arXiv subjects

Simin Ma

Publications and source records attributed to Simin Ma.

14 recordsLinked to original sources

arXiv2026

Understanding Knowledge Distillation in Post-Training: When It Helps and When It Fails

Large language models (LLMs) achieve strong performance across many tasks, but their high computational cost limits deployment in resource-constrained environments. Knowledge Distillation (KD) offers a practical solution by transferring knowledge from a teacher model of a larger size to a smaller student model. While prior work has mainly examined task-specific or small-scale settings, the post-training stage for building general instruction-following models has received limited attention. In this paper, we conduct a systematic study of KD in post-training using the large-scale Tulu 3 dataset. We find that KD outperforms supervised fine-tuning (SFT) in low-data regimes, but its advantage diminishes as more training data is added. Distilling from a stronger instruction-tuned teacher restores substantial gains even with abundant data, indicating that KD remains effective when the teacher provides knowledge that the student cannot easily acquire from the training data alone. We further study domain-specific, low-resource scenarios and propose a two-stage KD strategy that leverages synthetic teacher-labeled data followed by refinement on human annotations. This method consistently improves student performance, providing practical guidance for building compact models in data-scarce environments.

cs.CL↗

arXiv2026

CM2: Reinforcement Learning with Checklist Rewards for Multi-Turn and Multi-Step Agentic Tool Use

AI agents are increasingly used to solve real-world tasks by reasoning over multi-turn user interactions and invoking external tools. However, applying reinforcement learning to such settings remains difficult: realistic objectives often lack verifiable rewards and instead emphasize open-ended behaviors; moreover, RL for multi-turn, multi-step agentic tool use is still underexplored; and building and maintaining executable tool environments is costly, limiting scale and coverage. We propose CM2, an RL framework that replaces verifiable outcome rewards with checklist rewards. CM2 decomposes each turn's intended behavior into fine-grained binary criteria with explicit evidence grounding and structured metadata, turning open-ended judging into more stable classification-style decisions. To balance stability and informativeness, our method adopts a strategy of sparse reward assignment but dense evaluation criteria. Training is performed in a scalable LLM-simulated tool environment, avoiding heavy engineering for large tool sets. Experiments show that CM2 consistently improves over supervised fine-tuning. Starting from an 8B Base model and training on an 8k-example RL dataset, CM2 improves over the SFT counterpart by 8 points on tau^-Bench, by 10 points on BFCL-V4, and by 12 points on ToolSandbox. The results match or even outperform similarly sized open-source baselines, including the judging model. CM2 thus provides a scalable recipe for optimizing multi-turn, multi-step tool-using agents without relying on verifiable rewards. Code provided by the open-source community: https://github.com/namezhenzhang/CM2-RLCR-Tool-Agent.

cs.AI↗

arXiv2025

Communication to Completion: Modeling Collaborative Workflows with Intelligent Multi-Agent Communication

Multi-agent LLM systems have demonstrated impressive capabilities in complex collaborative tasks, yet most frameworks treat communication as instantaneous and free, overlooking a fundamental constraint in real world teamwork, collaboration cost. We propose a scalable framework implemented via Communication to Completion (C2C), which explicitly models communication as a constrained resource with realistic temporal costs. We introduce the Alignment Factor (AF), a dynamic metric inspired by Shared Mental Models, to quantify the link between task understanding and work efficiency. Through experiments on 15 software engineering workflows spanning three complexity tiers and team sizes from 5 to 17 agents, we demonstrate that cost-aware strategies achieve over 40% higher efficiency compared to unconstrained interaction. Our analysis reveals emergent coordination patterns: agents naturally adopt manager centric hub-and-spoke topologies, strategically escalate from asynchronous to synchronous channels based on complexity, and prioritize high value help requests. These patterns remain consistent across multiple frontier models (GPT-5.2, Claude Sonnet 4.5, Gemini 2.5 Pro). This study moves beyond simple agent construction, offering a theoretical foundation for quantifying and optimizing the dynamics of collaboration in future digital workplaces.

cs.MA↗

arXiv2025

Aligning Multilingual Reasoning with Verifiable Semantics from a High-Resource Expert Model

While reinforcement learning has advanced the reasoning abilities of Large Language Models (LLMs), these gains are largely confined to English, creating a significant performance disparity across languages. To address this, we introduce Pivot-Based Reinforcement Learning with Semantically Verifiable Rewards (PB-RLSVR), a novel framework that enhances multilingual reasoning by circumventing the need for human-annotated data in target languages. Our approach employs a high-performing English LLM as a "pivot" model to generate reference responses for reasoning tasks. A multilingual model is then rewarded based on the semantic equivalence of its responses to the English reference, effectively transferring the pivot model's reasoning capabilities across languages. We investigate several cross-lingual semantic reward functions, including those based on embeddings and machine translation. Extensive experiments on a suite of multilingual reasoning benchmarks show that our method significantly narrows the performance gap between English and other languages, substantially outperforming traditional PPO baselines. Specifically, our PB-RLSVR framework improves the average multilingual performance of Llama-3.1-8B-Instruct and Qwen3-32B by 16.41% and 10.17%, respectively, demonstrating a powerful and data-efficient approach to building truly multilingual reasoning agents.

cs.CL↗

arXiv2025

TCIA: A Task-Centric Instruction Augmentation Method for Instruction Finetuning

Diverse instruction data is vital for effective instruction tuning of large language models, as it enables the model to generalize across different types of inputs . Building such diversified instruction dataset is an essential step in this process. Existing approaches often leverage large language models to automatically explore and generate diverse instructions, ensuring both data diversity and quality. However, they tend to overlook an important factor in real-world applications: on-task relevance. In practice, only a few real-world applications require a truly general-purpose model; most benefit from task-specific knowledge tailored to their particular use case. Therefore, it is vital to develop instruction augmentation methods that not only maintain diversity but are also optimized for specific, real-world scenarios. We thus introduce Task Centric Instruction Augmentation (TCIA), a framework that systematically expands instructions while preserving both diversity and task alignment. By representing instructions in a discrete query-constraints space, TCIA creates a rich set of task-relevant instructions and enables models to generalize to these task-specific instructions without sacrificing overall performance. Experiments show that TCIA improves open-source LLMs' performance by an average of 8.7% across four real-world, task-specific applications, and in some cases outperforming leading closed-source models. These improvements do not compromise general instruction-following ability, making TCIA a scalable and efficient solution for adapting LLMs to real-world, task-focused applications.

cs.AI↗

arXiv2025

LiveMCP-101: Stress Testing and Diagnosing MCP-enabled Agents on Challenging Queries

Tool calling has emerged as a critical capability for AI agents. In contrast to conventional tool calling frameworks that rely on static, provider-specific tool definitions, the Model Context Protocol (MCP) offers a unified interface to discover and invoke tools dynamically. However, there is a significant gap in benchmarking multi-step tasks using diverse MCP tools in realistic, dynamic scenarios. In this work, we present LiveMCP-101, a benchmark of 101 real-world queries that require coordinated use of multiple MCP tools. To address temporal variability in real-world tool responses, we introduce a parallel evaluation framework where a reference agent executes a validated plan simultaneously to produce real-time reference outputs. Experiments show that even frontier LLMs achieve a success rate below 60\%, highlighting challenges in multi-step tool use. Comprehensive error analysis identifies seven failure modes spanning tool planning, parameterization, and output handling, pointing to concrete directions for improving current models. LiveMCP-101 sets a rigorous standard for evaluating real-world agent capabilities, advancing toward autonomous agent systems that reliably execute complex tasks through MCP tool orchestration.

cs.CL↗

arXiv2025

LogicIF: Towards Complex Logic Instruction Following

Instruction following has catalyzed the recent era of Large Language Models (LLMs) and is the foundational skill underpinning more advanced capabilities such as reasoning and agentic behaviors. As tasks grow more challenging, the logic structures embedded in natural language instructions becomes increasingly intricate. However, how well LLMs perform on such logic-intensive instructions remains under-explored. We propose Logic Instruction Following (LogicIF), a task requiring models to precisely follow and simulate every logic step within complex instructions. To generate these instructions, we construct LogicIFGen, a scalable, automated framework for generating verifiable instructions from code functions, which can naturally express rich logic such as conditions, loops, nesting, and function calls. Using this framework, we curate a collection of complex code functions to construct LogicIFEval, a benchmark comprising 426 verifiable, human-examined, logic-rich instructions. Our experiments demonstrate that current state-of-the-art LLMs still struggle to correctly follow the instructions in LogicIFEval. Most LLMs can only follow fewer than 60% of the instructions, revealing significant deficiencies in their capacity to handle instructions that involve complex logic structures. To further demonstrate the efficacy of our framework, we collect an extensive set of code functions across a wide range of logic difficulties to generate LogicIFTrain, a fully synthesized training set featuring 27,324 verifiable instructions. Our results show that a model trained on LogicIFTrain using reinforcement learning significantly outperforms the base model on both in-domain and out-of-domain benchmarks. All the data, codes and models are released at https://github.com/mianzhang/LogicIF.

cs.CL↗

arXiv2025

Taming SQL Complexity: LLM-Based Equivalence Evaluation for Text-to-SQL

The rise of Large Language Models (LLMs) has significantly advanced Text-to-SQL (NL2SQL) systems, yet evaluating the semantic equivalence of generated SQL remains a challenge, especially given ambiguous user queries and multiple valid SQL interpretations. This paper explores using LLMs to assess both semantic and a more practical "weak" semantic equivalence. We analyze common patterns of SQL equivalence and inequivalence, discuss challenges in LLM-based evaluation.

cs.CL↗

arXiv2024

Inverse Probability of Treatment Weighting with Deep Sequence Models Enables Accurate treatment effect Estimation from Electronic Health Records

Observational data have been actively used to estimate treatment effect, driven by the growing availability of electronic health records (EHRs). However, EHRs typically consist of longitudinal records, often introducing time-dependent confoundings that hinder the unbiased estimation of treatment effect. Inverse probability of treatment weighting (IPTW) is a widely used propensity score method since it provides unbiased treatment effect estimation and its derivation is straightforward. In this study, we aim to utilize IPTW to estimate treatment effect in the presence of time-dependent confounding using claims records. Previous studies have utilized propensity score methods with features derived from claims records through feature processing, which generally requires domain knowledge and additional resources to extract information to accurately estimate propensity scores. Deep sequence models, particularly recurrent neural networks and self-attention-based architectures, have demonstrated good performance in modeling EHRs for various downstream tasks. We propose that these deep sequence models can provide accurate IPTW estimation of treatment effect by directly estimating the propensity scores from claims records without the need for feature processing. We empirically demonstrate this by conducting comprehensive evaluations using synthetic and semi-synthetic datasets.

cs.LG↗

arXiv2023

Deep Attention Q-Network for Personalized Treatment Recommendation

Tailoring treatment for individual patients is crucial yet challenging in order to achieve optimal healthcare outcomes. Recent advances in reinforcement learning offer promising personalized treatment recommendations; however, they rely solely on current patient observations (vital signs, demographics) as the patient's state, which may not accurately represent the true health status of the patient. This limitation hampers policy learning and evaluation, ultimately limiting treatment effectiveness. In this study, we propose the Deep Attention Q-Network for personalized treatment recommendations, utilizing the Transformer architecture within a deep reinforcement learning framework to efficiently incorporate all past patient observations. We evaluated the model on real-world sepsis and acute hypotension cohorts, demonstrating its superiority to state-of-the-art models. The source code for our model is available at https://github.com/stevenmsm/RL-ICU-DAQN.

cs.LG↗

arXiv2022

COVID-19 and Influenza Joint Forecasts Using Internet Search Information in the United States

As COVID-19 pandemic progresses, severe flu seasons may happen alongside an increase in cases in cases and death of COVID-19, causing severe burdens on health care resources and public safety. A consequence of a twindemic may be a mixture of two different infections in the same person at the same time, "flurona". Admist the raising trend of "flurona", forecasting both influenza outbreaks and COVID-19 waves in a timely manner is more urgent than ever, as accurate joint real-time tracking of the twindemic aids health organizations and policymakers in adequate preparation and decision making. Under the current pandemic, state-of-art influenza and COVID-19 forecasting models carry valuable domain information but face shortcomings under current complex disease dynamics, such as similarities in symptoms and public healthcare seeking patterns of the two diseases. Inspired by the inner-connection between influenza and COVID-19 activities, we propose ARGOX-Joint-Ensemble which allows us to combine historical influenza and COVID-19 disease forecasting models to a new ensemble framework that handles scenarios where flu and COVID co-exist. Our framework is able to emphasize learning from COVID-related or influenza signals, through a winner-takes-all ensemble fashion. Moreover, our experiments demonstrate that our approach is successful in adapting past influenza forecasting models to the current pandemic, while improving upon previous COVID-19 forecasting models, by steadily outperforming alternative benchmark methods, and remaining competitive with publicly available models.

stat.AP↗

arXiv2022

COVID-19 Hospitalizations Forecasts Using Internet Search Data

As the COVID-19 spread over the globe and new variants of COVID-19 keep occurring, reliable real-time forecasts of COVID-19 hospitalizations are critical for public health decision on medical resources allocations such as ICU beds, ventilators, and personnel to prepare for the surge of COVID-19 pandemics. Inspired by the strong association between public search behavior and hospitalization admission, we extended previously-proposed influenza tracking model, ARGO (AutoRegression with GOogle search data), to predict future 2-week national and state-level COVID-19 new hospital admissions. Leveraging the COVID-19 related time series information and Google search data, our method is able to robustly capture new COVID-19 variants' surges, and self-correct at both national and state level. Based on our retrospective out-of-sample evaluation over 12-month comparison period, our method achieves on average 15\% error reduction over the best alternative models collected from COVID-19 forecast hub. Overall, we showed that our method is flexible, self-correcting, robust, accurate, and interpretable, making it a potentially powerful tool to assist health-care officials and decision making for the current and future infectious disease outbreak.

cs.LG↗

arXiv2021

COVID-19 Forecasts Using Internet Search Information in the United States

As the COVID-19 ravaging through the globe, accurate forecasts of the disease spread is crucial for situational awareness, resource allocation, and public health decision-making. Alternative to the traditional disease surveillance data collected by the United States (US) Centers for Disease Control and Prevention (CDC), big data from Internet such as online search volumes has been previously shown to contain valuable information for tracking infectious disease dynamics. In this study, we evaluate the feasibility of using Internet search volume of relevant queries to track and predict COVID-19 pandemic. We found strong association between COVID-19 death trend and the search volume of symptom-related queries such as "loss of taste". Then, we further develop an influenza-tracking model to predict future 4-week COVID-19 deaths on the US national level, by combining search volume information with COVID-19 time series information. Encouraged by the 20% error reduction on national level comparing to the baseline time series model, we additionally build state-level COVID-19 deaths models, leveraging the cross-state cross-resolution spatial temporal framework that pools information from search volume and COVID-19 reports across states, regions and the nation. These variants of ARGOX are then aggregated in a winner-takes-all ensemble fashion to produce the final state-level 4-week forecasts. Numerical experiments demonstrate that our method steadily outperforms time series baseline models, and achieves the state-of-the-art performance among the publicly available benchmark models. Overall, we show that disease dynamics and relevant public search behaviors co-evolve during the COVID-19 pandemic, and capturing their dependencies while leveraging historical cases/deaths as well as spatial-temporal cross-region information will enable stable and accurate US national and state-level forecasts.

stat.AP↗

arXiv2021

MAGI-X: Manifold-Constrained Gaussian Process Inference for Unknown System Dynamics

Ordinary differential equations (ODEs), commonly used to characterize the dynamic systems, are difficult to propose in closed-form for many complicated scientific applications, even with the help of domain expert. We propose a fast and accurate data-driven method, MAGI-X, to learn the unknown dynamic from the observation data in a non-parametric fashion, without the need of any domain knowledge. Unlike the existing methods that mainly rely on the costly numerical integration, MAGI-X utilizes the powerful functional approximator of neural network to learn the unknown nonlinear dynamic within the MAnifold-constrained Gaussian process Inference (MAGI) framework that completely circumvents the numerical integration. Comparing against the state-of-the-art methods on three realistic examples, MAGI-X achieves competitive accuracy in both fitting and forecasting while only taking a fraction of computational time. Moreover, MAGI-X provides practical solution for the inference of partial observed systems, which no previous method is able to handle.

stat.ML↗