Enpei Zhang

Enpei Zhang, Shuo Tang, Xiaowen Dong et al.

Multi-agent learning has gained increasing attention to tackle distributed machine learning scenarios under constrictions of data exchanging. However, existing multi-agent learning models usually consider data fusion under fixed and compulsory collaborative relations among agents, which is not as flexible and autonomous as human collaboration. To fill this gap, we propose a distributed multi-agent learning model inspired by human collaboration, in which the agents can autonomously detect suitable collaborators and refer to collaborators' model for better performance. To implement such adaptive collaboration, we use a collaboration graph to indicate the pairwise collaborative relation. The collaboration graph can be obtained by graph learning techniques based on model similarity between different agents. Since model similarity can not be formulated by a fixed graphical optimization, we design a graph learning network by unrolling, which can learn underlying similar features among potential collaborators. By testing on both regression and classification tasks, we validate that our proposed collaboration model can figure out accurate collaborative relationship and greatly improve agents' learning performance.

Jiayu Li, Enpei Zhang, Dawei Zhou et al.

We study workflow learning in a setting where specialized agents hand off control through a shared artifact, each agent observes only a local function of that artifact and its own private state, and no centralized learner accesses joint trajectories -- the operating regime of multi-agent LLM pipelines that span organizational, vendor, or trust boundaries. We formalize this regime as an interface-constrained semi-Markov decision process (IC-SMDP), whose decision epochs occur at handoff times, and design IC-$Q$, an asynchronous decentralized $Q$-learning algorithm in which cross-agent coordination at every handoff is exactly one scalar. Our main result is a finite-sample bound for neural IC-$Q$ that decomposes into three independently controllable error sources: neural function-approximation error, interface representation gap, and a mixing-time residual, under the random option-duration discount. Establishing this bound requires lifting the approximate information state (AIS) framework from single-agent primitive-step MDPs to multi-agent SMDPs and controlling Markovian noise under random duration, neither of which has been done in prior work. To our knowledge this is the first finite-sample guarantee for neural $Q$-learning under decentralized partial observability. Four experiments: a controlled synthetic IC-SMDP that validates the bound term-by-term, multi-LLM mathematical reasoning, multi-agent routing, and multi-agent CPU programming, show that IC-$Q$ matches a centralized oracle without any agent observing joint trajectories, with each of the three error sources scaling along its corresponding axis as the bound predicts.

Jinhang Chai, Enpei Zhang, Elynn Chen et al.

We study online transfer reinforcement learning (RL) in episodic Markov decision processes, where experience from related source tasks is available during learning on a target task. A fundamental difficulty is that task similarity is typically defined in terms of rewards or transitions, whereas online RL algorithms operate on Bellman regression targets. As a result, naively reusing source Bellman updates introduces systematic bias and invalidates regret guarantees. We identify one-step Bellman alignment as the correct abstraction for transfer in online RL and propose re-weighted targeting (RWT), an operator-level correction that retargets continuation values and compensates for transition mismatch via a change of measure. RWT reduces task mismatch to a fixed one-step correction and enables statistically sound reuse of source data. This alignment yields a two-stage RWT $Q$-learning framework that separates variance reduction from bias correction. Under RKHS function approximation, we establish regret bounds that scale with the complexity of the task shift rather than the target MDP. Empirical results in both tabular and neural network settings demonstrate consistent improvements over single-task learning and naïve pooling, highlighting Bellman alignment as a model-agnostic transfer principle for online RL.

Enpei Zhang, Jingyi Chai, Rui Ye et al.

While data plays a crucial role in training contemporary AI models, it is acknowledged that valuable public data will be exhausted in a few years, directing the world's attention towards the massive decentralized private data. However, the privacy-sensitive nature of raw data and lack of incentive mechanism prevent these valuable data from being fully exploited. Addressing these challenges, this paper proposes inclusive and incentivized personalized federated learning (iPFL), which incentivizes data holders with diverse purposes to collaboratively train personalized models without revealing raw data. iPFL constructs a model-sharing market by solving a graph-based training optimization and incorporates an incentive mechanism based on game theory principles. Theoretical analysis shows that iPFL adheres to two key incentive properties: individual rationality and truthfulness. Empirical studies on eleven AI tasks (e.g., large language models' instruction-following tasks) demonstrate that iPFL consistently achieves the highest economic utility, and better or comparable model performance compared to baseline methods. We anticipate that our iPFL can serve as a valuable technique for boosting future AI models on decentralized private data while making everyone satisfied.

Zheng Huang, Enpei Zhang, Yinghao Cai et al.

Understanding how the brain encodes visual information is a central challenge in neuroscience and machine learning. A promising approach is to reconstruct visual stimuli, essentially images, from functional Magnetic Resonance Imaging (fMRI) signals. This involves two stages: transforming fMRI signals into a latent space and then using a pretrained generative model to reconstruct images. The reconstruction quality depends on how similar the latent space is to the structure of neural activity and how well the generative model produces images from that space. Yet, it remains unclear which type of latent space best supports this transformation and how it should be organized to represent visual stimuli effectively. We present two key findings. First, fMRI signals are more similar to the text space of a language model than to either a vision based space or a joint text image space. Second, text representations and the generative model should be adapted to capture the compositional nature of visual stimuli, including objects, their detailed attributes, and relationships. Building on these insights, we propose PRISM, a model that Projects fMRI sIgnals into a Structured text space as an interMediate representation for visual stimuli reconstruction. It includes an object centric diffusion module that generates images by composing individual objects to reduce object detection errors, and an attribute relationship search module that automatically identifies key attributes and relationships that best align with the neural activity. Extensive experiments on real world datasets demonstrate that our framework outperforms existing methods, achieving up to an 8% reduction in perceptual loss. These results highlight the importance of using structured text as the intermediate space to bridge fMRI signals and image reconstruction.

Chiyu Ma, Enpei Zhang, Yilun Zhao et al.

LLM-as-Judge has emerged as a scalable alternative to human evaluation, enabling large language models (LLMs) to provide reward signals in trainings. While recent work has explored multi-agent extensions such as multi-agent debate and meta-judging to enhance evaluation quality, the question of how intrinsic biases manifest in these settings remains underexplored. In this study, we conduct a systematic analysis of four diverse bias types: position bias, verbosity bias, chain-of-thought bias, and bandwagon bias. We evaluate these biases across two widely adopted multi-agent LLM-as-Judge frameworks: Multi-Agent-Debate and LLM-as-Meta-Judge. Our results show that debate framework amplifies biases sharply after the initial debate, and this increased bias is sustained in subsequent rounds, while meta-judge approaches exhibit greater resistance. We further investigate the incorporation of PINE, a leading single-agent debiasing method, as a bias-free agent within these systems. The results reveal that this bias-free agent effectively reduces biases in debate settings but provides less benefit in meta-judge scenarios. Our work provides a comprehensive study of bias behavior in multi-agent LLM-as-Judge systems and highlights the need for targeted bias mitigation strategies in collaborative evaluation settings.