arXiv subjects

Jin Miao

Publications and source records attributed to Jin Miao.

6 recordsLinked to original sources

arXiv2025

Budget-Aware Tool Use Enables Effective Agent Scaling

Scaling test-time computation has been extended from language model reasoning to tool-augmented agents, where scaling involves not only thinking in tokens but also acting via tool calls that directly constrain environmental interaction. However, we found that simply increasing the tool-call budget fails to improve performance, as agents lack "budget awareness" and quickly hit a performance ceiling. We study how to scale such agents effectively under explicit tool-call budgets, focusing on web search agents. We first introduce the Budget Tracker, a lightweight plug-in that provides the agent with continuous budget awareness, enabling simple yet effective scaling. We further develop BATS (Budget-Aware Test-time Scaling), an advanced framework that leverages this awareness to dynamically adapt its planning and verification strategy. To analyze cost-performance scaling in a controlled manner, we formalize a unified cost metric that jointly accounts for token and tool consumption. We provide the first systematic study on budget-constrained agents, showing that budget-aware methods produce more favorable scaling curves and push the cost-performance Pareto frontier. Our work offers empirical insights toward a more transparent and principled understanding of scaling in tool-augmented agents. Our code is available at https://github.com/google-research/budget-aware-agent.

cs.AI↗

arXiv2024

CodecLM: Aligning Language Models with Tailored Synthetic Data

Instruction tuning has emerged as the key in aligning large language models (LLMs) with specific task instructions, thereby mitigating the discrepancy between the next-token prediction objective and users' actual goals. To reduce the labor and time cost to collect or annotate data by humans, researchers start to explore the use of LLMs to generate instruction-aligned synthetic data. Recent works focus on generating diverse instructions and applying LLM to increase instruction complexity, often neglecting downstream use cases. It remains unclear how to tailor high-quality data to elicit better instruction-following abilities in different target instruction distributions and LLMs. To this end, we introduce CodecLM, a general framework for adaptively generating high-quality synthetic data for LLM alignment with different downstream instruction distributions and LLMs. Drawing on the Encode-Decode principles, we use LLMs as codecs to guide the data generation process. We first encode seed instructions into metadata, which are concise keywords generated on-the-fly to capture the target instruction distribution, and then decode metadata to create tailored instructions. We also introduce Self-Rubrics and Contrastive Filtering during decoding to tailor data-efficient samples. Extensive experiments on four open-domain instruction following benchmarks validate the effectiveness of CodecLM over the current state-of-the-arts.

cs.CL↗

arXiv2023

DocumentNet: Bridging the Data Gap in Document Pre-Training

Document understanding tasks, in particular, Visually-rich Document Entity Retrieval (VDER), have gained significant attention in recent years thanks to their broad applications in enterprise AI. However, publicly available data have been scarce for these tasks due to strict privacy constraints and high annotation costs. To make things worse, the non-overlapping entity spaces from different datasets hinder the knowledge transfer between document types. In this paper, we propose a method to collect massive-scale and weakly labeled data from the web to benefit the training of VDER models. The collected dataset, named DocumentNet, does not depend on specific document types or entity sets, making it universally applicable to all VDER tasks. The current DocumentNet consists of 30M documents spanning nearly 400 document types organized in a four-level ontology. Experiments on a set of broadly adopted VDER tasks show significant improvements when DocumentNet is incorporated into the pre-training for both classic and few-shot learning settings. With the recent emergence of large language models (LLMs), DocumentNet provides a large data source to extend their multi-modal capabilities for VDER.

cs.CL↗

arXiv2022

Dimensionality-controlled evolution of charge-transfer energy in digital nickelates superlattices

Fundamental understanding and control of the electronic structure evolution in rare-earth nickelates is a fascinating and meaningful issue, as well as being helpful to understand the mechanism of recently discovered superconductivity. Here we systematically study the dimensionality effect on the ground electronic state in high-quality (NdNiO3)m/(SrTiO3)1 superlattices through transport and soft x-ray absorption spectroscopy. The metal-to-insulator transition temperature decreases with the thickness of the NdNiO3 slab decreasing from bulk to 7 unit cells, then increases gradually as m further reduces to 1 unit cell. Spectral evidence demonstrates that the stabilization of insulating phase can be attributed to the increase of the charge-transfer energy between O 2p and Ni 3d bands. The prominent multiplet feature on the Ni L3 edge develops with the decrease of NdNiO3 slab thickness, suggesting the strengthening of the charge disproportionate state under the dimensional confinement. Our work provides convincing evidence that dimensionality is an effective knob to modulate the charge-transfer energy and thus the collective ground state in nickelates.

cond-mat.mtrl-sci↗

arXiv2019

Magnetic and Electronic Properties of Single-Crystalline BaCoSO

Doped BaCoSO was recently predicted to be a high-temperature superconductor in a new class based on Co and Ni. Using a Co-S self flux method, we synthesized single crystals of the antiferromagnetic insulator BaCoSO. Our magnetic and specific heat measurements and neutron diffraction provide details of its magnetic anisotropy and order. Its band gap was determined to be about 1.3 eV by our measurements of its photoemission spectrum and infrared optical conductivity. Our results can pave the way to exploring the predicted superconductivity in this Co-based material.

cond-mat.supr-con↗

arXiv2018

Cross-layer Optimization for High Speed Adders: A Pareto Driven Machine Learning Approach

In spite of maturity to the modern electronic design automation (EDA) tools, optimized designs at architectural stage may become sub-optimal after going through physical design flow. Adder design has been such a long studied fundamental problem in VLSI industry yet designers cannot achieve optimal solutions by running EDA tools on the set of available prefix adder architectures. In this paper, we enhance a state-of-the-art prefix adder synthesis algorithm to obtain a much wider solution space in architectural domain. On top of that, a machine learning-based design space exploration methodology is applied to predict the Pareto frontier of the adders in physical domain, which is infeasible by exhaustively running EDA tools for innumerable architectural solutions. Considering the high cost of obtaining the true values for learning, an active learning algorithm is utilized to select the representative data during learning process, which uses less labeled data while achieving better quality of Pareto frontier. Experimental results demonstrate that our framework can achieve Pareto frontier of high quality over a wide design space, bridging the gap between architectural and physical designs.

cs.AR↗