arXiv subjects

Yiwen Song

Publications and source records attributed to Yiwen Song.

35 records · Page 2Linked to original sources

arXiv2025

Judging with Confidence: Calibrating Autoraters to Preference Distributions

The alignment of large language models (LLMs) with human values increasingly relies on using other LLMs as automated judges, or ``autoraters''. However, their reliability is limited by a foundational issue: they are trained on discrete preference labels, forcing a single ground truth onto tasks that are often subjective, ambiguous, or nuanced. We argue that a reliable autorater must learn to model the full distribution of preferences defined by a target population. In this paper, we propose a general framework for calibrating probabilistic autoraters to any given preference distribution. We formalize the problem and present two learning methods tailored to different data conditions: 1) a direct supervised fine-tuning for dense, probabilistic labels, and 2) a reinforcement learning approach for sparse, binary labels. Our empirical results show that finetuning autoraters with a distribution-matching objective leads to verbalized probability predictions that are better aligned with the target preference distribution, with improved calibration and significantly lower positional bias, all while preserving performance on objective tasks.

cs.CL↗

arXiv2025

Towards Seeing Bones at Radio Frequency

Wireless sensing literature has long aspired to achieve X-ray-like vision at radio frequencies. Yet, state-of-the-art wireless sensing literature has yet to generate the archetypal X-ray image: one of the bones beneath flesh. In this paper, we explore MCT, a penetration-based RF-imaging system for imaging bones at mm-resolution, one that significantly exceeds prior penetration-based RF imaging literature. Indeed the long wavelength, significant attenuation and complex diffraction that occur as RF propagates through flesh, have long limited imaging resolution (to several centimeters at best). We address these concerns through a novel penetration-based synthetic aperture algorithm, coupled with a learning-based pipeline to correct for diffraction-induced artifacts. A detailed evaluation of meat models demonstrates a resolution improvement from sub-decimeter to sub-centimeter over prior art in RF penetrative imaging.

cs.GR↗

arXiv2025

In Prospect and Retrospect: Reflective Memory Management for Long-term Personalized Dialogue Agents

Large Language Models (LLMs) have made significant progress in open-ended dialogue, yet their inability to retain and retrieve relevant information from long-term interactions limits their effectiveness in applications requiring sustained personalization. External memory mechanisms have been proposed to address this limitation, enabling LLMs to maintain conversational continuity. However, existing approaches struggle with two key challenges. First, rigid memory granularity fails to capture the natural semantic structure of conversations, leading to fragmented and incomplete representations. Second, fixed retrieval mechanisms cannot adapt to diverse dialogue contexts and user interaction patterns. In this work, we propose Reflective Memory Management (RMM), a novel mechanism for long-term dialogue agents, integrating forward- and backward-looking reflections: (1) Prospective Reflection, which dynamically summarizes interactions across granularities-utterances, turns, and sessions-into a personalized memory bank for effective future retrieval, and (2) Retrospective Reflection, which iteratively refines the retrieval in an online reinforcement learning (RL) manner based on LLMs' cited evidence. Experiments show that RMM demonstrates consistent improvement across various metrics and benchmarks. For example, RMM shows more than 10% accuracy improvement over the baseline without memory management on the LongMemEval dataset.

cs.CL↗

arXiv2025

PLAN-TUNING: Post-Training Language Models to Learn Step-by-Step Planning for Complex Problem Solving

Recently, decomposing complex problems into simple subtasks--a crucial part of human-like natural planning--to solve the given problem has significantly boosted the performance of large language models (LLMs). However, leveraging such planning structures during post-training to boost the performance of smaller open-source LLMs remains underexplored. Motivated by this, we introduce PLAN-TUNING, a unified post-training framework that (i) distills synthetic task decompositions (termed "planning trajectories") from large-scale LLMs and (ii) fine-tunes smaller models via supervised and reinforcement-learning objectives designed to mimic these planning processes to improve complex reasoning. On GSM8k and the MATH benchmarks, plan-tuned models outperform strong baselines by an average $\sim7\%$. Furthermore, plan-tuned models show better generalization capabilities on out-of-domain datasets, with average $\sim10\%$ and $\sim12\%$ performance improvements on OlympiadBench and AIME 2024, respectively. Our detailed analysis demonstrates how planning trajectories improves complex reasoning capabilities, showing that PLAN-TUNING is an effective strategy for improving task-specific performance of smaller LLMs.

cs.CL↗

arXiv2025

On the Role of Feedback in Test-Time Scaling of Agentic AI Workflows

Agentic AI workflows (systems that autonomously plan and act) are becoming widespread, yet their task success rate on complex tasks remains low. A promising solution is inference-time alignment, which uses extra compute at test time to improve performance. Inference-time alignment relies on three components: sampling, evaluation, and feedback. While most prior work studies sampling and automatic evaluation, feedback remains underexplored. To study the role of feedback, we introduce Iterative Agent Decoding (IAD), a procedure that repeatedly inserts feedback extracted from different forms of critiques (reward models or AI-generated textual feedback) between decoding steps. Through IAD, we analyze feedback along four dimensions: (1) its role in the accuracy-compute trade-offs with limited inference budget, (2) quantifying the gains over diversity-only baselines such as best-of-N sampling, (3) effectiveness of composing feedback from reward models versus textual critique, and (4) robustness to noisy or low-quality feedback. Across Sketch2Code, Text2SQL, Intercode, and WebShop, we show that IAD with proper integration of high fidelity feedback leads to consistent gains up to 10 percent absolute performance improvement over various baselines such as best-of-N. Our findings underscore feedback as a crucial knob for inference-time alignment of agentic AI workflows with limited inference budget.

cs.CL↗

arXiv2025

Multiple Weaks Win Single Strong: Large Language Models Ensemble Weak Reinforcement Learning Agents into a Supreme One

Model ensemble is a useful approach in reinforcement learning (RL) for training effective agents. Despite wide success of RL, training effective agents remains difficult due to the multitude of factors requiring careful tuning, such as algorithm selection, hyperparameter settings, and even random seed choices, all of which can significantly influence an agent's performance. Model ensemble helps overcome this challenge by combining multiple weak agents into a single, more powerful one, enhancing overall performance. However, existing ensemble methods, such as majority voting and Boltzmann addition, are designed as fixed strategies and lack a semantic understanding of specific tasks, limiting their adaptability and effectiveness. To address this, we propose LLM-Ens, a novel approach that enhances RL model ensemble with task-specific semantic understandings driven by large language models (LLMs). Given a task, we first design an LLM to categorize states in this task into distinct 'situations', incorporating high-level descriptions of the task conditions. Then, we statistically analyze the strengths and weaknesses of each individual agent to be used in the ensemble in each situation. During the inference time, LLM-Ens dynamically identifies the changing task situation and switches to the agent that performs best in the current situation, ensuring dynamic model selection in the evolving task condition. Our approach is designed to be compatible with agents trained with different random seeds, hyperparameter settings, and various RL algorithms. Extensive experiments on the Atari benchmark show that LLM-Ens significantly improves the RL model ensemble, surpassing well-known baselines by up to 20.9%. For reproducibility, our code is open-source at https://anonymous.4open.science/r/LLM4RLensemble-F7EE.

cs.LG↗

arXiv2025

ShieldGemma 2: Robust and Tractable Image Content Moderation

We introduce ShieldGemma 2, a 4B parameter image content moderation model built on Gemma 3. This model provides robust safety risk predictions across the following key harm categories: Sexually Explicit, Violence \& Gore, and Dangerous Content for synthetic images (e.g. output of any image generation model) and natural images (e.g. any image input to a Vision-Language Model). We evaluated on both internal and external benchmarks to demonstrate state-of-the-art performance compared to LlavaGuard \citep{helff2024llavaguard}, GPT-4o mini \citep{hurst2024gpt}, and the base Gemma 3 model \citep{gemma_2025} based on our policies. Additionally, we present a novel adversarial data generation pipeline which enables a controlled, diverse, and robust image generation. ShieldGemma 2 provides an open image moderation tool to advance multimodal safety and responsible AI development.

cs.CV↗

arXiv2025

Towards Large Reasoning Models: A Survey of Reinforced Reasoning with Large Language Models

Language has long been conceived as an essential tool for human reasoning. The breakthrough of Large Language Models (LLMs) has sparked significant research interest in leveraging these models to tackle complex reasoning tasks. Researchers have moved beyond simple autoregressive token generation by introducing the concept of "thought" -- a sequence of tokens representing intermediate steps in the reasoning process. This innovative paradigm enables LLMs' to mimic complex human reasoning processes, such as tree search and reflective thinking. Recently, an emerging trend of learning to reason has applied reinforcement learning (RL) to train LLMs to master reasoning processes. This approach enables the automatic generation of high-quality reasoning trajectories through trial-and-error search algorithms, significantly expanding LLMs' reasoning capacity by providing substantially more training data. Furthermore, recent studies demonstrate that encouraging LLMs to "think" with more tokens during test-time inference can further significantly boost reasoning accuracy. Therefore, the train-time and test-time scaling combined to show a new research frontier -- a path toward Large Reasoning Model. The introduction of OpenAI's o1 series marks a significant milestone in this research direction. In this survey, we present a comprehensive review of recent progress in LLM reasoning. We begin by introducing the foundational background of LLMs and then explore the key technical components driving the development of large reasoning models, with a focus on automated data construction, learning-to-reason techniques, and test-time scaling. We also analyze popular open-source projects at building large reasoning models, and conclude with open challenges and future research directions.

cs.AI↗

arXiv2024

TDNetGen: Empowering Complex Network Resilience Prediction with Generative Augmentation of Topology and Dynamics

Predicting the resilience of complex networks, which represents the ability to retain fundamental functionality amidst external perturbations or internal failures, plays a critical role in understanding and improving real-world complex systems. Traditional theoretical approaches grounded in nonlinear dynamical systems rely on prior knowledge of network dynamics. On the other hand, data-driven approaches frequently encounter the challenge of insufficient labeled data, a predicament commonly observed in real-world scenarios. In this paper, we introduce a novel resilience prediction framework for complex networks, designed to tackle this issue through generative data augmentation of network topology and dynamics. The core idea is the strategic utilization of the inherent joint distribution present in unlabeled network data, facilitating the learning process of the resilience predictor by illuminating the relationship between network topology and dynamics. Experiment results on three network datasets demonstrate that our proposed framework TDNetGen can achieve high prediction accuracy up to 85%-95%. Furthermore, the framework still demonstrates a pronounced augmentation capability in extreme low-data regimes, thereby underscoring its utility and robustness in enhancing the prediction of network resilience. We have open-sourced our code in the following link, https://github.com/tsinghua-fib-lab/TDNetGen.

cs.AI↗

arXiv2024

ToMoBrush: Exploring Dental Health Sensing using a Sonic Toothbrush

Early detection of dental disease is crucial to prevent adverse outcomes. Today, dental X-rays are currently the most accurate gold standard for dental disease detection. Unfortunately, regular X-ray exam is still a privilege for billions of people around the world. In this paper, we ask: "Can we develop a low-cost sensing system that enables dental self-examination in the comfort of one's home?" This paper presents ToMoBrush, a dental health sensing system that explores using off-the-shelf sonic toothbrushes for dental condition detection. Our solution leverages the fact that a sonic toothbrush produces rich acoustic signals when in contact with teeth, which contain important information about each tooth's status. ToMoBrush extracts tooth resonance signatures from the acoustic signals to characterize varied dental health conditions of the teeth. We evaluate ToMoBrush on 19 participants and dental-standard models for detecting common dental problems including caries, calculus, and food impaction, achieving a detection ROC-AUC of 0.90, 0.83, and 0.88 respectively. Interviews with dental experts validate ToMoBrush's potential in enhancing at-home dental healthcare.

eess.AS↗

arXiv2023

Emu: Enhancing Image Generation Models Using Photogenic Needles in a Haystack

Training text-to-image models with web scale image-text pairs enables the generation of a wide range of visual concepts from text. However, these pre-trained models often face challenges when it comes to generating highly aesthetic images. This creates the need for aesthetic alignment post pre-training. In this paper, we propose quality-tuning to effectively guide a pre-trained model to exclusively generate highly visually appealing images, while maintaining generality across visual concepts. Our key insight is that supervised fine-tuning with a set of surprisingly small but extremely visually appealing images can significantly improve the generation quality. We pre-train a latent diffusion model on $1.1$ billion image-text pairs and fine-tune it with only a few thousand carefully selected high-quality images. The resulting model, Emu, achieves a win rate of $82.9\%$ compared with its pre-trained only counterpart. Compared to the state-of-the-art SDXLv1.0, Emu is preferred $68.4\%$ and $71.3\%$ of the time on visual appeal on the standard PartiPrompts and our Open User Input benchmark based on the real-world usage of text-to-image models. In addition, we show that quality-tuning is a generic approach that is also effective for other architectures, including pixel diffusion and masked generative transformer models.

cs.CV↗

arXiv2023

Controlling the 2D magnetism of CrBr$_3$ by van der Waals stacking engineering

The manipulation of two-dimensional (2D) magnetic order is of significant importance to facilitate future 2D magnets for low-power and high-speed spintronic devices. Van der Waals stacking engineering makes promises for controllable magnetism via interlayer magnetic coupling. However, directly examining the stacking order changes accompanying magnetic order transitions at the atomic scale and preparing device-ready 2D magnets with controllable magnetic orders remain elusive. Here, we demonstrate effective control of interlayer stacking in exfoliated CrBr$_3$ via thermally assisted strain engineering. The stable interlayer ferromagnetic (FM), antiferromagnetic (AFM), and FM-AFM coexistent ground states confirmed by the magnetic circular dichroism measurements are realized. Combined with the first-principles calculations, the atomically-resolved imaging technique reveals the correlation between magnetic order and interlay stacking order in the CrBr$_3$ flakes unambiguously. A tunable exchange bias effect is obtained in the mixed phase of FM and AFM states. This work will introduce new magnetic properties by controlling the stacking order, and sequence of 2D magnets, providing ample opportunities for their application in spintronic devices.

cond-mat.mtrl-sci↗

arXiv2023

Ultra-Wide Bandgap Ga$_2$O$_3$-on-SiC MOSFETs

Ulta-wide bandgap semiconductors based on $β$-Ga$_2$O$_3$ offer the potential to achieve higher power switching performance, efficiency, and lower manufacturing cost than today's wide bandgap power semiconductors. However, the most critical challenge to the commercialization of Ga$_2$O$_3$ electronics is overheating, which impacts the device's performance and reliability. We fabricated a Ga$_2$O$_3$/4H-SiC composite wafer using a fusion-bonding method. A low temperature ($\le$ 600 $^{\circ}$C) epitaxy and device processing approach based on low-temperature (LT) metalorganic vapor phase epitaxy is developed to grow a Ga$_2$O$_3$ epitaxial channel layer on the composite wafer and subsequently fabricate into Ga$_2$O$_3$ power MOSFETs. This LT approach is essential to preserve the structural integrity of the composite wafer. These LT-grown epitaxial Ga$_2$O$_3$ MOSFETs deliver high thermal performance (56% reduction in channel temperature), high voltage blocking capabilities up to 2.45 kV, and power figures of merit of $\sim$ 300 MW/cm$^2$, which is a record high for any heterogeneously integrated Ga$_2$O$_3$ devices reported to date. This work is the first realization of multi-kilovolt homoepitaxial Ga$_2$O$_3$ power MOSFETs fabricated on a composite substrate with high heat transfer performance which delivers state-of-the-art power density values while running much cooler than those on native substrates. Thermal characterization and modeling results reveal that a Ga$_2$O$_3$/diamond composite wafer with a reduced Ga$_2$O$_3$ thickness ($\sim$ 1 $μ$m) and thinner bonding interlayer ($<$ 10 nm) can reduce the device thermal impedance to a level lower than today's GaN-on-SiC power switches.

cond-mat.mtrl-sci↗

arXiv2023

PACO: Parts and Attributes of Common Objects

Object models are gradually progressing from predicting just category labels to providing detailed descriptions of object instances. This motivates the need for large datasets which go beyond traditional object masks and provide richer annotations such as part masks and attributes. Hence, we introduce PACO: Parts and Attributes of Common Objects. It spans 75 object categories, 456 object-part categories and 55 attributes across image (LVIS) and video (Ego4D) datasets. We provide 641K part masks annotated across 260K object boxes, with roughly half of them exhaustively annotated with attributes as well. We design evaluation metrics and provide benchmark results for three tasks on the dataset: part mask segmentation, object and part attribute prediction and zero-shot instance detection. Dataset, models, and code are open-sourced at https://github.com/facebookresearch/paco.

cs.CV↗

arXiv2021

130 mA/mm $β$-Ga$_2$O$_3$ MESFET with Low-Temperature MOVPE-Regrown Ohmic Contacts

We report on the demonstration of metalorganic vapor phase epitaxy-regrown (MOVPE) ohmic contacts in an all MOVPE-grown $β$-Ga$_2$O$_3$ metal-semiconductor field effect transistor (MESFET). The low-temperature (600$^{\circ}$C) heavy (n$^{+}$) Si-doped regrown layers exhibit extremely high conductivity with sheet resistance of 73 $Ω$/$\square$ and record low metal/n$^{+}$-Ga$_2$O$_3$ contact resistance of 80 m$Ω$.mm and specific contact resistivity of 8.3$\times$10$^{-7}$ $Ω$.cm$^{2}$ were achieved. The fabricated MESFETs exhibit a maximum drain-to-source current of 130 mA/mm, a high I$_{ON}$/I$_{OFF}$ of $>$10$^{10}$ with a high power FOM of 25 MW/cm$^{2}$ were achieved without any field plates. Nanoparticle-assisted Raman thermometry, thermal modeling, and infrared thermography were performed to assess the device self-heating under the high current and power conditions. This demonstration shows the promise of MOVPE technique for the realization of high-performance lateral $β$-Ga$_2$O$_3$ devices and also highlights the need for device-level thermal management.

physics.app-ph↗

arXiv2021

Clusterability as an Alternative to Anchor Points When Learning with Noisy Labels

The label noise transition matrix, characterizing the probabilities of a training instance being wrongly annotated, is crucial to designing popular solutions to learning with noisy labels. Existing works heavily rely on finding "anchor points" or their approximates, defined as instances belonging to a particular class almost surely. Nonetheless, finding anchor points remains a non-trivial task, and the estimation accuracy is also often throttled by the number of available anchor points. In this paper, we propose an alternative option to the above task. Our main contribution is the discovery of an efficient estimation procedure based on a clusterability condition. We prove that with clusterable representations of features, using up to third-order consensuses of noisy labels among neighbor representations is sufficient to estimate a unique transition matrix. Compared with methods using anchor points, our approach uses substantially more instances and benefits from a much better sample complexity. We demonstrate the estimation accuracy and advantages of our estimates using both synthetic noisy labels (on CIFAR-10/100) and real human-level noisy labels (on Clothing1M and our self-collected human-annotated CIFAR-10). Our code and human-level noisy CIFAR-10 labels are available at https://github.com/UCSC-REAL/HOC.

cs.LG↗

arXiv2019

Demand Adaptive Multi-Objective Electric Taxi Fleet Dispatching with Carbon Emission Analysis

As a foreseeable future mode of transport with lower emissions and higher efficiencies, electric vehicles have received worldwide attention. For convenient centralized management, taxis are considered as the fleet with electrification priority. In this work, we focus on the study on electric taxis dispatching, with consideration of picking up customers and recharging, based on real world traffic data of a large number of taxis in Beijing. First, the assumed electric taxi charging stations are located using the K mean method. Second, based on the station locations and the order demands, which are in form of origin-destination pairs and extracted from the trajectory data, a dispatching strategy as well as the simulation framework is developed with consideration of reducing customer waiting time, mitigating electric taxi charging congestion, and balancing order number distribution among electric taxis. The proposed method models the electric taxi charging behaviors temporally discretely from the aspects of charging demands and availability of chargers, and further incorporates a centralized and intelligent fleet dispatching platform, which is capable of handling taxi service requests and arranging electric taxis' recharging in real time. The methodology in this paper is readily applicable to dispatching of different types of electric vehicle fleet with similar dataset available. Among the method, we use queueing theory to model the electric vehicle charging station waiting phenomena and include this factor into dispatching platform. Carbon emission is also surveyed and analyzed.

math.OC↗