Jiacheng Xu

Kexun Zhang, Weiran Yao, Zuxin Liu et al. · cmu

Large language model (LLM) agents have shown great potential in solving real-world software engineering (SWE) problems. The most advanced open-source SWE agent can resolve over 27% of real GitHub issues in SWE-Bench Lite. However, these sophisticated agent frameworks exhibit varying strengths, excelling in certain tasks while underperforming in others. To fully harness the diversity of these agents, we propose DEI (Diversity Empowered Intelligence), a framework that leverages their unique expertise. DEI functions as a meta-module atop existing SWE agent frameworks, managing agent collectives for enhanced problem-solving. Experimental results show that a DEI-guided committee of agents is able to surpass the best individual agent's performance by a large margin. For instance, a group of open-source SWE agents, with a maximum individual resolve rate of 27.3% on SWE-Bench Lite, can achieve a 34.3% resolve rate with DEI, making a 25% improvement and beating most closed-source solutions. Our best-performing group excels with a 55% resolve rate, securing the highest ranking on SWE-Bench Lite. Our findings contribute to the growing body of research on collaborative AI systems and their potential to solve complex software engineering challenges.

Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao et al. · allen-ai, amazon-science

Language models demonstrate both quantitative improvement and new qualitative capabilities with increasing scale. Despite their potentially transformative impact, these new capabilities are as yet poorly characterized. In order to inform future research, prepare for disruptive new model capabilities, and ameliorate socially harmful effects, it is vital that we understand the present and near-future capabilities and limitations of language models. To address this challenge, we introduce the Beyond the Imitation Game benchmark (BIG-bench). BIG-bench currently consists of 204 tasks, contributed by 450 authors across 132 institutions. Task topics are diverse, drawing problems from linguistics, childhood development, math, common-sense reasoning, biology, physics, social bias, software development, and beyond. BIG-bench focuses on tasks that are believed to be beyond the capabilities of current language models. We evaluate the behavior of OpenAI's GPT models, Google-internal dense transformer architectures, and Switch-style sparse transformers on BIG-bench, across model sizes spanning millions to hundreds of billions of parameters. In addition, a team of human expert raters performed all tasks in order to provide a strong baseline. Findings include: model performance and calibration both improve with scale, but are poor in absolute terms (and when compared with rater performance); performance is remarkably similar across model classes, though with benefits from sparsity; tasks that improve gradually and predictably commonly involve a large knowledge or memorization component, whereas tasks that exhibit "breakthrough" behavior at a critical scale often involve multiple steps or components, or brittle metrics; social bias typically increases with scale in settings with ambiguous context, but this can be improved with prompting.

Aakshita Chandiramani, Aaron Blakeman, Abdullahi Olaoye et al. · amazon-science, cmu

We describe the pre-training, post-training, and quantization of Nemotron 3 Super, a 120 billion (active 12 billion) parameter hybrid Mamba-Attention Mixture-of-Experts model. Nemotron 3 Super is the first model in the Nemotron 3 family to 1) be pre-trained in NVFP4, 2) leverage LatentMoE, a new Mixture-of-Experts architecture that optimizes for both accuracy per FLOP and accuracy per parameter, and 3) include MTP layers for inference acceleration through native speculative decoding. We pre-trained Nemotron 3 Super on 25 trillion tokens followed by post-training using supervised fine tuning (SFT) and reinforcement learning (RL). The final model supports up to 1M context length and achieves comparable accuracy on common benchmarks, while also achieving up to 2.2x and 7.5x higher inference throughput compared to GPT-OSS-120B and Qwen3.5-122B, respectively. Nemotron 3 Super datasets, along with the base, post-trained, and quantized checkpoints, are open-sourced on HuggingFace.

Liyan Tang, Tanya Goyal, Alexander R. Fabbri et al. · microsoft-research, salesforce

The propensity of abstractive summarization models to make factual errors has been studied extensively, including design of metrics to detect factual errors and annotation of errors in current systems' outputs. However, the ever-evolving nature of summarization systems, metrics, and annotated benchmarks makes factuality evaluation a moving target, and drawing clear comparisons among metrics has become increasingly difficult. In this work, we aggregate factuality error annotations from nine existing datasets and stratify them according to the underlying summarization model. We compare performance of state-of-the-art factuality metrics, including recent ChatGPT-based metrics, on this stratified benchmark and show that their performance varies significantly across different types of summarization models. Critically, our analysis shows that much of the recent improvement in the factuality detection space has been on summaries from older (pre-Transformer) models instead of more relevant recent summarization models. We further perform a finer-grained analysis per error-type and find similar performance variance across error types for different factuality metrics. Our results show that no one metric is superior in all settings or for all error types, and we provide recommendations for best practices given these insights.

Prasann Singhal, Tanya Goyal, Jiacheng Xu et al.

Great success has been reported using Reinforcement Learning from Human Feedback (RLHF) to align large language models, with open preference datasets enabling wider experimentation, particularly for "helpfulness" in tasks like dialogue and web question answering. Alongside these improvements, however, RLHF also often drives models to produce longer outputs. This paper demonstrates, on three diverse settings, that optimizing for response length is, much more than previously thought, a significant factor behind RLHF. Studying the strategies RL optimization uses to maximize reward, we find improvements in reward to largely be driven by increasing response length, instead of other features. Indeed, we find that even a purely length-based reward reproduces most downstream RLHF improvements over supervised fine-tuned models. Testing a comprehensive set of length-countering interventions, we identify the dominant source of these biases to be reward models, which, by studying training dynamics, we find are non-robust and easily influenced by length biases in preference data.

Jiacheng Xu, Caiming Xiong, Silvio Savarese et al.

Modern natural language generation paradigms require a good decoding strategy to obtain quality sequences out of the model. Beam search yields high-quality but low diversity outputs; stochastic approaches suffer from high variance and sometimes low quality, but the outputs tend to be more natural and creative. In this work, we propose a deterministic search algorithm balancing both quality and diversity. We first investigate the vanilla best-first search (BFS) algorithm and then propose the Best-$k$ Search algorithm. Inspired by BFS, we greedily expand the top $k$ nodes, instead of only the first node, to boost efficiency and diversity. Upweighting recently discovered nodes accompanied by heap pruning ensures the completeness of the search procedure. Experiments on four NLG tasks, including question generation, commonsense generation, text summarization, and translation, show that best-$k$ search yields more diverse and natural outputs compared to strong baselines, while our approach maintains high text quality. The proposed algorithm is parameter-free, lightweight, efficient, and easy to use.

Yik-Cheung Tam, Jiacheng Xu, Jiakai Zou et al.

Performance of spoken language understanding (SLU) can be degraded with automatic speech recognition (ASR) errors. We propose a novel approach to improve SLU robustness by randomly corrupting clean training text with an ASR error simulator, followed by self-correcting the errors and minimizing the target classification loss in a joint manner. In the proposed error simulator, we leverage confusion networks generated from an ASR decoder without human transcriptions to generate a variety of error patterns for model training. We evaluate our approach on the DSTC10 challenge targeted for knowledge-grounded task-oriented conversational dialogues with ASR errors. Experimental results show the effectiveness of our proposed approach, boosting the knowledge-seeking turn detection (KTD) F1 significantly from 0.9433 to 0.9904. Knowledge cluster classification is boosted from 0.7924 to 0.9333 in Recall@1. After knowledge document re-ranking, our approach shows significant improvement in all knowledge selection metrics, from 0.7358 to 0.7806 in Recall@1, from 0.8301 to 0.9333 in Recall@5, and from 0.7798 to 0.8460 in MRR@5 on the test set. In the recent DSTC10 evaluation, our approach demonstrates significant improvement in knowledge selection, boosting Recall@1 from 0.495 to 0.7144 compared to the official baseline. Our source code is released in GitHub https://github.com/yctam/dstc10_track2_task2.git.

Ke Xue, Jiacheng Xu, Lei Yuan et al.

Automated algorithm configuration relieves users from tedious, trial-and-error tuning tasks. A popular algorithm configuration tuning paradigm is dynamic algorithm configuration (DAC), in which an agent learns dynamic configuration policies across instances by reinforcement learning (RL). However, in many complex algorithms, there may exist different types of configuration hyperparameters, and such heterogeneity may bring difficulties for classic DAC which uses a single-agent RL policy. In this paper, we aim to address this issue and propose multi-agent DAC (MA-DAC), with one agent working for one type of configuration hyperparameter. MA-DAC formulates the dynamic configuration of a complex algorithm with multiple types of hyperparameters as a contextual multi-agent Markov decision process and solves it by a cooperative multi-agent RL (MARL) algorithm. To instantiate, we apply MA-DAC to a well-known optimization algorithm for multi-objective optimization problems. Experimental results show the effectiveness of MA-DAC in not only achieving superior performance compared with other configuration tuning approaches based on heuristic rules, multi-armed bandits, and single-agent RL, but also being capable of generalizing to different problem classes. Furthermore, we release the environments in this paper as a benchmark for testing MARL algorithms, with the hope of facilitating the application of MARL.

Zexiang Liu, Dong Li, Kaiyue Lu et al.

Recently, numerous efficient Transformers have been proposed to reduce the quadratic computational complexity of standard Transformers caused by the Softmax attention. However, most of them simply swap Softmax with an efficient attention mechanism without considering the customized architectures specially for the efficient attention. In this paper, we argue that the handcrafted vanilla Transformer architectures for Softmax attention may not be suitable for efficient Transformers. To address this issue, we propose a new framework to find optimal architectures for efficient Transformers with the neural architecture search (NAS) technique. The proposed method is validated on popular machine translation and image classification tasks. We observe that the optimal architecture of the efficient Transformer has the reduced computation compared with that of the standard Transformer, but the general accuracy is less comparable. It indicates that the Softmax attention and efficient attention have their own distinctions but neither of them can simultaneously balance the accuracy and efficiency well. This motivates us to mix the two types of attention to reduce the performance imbalance. Besides the search spaces that commonly used in existing NAS Transformer approaches, we propose a new search space that allows the NAS algorithm to automatically search the attention variants along with architectures. Extensive experiments on WMT' 14 En-De and CIFAR-10 demonstrate that our searched architecture maintains comparable accuracy to the standard Transformer with notably improved computational efficiency.

Prasann Singhal, Jiacheng Xu, Xi Ye et al.

Standard decoding approaches for conditional text generation tasks typically search for an output hypothesis with high model probability, but this may not yield the best hypothesis according to human judgments of quality. Reranking to optimize for "downstream" metrics can better optimize for quality, but many metrics of interest are computed with pre-trained language models, which are slow to apply to large numbers of hypotheses. We explore an approach for reranking hypotheses by using Transformers to efficiently encode lattices of generated outputs, a method we call EEL. With a single Transformer pass over the entire lattice, we can approximately compute a contextualized representation of each token as if it were only part of a single hypothesis in isolation. We combine this approach with a new class of token-factored rerankers (TFRs) that allow for efficient extraction of high reranker-scoring hypotheses from the lattice. Empirically, our approach incurs minimal degradation error compared to the exponentially slower approach of encoding each hypothesis individually. When applying EEL with TFRs across three text generation tasks, our results show both substantial speedup compared to naive reranking and often better performance on downstream metrics than comparable approaches.

Hongjin Su, Shizhe Diao, Ximing Lu et al.

Large language models are powerful generalists, yet solving deep and complex problems such as those of the Humanity's Last Exam (HLE) remains both conceptually challenging and computationally expensive. We show that small orchestrators managing other models and a variety of tools can both push the upper bound of intelligence and improve efficiency in solving difficult agentic tasks. We introduce ToolOrchestra, a method for training small orchestrators that coordinate intelligent tools. ToolOrchestra explicitly uses reinforcement learning with outcome-, efficiency-, and user-preference-aware rewards. Using ToolOrchestra, we produce Orchestrator, an 8B model that achieves higher accuracy at lower cost than previous tool-use agents while aligning with user preferences on which tools are to be used for a given query. On HLE, Orchestrator achieves a score of 37.1%, outperforming GPT-5 (35.1%) while being 2.5x more efficient. On tau2-Bench and FRAMES, Orchestrator surpasses GPT-5 by a wide margin while using only about 30% of the cost. Extensive analysis shows that Orchestrator achieves the best trade-off between performance and cost under multiple metrics, and generalizes robustly to unseen tools. These results demonstrate that composing diverse tools with a lightweight orchestration model is both more efficient and more effective than existing methods, paving the way for practical and scalable tool-augmented reasoning systems.

Jujie He, Jiacai Liu, Chris Yuhao Liu et al.

The success of DeepSeek-R1 underscores the significant role of reinforcement learning (RL) in enhancing the reasoning capabilities of large language models (LLMs). In this work, we present Skywork-OR1, an effective and scalable RL implementation for long Chain-of-Thought (CoT) models. Building on the DeepSeek-R1-Distill model series, our RL approach achieves notable performance gains, increasing average accuracy across AIME24, AIME25, and LiveCodeBench from 57.8% to 72.8% (+15.0%) for the 32B model and from 43.6% to 57.5% (+13.9%) for the 7B model. Our Skywork-OR1-32B model surpasses both DeepSeek-R1 and Qwen3-32B on the AIME24 and AIME25 benchmarks, while achieving comparable results on LiveCodeBench. The Skywork-OR1-7B and Skywork-OR1-Math-7B models demonstrate competitive reasoning capabilities among models of similar size. We perform comprehensive ablation studies on the core components of our training pipeline to validate their effectiveness. Additionally, we thoroughly investigate the phenomenon of entropy collapse, identify key factors affecting entropy dynamics, and demonstrate that mitigating premature entropy collapse is critical for improved test performance. To support community research, we fully open-source our model weights, training code, and training datasets.

Jiacheng Xu, Heting Gao, Liufei Xie et al.

Human speech conveys expressiveness beyond linguistic content, including personality, mood, or performance elements, such as a comforting tone or humming a song, which we formalize as role-playing and singing. We present VITA-QinYu, the first expressive end-to-end (E2E) spoken language model (SLM) that goes beyond natural conversation to support both role-playing and singing generation. VITA-QinYu adopts a hybrid speech-text paradigm that extends interleaved text-audio modeling with multi-codebook audio tokens, a design enabling richer paralinguistic representation while preserving a clear separation between modalities to avoid interference. We further develop a comprehensive data generation pipeline to synthesize a total of 15.8K hours of natural conversation, role-playing, and singing data for training. VITA-QinYu demonstrates superior expressiveness, outperforming peer SLMs by 7 percentage points on objective role-playing benchmarks, and surpassing peer models by 0.13 points on a 5-point MOS scale for singing. Simultaneously, it achieves state-of-the-art conversational accuracy and fluency, exceeding prior SLMs by 1.38 and 4.98 percentage points on the C3 and URO benchmarks, respectively. We open-source our code and models and provide an easy-to-use demo with full-stack support for streaming and full-duplex interaction.

JiaCheng Xu, ChengXiang Tan

Membership inference attack is one of the most popular privacy attacks in machine learning, which aims to predict whether a given sample was contained in the target model's training set. Label-only membership inference attack is a variant that exploits sample robustness and attracts more attention since it assumes a practical scenario in which the adversary only has access to the predicted labels of the input samples. However, since the decision boundary distance, which measures robustness, is strongly affected by the random initial image, the adversary may get opposite results even for the same input samples. In this paper, we propose a new attack method, called muti-class adaptive membership inference attack in the label-only setting. All decision boundary distances for all target classes have been traversed in the early attack iterations, and the subsequent attack iterations continue with the shortest decision boundary distance to obtain a stable and optimal decision boundary distance. Instead of using a single boundary distance, the relative boundary distance between samples and neighboring points has also been employed as a new membership score to distinguish between member samples inside the training set and nonmember samples outside the training set. Experiments show that previous label-only membership inference attacks using the untargeted HopSkipJump algorithm fail to achieve optimal decision bounds in more than half of the samples, whereas our multi-targeted HopSkipJump algorithm succeeds in almost all samples. In addition, extensive experiments show that our multi-class adaptive MIA outperforms current label-only membership inference attacks in the CIFAR10, and CIFAR100 datasets, especially for the true positive rate at low false positive rates metric.

Jiafei Lyu, Kang Xu, Jiacheng Xu et al.

We consider off-dynamics reinforcement learning (RL) where one needs to transfer policies across different domains with dynamics mismatch. Despite the focus on developing dynamics-aware algorithms, this field is hindered due to the lack of a standard benchmark. To bridge this gap, we introduce ODRL, the first benchmark tailored for evaluating off-dynamics RL methods. ODRL contains four experimental settings where the source and target domains can be either online or offline, and provides diverse tasks and a broad spectrum of dynamics shifts, making it a reliable platform to comprehensively evaluate the agent's adaptation ability to the target domain. Furthermore, ODRL includes recent off-dynamics RL algorithms in a unified framework and introduces some extra baselines for different settings, all implemented in a single-file manner. To unpack the true adaptation capability of existing methods, we conduct extensive benchmarking experiments, which show that no method has universal advantages across varied dynamics shifts. We hope this benchmark can serve as a cornerstone for future research endeavors. Our code is publicly available at https://github.com/OffDynamicsRL/off-dynamics-rl.

Baining Zhao, Jiacheng Xu, Weicheng Feng et al.

Aerial vision-language navigation (VLN) requires agents to follow natural-language instructions through closed-loop perception and action in 3D environments. We argue that aerial VLN can be formulated as a prediction-driven world-action problem: the agent should anticipate latent world evolution and act according to the predicted consequences. To this end, we propose WorldVLN, the first autoregressive world action model for aerial VLN. Unlike full-sequence video-generation world models that generate an entire visual clip, WorldVLN adapts a latent autoregressive video backbone to predict short-horizon world-state transitions and directly decodes them into executable waypoint actions. After each action segment is executed, newly received observations are encoded back into the autoregressive context, enabling closed-loop world-action prediction. We further introduce a two-stage training framework that first grounds the video prior in instruction-conditioned navigation dynamics and then develops Action-aware GRPO, the first reinforcement learning method tailored to autoregressive WAMs, to optimize waypoint decisions through their downstream rollout consequences. On public outdoor and indoor benchmarks, WorldVLN consistently outperforms existing Vision-Language-Action baselines with 12\%+ success-rate gains and larger advantages on challenging cases. It further transfers zero-shot to real drone deployment, suggesting that the proposed WorldVLN offers a promising route for spatial action tasks. Demos and code are available at https://embodiedcity.github.io/WorldVLN/.

Baining Zhao, Ziyou Wang, Jianjie Fang et al.

Large multimodal models (LMMs) show strong visual-linguistic reasoning but their capacity for spatial decision-making and action remains unclear. In this work, we investigate whether LMMs can achieve embodied spatial action like human through a challenging scenario: goal-oriented navigation in urban 3D spaces. We first spend over 500 hours constructing a dataset comprising 5,037 high-quality goal-oriented navigation samples, with an emphasis on 3D vertical actions and rich urban semantic information. Then, we comprehensively assess 17 representative models, including non-reasoning LMMs, reasoning LMMs, agent-based methods, and vision-language-action models. Experiments show that current LMMs exhibit emerging action capabilities, yet remain far from human-level performance. Furthermore, we reveal an intriguing phenomenon: navigation errors do not accumulate linearly but instead diverge rapidly from the destination after a critical decision bifurcation. The limitations of LMMs are investigated by analyzing their behavior at these critical decision bifurcations. Finally, we experimentally explore four promising directions for improvement: geometric perception, cross-view understanding, spatial imagination, and long-term memory. The project is available at: https://github.com/serenditipy-AC/Embodied-Navigation-Bench.

Fuying Wang, Jiacheng Xu, Lequan Yu

Electrocardiograms (ECGs) play a vital role in monitoring cardiac health and diagnosing heart diseases. However, traditional deep learning approaches for ECG analysis rely heavily on large-scale manual annotations, which are both time-consuming and resource-intensive to obtain. To overcome this limitation, self-supervised learning (SSL) has emerged as a promising alternative, enabling the extraction of robust ECG representations that can be efficiently transferred to various downstream tasks. While previous studies have explored SSL for ECG pretraining and multi-modal ECG-language alignment, they often fail to capture the multi-scale nature of ECG signals. As a result, these methods struggle to learn generalized representations due to their inability to model the hierarchical structure of ECG data. To address this gap, we introduce MELP, a novel Multi-scale ECG-Language Pretraining (MELP) model that fully leverages hierarchical supervision from ECG-text pairs. MELP first pretrains a cardiology-specific language model to enhance its understanding of clinical text. It then applies three levels of cross-modal supervision-at the token, beat, and rhythm levels-to align ECG signals with textual reports, capturing structured information across different time scales. We evaluate MELP on three public ECG datasets across multiple tasks, including zero-shot ECG classification, linear probing, and transfer learning. Experimental results demonstrate that MELP outperforms existing SSL methods, underscoring its effectiveness and adaptability across diverse clinical applications. Our code is available at https://github.com/HKU-MedAI/MELP.

Kaixin Jia, Jiacheng Xu

We introduce Forecast-aware Gaussian Splatting (Forecast-GS), a predictive 3D representation framework for language-conditioned robotic manipulation. While recent manipulation systems have made progress by grounding language instructions into robot affordances, value maps, or relational keypoint constraints, they usually reason over the current scene and do not explicitly model the task-completed state. This limitation is critical when success depends on satisfying spatial and semantic goals under partial observations, where the robot must evaluate whether a candidate action leads to a feasible task-consistent outcome. We validate Forecast-GS on real-world pick-and-place manipulation tasks, including Cutter-to-Box, Apple-to-Bowl, and Sponge-to-Tray. For each task, we conduct 25 real-world trials under varied initial object configurations using the same robot platform and sensing setup. Forecast-GS with automatic candidate selection achieves success rates of 21/25, 23/25, and 16/25 on the three tasks, respectively, outperforming the ReKep baseline, which achieves 15/25, 19/25, and 10/25. A diagnostic human-assisted setting further improves success rates to 23/25, 24/25, and 19/25, suggesting that candidate generation is effective while automatic ranking remains imperfect. These results suggest that explicitly forecasting task-completed 3D states enables more reliable action evaluation, while the gap between automatic and human-assisted selection indicates that robust final-state ranking remains an important challenge for fully autonomous manipulation. Overall, Forecast-GS provides an interpretable bridge between language understanding, 3D perception, and robotic manipulation planning.

Jia Li, Yichao He, Jiacheng Xu et al.

Accurate and reliable personality assessment plays a vital role in many fields, such as emotional intelligence, mental health diagnostics, and personalized education. Unlike fleeting emotions, personality traits are stable, often subconsciously leaked through language, facial expressions, and body behaviors, with asynchronous patterns across modalities. It was hard to model personality semantics with traditional superficial features and seemed impossible to achieve effective cross-modal understanding. To address these challenges, we propose a novel personality assessment framework called \textit{\textbf{Traits Run Deep}}. It employs \textit{\textbf{psychology-informed prompts}} to elicit high-level personality-relevant semantic representations. Besides, it devises a \textit{\textbf{Text-Centric Trait Fusion Network}} that anchors rich text semantics to align and integrate asynchronous signals from other modalities. To be specific, such fusion module includes a Chunk-Wise Projector to decrease dimensionality, a Cross-Modal Connector and a Text Feature Enhancer for effective modality fusion and an ensemble regression head to improve generalization in data-scarce situations. To our knowledge, we are the first to apply personality-specific prompts to guide large language models (LLMs) in extracting personality-aware semantics for improved representation quality. Furthermore, extracting and fusing audio-visual apparent behavior features further improves the accuracy. Experimental results on the AVI validation set have demonstrated the effectiveness of the proposed components, i.e., approximately a 45\% reduction in mean squared error (MSE). Final evaluations on the test set of the AVI Challenge 2025 confirm our method's superiority, ranking first in the Personality Assessment track. The source code will be made available at https://github.com/MSA-LMC/TraitsRunDeep.

Aditya Gupta, Jiacheng Xu, Shyam Upadhyay et al.

Disfluencies is an under-studied topic in NLP, even though it is ubiquitous in human conversation. This is largely due to the lack of datasets containing disfluencies. In this paper, we present a new challenge question answering dataset, Disfl-QA, a derivative of SQuAD, where humans introduce contextual disfluencies in previously fluent questions. Disfl-QA contains a variety of challenging disfluencies that require a more comprehensive understanding of the text than what was necessary in prior datasets. Experiments show that the performance of existing state-of-the-art question answering models degrades significantly when tested on Disfl-QA in a zero-shot setting.We show data augmentation methods partially recover the loss in performance and also demonstrate the efficacy of using gold data for fine-tuning. We argue that we need large-scale disfluency datasets in order for NLP models to be robust to them. The dataset is publicly available at: https://github.com/google-research-datasets/disfl-qa.

Chris Yuhao Liu, Liang Zeng, Yuzhen Xiao et al.

Despite the critical role of reward models (RMs) in reinforcement learning from human feedback (RLHF), current state-of-the-art open RMs perform poorly on most existing evaluation benchmarks, failing to capture the spectrum of nuanced and sophisticated human preferences. Even approaches that incorporate advanced training techniques have not yielded meaningful performance improvements. We hypothesize that this brittleness stems primarily from limitations in preference datasets, which are often narrowly scoped, synthetically labeled, or lack rigorous quality control. To address these challenges, we present a large-scale preference dataset comprising 40 million preference pairs, named SynPref-40M. To enable data curation at scale, we design a human-AI synergistic two-stage pipeline that leverages the complementary strengths of human annotation quality and AI scalability. In this pipeline, humans provide verified annotations, while large language models perform automatic curation based on human guidance. Training on this preference mixture, we introduce Skywork-Reward-V2, a suite of eight reward models ranging from 0.6B to 8B parameters, trained on a carefully curated subset of 26 million preference pairs from SynPref-40M. We demonstrate that Skywork-Reward-V2 is versatile across a wide range of capabilities, including alignment with human preferences, objective correctness, safety, resistance to stylistic biases, and best-of-N scaling, achieving state-of-the-art performance across seven major reward model benchmarks. Ablation studies confirm that the effectiveness of our approach stems not only from data scale but also from high-quality curation. The Skywork-Reward-V2 series represents substantial progress in open reward models, highlighting the untapped potential of existing preference datasets and demonstrating how human-AI curation synergy can unlock significantly higher data quality.

Bo Pang, Hanze Dong, Jiacheng Xu et al.

Large language models (LLMs), such as o1 from OpenAI, have demonstrated remarkable reasoning capabilities. o1 generates a long chain-of-thought (LongCoT) before answering a question. LongCoT allows LLMs to analyze problems, devise plans, reflect, and backtrack effectively. These actions empower LLM to solve complex problems. After the release of o1, many teams have attempted to replicate its LongCoT and reasoning capabilities. In terms of methods, they primarily rely on knowledge distillation with data from existing models with LongCoT capacities (e.g., OpenAI-o1, Qwen-QwQ, DeepSeek-R1-Preview), leaving significant uncertainties on systematically developing such reasoning abilities. In terms of data domains, these works focus narrowly on math while a few others include coding, limiting their generalizability. This paper introduces a novel approach to enable LLM's LongCoT capacity without distillation from o1-like models or expensive human annotations, where we bootstrap LongCoT (BOLT) from a standard instruct model. BOLT involves three stages: 1) LongCoT data bootstrapping with in-context learning on a standard instruct model; 2) LongCoT supervised finetuning; 3) online training to further refine LongCoT capacities. In BOLT, only a few in-context examples need to be constructed during the bootstrapping stage; in our experiments, we created 10 examples, demonstrating the feasibility of this approach. We use Llama-3.1-70B-Instruct to bootstrap LongCoT and apply our method to various model scales (7B, 8B, 70B). We achieve impressive performance on a variety of benchmarks, Arena-Hard, MT-Bench, WildBench, ZebraLogic, MATH500, which evaluate diverse task-solving and reasoning capabilities.

Jiacheng Xu, Bo Pang, Jin Qu et al.

Software testing is a critical aspect of software development, yet generating test cases remains a routine task for engineers. This paper presents a benchmark, CLOVER, to evaluate models' capabilities in generating and completing test cases under specific conditions. Spanning from simple assertion completions to writing test cases that cover specific code blocks across multiple files, these tasks are based on 12 python repositories, analyzing 845 problems with context lengths ranging from 4k to 128k tokens. Utilizing code testing frameworks, we propose a method to construct retrieval contexts using coverage information. While models exhibit comparable performance with short contexts, notable differences emerge with 16k contexts. Notably, models like GPT-4o and Claude 3.5 can effectively leverage relevant snippets; however, all models score below 35\% on the complex Task III, even with the oracle context provided, underscoring the benchmark's significance and the potential for model improvement. The benchmark is containerized for code execution across tasks, and we will release the code, data, and construction methodologies.

Lifu Tu, Semih Yavuz, Jin Qu et al.

Large Language Models (LLMs) have demonstrated a powerful ability for text generation. However, achieving optimal results with a given prompt or instruction can be challenging, especially for billion-sized models. Additionally, undesired behaviors such as toxicity or hallucinations can manifest. While much larger models (e.g., ChatGPT) may demonstrate strength in mitigating these issues, there is still no guarantee of complete prevention. In this work, we propose formalizing text generation as a future-constrained generation problem to minimize undesirable behaviors and enforce faithfulness to instructions. The estimation of future constraint satisfaction, accomplished using LLMs, guides the text generation process. Our extensive experiments demonstrate the effectiveness of the proposed approach across three distinct text generation tasks: keyword-constrained generation (Lin et al., 2020), toxicity reduction (Gehman et al., 2020), and factual correctness in question-answering (Gao et al., 2023).

Chao Tang, Jiacheng Xu, Haofei Lu et al.

Building generalist robots capable of performing functional grasping in everyday, open-world environments remains a significant challenge due to the vast diversity of objects and tasks. Existing methods are either constrained to narrow object/task sets or rely on prohibitively large-scale data collection to capture real-world variability. In this work, we present an alternative approach, GraspDreamer, a method that leverages human demonstrations synthesized by visual generative models (VGMs) (e.g., video generation models) to enable zero-shot functional grasping without labor-intensive data collection. The key idea is that VGMs pre-trained on internet-scale human data implicitly encode generalized priors about how humans interact with the physical world, which can be combined with embodiment-specific action optimization to enable functional grasping with minimal effort. Extensive experiments on the public benchmarks with different robot hands demonstrate the superior data efficiency and generalization performance of GraspDreamer compared to previous methods. Real-world evaluations further validate the effectiveness on real robots. Additionally, we showcase that GraspDreamer can (1) be naturally extended to downstream manipulation tasks, and (2) can generate data to support visuomotor policy learning.

Fuxiang Zhang, Jiacheng Xu, Chaojie Wang et al.

Large Language Models (LLMs) have demonstrated remarkable progress in complex reasoning tasks through both post-training and test-time scaling laws. While prevalent test-time scaling approaches are often realized by using external reward models to guide the model generation process, we find that only marginal gains can be acquired when scaling a model post-trained on specific reasoning tasks. We identify that the limited improvement stems from distribution discrepancies between the specific post-trained generator and the general reward model. To address this, we propose a framework that incentivizes LLMs to self-verify their own answers. By unifying answer generation and verification within a single reinforcement learning (RL) process, we train models that can effectively assess the correctness of their own solutions. The trained model can further scale its performance at inference time by verifying its generations, without the need for external verifiers. We train our self-verification models based on Qwen2.5-Math-7B and DeepSeek-R1-Distill-Qwen-1.5B, demonstrating their capabilities across varying reasoning context lengths. Experiments on multiple mathematical reasoning benchmarks show that our models can not only improve post-training performance but also enable effective test-time scaling.

Zhenghai Xue, Lang Feng, Jiacheng Xu et al.

To learn from data collected in diverse dynamics, Imitation from Observation (IfO) methods leverage expert state trajectories based on the premise that recovering expert state distributions in other dynamics facilitates policy learning in the current one. However, Imitation Learning inherently imposes a performance upper bound of learned policies. Additionally, as the environment dynamics change, certain expert states may become inaccessible, rendering their distributions less valuable for imitation. To address this, we propose a novel framework that integrates reward maximization with IfO, employing F-distance regularized policy optimization. This framework enforces constraints on globally accessible states--those with nonzero visitation frequency across all considered dynamics--mitigating the challenge posed by inaccessible states. By instantiating F-distance in different ways, we derive two theoretical analysis and develop a practical algorithm called Accessible State Oriented Policy Regularization (ASOR). ASOR serves as a general add-on module that can be incorporated into various RL approaches, including offline RL and off-policy RL. Extensive experiments across multiple benchmarks demonstrate ASOR's effectiveness in enhancing state-of-the-art cross-domain policy transfer algorithms, significantly improving their performance.

Jing-Cheng Pang, Pengyuan Wang, Kaiyuan Li et al.

Large Language Models (LLMs) have exhibited remarkable performance across various natural language processing (NLP) tasks. However, fine-tuning these models often necessitates substantial supervision, which can be expensive and time-consuming to obtain. This paper introduces a novel unsupervised method called LanguageModel Self-Improvement by Reinforcement Learning Contemplation (SIRLC) that improves LLMs without reliance on external labels. Our approach is grounded in the observation that it is simpler for language models to assess text quality than to generate text. Building on this insight, SIRLC assigns LLMs dual roles as both student and teacher. As a student, the LLM generates answers to unlabeled questions, while as a teacher, it evaluates the generated text and assigns scores accordingly. The model parameters are updated using reinforcement learning to maximize the evaluation score. We demonstrate that SIRLC can be applied to various NLP tasks, such as reasoning problems, text generation, and machine translation. Our experiments show that SIRLC effectively improves LLM performance without external supervision, resulting in a 5.6% increase in answering accuracy for reasoning tasks and a rise in BERTScore from 0.82 to 0.86 for translation tasks. Furthermore, SIRLC can be applied to models of different sizes, showcasing its broad applicability.

Jiacheng Xu, Siddhartha Reddy Jonnalagadda, Greg Durrett

Conditional neural text generation models generate high-quality outputs, but often concentrate around a mode when what we really want is a diverse set of options. We present a search algorithm to construct lattices encoding a massive number of generation options. First, we restructure decoding as a best-first search, which explores the space differently than beam search and improves efficiency by avoiding pruning paths. Second, we revisit the idea of hypothesis recombination: we can identify pairs of similar generation candidates during search and merge them as an approximation. On both summarization and machine translation, we show that our algorithm encodes thousands of diverse options that remain grammatical and high-quality into one lattice. This algorithm provides a foundation for building downstream generation applications on top of massive-scale diverse outputs.

Tanya Goyal, Jiacheng Xu, Junyi Jessy Li et al.

Pre-trained language models (e.g. BART) have shown impressive results when fine-tuned on large summarization datasets. However, little is understood about this fine-tuning process, including what knowledge is retained from pre-training time or how content selection and generation strategies are learnt across iterations. In this work, we analyze the training dynamics for generation models, focusing on summarization. Across different datasets (CNN/DM, XSum, MediaSum) and summary properties, such as abstractiveness and hallucination, we study what the model learns at different stages of its fine-tuning process. We find that a propensity to copy the input is learned early in the training process consistently across all datasets studied. On the other hand, factual errors, such as hallucination of unsupported facts, are learnt in the later stages, though this behavior is more varied across domains. Based on these observations, we explore complementary approaches for modifying training: first, disregarding high-loss tokens that are challenging to learn and second, disregarding low-loss tokens that are learnt very quickly in the latter stages of the training process. We show that these simple training modifications allow us to configure our model to achieve different goals, such as improving factuality or improving abstractiveness.

Ojas Ahuja, Jiacheng Xu, Akshay Gupta et al.

Generic summaries try to cover an entire document and query-based summaries try to answer document-specific questions. But real users' needs often fall in between these extremes and correspond to aspects, high-level topics discussed among similar types of documents. In this paper, we collect a dataset of realistic aspect-oriented summaries, AspectNews, which covers different subtopics about articles in news sub-domains. We annotate data across two domains of articles, earthquakes and fraud investigations, where each article is annotated with two distinct summaries focusing on different aspects for each domain. A system producing a single generic summary cannot concisely satisfy both aspects. Our focus in evaluation is how well existing techniques can generalize to these domains without seeing in-domain training data, so we turn to techniques to construct synthetic training data that have been used in query-focused summarization work. We compare several training schemes that differ in how strongly keywords are used and how oracle summaries are extracted. Our evaluation shows that our final approach yields (a) focused summaries, better than those from a generic summarization system or from keyword matching; (b) a system sensitive to the choice of keywords.

Jiacheng Xu, Greg Durrett

Despite the prominence of neural abstractive summarization models, we know little about how they actually form summaries and how to understand where their decisions come from. We propose a two-step method to interpret summarization model decisions. We first analyze the model's behavior by ablating the full model to categorize each decoder decision into one of several generation modes: roughly, is the model behaving like a language model, is it relying heavily on the input, or is it somewhere in between? After isolating decisions that do depend on the input, we explore interpreting these decisions using several different attribution methods. We compare these techniques based on their ability to select content and reconstruct the model's predicted token from perturbations of the input, thus revealing whether highlighted attributions are truly important for the generation of the next token. While this machinery can be broadly useful even beyond summarization, we specifically demonstrate its capability to identify phrases the summarization model has memorized and determine where in the training pipeline this memorization happened, as well as study complex generation phenomena like sentence fusion on a per-instance basis.

Shrey Desai, Jiacheng Xu, Greg Durrett

Compressive summarization systems typically rely on a crafted set of syntactic rules to determine what spans of possible summary sentences can be deleted, then learn a model of what to actually delete by optimizing for content selection (ROUGE). In this work, we propose to relax the rigid syntactic constraints on candidate spans and instead leave compression decisions to two data-driven criteria: plausibility and salience. Deleting a span is plausible if removing it maintains the grammaticality and factuality of a sentence, and spans are salient if they contain important information from the summary. Each of these is judged by a pre-trained Transformer model, and only deletions that are both plausible and not salient can be applied. When integrated into a simple extraction-compression pipeline, our method achieves strong in-domain results on benchmark summarization datasets, and human evaluation shows that the plausibility model generally selects for grammatical and factual deletions. Furthermore, the flexibility of our approach allows it to generalize cross-domain: our system fine-tuned on only 500 samples from a new domain can match or exceed an in-domain extractive model trained on much more data.

Jiacheng Xu, Shrey Desai, Greg Durrett

An advantage of seq2seq abstractive summarization models is that they generate text in a free-form manner, but this flexibility makes it difficult to interpret model behavior. In this work, we analyze summarization decoders in both blackbox and whitebox ways by studying on the entropy, or uncertainty, of the model's token-level predictions. For two strong pre-trained models, PEGASUS and BART on two summarization datasets, we find a strong correlation between low prediction entropy and where the model copies tokens rather than generating novel text. The decoder's uncertainty also connects to factors like sentence position and syntactic distance between adjacent pairs of tokens, giving a sense of what factors make a context particularly selective for the model's next output token. Finally, we study the relationship of decoder uncertainty and attention behavior to understand how attention gives rise to these observed effects in the model. We show that uncertainty is a useful perspective for analyzing summarization and text generation models more broadly.

Jiacheng Xu, Zhe Gan, Yu Cheng et al.

Recently BERT has been adopted for document encoding in state-of-the-art text summarization models. However, sentence-based extractive models often result in redundant or uninformative phrases in the extracted summaries. Also, long-range dependencies throughout a document are not well captured by BERT, which is pre-trained on sentence pairs instead of documents. To address these issues, we present a discourse-aware neural summarization model - DiscoBert. DiscoBert extracts sub-sentential discourse units (instead of sentences) as candidates for extractive selection on a finer granularity. To capture the long-range dependencies among discourse units, structural discourse graphs are constructed based on RST trees and coreference mentions, encoded with Graph Convolutional Networks. Experiments show that the proposed model outperforms state-of-the-art methods by a significant margin on popular summarization benchmarks compared to other BERT-base models.

Jiacheng Xu, Greg Durrett

Recent neural network approaches to summarization are largely either selection-based extraction or generation-based abstraction. In this work, we present a neural model for single-document summarization based on joint extraction and syntactic compression. Our model chooses sentences from the document, identifies possible compressions based on constituency parses, and scores those compressions with a neural model to produce the final summary. For learning, we construct oracle extractive-compressive summaries, then learn both of our components jointly with this supervision. Experimental results on the CNN/Daily Mail and New York Times datasets show that our model achieves strong performance (comparable to state-of-the-art systems) as evaluated by ROUGE. Moreover, our approach outperforms an off-the-shelf compression module, and human and manual evaluation shows that our model's output generally remains grammatical.

Jiacheng Xu, Greg Durrett

A hallmark of variational autoencoders (VAEs) for text processing is their combination of powerful encoder-decoder models, such as LSTMs, with simple latent distributions, typically multivariate Gaussians. These models pose a difficult optimization problem: there is an especially bad local optimum where the variational posterior always equals the prior and the model does not use the latent variable at all, a kind of "collapse" which is encouraged by the KL divergence term of the objective. In this work, we experiment with another choice of latent distribution, namely the von Mises-Fisher (vMF) distribution, which places mass on the surface of the unit hypersphere. With this choice of prior and posterior, the KL divergence term now only depends on the variance of the vMF distribution, giving us the ability to treat it as a fixed hyperparameter. We show that doing so not only averts the KL collapse, but consistently gives better likelihoods than Gaussians across a range of modeling conditions, including recurrent language modeling and bag-of-words document modeling. An analysis of the properties of our vMF representations shows that they learn richer and more nuanced structures in their latent representations than their Gaussian counterparts.

Jinyue Su, Jiacheng Xu, Xipeng Qiu et al.

Generating plausible and fluent sentence with desired properties has long been a challenge. Most of the recent works use recurrent neural networks (RNNs) and their variants to predict following words given previous sequence and target label. In this paper, we propose a novel framework to generate constrained sentences via Gibbs Sampling. The candidate sentences are revised and updated iteratively, with sampled new words replacing old ones. Our experiments show the effectiveness of the proposed method to generate plausible and diverse sentences.

Jiacheng Xu, Kan Chen, Xipeng Qiu et al.

The objective of knowledge graph embedding is to encode both entities and relations of knowledge graphs into continuous low-dimensional vector spaces. Previously, most works focused on symbolic representation of knowledge graph with structure information, which can not handle new entities or entities with few facts well. In this paper, we propose a novel deep architecture to utilize both structural and textual information of entities. Specifically, we introduce three neural models to encode the valuable information from text description of entity, among which an attentive model can select related information as needed. Then, a gating mechanism is applied to integrate representations of structure and text into a unified architecture. Experiments show that our models outperform baseline by margin on link prediction and triplet classification tasks. Source codes of this paper will be available on Github.

Jiacheng Xu, Danlu Chen, Xipeng Qiu et al.

Recently, neural networks have achieved great success on sentiment classification due to their ability to alleviate feature engineering. However, one of the remaining challenges is to model long texts in document-level sentiment classification under a recurrent architecture because of the deficiency of the memory unit. To address this problem, we present a Cached Long Short-Term Memory neural networks (CLSTM) to capture the overall semantic information in long texts. CLSTM introduces a cache mechanism, which divides memory into several groups with different forgetting rates and thus enables the network to keep sentiment information better within a recurrent unit. The proposed CLSTM outperforms the state-of-the-art models on three publicly available document-level sentiment analysis datasets.