Hao Ma

Wenhan Xiong, Jingyu Liu, Igor Molybog et al. · meta-ai

We present a series of long-context LLMs that support effective context windows of up to 32,768 tokens. Our model series are built through continual pretraining from Llama 2 with longer training sequences and on a dataset where long texts are upsampled. We perform extensive evaluation on language modeling, synthetic context probing tasks, and a wide range of research benchmarks. On research benchmarks, our models achieve consistent improvements on most regular tasks and significant improvements on long-context tasks over Llama 2. Notably, with a cost-effective instruction tuning procedure that does not require human-annotated long instruction data, the 70B variant can already surpass gpt-3.5-turbo-16k's overall performance on a suite of long-context tasks. Alongside these results, we provide an in-depth analysis on the individual components of our method. We delve into Llama's position encodings and discuss its limitation in modeling long dependencies. We also examine the impact of various design choices in the pretraining process, including the data mix and the training curriculum of sequence lengths -- our ablation experiments suggest that having abundant long texts in the pretrain dataset is not the key to achieving strong performance, and we empirically verify that long context continual pretraining is more efficient and similarly effective compared to pretraining from scratch with long sequences.

Zhuofeng Wu, Sinong Wang, Jiatao Gu et al. · meta-ai

Prompt tuning is a new, efficient NLP transfer learning paradigm that adds a task-specific prompt in each input instance during the model training stage. It freezes the pre-trained language model and only optimizes a few task-specific prompts. In this paper, we propose a conditional prompt generation method to generate prompts for each input instance, referred to as the Instance-Dependent Prompt Generation (IDPG). Unlike traditional prompt tuning methods that use a fixed prompt, IDPG introduces a lightweight and trainable component to generate prompts based on each input sentence. Extensive experiments on ten natural language understanding (NLU) tasks show that the proposed strategy consistently outperforms various prompt tuning baselines and is on par with other efficient transfer learning methods such as Compacter while tuning far fewer model parameters.

Fang Nan, Hao Ma, Qinghua Guan et al.

We present an online model-based reinforcement learning algorithm suitable for controlling complex robotic systems directly in the real world. Unlike prevailing sim-to-real pipelines that rely on extensive offline simulation and model-free policy optimization, our method builds a dynamics model from real-time interaction data and performs policy updates guided by the learned dynamics model. This efficient model-based reinforcement learning scheme significantly reduces the number of samples to train control policies, enabling direct training on real-world rollout data. This significantly reduces the influence of bias in the simulated data, and facilitates the search for high-performance control policies. We adopt online optimization analysis to derive sublinear regret bounds under stochastic online optimization assumptions, providing formal guarantees on performance improvement as more interaction data are collected. Experimental evaluations were performed on a hydraulic excavator arm and a soft robot arm, where the algorithm demonstrates strong sample efficiency compared to model-free reinforcement learning methods, reaching comparable performance within hours. Robust adaptation to shifting dynamics was also observed when the payload condition was randomized. Our approach paves the way toward efficient and reliable on-robot learning for a broad class of challenging control tasks.

Tengyu Xu, Eryk Helenowski, Karthik Abinav Sankararaman et al.

Reinforcement learning from human feedback (RLHF) has become the leading approach for fine-tuning large language models (LLM). However, RLHF has limitations in multi-task learning (MTL) due to challenges of reward hacking and extreme multi-objective optimization (i.e., trade-off of multiple and/or sometimes conflicting objectives). Applying RLHF for MTL currently requires careful tuning of the weights for reward model and data combinations. This is often done via human intuition and does not generalize. In this work, we introduce a novel post-training paradigm which we called Constrained Generative Policy Optimization (CGPO). The core of CGPO is Mixture of Judges (MoJ) with cost-efficient constrained policy optimization with stratification, which can identify the perfect blend in RLHF in a principled manner. It shows strong empirical results with theoretical guarantees, does not require extensive hyper-parameter tuning, and is plug-and-play in common post-training pipelines. Together, this can detect and mitigate reward hacking behaviors while reaching a pareto-optimal point across an extremely large number of objectives. Our empirical evaluations demonstrate that CGPO significantly outperforms standard RLHF algorithms like PPO and DPO across various tasks including general chat, STEM questions, instruction following, and coding. Specifically, CGPO shows improvements of 7.4% in AlpacaEval-2 (general chat), 12.5% in Arena-Hard (STEM & reasoning), and consistent gains in other domains like math and coding. Notably, PPO, while commonly used, is prone to severe reward hacking in popular coding benchmarks, which CGPO successfully addresses. This breakthrough in RLHF not only tackles reward hacking and extreme multi-objective optimization challenges but also advances the state-of-the-art in aligning general-purpose LLMs for diverse applications.

Jiawei Chen, Zhe Chen, Chaoqun Du et al.

Applications such as embodied intelligence rely on a real-time perception-decision-action closed loop, posing stringent challenges for streaming video understanding. However, current agents suffer from fragmented capabilities, such as supporting only offline video understanding, lacking long-term multimodal memory mechanisms, or struggling to achieve real-time reasoning and proactive interaction under streaming inputs. These shortcomings have become a key bottleneck for preventing them from sustaining perception, making real-time decisions, and executing actions in real-world environments. To alleviate these issues, we propose StreamingClaw, a unified agent framework for streaming video understanding and embodied intelligence. It is also an OpenClaw-compatible framework that supports real-time, multimodal streaming interaction. StreamingClaw integrates five core capabilities: (1) It supports real-time streaming reasoning. (2) It supports reasoning about future events and proactive interaction under the online evolution of interaction objectives. (3) It supports multimodal long-term storage, hierarchical evolution, and efficient retrieval of shared memory across multiple agents. (4) It supports a closed-loop of perception-decision-action. In addition to conventional tools and skills, it also provides streaming tools and action-centric skills tailored for real-world physical environments. (5) It is compatible with the OpenClaw framework, allowing it to fully leverage the resources and support of the open-source community. With these designs, StreamingClaw integrates online real-time reasoning, multimodal long-term memory, and proactive interaction within a unified framework. Moreover, by translating decisions into executable actions, it enables direct control of the physical world, supporting practical deployment of embodied interaction.

Zehua Zhao, Zhixian Huang, Junren Li et al.

Current benchmarks for evaluating the chemical reasoning capabilities of Large Language Models (LLMs) are limited by oversimplified tasks, lack of process-level evaluation, and misalignment with expert-level chemistry skills. To address these issues, we introduce SUPERChem, a benchmark of 500 expert-curated reasoning-intensive chemistry problems, covering diverse subfields and provided in both multimodal and text-only formats. Original content and an iterative curation pipeline eliminate flawed items and mitigate data contamination. Each problem is paired with an expert-authored solution path, enabling Reasoning Path Fidelity (RPF) scoring to evaluate reasoning quality beyond final-answer accuracy. Evaluations against a human baseline of 40.3% accuracy show that even the best-performing model, GPT-5 (High), reaches only 38.5%, followed closely by Gemini 2.5 Pro (37.9%) and DeepSeek-V3.1-Think (37.3%). SUPERChem elicits multi-step, multimodal reasoning, reveals model-dependent effects of visual information, and distinguishes high-fidelity reasoners from heuristic ones. By providing a challenging benchmark and a reliable evaluation framework, SUPERChem aims to facilitate the advancement of LLMs toward expert-level chemical intelligence. The dataset of the benchmark is available at https://huggingface.co/datasets/ZehuaZhao/SUPERChem.

Zhen Liu, Hang Gao, Hao Ma et al.

Autonomous vehicle (AV) evaluation has been the subject of increased interest in recent years both in industry and in academia. This paper focuses on the development of a novel framework for generating adversarial driving behavior of background vehicle interfering against the AV to expose effective and rational risky events. Specifically, the adversarial behavior is learned by a reinforcement learning (RL) approach incorporated with the cumulative prospect theory (CPT) which allows representation of human risk cognition. Then, the extended version of deep deterministic policy gradient (DDPG) technique is proposed for training the adversarial policy while ensuring training stability as the CPT action-value function is leveraged. A comparative case study regarding the cut-in scenario is conducted on a high fidelity Hardware-in-the-Loop (HiL) platform and the results demonstrate the adversarial effectiveness to infer the weakness of the tested AV.

Lihao Zheng, Zhenwei Shao, Yu Zhou et al.

Although Multimodal Large Language Models (MLLMs) have advanced rapidly, they still face notable challenges in fine-grained multi-image understanding, often exhibiting spatial hallucination, attention leakage, and failures in object constancy. In addition, existing approaches typically rely on expensive human annotations or large-scale chain-of-thought (CoT) data generation. We propose Compositional Grounded Contrast (abbr. CGC), a low-cost full framework for boosting fine-grained multi-image understanding of MLLMs. Built on existing single-image grounding annotations, CGC constructs compositional multi-image training instances through Inter-Image Contrast and Intra-Image Contrast, which introduce semantically decoupled distractor contexts for cross-image discrimination and correlated cross-view samples for object constancy, respectively. CGC further introduces a Rule-Based Spatial Reward within the GRPO framework to improve source-image attribution, spatial alignment, and structured output validity under a Think-before-Grounding paradigm. Experiments show that CGC achieves state-of-the-art results on fine-grained multi-image benchmarks, including MIG-Bench and VLM2-Bench. The learned multi-image understanding capability also transfers to broader multimodal understanding and reasoning tasks, yielding consistent gains over the Qwen3-VL-8B base model on MathVista (+2.90), MuirBench (+2.88), MMStar (+1.93), MMMU (+1.77), and BLINK (+1.69).

Hao Ma, Zhiqiang Pu, Xiaolin Ai et al.

We present GuidedSAC, a novel reinforcement learning (RL) algorithm that facilitates efficient exploration in vast state-action spaces. GuidedSAC leverages large language models (LLMs) as intelligent supervisors that provide action-level guidance for the Soft Actor-Critic (SAC) algorithm. The LLM-based supervisor analyzes the most recent trajectory using state information and visual replays, offering action-level interventions that enable targeted exploration. Furthermore, we provide a theoretical analysis of GuidedSAC, proving that it preserves the convergence guarantees of SAC while improving convergence speed. Through experiments in both discrete and continuous control environments, including toy text tasks and complex MuJoCo benchmarks, we demonstrate that GuidedSAC consistently outperforms standard SAC and state-of-the-art exploration-enhanced variants (e.g., RND, ICM, and E3B) in terms of sample efficiency and final performance.

Keke Liu, Hao Ma, Zemin Wang

Scan undistortion is a key module for LiDAR odometry in high dynamic environment with high rotation and translation speed. The existing line of studies mostly focuses on one pass undistortion, which means undistortion for each point is conducted only once in the whole LiDAR-IMU odometry pipeline. In this paper, we propose an optimization based tightly coupled LiDAR-IMU odometry addressing iterated point-level undistortion. By jointly minimizing the cost derived from LiDAR and IMU measurements, our LiDAR-IMU odometry method performs more accurate and robust in high dynamic environment. Besides, the method characters good computation efficiency by limiting the quantity of parameters.

Hao Ma, Zhiqiang Pu, Yang Liu et al.

Constraints are essential for stabilizing reinforcement learning fine-tuning (RFT) and preventing degenerate outputs, yet they inherently conflict with the optimization objective because stronger constraints limit the ability of a fine-tuned model to discover better solutions. We propose \textit{dynamic constraints} that resolve this tension by adapting to the evolving capabilities of the fine-tuned model based on the insight that constraints should only intervene when degenerate outputs occur. We implement this by using a reference model as an \textit{online refiner} that takes the response from the fine-tuned model and generates a minimally corrected version which preserves correct content verbatim while fixing errors. A supervised fine-tuning loss then trains the fine-tuned model to produce the refined output. This mechanism yields a constraint that automatically strengthens or relaxes based on output quality. Experiments on dialogue and code generation show that dynamic constraints outperform both KL regularization and unconstrained baselines, achieving substantially higher task rewards while maintaining training stability.

Lihao Zheng, Jiawei Chen, Xintian Shen et al.

Multi-image reasoning and grounding require understanding complex cross-image relationships at both object levels and image levels. Current Large Visual Language Models (LVLMs) face two critical challenges: the lack of cross-image reasoning capabilities and insufficient cross-image reference reward modeling. To address these issues, we propose a unified framework - Multi-Image Reasoning and Grounding with Reinforcement Learning (MIRG-RL). Specifically, our two-stage training paradigm combines supervised fine-tuning with annotated trajectories and image-aware reinforcement learning optimization, progressively developing multi-image reasoning capabilities. Furthermore, we innovatively propose a method for constructing the trajectory data, which integrates object-level and image-level annotation information, and use this method to generate a lightweight reasoning-enhanced dataset. To effectively resolve cross-image ambiguities, we design an image-aware RL policy with dual reward functions for objects and images. Experiments demonstrate that MIRG-RL achieves state-of-the-art (SOTA) performance in multi-image grounding benchmarks, attaining 64.82% on cross-image reasoning tasks - exceeding the previous best method by 1%. The code and dataset have been released at https://github.com/ZEUS2035/MIRG-RL.

Jinghang Han, Jiawei Chen, Hang Shao et al.

Reinforcement learning has significantly enhanced the reasoning capabilities of Large Language Models (LLMs) in complex problem-solving tasks. Recently, the introduction of DeepSeek R1 has inspired a surge of interest in leveraging rule-based rewards as a low-cost alternative for computing advantage functions and guiding policy optimization. However, a common challenge observed across many replication and extension efforts is that when multiple sampled responses under a single prompt converge to identical outcomes, whether correct or incorrect, the group-based advantage degenerates to zero. This leads to vanishing gradients and renders the corresponding samples ineffective for learning, ultimately limiting training efficiency and downstream performance. To address this issue, we propose a consistency-aware policy optimization framework that introduces a structured global reward based on outcome consistency, the global loss based on it ensures that, even when model outputs show high intra-group consistency, the training process still receives meaningful learning signals, which encourages the generation of correct and self-consistent reasoning paths from a global perspective. Furthermore, we incorporate an entropy-based soft blending mechanism that adaptively balances local advantage estimation with global optimization, enabling dynamic transitions between exploration and convergence throughout training. Our method introduces several key innovations in both reward design and optimization strategy. We validate its effectiveness through substantial performance gains on multiple mathematical reasoning benchmarks, highlighting the proposed framework's robustness and general applicability. Code of this work has been released at https://github.com/hijih/copo-code.git.

Tianci Bu, Chuanrui Wang, Hao Ma et al.

Generating graphs with hierarchical structures remains a fundamental challenge due to the limitations of Euclidean geometry in capturing exponential complexity. Here we introduce \textbf{GGBall}, a novel hyperbolic framework for graph generation that integrates geometric inductive biases with modern generative paradigms. GGBall combines a Hyperbolic Vector-Quantized Autoencoder (HVQVAE) with a Riemannian flow matching prior defined via closed-form geodesics. This design enables flow-based priors to model complex latent distributions, while vector quantization helps preserve the curvature-aware structure of the hyperbolic space. We further develop a suite of hyperbolic GNN and Transformer layers that operate entirely within the manifold, ensuring stability and scalability. Empirically, our model reduces degree MMD by over 75\% on Community-Small and over 40\% on Ego-Small compared to state-of-the-art baselines, demonstrating an improved ability to preserve topological hierarchies. These results highlight the potential of hyperbolic geometry as a powerful foundation for the generative modeling of complex, structured, and hierarchical data domains. Our code is available at \href{https://github.com/AI4Science-WestlakeU/GGBall}{here}.

Jia-Qi Yang, Ke-Bin Fan, Hao Ma et al.

From an engineering perspective, a design should not only perform well in an ideal condition, but should also resist noises. Such a design methodology, namely robust design, has been widely implemented in the industry for product quality control. However, classic robust design requires a lot of evaluations for a single design target, while the results of these evaluations could not be reused for a new target. To achieve a data-efficient robust design, we propose Robust Inverse Design under Noise (RID-Noise), which can utilize existing noisy data to train a conditional invertible neural network (cINN). Specifically, we estimate the robustness of a design parameter by its predictability, measured by the prediction error of a forward neural network. We also define a sample-wise weight, which can be used in the maximum weighted likelihood estimation of an inverse model based on a cINN. With the visual results from experiments, we clearly justify how RID-Noise works by learning the distribution and robustness from data. Further experiments on several real-world benchmark tasks with noises confirm that our method is more effective than other state-of-the-art inverse design methods. Code and supplementary is publicly available at https://github.com/ThyrixYang/rid-noise-aaai22

Wei Song, Jiameng Ying, Sihao Shen et al.

Memory safety remains a critical and widely violated property in reality. Numerous defense techniques have been proposed and developed but most of them are not applied or enabled by default in production-ready environment due to their substantial running cost. The situation might change in the near future because the hardware supported defenses against these attacks are finally beginning to be adopted by commercial processors, operating systems and compilers. We then face a question as there is currently no suitable test suite to measure the memory safety extensions supported on different processors. In fact, the issue is not constrained only for memory safety but all aspect of processor security. All of the existing test suites related to processor security lack some of the key properties, such as comprehensiveness, distinguishability and portability. As an initial step, we propose an expandable test framework for measuring the processor security and open source a memory safety test suite utilizing this framework. The framework is deliberately designed to be flexible so it can be gradually extended to all types of hardware supported security extensions in processors. The initial test suite for memory safety currently contains 160 test cases covering spatial and temporal safety of memory, memory access control, pointer integrity and control-flow integrity. Each type of vulnerabilities and their related defenses have been individually evaluated by one or more test cases. The test suite has been ported to three different instruction set architectures (ISAs) and experimented on six different platforms. We have also utilized the test suite to explore the security benefits of applying different sets of compiler flags available on the latest GNU GCC and LLVM compilers.

Jiezhong Qiu, Yuxiao Dong, Hao Ma et al.

We study the problem of large-scale network embedding, which aims to learn latent representations for network mining applications. Previous research shows that 1) popular network embedding benchmarks, such as DeepWalk, are in essence implicitly factorizing a matrix with a closed form, and 2)the explicit factorization of such matrix generates more powerful embeddings than existing methods. However, directly constructing and factorizing this matrix---which is dense---is prohibitively expensive in terms of both time and space, making it not scalable for large networks. In this work, we present the algorithm of large-scale network embedding as sparse matrix factorization (NetSMF). NetSMF leverages theories from spectral sparsification to efficiently sparsify the aforementioned dense matrix, enabling significantly improved efficiency in embedding learning. The sparsified matrix is spectrally close to the original dense one with a theoretically bounded approximation error, which helps maintain the representation power of the learned embeddings. We conduct experiments on networks of various scales and types. Results show that among both popular benchmarks and factorization based methods, NetSMF is the only method that achieves both high efficiency and effectiveness. We show that NetSMF requires only 24 hours to generate effective embeddings for a large-scale academic collaboration network with tens of millions of nodes, while it would cost DeepWalk months and is computationally infeasible for the dense matrix factorization solution. The source code of NetSMF is publicly available (https://github.com/xptree/NetSMF).

Jiawei Chen, Xintian Shen, Lihao Zheng et al.

Traditional workflow-based agents exhibit limited intelligence when addressing real-world problems requiring tool invocation. Tool-integrated reasoning (TIR) agents capable of autonomous reasoning and tool invocation are rapidly emerging as a powerful approach for complex decision-making tasks involving multi-step interactions with external environments. In this work, we introduce MindWatcher, a TIR agent integrating interleaved thinking and multimodal chain-of-thought (CoT) reasoning. MindWatcher can autonomously decide whether and how to invoke diverse tools and coordinate their use, without relying on human prompts or workflows. The interleaved thinking paradigm enables the model to switch between thinking and tool calling at any intermediate stage, while its multimodal CoT capability allows manipulation of images during reasoning to yield more precise search results. We implement automated data auditing and evaluation pipelines, complemented by manually curated high-quality datasets for training, and we construct a benchmark, called MindWatcher-Evaluate Bench (MWE-Bench), to evaluate its performance. MindWatcher is equipped with a comprehensive suite of auxiliary reasoning tools, enabling it to address broad-domain multimodal problems. A large-scale, high-quality local image retrieval database, covering eight categories including cars, animals, and plants, endows model with robust object recognition despite its small size. Finally, we design a more efficient training infrastructure for MindWatcher, enhancing training speed and hardware utilization. Experiments not only demonstrate that MindWatcher matches or exceeds the performance of larger or more recent models through superior tool invocation, but also uncover critical insights for agent training, such as the genetic inheritance phenomenon in agentic RL.

Yun He, Di Jin, Chaoqi Wang et al.

Large Language Models (LLMs) have demonstrated impressive capabilities in various tasks, including instruction following, which is crucial for aligning model outputs with user expectations. However, evaluating LLMs' ability to follow instructions remains challenging due to the complexity and subjectivity of human language. Current benchmarks primarily focus on single-turn, monolingual instructions, which do not adequately reflect the complexities of real-world applications that require handling multi-turn and multilingual interactions. To address this gap, we introduce Multi-IF, a new benchmark designed to assess LLMs' proficiency in following multi-turn and multilingual instructions. Multi-IF, which utilizes a hybrid framework combining LLM and human annotators, expands upon the IFEval by incorporating multi-turn sequences and translating the English prompts into another 7 languages, resulting in a dataset of 4,501 multilingual conversations, where each has three turns. Our evaluation of 14 state-of-the-art LLMs on Multi-IF reveals that it presents a significantly more challenging task than existing benchmarks. All the models tested showed a higher rate of failure in executing instructions correctly with each additional turn. For example, o1-preview drops from 0.877 at the first turn to 0.707 at the third turn in terms of average accuracy over all languages. Moreover, languages with non-Latin scripts (Hindi, Russian, and Chinese) generally exhibit higher error rates, suggesting potential limitations in the models' multilingual capabilities. We release Multi-IF prompts and the evaluation code base to encourage further research in this critical area.

Hao Ma, Zhiyuan Peng, Mingjie Shao et al.

Target-speaker automatic speech recognition (ASR) aims to transcribe the desired speech of a target speaker from multi-talker overlapped utterances. Most of the existing target-speaker ASR (TS-ASR) methods involve either training from scratch or fully fine-tuning a pre-trained model, leading to significant training costs and becoming inapplicable to large foundation models. This work leverages prompt tuning, a parameter-efficient fine-tuning approach, to extend Whisper, a large-scale single-talker ASR model, to TS-ASR. Variants of prompt tuning approaches along with their configurations are explored and optimized for TS-ASR.Experimental results show that prompt tuning can achieve performance comparable to state-of-the-art full training approaches while only requiring about 1\% of task-specific model parameters. Notably, the original Whisper's features, such as inverse text normalization and timestamp tagging, are retained in target-speaker ASR, keeping the generated transcriptions natural and informative.

Xintian Shen, Jiawei Chen, Lihao Zheng et al.

Existing Tool-Integrated Reasoning (TIR) models have effectively extended the question-answering capabilities of LLMs by incorporating external tools. However, real-world scenarios present numerous open-ended problems where fixed tools often fail to meet task requirements. Furthermore, the lack of self-optimization mechanisms means that erroneous tool outputs can mislead the LLM's responses. Additionally, the construction of existing tools entails significant manual effort, which consequently constrains their applicability. Recognizing that the reasoning traces of LLMs encapsulate implicit problem-solving capabilities, we propose UCT, a novel training-free framework that transforms agents from tool users to tool creators. This approach harvests reasoning experiences and distills them into reusable assets. This method transforms the agent from a mere tool user into a tool creator, enabling adaptive tool creation and self-updating during the inference process. We also introduce a memory consolidation mechanism to maintain the tool library, ensuring high reusability of retained experiential memory for subsequent reasoning tasks. This novel automated tool construction paradigm continuously improves tool quality during reasoning, allowing the overall agent system to progress without additional training. Extensive experiments demonstrate that our method serves as a novel paradigm for enhancing the capabilities of TIR models. In particular, the significant performance gains achieved +20.86%$\uparrow$ and +23.04%$\uparrow$ on benchmarks across multi-domain mathematical and scientific reasoning tasks validate the self-evolving capability of the agent.

Zishun Yu, Tengyu Xu, Di Jin et al.

Solving mathematics problems has been an intriguing capability of large language models, and many efforts have been made to improve reasoning by extending reasoning length, such as through self-correction and extensive long chain-of-thoughts. While promising in problem-solving, advanced long reasoning chain models exhibit an undesired single-modal behavior, where trivial questions require unnecessarily tedious long chains of thought. In this work, we propose a way to allow models to be aware of inference budgets by formulating it as utility maximization with respect to an inference budget constraint, hence naming our algorithm Inference Budget-Constrained Policy Optimization (IBPO). In a nutshell, models fine-tuned through IBPO learn to ``understand'' the difficulty of queries and allocate inference budgets to harder ones. With different inference budgets, our best models are able to have a $4.14$\% and $5.74$\% absolute improvement ($8.08$\% and $11.2$\% relative improvement) on MATH500 using $2.16$x and $4.32$x inference budgets respectively, relative to LLaMA3.1 8B Instruct. These improvements are approximately $2$x those of self-consistency under the same budgets.

Chaoqi Wang, Zhuokai Zhao, Yibo Jiang et al.

Recent advances in large language models (LLMs) have demonstrated significant progress in performing complex tasks. While Reinforcement Learning from Human Feedback (RLHF) has been effective in aligning LLMs with human preferences, it is susceptible to spurious correlations in reward modeling. Consequently, it often introduces biases-such as length bias, sycophancy, conceptual bias, and discrimination-that hinder the model's ability to capture true causal relationships. To address this, we propose a novel causal reward modeling approach that integrates causality to mitigate these spurious correlations. Our method enforces counterfactual invariance, ensuring reward predictions remain consistent when irrelevant variables are altered. Through experiments on both synthetic and real-world datasets, we show that our approach mitigates various types of spurious correlations effectively, resulting in more reliable and fair alignment of LLMs with human preferences. As a drop-in enhancement to the existing RLHF workflow, our causal reward modeling provides a practical way to improve the trustworthiness and fairness of LLM finetuning.

Jiada Huang, Hao Ma, Zhibin Shen et al.

Local high strain in solid rocket motor grains is a primary cause of structural failure. However, traditional numerical simulations are computationally expensive, and existing surrogate models cannot explicitly establish geometric models and accurately capture high-strain regions. Therefore, this paper proposes an adaptive graph network, GrainGNet, which employs an adaptive pooling dynamic node selection mechanism to effectively preserve the key mechanical features of structurally critical regions, while concurrently utilising feature fusion to transmit deep features and enhance the model's representational capacity. In the joint prediction task involving four sequential conditions--curing and cooling, storage, overloading, and ignition--GrainGNet reduces the mean squared error by 62.8% compared to the baseline graph U-Net model, with only a 5.2% increase in parameter count and an approximately sevenfold improvement in training efficiency. Furthermore, in the high-strain regions of debonding seams, the prediction error is further reduced by 33% compared to the second-best method, offering a computationally efficient and high-fidelity approach to evaluate motor structural safety.

Chaoqi Wang, Zhuokai Zhao, Chen Zhu et al.

Recent advancements in generative models, particularly large language models (LLMs) and diffusion models, have been driven by extensive pretraining on large datasets followed by post-training. However, current post-training methods such as reinforcement learning from human feedback (RLHF) and direct alignment from preference methods (DAP) primarily utilize single-sample comparisons. These approaches often fail to capture critical characteristics such as generative diversity and bias, which are more accurately assessed through multiple samples. To address these limitations, we introduce a novel approach that extends post-training to include multi-sample comparisons. To achieve this, we propose Multi-sample Direct Preference Optimization (mDPO) and Multi-sample Identity Preference Optimization (mIPO). These methods improve traditional DAP methods by focusing on group-wise characteristics. Empirically, we demonstrate that multi-sample comparison is more effective in optimizing collective characteristics~(e.g., diversity and bias) for generative models than single-sample comparison. Additionally, our findings suggest that multi-sample comparisons provide a more robust optimization framework, particularly for dataset with label noise.

Yen-Ting Lin, Di Jin, Tengyu Xu et al.

Large language models (LLMs) have recently demonstrated remarkable success in mathematical reasoning. Despite progress in methods like chain-of-thought prompting and self-consistency sampling, these advances often focus on final correctness without ensuring that the underlying reasoning process is coherent and reliable. This paper introduces Step-KTO, a training framework that combines process-level and outcome-level binary feedback to guide LLMs toward more trustworthy reasoning trajectories. By providing binary evaluations for both the intermediate reasoning steps and the final answer, Step-KTO encourages the model to adhere to logical progressions rather than relying on superficial shortcuts. Our experiments on challenging mathematical benchmarks show that Step-KTO significantly improves both final answer accuracy and the quality of intermediate reasoning steps. For example, on the MATH-500 dataset, Step-KTO achieves a notable improvement in Pass@1 accuracy over strong baselines. These results highlight the promise of integrating stepwise process feedback into LLM training, paving the way toward more interpretable and dependable reasoning capabilities.

Hao Ma, Melanie Zeilinger, Michael Muehlebach

We propose a novel gradient-based online optimization framework for solving stochastic programming problems that frequently arise in the context of cyber-physical and robotic systems. Our problem formulation accommodates constraints that model the evolution of a cyber-physical system, which has, in general, a continuous state and action space, is nonlinear, and where the state is only partially observed. We also incorporate an approximate model of the dynamics as prior knowledge into the learning process and show that even rough estimates of the dynamics can significantly improve the convergence of our algorithms. Our online optimization framework encompasses both gradient descent and quasi-Newton methods, and we provide a unified convergence analysis of our algorithms in a non-convex setting. We also characterize the impact of modeling errors in the system dynamics on the convergence rate of the algorithms. Finally, we evaluate our algorithms in simulations of a flexible beam, a four-legged walking robot, and in real-world experiments with a ping-pong playing robot.

Jiawei Chen, Xintian Shen, Lihao Zheng et al.

Integrating web search tools has significantly extended the capability of LLMs to address open-world, real-time, and long-tail problems. However, evaluating these Search Agents presents formidable challenges. First, constructing high-quality deep search benchmarks is prohibitively expensive, while unverified synthetic data often suffers from unreliable sources. Second, static benchmarks face dynamic obsolescence: as internet information evolves, complex queries requiring deep research often degrade into simple retrieval tasks due to increased popularity, and ground truths become outdated due to temporal shifts. Third, attribution ambiguity confounds evaluation, as an agent's performance is often dominated by its parametric memory rather than its actual search and reasoning capabilities. Finally, reliance on specific commercial search engines introduces variability that hampers reproducibility. To address these issues, we propose a novel framework, Mind-ParaWorld, for evaluating Search Agents in a Parallel World. Specifically, MPW samples real-world entity names to synthesize future scenarios and questions situated beyond the model's knowledge cutoff. A ParaWorld Law Model then constructs a set of indivisible Atomic Facts and a unique ground-truth for each question. During evaluation, instead of retrieving real-world results, the agent interacts with a ParaWorld Engine Model that dynamically generates SERPs grounded in these inviolable Atomic Facts. We release MPW-Bench, an interactive benchmark spanning 19 domains with 1,608 instances. Experiments across three evaluation settings show that, while search agents are strong at evidence synthesis given complete information, their performance is limited not only by evidence collection and coverage in unfamiliar search environments, but also by unreliable evidence sufficiency judgment and when-to-stop decisions-bottlenecks.

Yuxin Yang, Haoyang Wu, Tao Wang et al.

The advent of Large Language Models (LLMs) has revolutionized natural language processing. However, these models face challenges in retrieving precise information from vast datasets. Retrieval-Augmented Generation (RAG) was developed to combining LLMs with external information retrieval systems to enhance the accuracy and context of responses. Despite improvements, RAG still struggles with comprehensive retrieval in high-volume, low-information-density databases and lacks relational awareness, leading to fragmented answers. To address this, this paper introduces the Pseudo-Knowledge Graph (PKG) framework, designed to overcome these limitations by integrating Meta-path Retrieval, In-graph Text and Vector Retrieval into LLMs. By preserving natural language text and leveraging various retrieval techniques, the PKG offers a richer knowledge representation and improves accuracy in information retrieval. Extensive evaluations using Open Compass and MultiHop-RAG benchmarks demonstrate the framework's effectiveness in managing large volumes of data and complex relationships.

Hao Ma, Melis Ilayda Bal, Liang Zhang et al.

Modern large language models are increasingly deployed under compute and memory constraints, making flexible control of model capacity a central challenge. While sparse and low-rank structures naturally trade off capacity and performance, existing approaches often rely on heuristic designs that ignore layer and matrix heterogeneity or require model-specific architectural modifications. We propose SALAAD, a plug-and-play framework applicable to different model architectures that induces sparse and low-rank structures during training. By formulating structured weight learning under an augmented Lagrangian framework and introducing an adaptive controller that dynamically balances the training loss and structural constraints, SALAAD preserves the stability of standard training dynamics while enabling explicit control over the evolution of effective model capacity during training. Experiments across model scales show that SALAAD substantially reduces memory consumption during deployment while achieving performance comparable to ad-hoc methods. Moreover, a single training run yields a continuous spectrum of model capacities, enabling smooth and elastic deployment across diverse memory budgets without the need for retraining.

Siyao Zhao, Hao Ma, Zhiqiang Pu et al.

Creating offensive advantages during open play is fundamental to football success. However, due to the highly dynamic and long-sequence nature of open play, the potential tactic space grows exponentially as the sequence progresses, making automated tactic discovery extremely challenging. To address this, we propose TacEleven, a generative framework for football open-play tactic discovery developed in close collaboration with domain experts from AJ Auxerre, designed to assist coaches and analysts in tactical decision-making. TacEleven consists of two core components: a language-controlled tactical generator that produces diverse tactical proposals, and a multimodal large language model-based tactical critic that selects the optimal proposal aligned with a high-level stylistic tactical instruction. The two components enables rapid exploration of tactical proposals and discovery of alternative open-play offensive tactics. We evaluate TacEleven across three tasks with progressive tactical complexity: counterfactual exploration, single-step discovery, and multi-step discovery, through both quantitative metrics and a questionnaire-based qualitative assessment. The results show that the TacEleven-discovered tactics exhibit strong realism and tactical creativity, with 52.50% of the multi-step tactical alternatives rated adoptable in real-world elite football scenarios, highlighting the framework's ability to rapidly generate numerous high-quality tactics for complex long-sequence open-play situations. TacEleven demonstrates the potential of creatively leveraging domain data and generative models to advance tactical analysis in sports.

Yinqiu Huang, Hao Ma, Wenshuai Chen et al.

Bid shading plays a crucial role in Real-Time Bidding~(RTB) by adaptively adjusting the bid to avoid advertisers overspending. Existing mainstream two-stage methods, which first model bid landscapes and then optimize surplus using operations research techniques, are constrained by unimodal assumptions that fail to adapt for non-convex surplus curves and are vulnerable to cascading errors in sequential workflows. Additionally, existing discretization models of continuous values ignore the dependence between discrete intervals, reducing the model's error correction ability, while sample selection bias in bidding scenarios presents further challenges for prediction. To address these issues, this paper introduces Generative Bid Shading~(GBS), which comprises two primary components: (1) an end-to-end generative model that utilizes an autoregressive approach to generate shading ratios by stepwise residuals, capturing complex value dependencies without relying on predefined priors; and (2) a reward preference alignment system, which incorporates a channel-aware hierarchical dynamic network~(CHNet) as the reward model to extract fine-grained features, along with modules for surplus optimization and exploration utility reward alignment, ultimately optimizing both short-term and long-term surplus using group relative policy optimization~(GRPO). Extensive experiments on both offline and online A/B tests validate GBS's effectiveness. Moreover, GBS has been deployed on the Meituan DSP platform, serving billions of bid requests daily.

Chengwei Qin, Wenxuan Zhou, Karthik Abinav Sankararaman et al.

Hallucination, the generation of factually incorrect information, remains a significant challenge for large language models (LLMs), especially in open-domain long-form generation. Existing approaches for detecting hallucination in long-form tasks either focus on limited domains or rely heavily on external fact-checking tools, which may not always be available. In this work, we systematically investigate reference-free hallucination detection in open-domain long-form responses. Our findings reveal that internal states (e.g., model's output probability and entropy) alone are insufficient for reliably (i.e., better than random guessing) distinguishing between factual and hallucinated content. To enhance detection, we explore various existing approaches, including prompting-based methods, probing, and fine-tuning, with fine-tuning proving the most effective. To further improve the accuracy, we introduce a new paradigm, named RATE-FT, that augments fine-tuning with an auxiliary task for the model to jointly learn with the main task of hallucination detection. With extensive experiments and analysis using a variety of model families & datasets, we demonstrate the effectiveness and generalizability of our method, e.g., +3% over general fine-tuning methods on LongFact.

Mingxu Tao, Bowen Tang, Mingxuan Ma et al.

The rise of Large Language Models~(LLMs) revolutionizes information retrieval, allowing users to obtain required answers through complex instructions within conversations. However, publicly available services remain inadequate in addressing the needs of faculty and students to search campus-specific information. It is primarily due to the LLM's lack of domain-specific knowledge and the limitation of search engines in supporting multilingual and timely scenarios. To tackle these challenges, we introduce ALOHA, a multilingual agent enhanced by hierarchical retrieval for university orientation. We also integrate external APIs into the front-end interface to provide interactive service. The human evaluation and case study show our proposed system has strong capabilities to yield correct, timely, and user-friendly responses to the queries in multiple languages, surpassing commercial chatbots and search engines. The system has been deployed and has provided service for more than 12,000 people.

Hao Ma, Zhiqiang Pu, Shijie Wang et al.

Trajectory prediction facilitates effective planning and decision-making, while constrained trajectory prediction integrates regulation into prediction. Recent advances in constrained trajectory prediction focus on structured constraints by constructing optimization objectives. However, handling unstructured constraints is challenging due to the lack of differentiable formal definitions. To address this, we propose a novel method for constrained trajectory prediction using a conditional generative paradigm, named Controllable Trajectory Diffusion (CTD). The key idea is that any trajectory corresponds to a degree of conformity to a constraint. By quantifying this degree and treating it as a condition, a model can implicitly learn to predict trajectories under unstructured constraints. CTD employs a pre-trained scoring model to predict the degree of conformity (i.e., a score), and uses this score as a condition for a conditional diffusion model to generate trajectories. Experimental results demonstrate that CTD achieves high accuracy on the ETH/UCY and SDD benchmarks. Qualitative analysis confirms that CTD ensures adherence to unstructured constraints and can predict trajectories that satisfy combinatorial constraints.

Hao Ma, Zhiqiang Pu, Yi Pan et al. · meta-ai, microsoft-research

Scalability remains a challenge in multi-agent reinforcement learning and is currently under active research. A framework named mean-field reinforcement learning (MFRL) could alleviate the scalability problem by employing the Mean Field Theory to turn a many-agent problem into a two-agent problem. However, this framework lacks the ability to identify essential interactions under nonstationary environments. Causality contains relatively invariant mechanisms behind interactions, though environments are nonstationary. Therefore, we propose an algorithm called causal mean-field Q-learning (CMFQ) to address the scalability problem. CMFQ is ever more robust toward the change of the number of agents though inheriting the compressed representation of MFRL's action-state space. Firstly, we model the causality behind the decision-making process of MFRL into a structural causal model (SCM). Then the essential degree of each interaction is quantified via intervening on the SCM. Furthermore, we design the causality-aware compact representation for behavioral information of agents as the weighted sum of all behavioral information according to their causal effects. We test CMFQ in a mixed cooperative-competitive game and a cooperative game. The result shows that our method has excellent scalability performance in both training in environments containing a large number of agents and testing in environments containing much more agents.

Hao Ma, Shijie Wang, Zhiqiang Pu et al.

Guiding the policy of multi-agent reinforcement learning to align with human common sense is a difficult problem, largely due to the complexity of modeling common sense as a reward, especially in complex and long-horizon multi-agent tasks. Recent works have shown the effectiveness of reward shaping, such as potential-based rewards, to enhance policy alignment. The existing works, however, primarily rely on experts to design rule-based rewards, which are often labor-intensive and lack a high-level semantic understanding of common sense. To solve this problem, we propose a hierarchical vision-based reward shaping method. At the bottom layer, a visual-language model (VLM) serves as a generic potential function, guiding the policy to align with human common sense through its intrinsic semantic understanding. To help the policy adapts to uncertainty and changes in long-horizon tasks, the top layer features an adaptive skill selection module based on a visual large language model (vLLM). The module uses instructions, video replays, and training records to dynamically select suitable potential function from a pre-designed pool. Besides, our method is theoretically proven to preserve the optimal policy. Extensive experiments conducted in the Google Research Football environment demonstrate that our method not only achieves a higher win rate but also effectively aligns the policy with human common sense.

Yudi Ruan, Hao Ma, Weikai Li et al.

Low-light image enhancement (LLIE) is critical in computer vision. Existing LLIE methods often fail to discover the underlying relationships between different sub-components, causing the loss of complementary information between multiple modules and network layers, ultimately resulting in the loss of image details. To beat this shortage, we design a hierarchical mutual Enhancement via a Cross Attention transformer (ECAFormer), which introduces an architecture that enables concurrent propagation and interaction of multiple features. The model preserves detailed information by introducing a Dual Multi-head self-attention (DMSA), which leverages visual and semantic features across different scales, allowing them to guide and complement each other. Besides, a Cross-Scale DMSA block is introduced to capture the residual connection, integrating cross-layer information to further enhance image detail. Experimental results show that ECAFormer reaches competitive performance across multiple benchmarks, yielding nearly a 3% improvement in PSNR over the suboptimal method, demonstrating the effectiveness of information interaction in LLIE.

Jan Achterhold, Philip Tobuschat, Hao Ma et al.

In this paper, we present a method for table tennis ball trajectory filtering and prediction. Our gray-box approach builds on a physical model. At the same time, we use data to learn parameters of the dynamics model, of an extended Kalman filter, and of a neural model that infers the ball's initial condition. We demonstrate superior prediction performance of our approach over two black-box approaches, which are not supplied with physical prior knowledge. We demonstrate that initializing the spin from parameters of the ball launcher using a neural network drastically improves long-time prediction performance over estimating the spin purely from measured ball positions. An accurate prediction of the ball trajectory is crucial for successful returns. We therefore evaluate the return performance with a pneumatic artificial muscular robot and achieve a return rate of 29/30 (97.7%).

Hao Ma, Yang Sun, Mario Chiarelli

The layout optimization of the heat conduction is essential during design in engineering, especially for thermal sensible products. When the optimization algorithm iteratively evaluates different loading cases, the traditional numerical simulation methods used usually lead to a substantial computational cost. To effectively reduce the computational effort, data-driven approaches are used to train a surrogate model as a mapping between the prescribed external loads and various geometry. However, the existing model are trained by data-driven methods which requires intensive training samples that from numerical simulations and not really effectively solve the problem. Choosing the steady heat conduction problems as examples, this paper proposes a Physics-driven Convolutional Neural Networks (PD-CNN) method to infer the physical field solutions for random varied loading cases. After that, the Particle Swarm Optimization (PSO) algorithm is used to optimize the sizes and the positions of the hole masks in the prescribed design domain, and the average temperature value of the entire heat conduction field is minimized, and the goal of minimizing heat transfer is achieved. Compared with the existing data-driven approaches, the proposed PD-CNN optimization framework not only predict field solutions that are highly consistent with conventional simulation results, but also generate the solution space with without any pre-obtained training data.

Darsh J Shah, Sinong Wang, Han Fang et al.

The ubiquity of offensive and hateful content on online fora necessitates the need for automatic solutions that detect such content competently across target groups. In this paper we show that text classification models trained on large publicly available datasets despite having a high overall performance, may significantly under-perform on several protected groups. On the \citet{vidgen2020learning} dataset, we find the accuracy to be 37\% lower on an under annotated Black Women target group and 12\% lower on Immigrants, where hate speech involves a distinct style. To address this, we propose to perform token-level hate sense disambiguation, and utilize tokens' hate sense representations for detection, modeling more general signals. On two publicly available datasets, we observe that the variance in model accuracy across target groups drops by at least 30\%, improving the average target group performance by 4\% and worst case performance by 13\%.

Shiyu Li, Hao Ma, Xiangyu Hu

Image beauty assessment is an important subject of computer vision. Therefore, building a model to mimic the image beauty assessment becomes an important task. To better imitate the behaviours of the human visual system (HVS), a complete survey about images of different categories should be implemented. This work focuses on image beauty assessment. In this study, the pairwise evaluation method was used, which is based on the Bradley-Terry model. We believe that this method is more accurate than other image rating methods within an image group. Additionally, Convolution neural network (CNN), which is fit for image quality assessment, is used in this work. The first part of this study is a survey about the image beauty comparison of different images. The Bradley-Terry model is used for the calculated scores, which are the target of CNN model. The second part of this work focuses on the results of the image beauty prediction, including landscape images, architecture images and portrait images. The models are pretrained by the AVA dataset to improve the performance later. Then, the CNN model is trained with the surveyed images and corresponding scores. Furthermore, this work compares the results of four CNN base networks, i.e., Alex net, VGG net, Squeeze net and LSiM net, as discussed in literature. In the end, the model is evaluated by the accuracy in pairs, correlation coefficient and relative error calculated by survey results. Satisfactory results are achieved by our proposed methods with about 70 percent accuracy in pairs. Our work sheds more light on the novel image beauty assessment method. While more studies should be conducted, this method is a promising step.

Yuning Mao, Lambert Mathias, Rui Hou et al.

Recent parameter-efficient language model tuning (PELT) methods manage to match the performance of fine-tuning with much fewer trainable parameters and perform especially well when training data is limited. However, different PELT methods may perform rather differently on the same task, making it nontrivial to select the most appropriate method for a specific task, especially considering the fast-growing number of new PELT methods and tasks. In light of model diversity and the difficulty of model selection, we propose a unified framework, UniPELT, which incorporates different PELT methods as submodules and learns to activate the ones that best suit the current data or task setup via gating mechanism. On the GLUE benchmark, UniPELT consistently achieves 1~4% gains compared to the best individual PELT method that it incorporates and even outperforms fine-tuning under different setups. Moreover, UniPELT generally surpasses the upper bound that takes the best performance of all its submodules used individually on each task, indicating that a mixture of multiple PELT methods may be inherently more effective than single methods.

Xuezhe Ma, Xiang Kong, Sinong Wang et al.

The quadratic computational and memory complexities of the Transformer's attention mechanism have limited its scalability for modeling long sequences. In this paper, we propose Luna, a linear unified nested attention mechanism that approximates softmax attention with two nested linear attention functions, yielding only linear (as opposed to quadratic) time and space complexity. Specifically, with the first attention function, Luna packs the input sequence into a sequence of fixed length. Then, the packed sequence is unpacked using the second attention function. As compared to a more traditional attention mechanism, Luna introduces an additional sequence with a fixed length as input and an additional corresponding output, which allows Luna to perform attention operation linearly, while also storing adequate contextual information. We perform extensive evaluations on three benchmarks of sequence modeling tasks: long-context sequence modeling, neural machine translation and masked language modeling for large-scale pretraining. Competitive or even better experimental results demonstrate both the effectiveness and efficiency of Luna compared to a variety

Sinong Wang, Han Fang, Madian Khabsa et al.

Large pre-trained language models (LMs) have demonstrated remarkable ability as few-shot learners. However, their success hinges largely on scaling model parameters to a degree that makes it challenging to train and serve. In this paper, we propose a new approach, named as EFL, that can turn small LMs into better few-shot learners. The key idea of this approach is to reformulate potential NLP task into an entailment one, and then fine-tune the model with as little as 8 examples. We further demonstrate our proposed method can be: (i) naturally combined with an unsupervised contrastive learning-based data augmentation method; (ii) easily extended to multilingual few-shot learning. A systematic evaluation on 18 standard NLP tasks demonstrates that this approach improves the various existing SOTA few-shot learning methods by 12\%, and yields competitive few-shot performance with 500 times larger models, such as GPT-3.

Qinyuan Ye, Belinda Z. Li, Sinong Wang et al.

Current NLP models are predominantly trained through a two-stage "pre-train then fine-tune" pipeline. Prior work has shown that inserting an intermediate pre-training stage, using heuristic masking policies for masked language modeling (MLM), can significantly improve final performance. However, it is still unclear (1) in what cases such intermediate pre-training is helpful, (2) whether hand-crafted heuristic objectives are optimal for a given task, and (3) whether a masking policy designed for one task is generalizable beyond that task. In this paper, we perform a large-scale empirical study to investigate the effect of various masking policies in intermediate pre-training with nine selected tasks across three categories. Crucially, we introduce methods to automate the discovery of optimal masking policies via direct supervision or meta-learning. We conclude that the success of intermediate pre-training is dependent on appropriate pre-train corpus, selection of output format (i.e., masked spans or full sentence), and clear understanding of the role that MLM plays for the downstream task. In addition, we find our learned masking policies outperform the heuristic of masking named entities on TriviaQA, and policies learned from one task can positively transfer to other tasks in certain cases, inviting future research in this direction.

Nayeon Lee, Belinda Z. Li, Sinong Wang et al.

In this paper, we introduce UnifiedM2, a general-purpose misinformation model that jointly models multiple domains of misinformation with a single, unified setup. The model is trained to handle four tasks: detecting news bias, clickbait, fake news, and verifying rumors. By grouping these tasks together, UnifiedM2learns a richer representation of misinformation, which leads to state-of-the-art or comparable performance across all tasks. Furthermore, we demonstrate that UnifiedM2's learned representation is helpful for few-shot learning of unseen misinformation tasks/datasets and model's generalizability to unseen events.

Qinyuan Ye, Belinda Z. Li, Sinong Wang et al.

Closed-book question-answering (QA) is a challenging task that requires a model to directly answer questions without access to external knowledge. It has been shown that directly fine-tuning pre-trained language models with (question, answer) examples yields surprisingly competitive performance, which is further improved upon through adding an intermediate pre-training stage between general pre-training and fine-tuning. Prior work used a heuristic during this intermediate stage, whereby named entities and dates are masked, and the model is trained to recover these tokens. In this paper, we aim to learn the optimal masking strategy for the intermediate pre-training stage. We first train our masking policy to extract spans that are likely to be tested, using supervision from the downstream task itself, then deploy the learned policy during intermediate pre-training. Thus, our policy packs task-relevant knowledge into the parameters of a language model. Our approach is particularly effective on TriviaQA, outperforming strong heuristics when used to pre-train BART.

Zhuofeng Wu, Sinong Wang, Jiatao Gu et al.

Pre-trained language models have proven their unique powers in capturing implicit language features. However, most pre-training approaches focus on the word-level training objective, while sentence-level objectives are rarely studied. In this paper, we propose Contrastive LEArning for sentence Representation (CLEAR), which employs multiple sentence-level augmentation strategies in order to learn a noise-invariant sentence representation. These augmentations include word and span deletion, reordering, and substitution. Furthermore, we investigate the key reasons that make contrastive learning effective through numerous experiments. We observe that different sentence augmentations during pre-training lead to different performance improvements on various downstream tasks. Our approach is shown to outperform multiple existing methods on both SentEval and GLUE benchmarks.

Hao Ma, Jingbin Liu, Zhirong Hu et al.

This paper designs a technique route to generate high-quality panoramic image with depth information, which involves two critical research hotspots: fusion of LiDAR and image data and image stitching. For the fusion of 3D points and image data, since a sparse depth map can be firstly generated by projecting LiDAR point onto the RGB image plane based on our reliable calibrated and synchronized sensors, we adopt a parameter self-adaptive framework to produce 2D dense depth map. For image stitching, optimal seamline for the overlapping area is searched using a graph-cuts-based method to alleviate the geometric influence and image blending based on the pyramid multi-band is utilized to eliminate the photometric effects near the stitching line. Since each pixel is associated with a depth value, we design this depth value as a radius in the spherical projection which can further project the panoramic image to the world coordinate and consequently produces a high-quality measurable panoramic image. The purposed method is tested on the data from our data collection platform and presents a satisfactory application prospects.