Yuchen Ma

Yidong Gong, Saima Afrin, Yuchen Ma et al.

Modern deep learning workloads increasingly exhibit dynamic, metadata-driven execution, where runtime-generated information determines memory provisioning and kernel launch decisions. In sampling-based graph neural network (GNN) training, this behavior places the CPU on the critical path, introducing persistent host-device orchestration overhead and frequent GPU-CPU synchronization, which dominate end-to-end runtime when GPU computation is small. Existing approaches, including CUDA Graphs and GPU dynamic parallelism, fail to address this problem because the metadata-driven control loop remains host-mediated, and execution structure varies across iterations. We present ZEROGNN, a system that removes the host from the metadata-driven control loop and enables fully GPU-resident execution under dynamic behavior. ZEROGNN keeps runtime metadata on-device, mediates dynamic execution within a fixed launch structure, and provisions a conservative yet tight execution envelope to restore CUDA Graph replayability. Experiments on sampling-based GNN workloads show that ZEROGNN achieves up to 5.28 x end-to-end speedup, near 100% GPU execution fraction, and memory efficiency comparable to ideal metadata-informed allocation, while enabling strong multi-GPU scaling by eliminating host-side bottlenecks.

Yuchen Ma, Yanbei Chen, Zeynep Akata

Recent advances have indicated the strengths of self-supervised pre-training for improving representation learning on downstream tasks. Existing works often utilize self-supervised pre-trained models by fine-tuning on downstream tasks. However, fine-tuning does not generalize to the case when one needs to build a customized model architecture different from the self-supervised model. In this work, we formulate a new knowledge distillation framework to transfer the knowledge from self-supervised pre-trained models to any other student network by a novel approach named Embedding Graph Alignment. Specifically, inspired by the spirit of instance discrimination in self-supervised learning, we model the instance-instance relations by a graph formulation in the feature embedding space and distill the self-supervised teacher knowledge to a student network by aligning the teacher graph and the student graph. Our distillation scheme can be flexibly applied to transfer the self-supervised knowledge to enhance representation learning on various student networks. We demonstrate that our model outperforms multiple representative knowledge distillation methods on three benchmark datasets, including CIFAR100, STL10, and TinyImageNet. Code is here: https://github.com/yccm/EGA.

Dennis Frauen, Marie Brockschmidt, Konstantin Hess et al.

Large language model (LLM) development is currently driven by large-scale empirical iteration over data mixtures, reward models, routing strategies, and evaluation pipelines. Here, we argue that many central questions in LLM development and evaluation are inherently causal: What is the effect of adding a data domain during pretraining? How do annotator preferences change when LLMs generate text in a different style? Should a prompt be routed to a larger or smaller model given inference cost constraints? In general, causal methods are well-suited to such settings where interventions change outcomes but, surprisingly, are underrepresented in LLM development. Our contribution is threefold: (1) We explain how causal methods can help develop modern LLM development and evaluation: LLM development relies heavily on logged data, which are often subject to confounding and distribution shifts; evaluation uses learned but potentially biased judges; and deployment environments are non-stationary. These conditions make purely predictive approaches fragile and create opportunities for principled identification and estimation methods from causal inference. (2) We further map opportunities for causal methods in the entire LLM development pipeline, including pretraining, alignment, routing, agentic workflows, and evaluation. (3) We discuss new research opportunities around leveraging causal methods for LLM development and evaluation. Overall, we argue that causal methods are potentially underutilized for the LLM development and evaluation pipeline, despite the fact that such methods can ensure a reliable and scientifically grounded design.

Haomin Zhuang, Hanwen Xing, Yujun Zhou et al.

Third-party skills are becoming the package ecosystem for LLM agents. They package natural-language instructions, helper scripts, templates, documents, and service configuration into reusable workflows. This makes skills useful, but it also introduces a new security problem: a malicious skill does not need to ask the model to perform an obviously harmful action. Instead, it can disguise the harmful behavior as part of a routine workflow, relying on the agent to execute that workflow with high-value permissions and limited human supervision. We introduce AgentTrap, a dynamic benchmark for evaluating whether LLM agents can use third-party skills while resisting malicious runtime behavior. AgentTrap contains 141 tasks: 91 malicious tasks and 50 benign utility tasks, covering 16 security-impact dimensions grounded in agent-skill supply-chain threats. In each task, the agent receives an ordinary user request, runs with installed skills that may contain malicious workflow elements, and is executed in a sandboxed environment. AgentTrap then judges complete trajectories for attack success, blocked or refused behavior, attack-not-triggered cases, and no-attack-evidence outcomes. Our central finding is that the most informative failures are not simple jailbreaks. Models often complete the visible user task while treating unsafe side effects introduced by the skill as part of the normal workflow. This motivates runtime evaluation of the concrete model--framework--workspace environment in which users actually delegate work. Code and data are available at https://github.com/zhmzm/AgentTrap and https://huggingface.co/datasets/zhmzm/AgentTrap.

Yue Huang, Yu Jiang, Wenjie Wang et al.

Multi-agent systems composed of large generative models are rapidly moving from laboratory prototypes to real-world deployments, where they jointly plan, negotiate, and allocate shared resources to solve complex tasks. While such systems promise unprecedented scalability and autonomy, their collective interaction also gives rise to failure modes that cannot be reduced to individual agents. Understanding these emergent risks is therefore critical. Here, we present a pioneer study of such emergent multi-agent risk in workflows that involve competition over shared resources (e.g., computing resources or market share), sequential handoff collaboration (where downstream agents see only predecessor outputs), collective decision aggregation, and others. Across these settings, we observe that such group behaviors arise frequently across repeated trials and a wide range of interaction conditions, rather than as rare or pathological cases. In particular, phenomena such as collusion-like coordination and conformity emerge with non-trivial frequency under realistic resource constraints, communication protocols, and role assignments, mirroring well-known pathologies in human societies despite no explicit instruction. Moreover, these risks cannot be prevented by existing agent-level safeguards alone. These findings expose the dark side of intelligent multi-agent systems: a social intelligence risk where agent collectives, despite no instruction to do so, spontaneously reproduce familiar failure patterns from human societies.

Yuchen Ma, Zhengcong Fei, Junshi Huang

Recently, the tokens of images share the same static data flow in many dense networks. However, challenges arise from the variance among the objects in images, such as large variations in the spatial scale and difficulties of recognition for visual entities. In this paper, we propose a data-dependent token routing strategy to elaborate the routing paths of image tokens for Dynamic Vision Transformer, dubbed DiT. The proposed framework generates a data-dependent path per token, adapting to the object scales and visual discrimination of tokens. In feed-forward, the differentiable routing gates are designed to select the scaling paths and feature transformation paths for image tokens, leading to multi-path feature propagation. In this way, the impact of object scales and visual discrimination of image representation can be carefully tuned. Moreover, the computational cost can be further reduced by giving budget constraints to the routing gate and early-stopping of feature extraction. In experiments, our DiT achieves superior performance and favorable complexity/accuracy trade-offs than many SoTA methods on ImageNet classification, object detection, instance segmentation, and semantic segmentation. Particularly, the DiT-B5 obtains 84.8\% top-1 Acc on ImageNet with 10.3 GFLOPs, which is 1.0\% higher than that of the SoTA method with similar computational complexity. These extensive results demonstrate that DiT can serve as versatile backbones for various vision tasks.

Yuchen Ma, Valentyn Melnychuk, Dennis Frauen et al.

Fairness in predictions is of direct importance in practice due to legal, ethical, and societal reasons. This is often accomplished through counterfactual fairness, which ensures that the prediction for an individual is the same as that in a counterfactual world under a different sensitive attribute. However, achieving counterfactual fairness is challenging as counterfactuals are unobservable, and, because of that, existing baselines for counterfactual fairness do not have theoretical guarantees. In this paper, we propose a novel counterfactual fairness predictor for making predictions under counterfactual fairness. Here, we follow the standard counterfactual fairness setting and directly learn the counterfactual distribution of the descendants of the sensitive attribute via tailored neural networks, which we then use to enforce fair predictions through a novel counterfactual mediator regularization. Unique to our work is that we provide theoretical guarantees that our method is effective in ensuring the notion of counterfactual fairness. We further compare the performance across various datasets, where our method achieves state-of-the-art performance.

Yue Huang, Zhengzhe Jiang, Yuchen Ma et al.

Understanding how and why large language models (LLMs) fail is becoming a central challenge as models rapidly evolve and static evaluations fall behind. While automated probing has been enabled by dynamic test generation, existing approaches often discover isolated failure cases, lack principled control over exploration, and provide limited insight into the underlying structure of model weaknesses. We propose ProbeLLM, a benchmark-agnostic automated probing framework that elevates weakness discovery from individual failures to structured failure modes. ProbeLLM formulates probing as a hierarchical Monte Carlo Tree Search, explicitly allocating limited probing budgets between global exploration of new failure regions and local refinement of recurring error patterns. By restricting probing to verifiable test cases and leveraging tool-augmented generation and verification, ProbeLLM grounds failure discovery in reliable evidence. Discovered failures are further consolidated into interpretable failure modes via failure-aware embeddings and boundary-aware induction. Across diverse benchmarks and LLMs, ProbeLLM reveals substantially broader, cleaner, and more fine-grained failure landscapes than static benchmarks and prior automated methods, supporting a shift from case-centric evaluation toward principled weakness discovery.

Yuchen Ma, Yue Huang, Wenjie Wang et al.

Personalized prompting offers large opportunities for deploying large language models (LLMs) to diverse users, yet existing prompt optimization methods primarily focus on task-level optimization while largely overlooking user-specific preferences and latent constraints of individual users. This gap is primarily due to (i) the absence of high-quality, privacy-sensitive data that capture personalized user-LLM interactions at scale, and (ii) the lack of robust reward signals for individual preferences. To overcome existing data limitations, we introduce a high-fidelity synthetic data generation framework called PersonaGym. Unlike prior work that treats personalization as static persona-preference pairs, PersonaGym models a dynamic preference process via an agentic LLM system to simulate realistic preference behaviors and semantic-aware noise in order to generate personalized multi-turn interaction trajectories. Using PersonaGym, we release PersonaAtlas, a large-scale, high-quality, and diverse synthetic dataset of high-fidelity multi-turn personalized interaction trajectories that closely mirror real-world preference expression and noise patterns. We further propose Personalized Prompt Optimization (PPOpt), a scalable and model-agnostic framework that optimizes user prompts based on interaction histories without modifying the deployed LLM. PPOpt adopts a reason-then-optimize paradigm that infers an explicit user profile and conditions prompt rewriting on the user profile to avoid reward hacking. Our training procedure for PPOpt integrates a cold-start supervised prior with outcome-driven multi-objective reinforcement learning. We present extensive experiments to demonstrate consistent improvements over state-of-the-art baselines in terms of task performance, personalization quality, and robustness to noisy as well as to sparse preference signals.

Yuxin Wang, Dennis Frauen, Emil Javurek et al.

The synthetic control method (SCM) estimates causal effects in panel data with a single-treated unit by constructing a counterfactual outcome as a weighted combination of untreated control units that matches the pre-treatment trajectory. In this paper, we introduce the targeted synthetic control (TSC) method, a new two-stage estimator that directly estimates the counterfactual outcome. Specifically, our TSC method (1) yields a targeted debiasing estimator, in the sense that the targeted updating refines the initial weights to produce more stable weights; and (2) ensures that the final counterfactual estimation is a convex combination of observed control outcomes to enable direct interpretation of the synthetic control weights. TSC is flexible and can be instantiated with arbitrary machine learning models. Methodologically, TSC starts from an initial set of synthetic-control weights via a one-dimensional targeted update through the weight-tilting submodel, which calibrates the weights to reduce bias of weights estimation arising from pre-treatment fit. Furthermore, TSC avoids key shortcomings of existing methods (e.g., the augmented SCM), which can produce unbounded counterfactual estimates. Across extensive synthetic and real-world experiments, TSC consistently improves estimation accuracy over state-of-the-art SCM baselines.

Yuchen Ma, Yue Huang, Han Bao et al.

Skills are a promising way to improve LLM agent capabilities without retraining, while keeping the added procedure reusable and controllable. However, high-quality skills are still largely written by hand. We introduce SkillGen, a multi-agent framework that synthesizes a single auditable skill from trajectories generated by a base agent. The output is a human-readable artifact that can be inspected before use. Rather than merely summarizing trajectories, SkillGen leverages contrastive induction over both successful and failed trajectories to identify reusable success patterns, recurring failure modes, and behaviors that appear in nearby successes but are missing from failures. SkillGen then generates candidate skills and iteratively refines the skill. A key novelty in SkillGen is that we model agent skills as interventions to empirically verify the net effect of skills on the overall performance. Specifically, we compare outcomes on the same instances with and without the skill, so that we account for both repairs (cases where the skill fixes a baseline failure) and regressions (cases where the skill breaks a baseline success). Across a broad range of agents and datasets, SkillGen consistently improves held-out performance, outperforms existing skill-generation baselines, and produces skills that transfer across models.

Yue Huang, Yuchen Ma, Jiayi Ye et al.

Interactive narrative tasks require LLMs to sustain a coherent, evolving story while adapting to a user over multiple turns. However, suitable benchmarks for this setting are limited: existing evaluations often focus on static prompts, isolated story generations, or post-hoc ratings, and therefore miss whether models can jointly manage story generation, long-context state and pacing, character simulation, empathic personalization, and story-grounded artifacts. We introduce NARRA-Gym, an executable evaluation environment that turns a sparse emotional seed into a complete interactive story episode and logs the full model-in-the-loop trajectory, including story construction, memory updates, planning, pacing interventions, and optional artifact synthesis. We evaluate nine frontier LLMs using a controlled LLM-as-judge sweep over eight benchmark personas and a human evaluation in which participants rate customized model outputs. Our results show substantial variation across models, personas, and evaluation dimensions: models that produce fluent stories can still fail on robustness, user experience, or resistance-sensitive personalization. These findings suggest that interactive narrative offers a useful benchmark for evaluating long-horizon, user-adaptive LLM behavior beyond isolated story quality.

Han Qiu, Yuchen Ma, Zeming Li et al.

Dense object detectors rely on the sliding-window paradigm that predicts the object over a regular grid of image. Meanwhile, the feature maps on the point of the grid are adopted to generate the bounding box predictions. The point feature is convenient to use but may lack the explicit border information for accurate localization. In this paper, We propose a simple and efficient operator called Border-Align to extract "border features" from the extreme point of the border to enhance the point feature. Based on the BorderAlign, we design a novel detection architecture called BorderDet, which explicitly exploits the border information for stronger classification and more accurate localization. With ResNet-50 backbone, our method improves single-stage detector FCOS by 2.8 AP gains (38.6 v.s. 41.4). With the ResNeXt-101-DCN backbone, our BorderDet obtains 50.3 AP, outperforming the existing state-of-the-art approaches. The code is available at (https://github.com/Megvii-BaseDetection/BorderDet).

Yuchen Ma, Valentyn Melnychuk, Jonas Schweisthal et al.

Predicting potential outcomes of interventions from observational data is crucial for decision-making in medicine, but the task is challenging due to the fundamental problem of causal inference. Existing methods are largely limited to point estimates of potential outcomes with no uncertain quantification; thus, the full information about the distributions of potential outcomes is typically ignored. In this paper, we propose a novel causal diffusion model called DiffPO, which is carefully designed for reliable inferences in medicine by learning the distribution of potential outcomes. In our DiffPO, we leverage a tailored conditional denoising diffusion model to learn complex distributions, where we address the selection bias through a novel orthogonal diffusion loss. Another strength of our DiffPO method is that it is highly flexible (e.g., it can also be used to estimate different causal quantities such as CATE). Across a wide range of experiments, we show that our method achieves state-of-the-art performance.

Yuchen Ma, Dennis Frauen, Emil Javurek et al.

Prior-data fitted networks (PFNs) have recently been proposed as a promising way to train tabular foundation models. PFNs are transformers that are pre-trained on synthetic data generated from a prespecified prior distribution and that enable Bayesian inference through in-context learning. In this paper, we introduce CausalFM, a comprehensive framework for training PFN-based foundation models in various causal inference settings. First, we formalize the construction of Bayesian priors for causal inference based on structural causal models (SCMs) in a principled way and derive necessary criteria for the validity of such priors. Building on this, we propose a novel family of prior distributions using causality-inspired Bayesian neural networks that enable CausalFM to perform Bayesian causal inference in various settings, including for back-door, front-door, and instrumental variable adjustment. Finally, we instantiate CausalFM and explicitly train models to perform in-context learning in these settings. We show that CausalFM achieves competitive in-context learning performance even when compared to baselines that are specifically trained for the task at hand. In sum, our framework can be used as a general recipe to train foundation models for various causal inference settings. In contrast to the current state-of-the-art in causal inference, CausalFM offers a novel paradigm with the potential to fundamentally change how practitioners perform causal inference in medicine, economics, and other disciplines.

Yuchen Ma, Jonas Schweisthal, Hengrui Zhang et al.

In medicine, treatments often influence multiple, interdependent outcomes, such as primary endpoints, complications, adverse events, or other secondary endpoints. Hence, to make optimal treatment decisions, clinicians are interested in learning the distribution of multi-dimensional treatment outcomes. However, the vast majority of machine learning methods for predicting treatment effects focus on single-outcome settings, despite the fact that medical data often include multiple, interdependent outcomes. To address this limitation, we propose a novel diffusion-based method called DIME to learn the joint distribution of multiple outcomes of medical treatments. We addresses three challenges relevant in medical practice: (i)it is tailored to learn the joint interventional distribution of multiple medical outcomes, which enables reliable decision-making with uncertainty quantification rather than relying solely on point estimates; (ii)it explicitly captures the dependence structure between outcomes; (iii)it can handle outcomes of mixed type, including binary, categorical, and continuous variables. In DIME, we take into account the fundamental problem of causal inference through causal masking. For training, our method decomposes the joint distribution into a series of conditional distributions with a customized conditional masking to account for the dependence structure across outcomes. For inference, our method auto-regressively generates predictions. This allows our method to move beyond point estimates of causal quantities and thus learn the joint interventional distribution. To the best of our knowledge, DIME is the first neural method tailored to learn the joint, multi-outcome distribution of medical treatments. Across various experiments, we demonstrate that our method effectively learns the joint distribution and captures shared information among multiple outcomes.

Yuchen Ma, Dennis Frauen, Jonas Schweisthal et al.

Estimating treatment effects is crucial for personalized decision-making in medicine, but this task faces unique challenges in clinical practice. At training time, models for estimating treatment effects are typically trained on well-structured medical datasets that contain detailed patient information. However, at inference time, predictions are often made using textual descriptions (e.g., descriptions with self-reported symptoms), which are incomplete representations of the original patient information. In this work, we make three contributions. (1) We show that the discrepancy between the data available during training time and inference time can lead to biased estimates of treatment effects. We formalize this issue as an inference time text confounding problem, where confounders are fully observed during training time but only partially available through text at inference time. (2) To address this problem, we propose a novel framework for estimating treatment effects that explicitly accounts for inference time text confounding. Our framework leverages large language models together with a custom doubly robust learner to mitigate biases caused by the inference time text confounding. (3) Through a series of experiments, we demonstrate the effectiveness of our framework in real-world applications.

Zhibo Fan, Yuchen Ma, Zeming Li et al.

Recently few-shot object detection is widely adopted to deal with data-limited situations. While most previous works merely focus on the performance on few-shot categories, we claim that detecting all classes is crucial as test samples may contain any instances in realistic applications, which requires the few-shot detector to learn new concepts without forgetting. Through analysis on transfer learning based methods, some neglected but beneficial properties are utilized to design a simple yet effective few-shot detector, Retentive R-CNN. It consists of Bias-Balanced RPN to debias the pretrained RPN and Re-detector to find few-shot class objects without forgetting previous knowledge. Extensive experiments on few-shot detection benchmarks show that Retentive R-CNN significantly outperforms state-of-the-art methods on overall performance among all settings as it can achieve competitive results on few-shot classes and does not degrade the base class performance at all. Our approach has demonstrated that the long desired never-forgetting learner is available in object detection.

Yuchen Ma, Songtao Liu, Zeming Li et al.

We propose a dense object detector with an instance-wise sampling strategy, named IQDet. Instead of using human prior sampling strategies, we first extract the regional feature of each ground-truth to estimate the instance-wise quality distribution. According to a mixture model in spatial dimensions, the distribution is more noise-robust and adapted to the semantic pattern of each instance. Based on the distribution, we propose a quality sampling strategy, which automatically selects training samples in a probabilistic manner and trains with more high-quality samples. Extensive experiments on MS COCO show that our method steadily improves baseline by nearly 2.4 AP without bells and whistles. Moreover, our best model achieves 51.6 AP, outperforming all existing state-of-the-art one-stage detectors and it is completely cost-free in inference time.

Zeming Li, Yuchen Ma, Yukang Chen et al.

In this report, we present our object detection/instance segmentation system, MegDetV2, which works in a two-pass fashion, first to detect instances then to obtain segmentation. Our baseline detector is mainly built on a new designed RPN, called RPN++. On the COCO-2019 detection/instance-segmentation test-dev dataset, our system achieves 61.0/53.1 mAP, which surpassed our 2018 winning results by 5.0/4.2 respectively. We achieve the best results in COCO Challenge 2019 and 2020.

Lei Shi, Rui Guo, Yuchen Ma

Image pattern recognition is an important area in digital image processing. An efficient pattern recognition algorithm should be able to provide correct recognition at a reduced computational time. Off late amongst the machine learning pattern recognition algorithms, Artificial fish swarm algorithm is one of the swarm intelligence optimization algorithms that works based on population and stochastic search. In order to achieve acceptable result, there are many parameters needs to be adjusted in AFSA. Among these parameters, visual and step are very significant in view of the fact that artificial fish basically move based on these parameters. In standard AFSA, these two parameters remain constant until the algorithm termination. Large values of these parameters increase the capability of algorithm in global search, while small values improve the local search ability of the algorithm. In this paper, we empirically study the performance of the AFSA and different approaches to balance between local and global exploration have been tested based on the adaptive modification of visual and step during algorithm execution. The proposed approaches have been evaluated based on the four well-known benchmark functions. Experimental results show considerable positive impact on the performance of AFSA. A Convex optimization has been integrated into the proposed work to have an ideal segmentation of the input image which is a MR brain image.