Boyuan Chen

Zhaofeng Wu, Linlu Qiu, Alexis Ross et al. · microsoft-research, mit

The impressive performance of recent language models across a wide range of tasks suggests that they possess a degree of abstract reasoning skills. Are these skills general and transferable, or specialized to specific tasks seen during pretraining? To disentangle these effects, we propose an evaluation framework based on "counterfactual" task variants that deviate from the default assumptions underlying standard tasks. Across a suite of 11 tasks, we observe nontrivial performance on the counterfactual variants, but nevertheless find that performance substantially and consistently degrades compared to the default conditions. This suggests that while current LMs may possess abstract task-solving skills to an extent, they often also rely on narrow, non-transferable procedures for task-solving. These results motivate a more careful interpretation of language model performance that teases apart these aspects of behavior.

Boyuan Chen, Diego Marti Monso, Yilun Du et al. · mit

This paper presents Diffusion Forcing, a new training paradigm where a diffusion model is trained to denoise a set of tokens with independent per-token noise levels. We apply Diffusion Forcing to sequence generative modeling by training a causal next-token prediction model to generate one or several future tokens without fully diffusing past ones. Our approach is shown to combine the strengths of next-token prediction models, such as variable-length generation, with the strengths of full-sequence diffusion models, such as the ability to guide sampling to desirable trajectories. Our method offers a range of additional capabilities, such as (1) rolling-out sequences of continuous tokens, such as video, with lengths past the training horizon, where baselines diverge and (2) new sampling and guiding schemes that uniquely profit from Diffusion Forcing's variable-horizon and causal architecture, and which lead to marked performance gains in decision-making and planning tasks. In addition to its empirical success, our method is proven to optimize a variational lower bound on the likelihoods of all subsequences of tokens drawn from the true joint distribution. Project website: https://boyuan.space/diffusion-forcing

Boyuan Chen, Fei Xia, Brian Ichter et al. · mit

Large language models (LLMs) have unlocked new capabilities of task planning from human instructions. However, prior attempts to apply LLMs to real-world robotic tasks are limited by the lack of grounding in the surrounding scene. In this paper, we develop NLMap, an open-vocabulary and queryable scene representation to address this problem. NLMap serves as a framework to gather and integrate contextual information into LLM planners, allowing them to see and query available objects in the scene before generating a context-conditioned plan. NLMap first establishes a natural language queryable scene representation with Visual Language models (VLMs). An LLM based object proposal module parses instructions and proposes involved objects to query the scene representation for object availability and location. An LLM planner then plans with such information about the scene. NLMap allows robots to operate without a fixed list of objects nor executable options, enabling real robot operation unachievable by previous methods. Project website: https://nlmap-saycan.github.io

Sitong Fang, Shiyi Hou, Kaile Wang et al.

Are frontier AI systems becoming more capable? Certainly. Yet such progress is not an unalloyed blessing but rather a Trojan horse: behind their performance leaps lie more insidious and destructive safety risks, namely deception. Unlike hallucination, which arises from insufficient capability and leads to mistakes, deception represents a deeper threat in which models deliberately mislead users through complex reasoning and insincere responses. As system capabilities advance, deceptive behaviours have spread from textual to multimodal settings, amplifying their potential harm. First and foremost, how can we monitor these covert multimodal deceptive behaviors? Nevertheless, current research remains almost entirely confined to text, leaving the deceptive risks of multimodal large language models unexplored. In this work, we systematically reveal and quantify multimodal deception risks, introducing MM-DeceptionBench, the first benchmark explicitly designed to evaluate multimodal deception. Covering six categories of deception, MM-DeceptionBench characterizes how models strategically manipulate and mislead through combined visual and textual modalities. On the other hand, multimodal deception evaluation is almost a blind spot in existing methods. Its stealth, compounded by visual-semantic ambiguity and the complexity of cross-modal reasoning, renders action monitoring and chain-of-thought monitoring largely ineffective. To tackle this challenge, we propose debate with images, a novel multi-agent debate monitor framework. By compelling models to ground their claims in visual evidence, this method substantially improves the detectability of deceptive strategies. Experiments show that it consistently increases agreement with human judgements across all tested models, boosting Cohen's kappa by 1.5x and accuracy by 1.25x on GPT-4o.

Yunxiang Zhang, Benjamin Liang, Boyuan Chen et al.

While tremendous advances in visual and auditory realism have been made for virtual and augmented reality (VR/AR), introducing a plausible sense of physicality into the virtual world remains challenging. Closing the gap between real-world physicality and immersive virtual experience requires a closed interaction loop: applying user-exerted physical forces to the virtual environment and generating haptic sensations back to the users. However, existing VR/AR solutions either completely ignore the force inputs from the users or rely on obtrusive sensing devices that compromise user experience. By identifying users' muscle activation patterns while engaging in VR/AR, we design a learning-based neural interface for natural and intuitive force inputs. Specifically, we show that lightweight electromyography sensors, resting non-invasively on users' forearm skin, inform and establish a robust understanding of their complex hand activities. Fuelled by a neural-network-based model, our interface can decode finger-wise forces in real-time with 3.3% mean error, and generalize to new users with little calibration. Through an interactive psychophysical study, we show that human perception of virtual objects' physical properties, such as stiffness, can be significantly enhanced by our interface. We further demonstrate that our interface enables ubiquitous control via finger tapping. Ultimately, we envision our findings to push forward research towards more realistic physicality in future VR/AR.

Jiaxun Liu, Boxi Xia, Boyuan Chen

Symmetry is a central organizing principle in natural systems, yet its use as a unifying design strategy in robotics has largely remained limited to geometric form. We show that symmetry can instead be leveraged at the level of dynamic actuation capability. We introduce dynamic symmetry, the uniformity of a robot's attainable center-of-mass accelerations, and formalize it through a measure coined as dynamic isotropy. Across more than 1000 simulated morphologies, we found that higher dynamic symmetry consistently improved trajectory tracking, task success, robustness, resiliency, and energy efficiency, with the benefits becoming most pronounced as dynamic isotropy approached its theoretical limit. To study this regime systematically, we developed Argus, a family of spherical robots designed to explore the effects of increasing dynamic symmetry. Members of the Argus family vary in their actuation geometry and dynamic symmetry level while sharing a common architectural principle: radially oriented linear actuators that directly shape the robot's center-of-mass dynamics. Among them, we built a physical 20-leg Argus variant that achieved near-extreme dynamic isotropy and demonstrated orientation-invariant locomotion, agile traversal of cluttered and deformable terrain, rapid self-stabilization, and resilience to partial actuator failures. Its distributed sensing further enabled omnidirectional perception and object interaction during continuous motion. These results show that designing robots for symmetry not only in morphology but also in their attainable dynamics provides a powerful and general pathway toward agility, robustness, and multifunctionality in uncertain terrestrial and extraterrestrial environments.

Jiaming Ji, Tianyi Qiu, Boyuan Chen et al.

AI alignment aims to make AI systems behave in line with human intentions and values. As AI systems grow more capable, so do risks from misalignment. To provide a comprehensive and up-to-date overview of the alignment field, in this survey, we delve into the core concepts, methodology, and practice of alignment. First, we identify four principles as the key objectives of AI alignment: Robustness, Interpretability, Controllability, and Ethicality (RICE). Guided by these four principles, we outline the landscape of current alignment research and decompose them into two key components: forward alignment and backward alignment. The former aims to make AI systems aligned via alignment training, while the latter aims to gain evidence about the systems' alignment and govern them appropriately to avoid exacerbating misalignment risks. On forward alignment, we discuss techniques for learning from feedback and learning under distribution shift. On backward alignment, we discuss assurance techniques and governance practices. We also release and continually update the website (www.alignmentsurvey.com) which features tutorials, collections of papers, blog posts, and other resources.

Ray Chen Zheng, Kaizhe Hu, Zhecheng Yuan et al. · mit

Visual imitation learning provides an effective framework to learn skills from demonstrations. However, the quality of the provided demonstrations usually significantly affects the ability of an agent to acquire desired skills. Therefore, the standard visual imitation learning assumes near-optimal demonstrations, which are expensive or sometimes prohibitive to collect. Previous works propose to learn from noisy demonstrations; however, the noise is usually assumed to follow a context-independent distribution such as a uniform or gaussian distribution. In this paper, we consider another crucial yet underexplored setting -- imitation learning with task-irrelevant yet locally consistent segments in the demonstrations (e.g., wiping sweat while cutting potatoes in a cooking tutorial). We argue that such noise is common in real world data and term them "extraneous" segments. To tackle this problem, we introduce Extraneousness-Aware Imitation Learning (EIL), a self-supervised approach that learns visuomotor policies from third-person demonstrations with extraneous subsequences. EIL learns action-conditioned observation embeddings in a self-supervised manner and retrieves task-relevant observations across visual demonstrations while excluding the extraneous ones. Experimental results show that EIL outperforms strong baselines and achieves comparable policies to those trained with perfect demonstration on both simulated and real-world robot control tasks. The project page can be found at https://sites.google.com/view/eil-website.

Sizhe Lester Li, Annan Zhang, Boyuan Chen et al. · mit

Mirroring the complex structures and diverse functions of natural organisms is a long-standing challenge in robotics. Modern fabrication techniques have greatly expanded the feasible hardware, but using these systems requires control software to translate the desired motions into actuator commands. Conventional robots can easily be modeled as rigid links connected by joints, but it remains an open challenge to model and control biologically inspired robots that are often soft or made of several materials, lack sensing capabilities, and may change their material properties with use. Here, we introduce a method that uses deep neural networks to map a video stream of a robot to its visuomotor Jacobian field (the sensitivity of all 3D points to the robot's actuators). Our method enables the control of robots from only a single camera, makes no assumptions about the robots' materials, actuation, or sensing, and is trained without expert intervention by observing the execution of random commands. We demonstrate our method on a diverse set of robot manipulators that vary in actuation, materials, fabrication, and cost. Our approach achieves accurate closed-loop control and recovers the causal dynamic structure of each robot. Because it enables robot control using a generic camera as the only sensor, we anticipate that our work will broaden the design space of robotic systems and serve as a starting point for lowering the barrier to robotic automation.

Robert Kwiatkowski, Yuhang Hu, Boyuan Chen et al.

Self-Modeling is the process by which an agent, such as an animal or machine, learns to create a predictive model of its own dynamics. Once captured, this self-model can then allow the agent to plan and evaluate various potential behaviors internally using the self-model, rather than using costly physical experimentation. Here, we quantify the benefits of such self-modeling against the complexity of the robot. We find a R2 =0.90 correlation between the number of degrees of freedom a robot has, and the added value of self-modeling as compared to a direct learning baseline. This result may help motivate self modeling in increasingly complex robotic systems, as well as shed light on the origins of self-modeling, and ultimately self-awareness, in animals and humans.

Boyuan Chen, Tianyuan Zhang, Haoran Geng et al. · mit

General-purpose robots require decision-making models that generalize across diverse tasks and environments. Recent works build robot foundation models by extending multimodal large language models (MLLMs) with action outputs, creating vision-language-action (VLA) systems. These efforts are motivated by the intuition that MLLMs' large-scale language and image pretraining can be effectively transferred to the action output modality. In this work, we explore an alternative paradigm of using large-scale video pretraining as a primary modality for building robot foundation models. Unlike static images and language, videos capture spatio-temporal sequences of states and actions in the physical world that are naturally aligned with robotic behavior. We curate an internet-scale video dataset of human activities and task demonstrations, and train, for the first time at a foundation-model scale, an open video model for generative robotics planning. The model produces zero-shot video plans for novel scenes and tasks, which we post-process to extract executable robot actions. We evaluate task-level generalization through third-party selected tasks in the wild and real-robot experiments, demonstrating successful physical execution. Together, these results show robust instruction following, strong generalization, and real-world feasibility. We release both the model and dataset to support open, reproducible video-based robot learning. Our website is available at https://www.boyuan.space/large-video-planner/.

Tianzhuo Yang, Zihan Shen, Zirui Mi et al.

Action-conditioned world models are increasingly used as scalable simulators for robot learning, yet current evaluations provide limited evidence that their predictions are reliable under the actions they condition on. Existing benchmarks largely emphasize visual fidelity, leaving unclear whether predicted futures are physically plausible, faithful to commanded actions, and calibrated to failure when actions should not succeed. We introduce \textsc{MiraBench}, a hierarchical benchmark that defines \emph{action-conditioned reliability} as a core evaluation target for robotic world models. MiraBench decomposes this target into three progressively demanding levels: \emph{Physics Adherence}, which evaluates reference-free physical consistency; \emph{Action-Following Fidelity}, which measures whether predictions respect task-relevant action inputs; and \emph{Optimism Bias Detection}, which probes the tendency to predict successful outcomes under failure-inducing actions. To support this evaluation, we curate a human-annotated corpus with over 16,000 judgments across tasks, failure categories, and leading world models. We evaluate 12 representative model configurations spanning vector-conditioned robotic world models, text-conditioned generative world models, open-weight systems, closed-source systems, and multiple model scales. Across this broad model landscape, MiraBench reveals three central findings: visual fidelity is a poor proxy for action fidelity; increasing model scale does not reliably improve action following; and optimism bias is pervasive across current systems. By shifting evaluation from appearance to action-conditioned reliability, MiraBench provides a diagnostic foundation for assessing and improving robotic world models as faithful simulators.

Zhengran Ji, Lingyu Zhang, Paul Sajda et al.

Learning collaborative behaviors is essential for multi-agent systems. Traditionally, multi-agent reinforcement learning solves this implicitly through a joint reward and centralized observations, assuming collaborative behavior will emerge. Other studies propose to learn from demonstrations of a group of collaborative experts. Instead, we propose an efficient and explicit way of learning collaborative behaviors in multi-agent systems by leveraging expertise from only a single human. Our insight is that humans can naturally take on various roles in a team. We show that agents can effectively learn to collaborate by allowing a human operator to dynamically switch between controlling agents for a short period and incorporating a human-like theory-of-mind model of teammates. Our experiments showed that our method improves the success rate of a challenging collaborative hide-and-seek task by up to 58% with only 40 minutes of human guidance. We further demonstrate our findings transfer to the real world by conducting multi-robot experiments.

Yanbaihui Liu, Erica Babusci, Claudia K. Gunsch et al.

While robotic perception has advanced rapidly in vision and touch, enabling robots to reason about indoor fungal contamination from weak, diffusion-dominated chemical signals remains an open challenge. We introduce Scensory, a learning-based robotic olfaction framework that simultaneously identifies fungal species and localizes their source from short time series measured by affordable, cross-sensitive VOC sensor arrays. Temporal VOC dynamics encode both chemical and spatial signatures, which we decode through neural networks trained on robot-automated data collection with spatial supervision. Across five fungal species, Scensory achieves up to 89.85% species accuracy and 87.31% source localization accuracy under ambient conditions with 3-7s sensor inputs. These results demonstrate real-time, spatially grounded perception from diffusion-dominated chemical signals, enabling scalable and low-cost source localization for robotic indoor environmental monitoring.

Kairui Ding, Boyuan Chen, Yuchen Su et al.

End-to-end architectures in autonomous driving (AD) face a significant challenge in interpretability, impeding human-AI trust. Human-friendly natural language has been explored for tasks such as driving explanation and 3D captioning. However, previous works primarily focused on the paradigm of declarative interpretability, where the natural language interpretations are not grounded in the intermediate outputs of AD systems, making the interpretations only declarative. In contrast, aligned interpretability establishes a connection between language and the intermediate outputs of AD systems. Here we introduce Hint-AD, an integrated AD-language system that generates language aligned with the holistic perception-prediction-planning outputs of the AD model. By incorporating the intermediate outputs and a holistic token mixer sub-network for effective feature adaptation, Hint-AD achieves desirable accuracy, achieving state-of-the-art results in driving language tasks including driving explanation, 3D dense captioning, and command prediction. To facilitate further study on driving explanation task on nuScenes, we also introduce a human-labeled dataset, Nu-X. Codes, dataset, and models will be publicly available.

Aya El Mir, Lukelo Thadei Luoga, Boyuan Chen et al.

Deploying Multi-Modal Large Language Models (MLLMs) in healthcare is hindered by their high computational demands and significant memory requirements, which are particularly challenging for resource-constrained devices like the Nvidia Jetson Xavier. This problem is particularly evident in remote medical settings where advanced diagnostics are needed but resources are limited. In this paper, we introduce an optimization method for the general-purpose MLLM, TinyLLaVA, which we have adapted and renamed TinyLLaVA-Med. This adaptation involves instruction-tuning and fine-tuning TinyLLaVA on a medical dataset by drawing inspiration from the LLaVA-Med training pipeline. Our approach successfully minimizes computational complexity and power consumption, with TinyLLaVA-Med operating at 18.9W and using 11.9GB of memory, while achieving accuracies of 64.54% on VQA-RAD and 70.70% on SLAKE for closed-ended questions. Therefore, TinyLLaVA-Med achieves deployment viability in hardware-constrained environments with low computational resources, maintaining essential functionalities and delivering accuracies close to state-of-the-art models.

Parteek Jamwal, Minghao Shao, Boyuan Chen et al.

Large Language Models (LLMs) have demonstrated remarkable capabilities across various cybersecurity tasks, including vulnerability classification, detection, and patching. However, their potential in automated vulnerability report documentation and analysis remains underexplored. We present RAVEN (Retrieval Augmented Vulnerability Exploration Network), a framework leveraging LLM agents and Retrieval Augmented Generation (RAG) to synthesize comprehensive vulnerability analysis reports. Given vulnerable source code, RAVEN generates reports following the Google Project Zero Root Cause Analysis template. The framework uses four modules: an Explorer agent for vulnerability identification, a RAG engine retrieving relevant knowledge from curated databases including Google Project Zero reports and CWE entries, an Analyst agent for impact and exploitation assessment, and a Reporter agent for structured report generation. To ensure quality, RAVEN includes a task specific LLM Judge evaluating reports across structural integrity, ground truth alignment, code reasoning quality, and remediation quality. We evaluate RAVEN on 105 vulnerable code samples covering 15 CWE types from the NIST-SARD dataset. Results show an average quality score of 54.21%, supporting the effectiveness of our approach for automated vulnerability documentation.

Jiaming Ji, Boyuan Chen, Hantao Lou et al.

With the rapid development of large language models (LLMs) and ever-evolving practical requirements, finding an efficient and effective alignment method has never been more critical. However, the tension between the complexity of current alignment methods and the need for rapid iteration in deployment scenarios necessitates the development of a model-agnostic alignment approach that can operate under these constraints. In this paper, we introduce Aligner, a novel and simple alignment paradigm that learns the correctional residuals between preferred and dispreferred answers using a small model. Designed as a model-agnostic, plug-and-play module, Aligner can be directly applied to various open-source and API-based models with only one-off training, making it suitable for rapid iteration. Notably, Aligner can be applied to any powerful, large-scale upstream models. Moreover, it can even iteratively bootstrap the upstream models using corrected responses as synthetic human preference data, breaking through the model's performance ceiling. Our experiments demonstrate performance improvements by deploying the same Aligner model across 11 different LLMs, evaluated on the 3H dimensions (helpfulness, harmlessness, and honesty). Specifically, Aligner-7B has achieved an average improvement of 68.9% in helpfulness and 23.8% in harmlessness across the tested LLMs while also effectively reducing hallucination. In the Alpaca-Eval leaderboard, stacking Aligner-2B on GPT-4 Turbo improved its LC Win Rate from 55.0% to 58.3%, surpassing GPT-4 Omni's 57.5% Win Rate (community report).

Haotian Xu, Xing Wu, Weinong Wang et al.

Can scaling transform reasoning? In this work, we explore the untapped potential of scaling Long Chain-of-Thought (Long-CoT) data to 1000k samples, pioneering the development of a slow-thinking model, RedStar. Through extensive experiments with various LLMs and different sizes, we uncover the ingredients for specialization and scale for Long-CoT training. Surprisingly, even smaller models show significant performance gains with limited data, revealing the sample efficiency of Long-CoT and the critical role of sample difficulty in the learning process. Our findings demonstrate that Long-CoT reasoning can be effectively triggered with just a few thousand examples, while larger models achieve unparalleled improvements. We also introduce reinforcement learning (RL)-scale training as a promising direction for advancing slow-thinking systems. RedStar shines across domains: on the MATH-Hard benchmark, RedStar-code-math boosts performance from 66.2\% to 81.6\%, and on the USA Math Olympiad (AIME), it solves 46.7\% of problems using only 21k mixed-code-math datasets. In multimodal tasks like GeoQA and MathVista-GEO, RedStar-Geo achieves competitive results with minimal Long-CoT data, outperforming other slow-thinking systems like QvQ-Preview. Compared to QwQ, RedStar strikes the perfect balance between reasoning and generalizability. Our work highlights that, with careful tuning, scaling Long-CoT can unlock extraordinary reasoning capabilities-even with limited dataset and set a new standard for slow-thinking models across diverse challenges. Our data and models are released at https://huggingface.co/RedStar-Reasoning.

Jiaming Ji, Jiayi Zhou, Hantao Lou et al.

Reinforcement learning from human feedback (RLHF) has proven effective in enhancing the instruction-following capabilities of large language models; however, it remains underexplored in the cross-modality domain. As the number of modalities increases, aligning all-modality models with human intentions -- such as instruction following -- becomes a pressing challenge. In this work, we make the first attempt to fine-tune all-modality models (i.e. input and output with any modality, also named any-to-any models) using human preference data across all modalities (including text, image, audio, and video), ensuring its behavior aligns with human intentions. This endeavor presents several challenges. First, there is no large-scale all-modality human preference data in existing open-source resources, as most datasets are limited to specific modalities, predominantly text and image. Secondly, the effectiveness of binary preferences in RLHF for post-training alignment in complex all-modality scenarios remains an unexplored area. Finally, there is a lack of a systematic framework to evaluate the capabilities of all-modality models, particularly regarding modality selection and synergy. To address these challenges, we propose the align-anything framework, which includes meticulously annotated 200k all-modality human preference data. Then, we introduce an alignment method that learns from unified language feedback, effectively capturing complex modality-specific human preferences and enhancing the model's instruction-following capabilities. Furthermore, to assess performance improvements in all-modality models after post-training alignment, we construct a challenging all-modality capability evaluation framework -- eval-anything. All data, models, and code frameworks have been open-sourced for the community. For more details, please refer to https://github.com/PKU-Alignment/align-anything.

Xinyuan Luo, Xingrui Chen, Xunjian Yin et al.

Humanoid robots have achieved impressive locomotion performance, yet contact-rich and long-horizon manipulation remains a major bottleneck. Manipulation is inherently contact-rich and demands compliant whole-body control for stable interaction, while its diversity and long-horizon nature favor modular, planner-compatible interfaces over joint-space tracking. We propose CEER, a compliant end-effector-root (EE-root) control abstraction for modular humanoid loco-manipulation within a hierarchical planning framework. CEER enables compliance-aware whole-body control in an interpretable task space defined by root motion commands and end-effector pose targets, and supports plug-and-play integration with heterogeneous high-level planners. A teacher-student framework is adopted to distill a general motion-tracking controller into a low-level policy that consumes only EE-root commands. We further construct a hierarchical system that integrates heterogeneous planners and task modules through the EE-root interface, enabling diverse manipulation tasks without retraining the underlying whole-body policy. Experiments in simulation and on hardware demonstrate 3.3 cm end-effector tracking accuracy with substantially reduced jerk compared to baselines, stable contact-rich manipulation under teleoperation, and up to 70% success in simulated single-object loco-manipulation tasks within a room-scale environment. These results indicate that compliant EE-root control provides a practical abstraction for humanoid loco-manipulation, enabling modular and scalable integration of diverse skills.

Haoran Zhu, Boyuan Chen, Carter Yang

Vision transformer (ViT) is an attention neural network architecture that is shown to be effective for computer vision tasks. However, compared to ResNet-18 with a similar number of parameters, ViT has a significantly lower evaluation accuracy when trained on small datasets. To facilitate studies in related fields, we provide a visual intuition to help understand why it is the case. We first compare the performance of the two models and confirm that ViT has less accuracy than ResNet-18 when trained on small datasets. We then interpret the results by showing attention map visualization for ViT and feature map visualization for ResNet-18. The difference is further analyzed through a representation similarity perspective. We conclude that the representation of ViT trained on small datasets is hugely different from ViT trained on large datasets, which may be the reason why the performance drops a lot on small datasets.

Boyuan Chen, Minghao Shao, Siddharth Garg et al.

Vision Language Models (VLMs) exhibit persistent hallucinations in counting tasks, with accuracy substantially lower than other visual reasoning tasks (excluding sentiment). This phenomenon persists even in state-of-the-art reasoning-capable VLMs. Conversely, CNN-based object detection models (ODMs) such as YOLO excel at spatial localization and instance counting with minimal computational overhead. We propose GroundCount, a framework that augments VLMs with explicit spatial grounding from ODMs to mitigate counting hallucinations. In the best case, our prompt-based augmentation strategy achieves 81.3% counting accuracy on the best-performing model (Ovis2.5-2B) - a 6.6pp improvement - while reducing inference time by 22% through elimination of hallucination-driven reasoning loops for stronger models. We conduct comprehensive ablation studies demonstrating that positional encoding is a critical component, being beneficial for stronger models but detrimental for weaker ones. Confidence scores, by contrast, introduce noise for most architectures and their removal improves performance in four of five evaluated models. We further evaluate feature-level fusion architectures, finding that explicit symbolic grounding via structured prompts outperforms implicit feature fusion despite sophisticated cross-attention mechanisms. Our approach yields consistent improvements across four of five evaluated VLM architectures (6.2--7.5pp), with one architecture exhibiting degraded performance due to incompatibility between its iterative reflection mechanisms and structured prompts. These results suggest that counting failures stem from fundamental spatial-semantic integration limitations rather than architecture-specific deficiencies, while highlighting the importance of architectural compatibility in augmentation strategies.

Jonathan Hyun, Nicholas R Waytowich, Boyuan Chen

Despite rapid progress in large language model (LLM)-based multi-agent systems, current benchmarks fall short in evaluating their scalability, robustness, and coordination capabilities in complex, dynamic, real-world tasks. Existing environments typically focus on small-scale, fully observable, or low-complexity domains, limiting their utility for developing and assessing next-generation multi-agent Agentic AI frameworks. We introduce CREW-Wildfire, an open-source benchmark designed to close this gap. Built atop the human-AI teaming CREW simulation platform, CREW-Wildfire offers procedurally generated wildfire response scenarios featuring large maps, heterogeneous agents, partial observability, stochastic dynamics, and long-horizon planning objectives. The environment supports both low-level control and high-level natural language interactions through modular Perception and Execution modules. We implement and evaluate several state-of-the-art LLM-based multi-agent Agentic AI frameworks, uncovering significant performance gaps that highlight the unsolved challenges in large-scale coordination, communication, spatial reasoning, and long-horizon planning under uncertainty. By providing more realistic complexity, scalable architecture, and behavioral evaluation metrics, CREW-Wildfire establishes a critical foundation for advancing research in scalable multi-agent Agentic intelligence. All code, environments, data, and baselines will be released to support future research in this emerging domain.

Jiayi Zhou, Jiaming Ji, Boyuan Chen et al.

Training multi-modal large language models (MLLMs) that align with human intentions is a long-term challenge. Traditional score-only reward models for alignment suffer from low accuracy, weak generalization, and poor interpretability, blocking the progress of alignment methods, e.g., reinforcement learning from human feedback (RLHF). Generative reward models (GRMs) leverage MLLMs' intrinsic reasoning capabilities to discriminate pair-wise responses, but their pair-wise paradigm makes it hard to generalize to learnable rewards. We introduce Generative RLHF-V, a novel alignment framework that integrates GRMs with multi-modal RLHF. We propose a two-stage pipeline: $\textbf{multi-modal generative reward modeling from RL}$, where RL guides GRMs to actively capture human intention, then predict the correct pair-wise scores; and $\textbf{RL optimization from grouped comparison}$, which enhances multi-modal RL scoring precision by grouped responses comparison. Experimental results demonstrate that, besides out-of-distribution generalization of RM discrimination, our framework improves 4 MLLMs' performance across 7 benchmarks by $18.1\%$, while the baseline RLHF is only $5.3\%$. We further validate that Generative RLHF-V achieves a near-linear improvement with an increasing number of candidate responses. Our code and models can be found at https://generative-rlhf-v.github.io.

Lingyu Zhang, Zhengran Ji, Boyuan Chen

With the increasing deployment of artificial intelligence (AI) technologies, the potential of humans working with AI agents has been growing at a great speed. Human-AI teaming is an important paradigm for studying various aspects when humans and AI agents work together. The unique aspect of Human-AI teaming research is the need to jointly study humans and AI agents, demanding multidisciplinary research efforts from machine learning to human-computer interaction, robotics, cognitive science, neuroscience, psychology, social science, and complex systems. However, existing platforms for Human-AI teaming research are limited, often supporting oversimplified scenarios and a single task, or specifically focusing on either human-teaming research or multi-agent AI algorithms. We introduce CREW, a platform to facilitate Human-AI teaming research in real-time decision-making scenarios and engage collaborations from multiple scientific disciplines, with a strong emphasis on human involvement. It includes pre-built tasks for cognitive studies and Human-AI teaming with expandable potentials from our modular design. Following conventional cognitive neuroscience research, CREW also supports multimodal human physiological signal recording for behavior analysis. Moreover, CREW benchmarks real-time human-guided reinforcement learning agents using state-of-the-art algorithms and well-tuned baselines. With CREW, we were able to conduct 50 human subject studies within a week to verify the effectiveness of our benchmark.

Boyuan Chen, Zhuo Xu, Sean Kirmani et al. · mit

Understanding and reasoning about spatial relationships is a fundamental capability for Visual Question Answering (VQA) and robotics. While Vision Language Models (VLM) have demonstrated remarkable performance in certain VQA benchmarks, they still lack capabilities in 3D spatial reasoning, such as recognizing quantitative relationships of physical objects like distances or size differences. We hypothesize that VLMs' limited spatial reasoning capability is due to the lack of 3D spatial knowledge in training data and aim to solve this problem by training VLMs with Internet-scale spatial reasoning data. To this end, we present a system to facilitate this approach. We first develop an automatic 3D spatial VQA data generation framework that scales up to 2 billion VQA examples on 10 million real-world images. We then investigate various factors in the training recipe, including data quality, training pipeline, and VLM architecture. Our work features the first internet-scale 3D spatial reasoning dataset in metric space. By training a VLM on such data, we significantly enhance its ability on both qualitative and quantitative spatial VQA. Finally, we demonstrate that this VLM unlocks novel downstream applications in chain-of-thought spatial reasoning and robotics due to its quantitative estimation capability. Project website: https://spatial-vlm.github.io/

Jiaming Ji, Xinyu Chen, Rui Pan et al.

Multimodal large language models (MLLMs) are essential for building general-purpose AI assistants; however, they pose increasing safety risks. How can we ensure safety alignment of MLLMs to prevent undesired behaviors? Going further, it is critical to explore how to fine-tune MLLMs to preserve capabilities while meeting safety constraints. Fundamentally, this challenge can be formulated as a min-max optimization problem. However, existing datasets have not yet disentangled single preference signals into explicit safety constraints, hindering systematic investigation in this direction. Moreover, it remains an open question whether such constraints can be effectively incorporated into the optimization process for multi-modal models. In this work, we present the first exploration of the Safe RLHF-V -- the first multimodal safety alignment framework. The framework consists of: $\mathbf{(I)}$ BeaverTails-V, the first open-source dataset featuring dual preference annotations for helpfulness and safety, supplemented with multi-level safety labels (minor, moderate, severe); $\mathbf{(II)}$ Beaver-Guard-V, a multi-level guardrail system to proactively defend against unsafe queries and adversarial attacks. Applying the guard model over five rounds of filtering and regeneration significantly enhances the precursor model's overall safety by an average of 40.9%. $\mathbf{(III)}$ Based on dual preference, we initiate the first exploration of multi-modal safety alignment within a constrained optimization. Experimental results demonstrate that Safe RLHF effectively improves both model helpfulness and safety. Specifically, Safe RLHF-V enhances model safety by 34.2% and helpfulness by 34.3%.

Gustavo A. Oliva, Gopi Krishnan Rajbahadur, Aaditya Bhatia et al.

High-quality labeled datasets are crucial for training and evaluating foundation models in software engineering, but creating them is often prohibitively expensive and labor-intensive. We introduce SPICE, a scalable, automated pipeline for labeling SWE-bench-style datasets with annotations for issue clarity, test coverage, and effort estimation. SPICE combines context-aware code navigation, rationale-driven prompting, and multi-pass consensus to produce labels that closely approximate expert annotations. SPICE's design was informed by our own experience and frustration in labeling more than 800 instances from SWE-Gym. SPICE achieves strong agreement with human-labeled SWE-bench Verified data while reducing the cost of labeling 1,000 instances from around \$100,000 (manual annotation) to just \$5.10. These results demonstrate SPICE's potential to enable cost-effective, large-scale dataset creation for SE-focused FMs. To support the community, we release both SPICE tool and SPICE Bench, a new dataset of 6,802 SPICE-labeled instances curated from 291 open-source projects in SWE-Gym (over 13x larger than SWE-bench Verified).

Michael Scherk, Boyuan Chen

Symbolic regression (SR) seeks to recover closed-form mathematical expressions that describe observed data. While existing methods have advanced the discovery of either explicit mappings (i.e., $y = f(\mathbf{x})$) or discovering implicit relations (i.e., $F(\mathbf{x}, y)=0$), few modern and accessible frameworks support both. Moreover, most approaches treat each expression candidate in isolation, without reusing recurring structural patterns that could accelerate search. We introduce SymMatika, a hybrid SR algorithm that combines multi-island genetic programming (GP) with a reusable motif library inspired by biological sequence analysis. SymMatika identifies high-impact substructures in top-performing candidates and reintroduces them to guide future generations. Additionally, it incorporates a feedback-driven evolutionary engine and supports both explicit and implicit relation discovery using implicit-derivative metrics. Across benchmarks, SymMatika achieves state-of-the-art recovery rates on the Nguyen and Feynman benchmark suites, an impressive recovery rate of 61\% on Nguyen-12 compared to the next best 2\%, and strong placement on the error-complexity Pareto fronts on the Feynman equations and on a subset of 57 SRBench Black-box problems. Our results demonstrate the power of structure-aware evolutionary search for scientific discovery. To support broader research in interpretable modeling and symbolic discovery, we have open-sourced the full SymMatika framework.

Boyuan Chen, Donghai Hong, Jiaming Ji et al.

As multimodal large models (MLLMs) continue to advance across challenging tasks, a key question emerges: What essential capabilities are still missing? A critical aspect of human learning is continuous interaction with the environment -- not limited to language, but also involving multimodal understanding and generation. To move closer to human-level intelligence, models must similarly support multi-turn, multimodal interaction. In particular, they should comprehend interleaved multimodal contexts and respond coherently in ongoing exchanges. In this work, we present an initial exploration through the InterMT -- the first preference dataset for multi-turn multimodal interaction, grounded in real human feedback. In this exploration, we particularly emphasize the importance of human oversight, introducing expert annotations to guide the process, motivated by the fact that current MLLMs lack such complex interactive capabilities. InterMT captures human preferences at both global and local levels into nine sub-dimensions, consists of 15.6k prompts, 52.6k multi-turn dialogue instances, and 32.4k human-labeled preference pairs. To compensate for the lack of capability for multi-modal understanding and generation, we introduce an agentic workflow that leverages tool-augmented MLLMs to construct multi-turn QA instances. To further this goal, we introduce InterMT-Bench to assess the ability of MLLMs in assisting judges with multi-turn, multimodal tasks. We demonstrate the utility of \InterMT through applications such as judge moderation and further reveal the multi-turn scaling law of judge model. We hope the open-source of our data can help facilitate further research on aligning current MLLMs to the next step. Our project website can be found at https://pku-intermt.github.io .

Ximing Dong, Shaowei Wang, Dayi Lin et al.

Large Language Models (LLMs) achieve strong performance across many tasks but suffer from high inference latency due to autoregressive decoding. The issue is exacerbated in Large Reasoning Models (LRMs), which generate lengthy chains of thought. While speculative decoding accelerates inference by drafting and verifying multiple tokens in parallel, existing methods operate at the token level and ignore semantic equivalence (i.e., different token sequences expressing the same meaning), leading to inefficient rejections. We propose SemanticSpec, a semantic-aware speculative decoding framework that verifies entire semantic sequences instead of tokens. SemanticSpec introduces a semantic probability estimation mechanism that probes the model's internal hidden states to assess the likelihood of generating sequences with specific meanings. Experiments on four benchmarks show that SemanticSpec achieves up to 2.7x speedup on DeepSeekR1-32B and 2.1x on QwQ-32B, consistently outperforming token-level and sequence-level baselines in both efficiency and effectiveness.

Han Zhu, Juntao Dai, Jiaming Ji et al.

With the widespread use of multi-modal Large Language models (MLLMs), safety issues have become a growing concern. Multi-turn dialogues, which are more common in everyday interactions, pose a greater risk than single prompts; however, existing benchmarks do not adequately consider this situation. To encourage the community to focus on the safety issues of these models in multi-turn dialogues, we introduce SafeMT, a benchmark that features dialogues of varying lengths generated from harmful queries accompanied by images. This benchmark consists of 10,000 samples in total, encompassing 17 different scenarios and four jailbreak methods. Additionally, we propose Safety Index (SI) to evaluate the general safety of MLLMs during conversations. We assess the safety of 17 models using this benchmark and discover that the risk of successful attacks on these models increases as the number of turns in harmful dialogues rises. This observation indicates that the safety mechanisms of these models are inadequate for recognizing the hazard in dialogue interactions. We propose a dialogue safety moderator capable of detecting malicious intent concealed within conversations and providing MLLMs with relevant safety policies. Experimental results from several open-source models indicate that this moderator is more effective in reducing multi-turn ASR compared to existed guard models.

Jiaming Ji, Donghai Hong, Borong Zhang et al.

In this study, we introduce the safety human preference dataset, PKU-SafeRLHF, designed to promote research on safety alignment in large language models (LLMs). As a sibling project to SafeRLHF and BeaverTails, we separate annotations of helpfulness and harmlessness for question-answering pairs, providing distinct perspectives on these coupled attributes. Overall, we provide 44.6k refined prompts and 265k question-answer pairs with safety meta-labels for 19 harm categories and three severity levels ranging from minor to severe, with answers generated by Llama-family models. Based on this, we collected 166.8k preference data, including dual-preference (helpfulness and harmlessness decoupled) and single-preference data (trade-off the helpfulness and harmlessness from scratch), respectively. Using the large-scale annotation data, we further train severity-sensitive moderation for the risk control of LLMs and safety-centric RLHF algorithms for the safety alignment of LLMs. We believe this dataset will be a valuable resource for the community, aiding in the safe deployment of LLMs. Data is available at https://huggingface.co/datasets/PKU-Alignment/PKU-SafeRLHF.

Minghao Shao, Sofija Jancheska, Meet Udeshi et al.

Large Language Models (LLMs) are being deployed across various domains today. However, their capacity to solve Capture the Flag (CTF) challenges in cybersecurity has not been thoroughly evaluated. To address this, we develop a novel method to assess LLMs in solving CTF challenges by creating a scalable, open-source benchmark database specifically designed for these applications. This database includes metadata for LLM testing and adaptive learning, compiling a diverse range of CTF challenges from popular competitions. Utilizing the advanced function calling capabilities of LLMs, we build a fully automated system with an enhanced workflow and support for external tool calls. Our benchmark dataset and automated framework allow us to evaluate the performance of five LLMs, encompassing both black-box and open-source models. This work lays the foundation for future research into improving the efficiency of LLMs in interactive cybersecurity tasks and automated task planning. By providing a specialized benchmark, our project offers an ideal platform for developing, testing, and refining LLM-based approaches to vulnerability detection and resolution. Evaluating LLMs on these challenges and comparing with human performance yields insights into their potential for AI-driven cybersecurity solutions to perform real-world threat management. We make our benchmark dataset open source to public https://github.com/NYU-LLM-CTF/NYU_CTF_Bench along with our playground automated framework https://github.com/NYU-LLM-CTF/llm_ctf_automation.

Jiawen Xiong, Yong Shi, Boyuan Chen et al.

Build verifiability refers to the property that the build of a software system can be verified by independent third parties and it is crucial for the trustworthiness of a software system. Various efforts towards build verifiability have been made to C/C++-based systems, yet the techniques for Java-based systems are not systematic and are often specific to a particular build tool (e.g., Maven). In this study, we present a systematic approach towards build verifiability on Java-based systems. Our approach consists of three parts: a unified build process, a tool that dynamically controls non-determinism during the build process, and another tool that eliminates non-equivalences by post-processing the build artifacts. We apply our approach on 46 unverified open source projects from Reproducible Central and 13 open source projects that are widely used by Huawei commercial products. As a result, 91% of the unverified Reproducible Central projects and 100% of the commercially adopted OSS projects are successfully verified with our approach. In addition, based on our experience in analyzing thousands of builds for both commercial and open source Java-based systems, we present 14 patterns that introduce non-equivalences in generated build artifacts and their respective mitigation strategies. Among these patterns, 11 (78%) are unique for Java-based system, whereas the remaining 3 (22%) are common with C/C++-based systems. The approach and the findings of this paper are useful for both practitioners and researchers who are interested in build verifiability.

Boyuan Chen, Mingzhi Wen, Yong Shi et al.

Reproducibility is an increasing concern in Artificial Intelligence (AI), particularly in the area of Deep Learning (DL). Being able to reproduce DL models is crucial for AI-based systems, as it is closely tied to various tasks like training, testing, debugging, and auditing. However, DL models are challenging to be reproduced due to issues like randomness in the software (e.g., DL algorithms) and non-determinism in the hardware (e.g., GPU). There are various practices to mitigate some of the aforementioned issues. However, many of them are either too intrusive or can only work for a specific usage context. In this paper, we propose a systematic approach to training reproducible DL models. Our approach includes three main parts: (1) a set of general criteria to thoroughly evaluate the reproducibility of DL models for two different domains, (2) a unified framework which leverages a record-and-replay technique to mitigate software-related randomness and a profile-and-patch technique to control hardware-related non-determinism, and (3) a reproducibility guideline which explains the rationales and the mitigation strategies on conducting a reproducible training process for DL models. Case study results show our approach can successfully reproduce six open source and one commercial DL models.

Gopi Krishnan Rajbahadur, Erika Tuck, Li Zi et al.

Publicly available datasets are one of the key drivers for commercial AI software. The use of publicly available datasets is governed by dataset licenses. These dataset licenses outline the rights one is entitled to on a given dataset and the obligations that one must fulfil to enjoy such rights without any license compliance violations. Unlike standardized Open Source Software (OSS) licenses, existing dataset licenses are defined in an ad-hoc manner and do not clearly outline the rights and obligations associated with their usage. Further, a public dataset may be hosted in multiple locations and created from multiple data sources each of which may have different licenses. Hence, existing approaches on checking OSS license compliance cannot be used. In this paper, we propose a new approach to assessing the potential license compliance violations if a given publicly available dataset were to be used for building commercial AI software. We conduct a case study with our approach on 6 commonly used publicly available image datasets. Our results show that there exists potential risks of license violations associated with all of the studied datasets if they were used for commercial purposes.

Yinsen Jia, Boyuan Chen

Temporal awareness plays a central role in intelligent behavior by shaping how actions are paced, coordinated, and adapted to changing goals and environments. In contrast, most robot learning algorithms treat time only as a fixed episode horizon or scheduling constraint. Here we introduce time-aware policy learning, a reinforcement learning framework that treats time as a control dimension of robot behavior. The approach augments policies with two temporal signals, the remaining time and a time ratio that modulates the policy's internal progression of time, allowing a single policy to regulate its execution strategy across temporal regimes. Across diverse manipulation tasks including long-horizon manipulation, granular-media pouring, articulated-object interaction, and multi-agent coordination, the resulting policies adapt their behavior continuously from dynamic execution under tight schedules to stable and deliberate interaction when more time is available. This temporal awareness improves efficiency, robustness under sim-to-real mismatch and disturbances, and controllability through human input without retraining. Treating time as a controllable variable provides a new framework for adaptive and human-aligned robot autonomy.

Lingyu Zhang, Mitchell Wang, Boyuan Chen

Scientific discovery begins with ideas, yet evaluating early-stage research concepts is a subtle and subjective human judgment. As large language models (LLMs) are increasingly tasked with generating scientific hypotheses, most systems implicitly assume that scientists' evaluations form a fixed gold standard, assuming that scientists' judgments do not change. Here we challenge this assumption. In a two-wave study with 7,938 ratings from 63 active researchers across six scientific departments, each participant repeatedly evaluated a constant "control" research idea alongside AI-generated ideas. We find that expert evaluations are not stable: test-retest reliability of overall quality is only moderate (ICC~0.59-0.74), indicating substantial within-participant variability even for identical ideas. Yet the internal structure of judgment remained stable, such as the relative importance placed on originality, feasibility, clarity, and other criteria. We then aligned an LLM-based ideation system to first-wave human ratings and used it to select new ideas. Although alignment improved agreement with Wave-1 evaluations, its apparent gains disappeared once drift in human standards was accounted for. Thus, tuning to a fixed human snapshot produced improvements that were transient rather than persistent. These findings reveal that human evaluation of scientific ideas is not static but a dynamic process with stable priorities and requires shifting calibration. Treating one-time human ratings as immutable ground truth risks overstating progress in AI-assisted ideation and obscuring the challenge of co-evolving with changing expert standards. Drift-aware evaluation protocols and longitudinal benchmarks may therefore be essential for building AI systems that reliably augment, rather than overfit to, human scientific judgment.

Dong Heon Cho, Boyuan Chen

Differentiable simulators enable gradient-based optimization of soft robots over material parameters, control, and morphology, but accurately modeling real systems remains challenging due to the sim-to-real gap. This issue becomes more pronounced when geometry is itself a design variable. System identification reduces discrepancies by fitting global material parameters to data; however, when constitutive models are misspecified or observations are sparse, identified parameters often absorb geometry-dependent effects rather than reflect intrinsic material behavior. More expressive constitutive models can improve accuracy but substantially increase computational cost, limiting practicality. We propose a residual acceleration field learning (RAFL) framework that augments a base simulator with a transferable, element-level corrective dynamics field. Operating on shared local features, the model is agnostic to global mesh topology and discretization. Trained end-to-end through a differentiable simulator using sparse marker observations, the learned residual generalizes across shapes. In both sim-to-sim and sim-to-real experiments, our method achieves consistent zero-shot improvements on unseen morphologies, while system identification frequently exhibits negative transfer. The framework also supports continual refinement, enabling simulation accuracy to accumulate during morphology optimization.

Kishanthan Thangarajah, Boyuan Chen, Ahmed E. Hassan

Enterprises want AI code completion that is both high-quality and private, but they face a tension: proprietary models yield better results yet risk exposing proprietary code, while self-hosting large models is expensive and hard to maintain. As a lighter alternative, small CodeLLMs (1B-3B) can run on a developer's workstation accelerator with code never leaving the machine, but they fail on harder tasks. A practical solution is to use the small model for most requests and selectively route difficult ones to a larger self-hosted model. In this study, we evaluate 29 code specialized LLMs (0.5B-480B) from 12 families on execution-based fill-in-the-middle (FIM) code completion benchmarks across Python, Java, and C++, and find that model family and code specialized training matter more than size: a 3B model matches a 32B model despite being 10x smaller. Analyzing the 3B model's failures, we discover that 46% of its incorrect completions are not valid code. To enable efficient code completion, we propose SynConfRoute, a training-free method that combines token confidence with syntax validation to automatically decide per-request whether to keep the local completion or escalate to a larger self-hosted model. SynConfRoute improves pass@1 by 6.4% over confidence only routing on routine completions and by up to 31% on harder multi-language tasks, and the resulting pipeline achieves 78.9% on routine completions, 7.4% higher than always using the 480B model alone, while reducing accelerator usage by 58%. SynConfRoute generalizes across Python, Java, and C++, improving over confidence only routing on all three languages without ever rejecting a correct local completion. The pipeline uses off-the-shelf models with no custom training, making it immediately deployable in practice.

Kiwhan Song, Boyuan Chen, Max Simchowitz et al. · mit

Classifier-free guidance (CFG) is a key technique for improving conditional generation in diffusion models, enabling more accurate control while enhancing sample quality. It is natural to extend this technique to video diffusion, which generates video conditioned on a variable number of context frames, collectively referred to as history. However, we find two key challenges to guiding with variable-length history: architectures that only support fixed-size conditioning, and the empirical observation that CFG-style history dropout performs poorly. To address this, we propose the Diffusion Forcing Transformer (DFoT), a video diffusion architecture and theoretically grounded training objective that jointly enable conditioning on a flexible number of history frames. We then introduce History Guidance, a family of guidance methods uniquely enabled by DFoT. We show that its simplest form, vanilla history guidance, already significantly improves video generation quality and temporal consistency. A more advanced method, history guidance across time and frequency further enhances motion dynamics, enables compositional generalization to out-of-distribution history, and can stably roll out extremely long videos. Project website: https://boyuan.space/history-guidance

Ahmed E. Hassan, Gustavo A. Oliva, Dayi Lin et al.

The advent of Foundation Models (FMs) and AI-powered copilots has transformed the landscape of software development, offering unprecedented code completion capabilities and enhancing developer productivity. However, the current task-driven nature of these copilots falls short in addressing the broader goals and complexities inherent in software engineering (SE). In this paper, we propose a paradigm shift towards goal-driven AI-powered pair programmers that collaborate with human developers in a more holistic and context-aware manner. We envision AI pair programmers that are goal-driven, human partners, SE-aware, and self-learning. These AI partners engage in iterative, conversation-driven development processes, aligning closely with human goals and facilitating informed decision-making. We discuss the desired attributes of such AI pair programmers and outline key challenges that must be addressed to realize this vision. Ultimately, our work represents a shift from AI-augmented SE to AI-transformed SE by replacing code completion with a collaborative partnership between humans and AI that enhances both productivity and software quality.

Boyuan Chen, Hanxiao Jiang, Shaowei Liu et al.

Envisioning physically plausible outcomes from a single image requires a deep understanding of the world's dynamics. To address this, we introduce PhysGen3D, a novel framework that transforms a single image into an amodal, camera-centric, interactive 3D scene. By combining advanced image-based geometric and semantic understanding with physics-based simulation, PhysGen3D creates an interactive 3D world from a static image, enabling us to "imagine" and simulate future scenarios based on user input. At its core, PhysGen3D estimates 3D shapes, poses, physical and lighting properties of objects, thereby capturing essential physical attributes that drive realistic object interactions. This framework allows users to specify precise initial conditions, such as object speed or material properties, for enhanced control over generated video outcomes. We evaluate PhysGen3D's performance against closed-source state-of-the-art (SOTA) image-to-video models, including Pika, Kling, and Gen-3, showing PhysGen3D's capacity to generate videos with realistic physics while offering greater flexibility and fine-grained control. Our results show that PhysGen3D achieves a unique balance of photorealism, physical plausibility, and user-driven interactivity, opening new possibilities for generating dynamic, physics-grounded video from an image.

Adam Wei, Abhinav Agarwal, Boyuan Chen et al. · mit

Cotraining with demonstration data generated both in simulation and on real hardware has emerged as a promising recipe for scaling imitation learning in robotics. This work seeks to elucidate basic principles of this sim-and-real cotraining to inform simulation design, sim-and-real dataset creation, and policy training. Our experiments confirm that cotraining with simulated data can dramatically improve performance, especially when real data is limited. We show that these performance gains scale with additional simulated data up to a plateau; adding more real-world data increases this performance ceiling. The results also suggest that reducing physical domain gaps may be more impactful than visual fidelity for non-prehensile or contact-rich tasks. Perhaps surprisingly, we find that some visual gap can help cotraining -- binary probes reveal that high-performing policies must learn to distinguish simulated domains from real. We conclude by investigating this nuance and mechanisms that facilitate positive transfer between sim-and-real. Focusing narrowly on the canonical task of planar pushing from pixels allows us to be thorough in our study. In total, our experiments span 50+ real-world policies (evaluated on 1000+ trials) and 250 simulated policies (evaluated on 50,000+ trials). Videos and code can be found at https://sim-and-real-cotraining.github.io/.

Jiaming Ji, Wenqi Chen, Kaile Wang et al.

Modern large language models rely on chain-of-thought (CoT) reasoning to achieve impressive performance, yet the same mechanism can amplify deceptive alignment, situations in which a model appears aligned while covertly pursuing misaligned goals. Existing safety pipelines treat deception as a black-box output to be filtered post-hoc, leaving the model free to scheme during its internal reasoning. We ask: Can deception be intercepted while the model is thinking? We answer this question, the first framework that embeds a Self-Monitor inside the CoT process itself, named CoT Monitor+. During generation, the model produces (i) ordinary reasoning steps and (ii) an internal self-evaluation signal trained to flag and suppress misaligned strategies. The signal is used as an auxiliary reward in reinforcement learning, creating a feedback loop that rewards honest reasoning and discourages hidden goals. To study deceptive alignment systematically, we introduce DeceptionBench, a five-category benchmark that probes covert alignment-faking, sycophancy, etc. We evaluate various LLMs and show that unrestricted CoT roughly aggravates the deceptive tendency. In contrast, CoT Monitor+ cuts deceptive behaviors by 43.8% on average while preserving task accuracy. Further, when the self-monitor signal replaces an external weak judge in RL fine-tuning, models exhibit substantially fewer obfuscated thoughts and retain transparency. Our project website can be found at cot-monitor-plus.github.io

Kairui Ding, Boyuan Chen, Ruihai Wu et al. · tsinghua

Robotic manipulation with two-finger grippers is challenged by objects lacking distinct graspable features. Traditional pre-grasping methods, which typically involve repositioning objects or utilizing external aids like table edges, are limited in their adaptability across different object categories and environments. To overcome these limitations, we introduce PreAfford, a novel pre-grasping planning framework incorporating a point-level affordance representation and a relay training approach. Our method significantly improves adaptability, allowing effective manipulation across a wide range of environments and object types. When evaluated on the ShapeNet-v2 dataset, PreAfford not only enhances grasping success rates by 69% but also demonstrates its practicality through successful real-world experiments. These improvements highlight PreAfford's potential to redefine standards for robotic handling of complex manipulation tasks in diverse settings.

Ahmed E. Hassan, Dayi Lin, Gopi Krishnan Rajbahadur et al.

Foundation models (FMs), such as Large Language Models (LLMs), have revolutionized software development by enabling new use cases and business models. We refer to software built using FMs as FMware. The unique properties of FMware (e.g., prompts, agents, and the need for orchestration), coupled with the intrinsic limitations of FMs (e.g., hallucination) lead to a completely new set of software engineering challenges. Based on our industrial experience, we identified 10 key SE4FMware challenges that have caused enterprise FMware development to be unproductive, costly, and risky. In this paper, we discuss these challenges in detail and state the path for innovation that we envision. Next, we present FMArts, which is our long-term effort towards creating a cradle-to-grave platform for the engineering of trustworthy FMware. Finally, we (i) show how the unique properties of FMArts enabled us to design and develop a complex FMware for a large customer in a timely manner and (ii) discuss the lessons that we learned in doing so. We hope that the disclosure of the aforementioned challenges and our associated efforts to tackle them will not only raise awareness but also promote deeper and further discussions, knowledge sharing, and innovative solutions across the software engineering discipline.

Lingyu Zhang, Zhengran Ji, Nicholas R Waytowich et al.

The recent rapid advancement of machine learning has been driven by increasingly powerful models with the growing availability of training data and computational resources. However, real-time decision-making tasks with limited time and sparse learning signals remain challenging. One way of improving the learning speed and performance of these agents is to leverage human guidance. In this work, we introduce GUIDE, a framework for real-time human-guided reinforcement learning by enabling continuous human feedback and grounding such feedback into dense rewards to accelerate policy learning. Additionally, our method features a simulated feedback module that learns and replicates human feedback patterns in an online fashion, effectively reducing the need for human input while allowing continual training. We demonstrate the performance of our framework on challenging tasks with sparse rewards and visual observations. Our human study involving 50 subjects offers strong quantitative and qualitative evidence of the effectiveness of our approach. With only 10 minutes of human feedback, our algorithm achieves up to 30% increase in success rate compared to its RL baseline.