Ziheng Liu

Chengdong Xu, Kaiqiang Ke, Ziheng Liu et al.

Large language model (LLM)-based multi-agent systems have shown strong potential on complex tasks through agent specialization, tool use, and collaborative reasoning. However, most automated multi-agent system design methods still follow a one-shot paradigm: a workflow is optimized or selected before execution and then reused unchanged throughout the task. This static coordination strategy is ill-suited for long-horizon tasks whose subgoals, intermediate evidence, and information needs evolve over multiple execution stages. We propose EvoMAS, a framework for execution-time multi-agent workflow construction. EvoMAS formulates workflow construction as a meta-level sequential decision problem along a single task trajectory. At each stage, it constructs an explicit task state through a Planner-Evaluator-Updater pipeline and uses a learned Workflow Adapter to instantiate a stage-specific layered workflow from a fixed pool of candidate agents. The adapter is trained with policy gradients using sparse, verifiable terminal task success as the main supervision signal, while evaluator-based process reward is analyzed separately under very-hard sparse-reward settings. Experiments on GAIA, HLE, and DeepResearcher show that EvoMAS outperforms single-agent baselines and recent automated multi-agent workflow design methods. Our analyses further show that explicit task-state construction and learned workflow adaptation provide complementary benefits. Additional results indicate that process reward is most useful when terminal success is extremely sparse, and qualitative case studies illustrate that EvoMAS adapts agent coordination as the task state evolves.

JiaWei Zhu, ZiHeng Liu

Hidden confounding remains a fundamental challenge in causal inference from observational data. Recent advances leverage Large Language Models (LLMs) to generate plausible hidden confounders based on domain knowledge, yet a critical gap exists: LLM-generated confounders often exhibit semantic plausibility without statistical utility. We propose VIGOR+ (Variational Information Gain for iterative cOnfounder Refinement), a novel framework that closes the loop between LLM-based confounder generation and CEVAE-based statistical validation. Unlike prior approaches that treat generation and validation as separate stages, VIGOR+ establishes an iterative feedback mechanism: validation signals from CEVAE (including information gain, latent consistency metrics, and diagnostic messages) are transformed into natural language feedback that guides subsequent LLM generation rounds. This iterative refinement continues until convergence criteria are met. We formalize the feedback mechanism, prove convergence properties under mild assumptions, and provide a complete algorithmic framework.

Ruanjun Li, Ziheng Liu, Yuanming Shi et al.

Large language models (LLMs) deliver impressive generation quality, but incur very high inference cost because each output token is generated auto-regressively through all model layers. Early-exit based self-speculative decoding (EESD) has emerged to mitigate this cost. However, in practice, many approaches struggle to achieve the expected acceleration in such draft-then-verify paradigm even with a well-aligned early-exit head and selected exit position. Our analysis reveals that EESD only pays off when the vast majority of draft tokens are accepted by the LLM. Otherwise, the draft cost may overcome the acceleration gain and lead to a negative speedup. To mitigate this, we propose Pipeline-Parallel Self-Speculative Decoding (PPSD) that fully pipelines the draft and verification work so that no effort is wasted on failed predictions. It has two key innovations. We configure the model layers as a pipeline in which early-exit (draft) computations and remaining-layer (verification) computations overlap. We interleave drafting and verification per token. While the LLM is verifying the current token in its final layers, the early-exit path simultaneously drafts the next token. Such a verify-while-draft scheme keeps all units busy and validates tokens on-the-fly analogous to pipelining the speculation and verification stages. Empirical results confirm that PPSD achieves state-of-the-art acceleration in self-speculative LLM inference. On diverse benchmarks, PPSD achieves speedup ratios in the range of 2.01x~3.81x, which gains almost the optimal acceleration at the fixed acceptance rate and exit position, showcasing its advancement in providing efficient self-speculation.

Zhuonan Liang, Ziheng Liu, Huaze Shi et al.

Recent years have witnessed the rapid growth of Small Private Online Courses (SPOC) which is able to highly customized and personalized to adapt variable educational requests, in which machine learning techniques are explored to summarize and predict the learner's performance, mostly focus on the final grade. However, the problem is that the final grade of learners on SPOC is generally seriously imbalance which handicaps the training of prediction model. To solve this problem, a sampling batch normalization embedded deep neural network (SBNEDNN) method is developed in this paper. First, a combined indicator is defined to measure the distribution of the data, then a rule is established to guide the sampling process. Second, the batch normalization (BN) modified layers are embedded into full connected neural network to solve the data imbalanced problem. Experimental results with other three deep learning methods demonstrates the superiority of the proposed method.