVerifier-TTS achieves improvements on T2I-ReasonBench(+3.7), and GenEval++(+4.3), outperforming existing parallel test-time scaling methods, such as Best-of-N. By endowing multimodal reasoning with reliable visual verification, OmniVerifier advances both reliable reflection during generation and scalable test-time refinement, marking a step toward more trustworthy and controllable next-generation reasoning systems.

Xiaoying Zhang

Developing adaptable, extensible, and accurate task bots with minimal or zero human intervention is a significant challenge in dialog research. This thesis examines the obstacles and potential solutions for creating such bots, focusing on innovative techniques that enable bots to learn and adapt autonomously in constantly changing environments.

Yipeng Zhang, Yifan Liu, Zonghao Guo et al.

Vision transformers (ViTs) are widely employed in multimodal large language models (MLLMs) for visual encoding. However, they exhibit inferior performance on tasks regarding fine-grained visual perception. We attribute this to the limitations of ViTs in capturing diverse multi-modal visual levels, such as low-level details. To address this issue, we present LLaVA-UHD v2, an MLLM with advanced perception abilities by introducing a well-designed vision-language projector, the Hierarchical window (Hiwin) transformer. Hiwin transformer enhances MLLM's ability to capture diverse multi-modal visual granularities, by incorporating our constructed high-resolution semantic pyramid. Specifically, Hiwin transformer comprises two key modules: (i) a visual detail injection module, which progressively injects low-level visual details into high-level language-aligned semantics features, thereby forming an inverse semantic pyramid (ISP), and (ii) a hierarchical window attention module, which leverages cross-scale windows to condense multi-level semantics from the ISP. Extensive experiments show that LLaVA-UHD v2 outperforms compared MLLMs on a wide range of benchmarks. Notably, our design achieves an average boost of 3.7% across 14 benchmarks compared with the baseline method, 9.3% on DocVQA for instance. All the data and code will be publicly available to facilitate future research.

Rui Zheng, Hongyi Guo, Zhihan Liu et al.

The standard Reinforcement Learning from Human Feedback (RLHF) framework primarily focuses on optimizing the performance of large language models using pre-collected prompts. However, collecting prompts that provide comprehensive coverage is both tedious and challenging, and often fails to include scenarios that LLMs need to improve on the most. In this paper, we investigate alignment through the lens of two-agent games, involving iterative interactions between an adversarial and a defensive agent. The adversarial agent's task at each step is to generate prompts that expose the weakness of the defensive agent. In return, the defensive agent seeks to improve its responses to these newly identified prompts it struggled with, based on feedback from the reward model. We theoretically demonstrate that this iterative reinforcement learning optimization converges to a Nash Equilibrium for the game induced by the agents. Experimental results in safety scenarios demonstrate that learning in such a competitive environment not only fully trains agents but also leads to policies with enhanced generalization capabilities for both adversarial and defensive agents.

Xiaoying Zhang, Baolin Peng, Ye Tian et al.

Large language models (LLMs) often struggle to provide up-to-date information due to their one-time training and the constantly evolving nature of the world. To keep LLMs current, existing approaches typically involve continued pre-training on new documents. However, they frequently face difficulties in extracting stored knowledge. Motivated by the remarkable success of the Feynman Technique in efficient human learning, we introduce Self-Tuning, a learning framework aimed at improving an LLM's ability to effectively acquire new knowledge from unseen raw documents through self-teaching. Specifically, we develop a Self-Teaching strategy that augments the documents with a set of knowledge-intensive tasks created in a self-supervised manner, focusing on three crucial aspects: memorization, comprehension, and self-reflection. Additionally, we introduce three Wiki-Newpages-2023-QA datasets to facilitate an in-depth analysis of an LLM's knowledge acquisition ability concerning memorization, extraction, and reasoning. Extensive experimental results on various models, e.g., Llama2-7B reveal that Self-Tuning consistently exhibits superior performance across all knowledge acquisition tasks and excels in preserving previous knowledge.

Xiaoying Zhang, Baolin Peng, Kun Li et al.

Building end-to-end task bots and maintaining their integration with new functionalities using minimal human efforts is a long-standing challenge in dialog research. Recently large language models (LLMs) have demonstrated exceptional proficiency in conversational engagement and adherence to instructions across various downstream tasks. In this work, we introduce SGP-TOD, Schema-Guided Prompting for building Task-Oriented Dialog systems effortlessly based on LLMs. Utilizing the symbolic knowledge -- task schema, we instruct fixed LLMs to generate appropriate responses on novel tasks, circumventing the need for training data. Specifically, SGP-TOD comprises three components: a LLM for engaging with users, a DST Prompter to aid the LLM with dialog state tracking, which is then used to retrieve database items, and a Policy Prompter to elicit proper responses adhering to the provided dialog policy. Experimental results on Multiwoz, RADDLE and STAR datasets show that our training-free strategy SGP-TOD, without any task-specific data, yields state-of-the-art (SOTA) zero-shot performance, greatly surpasses the few-shot approaches. In a domain-extension setting, SGP-TOD aptly adapts to new functionalities by merely adding supplementary schema rules. We make our code and data publicly available.

Haidong Xie, Jia Tan, Xiaoying Zhang et al.

Deep learning is applied to many complex tasks in the field of wireless communication, such as modulation recognition of spectrum waveforms, because of its convenience and efficiency. This leads to the problem of a malicious third party using a deep learning model to easily recognize the modulation format of the transmitted waveform. Some existing works address this problem directly using the concept of adversarial examples in the image domain without fully considering the characteristics of the waveform transmission in the physical world. Therefore, we propose a low-intercept waveform~(LIW) generation method that can reduce the probability of the modulation being recognized by a third party without affecting the reliable communication of the friendly party. Our LIW exhibits significant low-interception performance even in the physical hardware experiment, decreasing the accuracy of the state of the art model to approximately $15\%$ with small perturbations.

Xiaoying Zhang, Baolin Peng, Jianfeng Gao et al.

End-to-end task bots are typically learned over a static and usually limited-size corpus. However, when deployed in dynamic, changing, and open environments to interact with users, task bots tend to fail when confronted with data that deviate from the training corpus, i.e., out-of-distribution samples. In this paper, we study the problem of automatically adapting task bots to changing environments by learning from human-bot interactions with minimum or zero human annotations. We propose SL-AGENT, a novel self-learning framework for building end-to-end task bots. SL-AGENT consists of a dialog model and a pre-trained reward model to predict the quality of an agent response. It enables task bots to automatically adapt to changing environments by learning from the unlabeled human-bot dialog logs accumulated after deployment via reinforcement learning with the incorporated reward model. Experimental results on four well-studied dialog tasks show the effectiveness of SL-AGENT to automatically adapt to changing environments, using both automatic and human evaluations. We will release code and data for further research.

Mudit Chaudhary, Borislav Dzodzo, Sida Huang et al.

Dialog systems enriched with external knowledge can handle user queries that are outside the scope of the supporting databases/APIs. In this paper, we follow the baseline provided in DSTC9 Track 1 and propose three subsystems, KDEAK, KnowleDgEFactor, and Ens-GPT, which form the pipeline for a task-oriented dialog system capable of accessing unstructured knowledge. Specifically, KDEAK performs knowledge-seeking turn detection by formulating the problem as natural language inference using knowledge from dialogs, databases and FAQs. KnowleDgEFactor accomplishes the knowledge selection task by formulating a factorized knowledge/document retrieval problem with three modules performing domain, entity and knowledge level analyses. Ens-GPT generates a response by first processing multiple knowledge snippets, followed by an ensemble algorithm that decides if the response should be solely derived from a GPT2-XL model, or regenerated in combination with the top-ranking knowledge snippet. Experimental results demonstrate that the proposed pipeline system outperforms the baseline and generates high-quality responses, achieving at least 58.77% improvement on BLEU-4 score.

Xiaoying Zhang, Hong Xie, Hang Li et al.

Contextual bandit algorithms provide principled online learning solutions to balance the exploitation-exploration trade-off in various applications such as recommender systems. However, the learning speed of the traditional contextual bandit algorithms is often slow due to the need for extensive exploration. This poses a critical issue in applications like recommender systems, since users may need to provide feedbacks on a lot of uninterested items. To accelerate the learning speed, we generalize contextual bandit to conversational contextual bandit. Conversational contextual bandit leverages not only behavioral feedbacks on arms (e.g., articles in news recommendation), but also occasional conversational feedbacks on key-terms from the user. Here, a key-term can relate to a subset of arms, for example, a category of articles in news recommendation. We then design the Conversational UCB algorithm (ConUCB) to address two challenges in conversational contextual bandit: (1) which key-terms to select to conduct conversation, (2) how to leverage conversational feedbacks to accelerate the speed of bandit learning. We theoretically prove that ConUCB can achieve a smaller regret upper bound than the traditional contextual bandit algorithm LinUCB, which implies a faster learning speed. Experiments on synthetic data, as well as real datasets from Yelp and Toutiao, demonstrate the efficacy of the ConUCB algorithm.

Yao Wu, Tianzhen Wu, Ziyi Xiong et al.

Proliferation of ubiquitous mobile devices makes location based services prevalent. Mobile users are able to volunteer as providers of specific services and in the meanwhile to search these services. For example, drivers may be interested in tracking available nearby users who are willing to help with motor repair or are willing to provide travel directions or first aid. With the diffusion of mobile users, it is necessary to provide scalable means of enabling such users to connect with other nearby users so that they can help each other with specific services. Motivated by these observations, we design and implement a general location based system HelPal for mobile users to provide and enjoy instant service, which is called mobile crowd service. In this demo, we introduce a mobile crowd service system featured with several novel techniques. We sketch the system architecture and illustrate scenarios via several cases. Demonstration shows the user-friendly search interface for users to conveniently find skilled and qualified nearby service providers.