arXiv subjects

Liang Qiu

Publications and source records attributed to Liang Qiu.

At least 19 recordsLinked to original sources

arXiv2025

Self-Rewarding PPO: Aligning Large Language Models with Demonstrations Only

Supervised fine-tuning (SFT) has emerged as a crucial method for aligning large language models (LLMs) with human-annotated demonstrations. However, SFT, being an off-policy approach similar to behavior cloning, often struggles with overfitting and poor out-of-domain generalization, especially in limited-data scenarios. To address these limitations, we propose Self-Rewarding PPO, a novel fine-tuning method that leverages on-policy techniques to enhance generalization performance. Our approach combines the strengths of SFT and proximal policy optimization (PPO) to achieve more effective alignment from demonstration data. At its core is a reward function designed as the log policy ratio between the SFT model and the pretrained base model. This function serves as an implicit reward signal, using the pretrained policy as a baseline and the SFT policy as a target. By doing so, it enables on-policy fine-tuning without relying on human preference annotations. The integration of this self-rewarding mechanism with PPO addresses key limitations of SFT, improving generalization, data efficiency, and robustness. Our empirical evaluation across a range of natural language processing tasks demonstrates that Self-Rewarding PPO consistently outperforms traditional SFT methods. The results highlight the effectiveness of our approach in aligning LLMs using demonstration data, particularly in scenarios where high-quality annotated data is scarce.

cs.CL↗

arXiv2025

Ask a Strong LLM Judge when Your Reward Model is Uncertain

Reward model (RM) plays a pivotal role in reinforcement learning with human feedback (RLHF) for aligning large language models (LLMs). However, classical RMs trained on human preferences are vulnerable to reward hacking and generalize poorly to out-of-distribution (OOD) inputs. By contrast, strong LLM judges equipped with reasoning capabilities demonstrate superior generalization, even without additional training, but incur significantly higher inference costs, limiting their applicability in online RLHF. In this work, we propose an uncertainty-based routing framework that efficiently complements a fast RM with a strong but costly LLM judge. Our approach formulates advantage estimation in policy gradient (PG) methods as pairwise preference classification, enabling principled uncertainty quantification to guide routing. Uncertain pairs are forwarded to the LLM judge, while confident ones are evaluated by the RM. Experiments on RM benchmarks demonstrate that our uncertainty-based routing strategy significantly outperforms random judge calling at the same cost, and downstream alignment results showcase its effectiveness in improving online RLHF.

cs.LG↗

arXiv2025

DeepPlanner: Scaling Planning Capability for Deep Research Agents via Advantage Shaping

Large language models (LLMs) augmented with multi-step reasoning and action generation abilities have shown promise in leveraging external tools to tackle complex tasks that require long-horizon planning. However, existing approaches either rely on implicit planning in the reasoning stage or introduce explicit planners without systematically addressing how to optimize the planning stage. As evidence, we observe that under vanilla reinforcement learning (RL), planning tokens exhibit significantly higher entropy than other action tokens, revealing uncertain decision points that remain under-optimized. To address this, we propose DeepPlanner, an end-to-end RL framework that effectively enhances the planning capabilities of deep research agents. Our approach shapes token-level advantage with an entropy-based term to allocate larger updates to high entropy tokens, and selectively upweights sample-level advantages for planning-intensive rollouts. Extensive experiments across seven deep research benchmarks demonstrate that DeepPlanner improves planning quality and achieves state-of-the-art results under a substantially lower training budget.

cs.AI↗

arXiv2025

SAGE: A Top-Down Bottom-Up Knowledge-Grounded User Simulator for Multi-turn AGent Evaluation

Evaluating multi-turn interactive agents is challenging due to the need for human assessment. Evaluation with simulated users has been introduced as an alternative, however existing approaches typically model generic users and overlook the domain-specific principles required to capture realistic behavior. We propose SAGE, a novel user Simulation framework for multi-turn AGent Evaluation that integrates knowledge from business contexts. SAGE incorporates top-down knowledge rooted in business logic, such as ideal customer profiles, grounding user behavior in realistic customer personas. We further integrate bottom-up knowledge taken from business agent infrastructure (e.g., product catalogs, FAQs, and knowledge bases), allowing the simulator to generate interactions that reflect users' information needs and expectations in a company's target market. Through empirical evaluation, we find that this approach produces interactions that are more realistic and diverse, while also identifying up to 33% more agent errors, highlighting its effectiveness as an evaluation tool to support bug-finding and iterative agent improvement.

cs.CL↗

arXiv2025

Think-RM: Enabling Long-Horizon Reasoning in Generative Reward Models

Reinforcement learning from human feedback (RLHF) has become a powerful post-training paradigm for aligning large language models with human preferences. A core challenge in RLHF is constructing accurate reward signals, where the conventional Bradley-Terry reward models (BT RMs) often suffer from sensitivity to data size and coverage, as well as vulnerability to reward hacking. Generative reward models (GenRMs) offer a more robust alternative by generating chain-of-thought (CoT) rationales followed by a final reward. However, existing GenRMs rely on shallow, vertically scaled reasoning, limiting their capacity to handle nuanced or complex (e.g., reasoning-intensive) tasks. Moreover, their pairwise preference outputs are incompatible with standard RLHF algorithms that require pointwise reward signals. In this work, we introduce Think-RM, a training framework that enables long-horizon reasoning in GenRMs by modeling an internal thinking process. Rather than producing structured, externally provided rationales, Think-RM generates flexible, self-guided reasoning traces that support advanced capabilities such as self-reflection, hypothetical reasoning, and divergent reasoning. To elicit these reasoning abilities, we first warm-up the models by supervised fine-tuning (SFT) over long CoT data. We then further improve the model's long-horizon abilities by rule-based reinforcement learning (RL). In addition, we propose a novel pairwise RLHF pipeline that directly optimizes policies using pairwise preference rewards, eliminating the need for pointwise reward conversion and enabling more effective use of Think-RM outputs. Experiments show that Think-RM achieves state-of-the-art results on RM-Bench, outperforming both BT RM and vertically scaled GenRM by 8%. When combined with our pairwise RLHF pipeline, it demonstrates superior end-policy performance compared to traditional approaches.

cs.LG↗

arXiv2025

WebAgent-R1: Training Web Agents via End-to-End Multi-Turn Reinforcement Learning

While reinforcement learning (RL) has demonstrated remarkable success in enhancing large language models (LLMs), it has primarily focused on single-turn tasks such as solving math problems. Training effective web agents for multi-turn interactions remains challenging due to the complexity of long-horizon decision-making across dynamic web interfaces. In this work, we present WebAgent-R1, a simple yet effective end-to-end multi-turn RL framework for training web agents. It learns directly from online interactions with web environments by asynchronously generating diverse trajectories, entirely guided by binary rewards depending on task success. Experiments on the WebArena-Lite benchmark demonstrate the effectiveness of WebAgent-R1, boosting the task success rate of Qwen-2.5-3B from 6.1% to 33.9% and Llama-3.1-8B from 8.5% to 44.8%, significantly outperforming existing state-of-the-art methods and strong proprietary models such as OpenAI o3. In-depth analyses reveal the effectiveness of the thinking-based prompting strategy and test-time scaling through increased interactions for web tasks. We further investigate different RL initialization policies by introducing two variants, namely WebAgent-R1-Zero and WebAgent-R1-CoT, which highlight the importance of the warm-up training stage (i.e., behavior cloning) and provide insights on incorporating long chain-of-thought (CoT) reasoning in web agents.

cs.CL↗

arXiv2025

Can Language Models Follow Multiple Turns of Entangled Instructions?

Despite significant achievements in improving the instruction-following capabilities of large language models (LLMs), the ability to process multiple potentially entangled or conflicting instructions remains a considerable challenge. Real-world scenarios often require consistency across multiple instructions over time, such as secret privacy, personal preferences, and prioritization, which demand sophisticated abilities to integrate multiple turns and carefully balance competing objectives when instructions intersect or conflict. This work presents a systematic investigation of LLMs' capabilities in handling multiple turns of instructions, covering three levels of difficulty: (1) retrieving information from instructions, (2) tracking and reasoning across turns, and (3) resolving conflicts among instructions. We construct MultiTurnInstruct~with $\sim$1.1K high-quality multi-turn conversations through the human-in-the-loop approach and result in nine capability categories, including statics and dynamics, reasoning, and multitasking. Our finding reveals an intriguing trade-off between different capabilities. While GPT models demonstrate superior memorization, they show reduced effectiveness in privacy-protection tasks requiring selective information withholding. Larger models exhibit stronger reasoning capabilities but still struggle with resolving conflicting instructions. Importantly, these performance gaps cannot be attributed solely to information loss, as models demonstrate strong BLEU scores on memorization tasks. Still, their attention mechanisms fail to integrate multiple related instructions effectively. These findings highlight critical areas for improvement in complex real-world tasks involving multi-turn instructions. Data and codes are released at https://github.com/Glaciohound/Multi-Turn-Instruct.

cs.CL↗

arXiv2025

Discriminative Finetuning of Generative Large Language Models without Reward Models and Human Preference Data

Supervised fine-tuning (SFT) has become a crucial step for aligning pretrained large language models (LLMs) using supervised datasets of input-output pairs. However, despite being supervised, SFT is inherently limited by its generative training objective. To address its limitations, the existing common strategy is to follow SFT with a separate phase of preference optimization (PO), which relies on either human-labeled preference data or a strong reward model to guide the learning process. In this paper, we address the limitations of SFT by exploring one of the most successful techniques in conventional supervised learning: discriminative learning. We introduce Discriminative Fine-Tuning (DFT), an improved variant of SFT, which mitigates the burden of collecting human-labeled preference data or training strong reward models. Unlike SFT that employs a generative approach and overlooks negative data, DFT adopts a discriminative paradigm that increases the probability of positive answers while suppressing potentially negative ones, aiming for data prediction instead of token prediction. Our contributions include: (i) a discriminative probabilistic framework for fine-tuning LLMs by explicitly modeling the discriminative likelihood of an answer among all possible outputs given an input; (ii) efficient algorithms to optimize this discriminative likelihood; and (iii) extensive experiments demonstrating DFT's effectiveness, achieving performance better than SFT and comparable to if not better than SFT$\rightarrow$PO. The code can be found at https://github.com/Optimization-AI/DFT.

cs.CL↗

arXiv2024

Artificial Intelligence-Enhanced Couinaud Segmentation for Precision Liver Cancer Therapy

Precision therapy for liver cancer necessitates accurately delineating liver sub-regions to protect healthy tissue while targeting tumors, which is essential for reducing recurrence and improving survival rates. However, the segmentation of hepatic segments, known as Couinaud segmentation, is challenging due to indistinct sub-region boundaries and the need for extensive annotated datasets. This study introduces LiverFormer, a novel Couinaud segmentation model that effectively integrates global context with low-level local features based on a 3D hybrid CNN-Transformer architecture. Additionally, a registration-based data augmentation strategy is equipped to enhance the segmentation performance with limited labeled data. Evaluated on CT images from 123 patients, LiverFormer demonstrated high accuracy and strong concordance with expert annotations across various metrics, allowing for enhanced treatment planning for surgery and radiation therapy. It has great potential to reduces complications and minimizes potential damages to surrounding tissue, leading to improved outcomes for patients undergoing complex liver cancer treatments.

eess.IV↗

arXiv2024

ST-NeRP: Spatial-Temporal Neural Representation Learning with Prior Embedding for Patient-specific Imaging Study

During and after a course of therapy, imaging is routinely used to monitor the disease progression and assess the treatment responses. Despite of its significance, reliably capturing and predicting the spatial-temporal anatomic changes from a sequence of patient-specific image series presents a considerable challenge. Thus, the development of a computational framework becomes highly desirable for a multitude of practical applications. In this context, we propose a strategy of Spatial-Temporal Neural Representation learning with Prior embedding (ST-NeRP) for patient-specific imaging study. Our strategy involves leveraging an Implicit Neural Representation (INR) network to encode the image at the reference time point into a prior embedding. Subsequently, a spatial-temporally continuous deformation function is learned through another INR network. This network is trained using the whole patient-specific image sequence, enabling the prediction of deformation fields at various target time points. The efficacy of the ST-NeRP model is demonstrated through its application to diverse sequential image series, including 4D CT and longitudinal CT datasets within thoracic and abdominal imaging. The proposed ST-NeRP model exhibits substantial potential in enabling the monitoring of anatomical changes within a patient throughout the therapeutic journey.

eess.IV↗

arXiv2024

Contrastive Learning with Counterfactual Explanations for Radiology Report Generation

Due to the common content of anatomy, radiology images with their corresponding reports exhibit high similarity. Such inherent data bias can predispose automatic report generation models to learn entangled and spurious representations resulting in misdiagnostic reports. To tackle these, we propose a novel \textbf{Co}unter\textbf{F}actual \textbf{E}xplanations-based framework (CoFE) for radiology report generation. Counterfactual explanations serve as a potent tool for understanding how decisions made by algorithms can be changed by asking ``what if'' scenarios. By leveraging this concept, CoFE can learn non-spurious visual representations by contrasting the representations between factual and counterfactual images. Specifically, we derive counterfactual images by swapping a patch between positive and negative samples until a predicted diagnosis shift occurs. Here, positive and negative samples are the most semantically similar but have different diagnosis labels. Additionally, CoFE employs a learnable prompt to efficiently fine-tune the pre-trained large language model, encapsulating both factual and counterfactual content to provide a more generalizable prompt representation. Extensive experiments on two benchmarks demonstrate that leveraging the counterfactual explanations enables CoFE to generate semantically coherent and factually complete reports and outperform in terms of language generation and clinical efficacy metrics.

cs.CV↗

arXiv2024

Aligning Large Language Models via Fine-grained Supervision

Pre-trained large-scale language models (LLMs) excel at producing coherent articles, yet their outputs may be untruthful, toxic, or fail to align with user expectations. Current approaches focus on using reinforcement learning with human feedback (RLHF) to improve model alignment, which works by transforming coarse human preferences of LLM outputs into a feedback signal that guides the model learning process. However, because this approach operates on sequence-level feedback, it lacks the precision to identify the exact parts of the output affecting user preferences. To address this gap, we propose a method to enhance LLM alignment through fine-grained token-level supervision. Specifically, we ask annotators to minimally edit less preferred responses within the standard reward modeling dataset to make them more favorable, ensuring changes are made only where necessary while retaining most of the original content. The refined dataset is used to train a token-level reward model, which is then used for training our fine-grained Proximal Policy Optimization (PPO) model. Our experiment results demonstrate that this approach can achieve up to an absolute improvement of $5.1\%$ in LLM performance, in terms of win rate against the reference model, compared with the traditional PPO model.

cs.CL↗

arXiv2024

Switching intrinsic magnetic skyrmions with controllable magnetic anisotropy in van der Waals multiferroic heterostructures

Magnetic skyrmions, topologically nontrivial whirling spin textures at nanometer scales, have emerged as potential information carriers for spintronic devices. The ability to efficiently create and erase magnetic skyrmions is vital yet challenging for such applications. Based on first-principles studies, we find that switching between intrinsic magnetic skyrmion and high-temperature ferromagnetic states can be achieved in two-dimensional van der Waals (vdW) multiferroic heterostructure CrSeI/In2Te3 by reversing the ferroelectric polarization of In2Te3. The core mechanism of this switching is traced to the controllable magnetic anisotropy of CrSeI influenced by the ferroelectric polarization of In2Te3. We propose a useful descriptor linking the presence of magnetic skyrmions to magnetic parameters, and validate this connection through studies of a variety of similar vdW multiferroic heterostructures. Our work demonstrates that manipulating magnetic skyrmions via tunable magnetic anisotropies in vdW multiferroic heterostructures represents a highly promising and energy-efficient strategy for future development of spintronics.

cond-mat.mtrl-sci↗

arXiv2023

Quantitative determination of the orbital-selective Mott transition and quantum entanglement in the orbital-selective Mott phase

We examine the orbital-selective Mott transition in the non-hybridized two-band Hubbard model using the dynamical mean-field theory. We find that the orbital-selective Mott transition could be quantitatively depicted by the {local two-qubit fidelity}. Furthermore, within the orbital-selective Mott phase, the combined characteristics of the two orbitals lead to the presence of quantum entanglement, which is characterized by the non-semi-integer values of local two-qubit fidelity. It is demonstrated that the Hund's coupling results in the ground states of both wide and narrow bands exhibiting the specific superposition states, indicating the existence of quantum entanglement within orbital-selective Mott phase. Without Hund's coupling, there are no specific superposition states, nor does quantum entanglement occur within the orbital-selective Mott phase. The mechanisms underlying the orbital-selective Mott transition show prominent variations depending on the presence or absence of Hund's coupling and its transverse terms.

cond-mat.str-el↗

arXiv2023

Electrically tunable Gilbert damping in van der Waals heterostructures of two-dimensional ferromagnetic metals and ferroelectrics

Tuning the Gilbert damping of ferromagnetic (FM) metals via a nonvolatile way is of importance to exploit and design next-generation novel spintronic devices. Through systematical first-principles calculations, we study the magnetic properties of the van der Waals heterostructure of two-dimensional FM metal CrTe2 and ferroelectric (FE) In2Te3 monolayers. The ferromagnetism of CrTe2 is maintained in CrTe2/In2Te3 and its magnetic easy axis can be switched from in-plane to out-of-plane by reversing the FE polarization of In2Te3. Excitingly, we find that the Gilbert damping of CrTe2 is tunable when the FE polarization of In2Te3 is reversed from upward to downward. By analyzing the k-dependent contributions to the Gilbert damping, we unravel that such tunability results from the changed intersections between the bands of CrTe2 and Fermi level on the reversal of the FE polarizations of In2Te3 in CrTe2/In2Te3. Our work provides an appealing way to electrically tailor Gilbert dampings of two-dimensional FM metals by contacting them with ferroelectrics.

cond-mat.mtrl-sci↗

arXiv2023

Coexistence of zigzag antiferromagnetic order and superconductivity in compressed NiPSe3

NiPSe3 is regarded as a bandwidth-controlled Mott insulator, distinct from the widely studied Mott insulating magnetic graphene MPSe3 (M = Mn and Fe) family. By employing high-pressure synchrotron X-ray diffraction, we observe two structural transitions as a function of pressure. With the help of first-principles calculations, we discover the antiferromagnetic (AFM) moment directions of NiPSe3 switch from out-of-plane to in-plane and the honeycomb layers slide relative to each other at the first structural transition. The in-plane AFM order persists until the second structural transition, whereupon the two-dimensional (2D) structure assumes a more three-dimensional (3D) character. A bandwidth-controlled Mott insulator-metal transition (IMT) occurs between the two structural transitions at 8.0 GPa, concomitant with the emergence of superconductivity with 4.8 K. The superconductivity in NiPSe3 emerging in the 2D monoclinic phase coexists with the in-plane AFM order and continues into the 3D trigonal phase. Our electronic structure calculations reveal that the Mott IMT and superconductivity in NiPSe3 are both closely related to the enhanced Se2- 4p and Ni2+ 3d electronic hybridizations under pressure. From these results, we construct the temperature-pressure electronic phase diagram of NiPSe3, revealing rich physics and many similarities with copper oxide and iron-based superconductors.

cond-mat.supr-con↗

arXiv2022

A Survey of Deep Learning for Mathematical Reasoning

Mathematical reasoning is a fundamental aspect of human intelligence and is applicable in various fields, including science, engineering, finance, and everyday life. The development of artificial intelligence (AI) systems capable of solving math problems and proving theorems has garnered significant interest in the fields of machine learning and natural language processing. For example, mathematics serves as a testbed for aspects of reasoning that are challenging for powerful deep learning models, driving new algorithmic and modeling advances. On the other hand, recent advances in large-scale neural language models have opened up new benchmarks and opportunities to use deep learning for mathematical reasoning. In this survey paper, we review the key tasks, datasets, and methods at the intersection of mathematical reasoning and deep learning over the past decade. We also evaluate existing benchmarks and methods, and discuss future research directions in this domain.

cs.AI↗

arXiv2022

OpenD: A Benchmark for Language-Driven Door and Drawer Opening

We introduce OPEND, a benchmark for learning how to use a hand to open cabinet doors or drawers in a photo-realistic and physics-reliable simulation environment driven by language instruction. To solve the task, we propose a multi-step planner composed of a deep neural network and rule-base controllers. The network is utilized to capture spatial relationships from images and understand semantic meaning from language instructions. Controllers efficiently execute the plan based on the spatial and semantic understanding. We evaluate our system by measuring its zero-shot performance in test data set. Experimental results demonstrate the effectiveness of decision planning by our multi-step planner for different hands, while suggesting that there is significant room for developing better models to address the challenge brought by language understanding, spatial reasoning, and long-term manipulation. We will release OPEND and host challenges to promote future research in this area.

cs.LG↗