arXiv subjects

Peng Fu

Publications and source records attributed to Peng Fu.

At least 37 records · Page 2Linked to original sources

arXiv2026

Near-Future Policy Optimization

Reinforcement learning with verifiable rewards (RLVR) has become a core post-training recipe. Introducing suitable off-policy trajectories into on-policy exploration accelerates RLVR convergence and raises the performance ceiling, yet finding a source of such trajectories remains the key challenge. Existing mixed-policy methods either import trajectories from external teachers (high-quality but distributionally far) or replay past training trajectories (close but capped in quality), and neither simultaneously satisfies the strong enough (higher $Q$ , more new knowledge to learn) and close enough (lower $V$ , more readily absorbed) conditions required to maximize the effective learning signal $\mathcal{S} = Q/V$. We propose \textbf{N}ear-Future \textbf{P}olicy \textbf{O}ptimization (\textbf{NPO}), a simple mixed-policy scheme that learns from a policy's own near-future self: a later checkpoint from the same training run is a natural source of auxiliary trajectories that is both stronger than the current policy and closer than any external source, directly balancing trajectory quality against variance cost. We validate NPO through two manual interventions, early-stage bootstrapping and late-stage plateau breakthrough, and further propose \textbf{AutoNPO},an adaptive variant that automatically triggers interventions from online training signals and selects the guide checkpoint that maximizes $S$. On Qwen3-VL-8B-Instruct with GRPO, NPO improves average performance from 57.88 to 62.84, and AutoNPO pushes it to 63.15, raising the final performance ceiling while accelerating convergence.

cs.LG↗

arXiv2026

EasyVideoR1: Easier RL for Video Understanding

Reinforcement learning from verifiable rewards (RLVR) has demonstrated remarkable effectiveness in improving the reasoning capabilities of large language models. As models evolve into natively multimodal architectures, extending RLVR to video understanding becomes increasingly important yet remains largely unexplored, due to the diversity of video task types, the computational overhead of repeatedly decoding and preprocessing high-dimensional visual inputs, and the difficulty of reproducible evaluation across numerous sensitive hyperparameters. Existing open-source RL training frameworks provide solid infrastructure for text and image scenarios but lack systematic optimizations tailored for video modality. In this work, we present \textbf{EasyVideoR1}, a complete and efficient reinforcement learning framework specifically designed for training large vision-language models on video understanding tasks. EasyVideoR1 makes the following contributions: (1) a full video RL training pipeline with offline preprocessing and tensor caching that eliminates redundant video decoding and yields a 1.47 $\times$ throughput improvement; (2) a comprehensive, task-aware reward system covering 11 distinct video and image problem types with unified routing and modular extension; (3) a mixed offline-online data training paradigm that combines curated high-quality trajectories with on-policy exploration, benefiting the learning of more challenging tasks; (4) joint image-video training with independently configurable pixel budgets, allowing the two modalities to mutually reinforce each other; and (5) an asynchronous multi-benchmark evaluation framework covering 22 mainstream video understanding benchmarks, with reproduced accuracy closely aligned with officially reported scores.

cs.CV↗

arXiv2026

Superconductivity and Electronic Structures of Nickelate Thin Film Superstructures

Ruddlesden-Popper (RP) nickelates have emerged as a crucial platform for exploring the mechanisms of high-temperature superconductivity. However, the Fermi surface topology required for superconductivity remains elusive. Here, beyond the superconducting pure bilayer (2222) phase, we report the thin film growth and ambient-pressure superconductivity of monolayer-bilayer (1212) and bilayer-trilayer (2323) superstructures, together with the absence of superconductivity in monolayer-trilayer (1313) superstructure, under identical compressive epitaxial strain. The onset superconducting transition temperatures range from 46 to 50 K, exceeding the McMillan limit. Angle-resolved photoemission spectroscopy reveals key Fermi surface differences in these atomically-engineered structures. In superconducting 1212 and 2222 films, a dispersive hole-like band ($γ^{\mathrm{II}}$) forms an underlying Fermi pocket, surrounding the Brillouin zone corner. In contrast, the top of the flat band ($γ^{\mathrm{III}}$) is observed ~70 meV below $E_\text{F}$ in the non-superconducting 1313 films. Particularly, the superconducting 2323 films host both $γ^{\mathrm{II}}$ and $γ^{\mathrm{III}}$ bands. The polarization dependence of the $γ$ bands reveals their Ni $d_{z^2}$ origin. Our findings expand the family of ambient-pressure nickelate superconductors and establish a connection between structural configuration, electronic structure, and the emergence of superconductivity in nickelates.

cond-mat.supr-con↗

arXiv2026

Mitigating Overthinking in Large Reasoning Language Models via Reasoning Path Deviation Monitoring

Large Reasoning Language Models (LRLMs) demonstrate impressive capabilities on complex tasks by utilizing long Chain-of-Thought reasoning. However, they are prone to overthinking, which generates redundant reasoning steps that degrade both performance and efficiency. Recently, early-exit strategies are proposed to mitigate overthinking by dynamically and adaptively terminating redundant reasoning. However, current early-exit methods either introduce extra training overhead by relying on proxy models or limit inference throughput due to the frequent content switching between reasoning and generating probing answers. Moreover, most early-exit methods harm LRLMs performance due to over-truncation. Our insight stems from an observation: overthinking often causes LRLMs to deviate from the correct reasoning path, which is frequently accompanied by high-entropy transition tokens. Given this, we propose an early-exit method deeply coupled with the native reasoning process, which leverages the path deviation index as a dedicated monitoring metric for the frequent occurrence of high-entropy transition tokens to dynamically detect and terminate overthinking trajectories. We conduct experiments across multiple benchmarks using LRLMs of different types and scales, and the results indicate that our method delivers the largest performance improvement over vanilla CoT compared to existing early-exit methods.

cs.CL↗

arXiv2026

EXaMCaP: Subset Selection with Entropy Gain Maximization for Probing Capability Gains of Large Chart Understanding Training Sets

Recent works focus on synthesizing Chart Understanding (ChartU) training sets to inject advanced chart knowledge into Multimodal Large Language Models (MLLMs), where the sufficiency of the knowledge is typically verified by quantifying capability gains via the fine-tune-then-evaluate paradigm. However, full-set fine-tuning MLLMs to assess such gains incurs significant time costs, hindering the iterative refinement cycles of the ChartU dataset. Reviewing the ChartU dataset synthesis and data selection domains, we find that subsets can potentially probe the MLLMs' capability gains from full-set fine-tuning. Given that data diversity is vital for boosting MLLMs' performance and entropy reflects this feature, we propose EXaMCaP, which uses entropy gain maximization to select a subset. To obtain a high-diversity subset, EXaMCaP chooses the maximum-entropy subset from the large ChartU dataset. As enumerating all possible subsets is impractical, EXaMCaP iteratively selects samples to maximize the gain in set entropy relative to the current set, approximating the maximum-entropy subset of the full dataset. Experiments show that EXaMCaP outperforms baselines in probing the capability gains of the ChartU training set, along with its strong effectiveness across diverse subset sizes and compatibility with various MLLM architectures.

cs.LG↗

arXiv2025

Reconstruction of Differentially Private Text Sanitization via Large Language Models

Differential privacy (DP) is the de facto privacy standard against privacy leakage attacks, including many recently discovered ones against large language models (LLMs). However, we discovered that LLMs could reconstruct the altered/removed privacy from given DP-sanitized prompts. We propose two attacks (black-box and white-box) based on the accessibility to LLMs and show that LLMs could connect the pair of DP-sanitized text and the corresponding private training data of LLMs by giving sample text pairs as instructions (in the black-box attacks) or fine-tuning data (in the white-box attacks). To illustrate our findings, we conduct comprehensive experiments on modern LLMs (e.g., LLaMA-2, LLaMA-3, ChatGPT-3.5, ChatGPT-4, ChatGPT-4o, Claude-3, Claude-3.5, OPT, GPT-Neo, GPT-J, Gemma-2, and Pythia) using commonly used datasets (such as WikiMIA, Pile-CC, and Pile-Wiki) against both word-level and sentence-level DP. The experimental results show promising recovery rates, e.g., the black-box attacks against the word-level DP over WikiMIA dataset gave 72.18% on LLaMA-2 (70B), 82.39% on LLaMA-3 (70B), 75.35% on Gemma-2, 91.2% on ChatGPT-4o, and 94.01% on Claude-3.5 (Sonnet). More urgently, this study indicates that these well-known LLMs have emerged as a new security risk for existing DP text sanitization approaches in the current environment.

cs.CR↗

arXiv2025

CBP-Tuning: Efficient Local Customization for Black-box Large Language Models

The high costs of customizing large language models (LLMs) fundamentally limit their adaptability to user-specific needs. Consequently, LLMs are increasingly offered as cloud-based services, a paradigm that introduces critical limitations: providers struggle to support personalized customization at scale, while users face privacy risks when exposing sensitive data. To address this dual challenge, we propose Customized Black-box Prompt Tuning (CBP-Tuning), a novel framework that facilitates efficient local customization while preserving bidirectional privacy. Specifically, we design a two-stage framework: (1) a prompt generator trained on the server-side to capture domain-specific and task-agnostic capabilities, and (2) user-side gradient-free optimization that tailors soft prompts for individual tasks. This approach eliminates the need for users to access model weights or upload private data, requiring only a single customized vector per task while achieving effective adaptation. Furthermore, the evaluation of CBP-Tuning in the commonsense reasoning, medical and financial domain settings demonstrates superior performance compared to baselines, showcasing its advantages in task-agnostic processing and privacy preservation.

cs.CL↗

arXiv2025

Proto-Quipper with Reversing and Control

The quantum programming language Quipper supports circuit operations such as reversing and controlling certain quantum circuits. Additionally, Quipper provides a function called with-computed, which can be used to program circuits of the form g; f; g-dagger. The latter is a common pattern in quantum circuit design. One benefit of using with-computed, as opposed to constructing the circuit g ; f; g-dagger directly from g, f, and g-dagger, is that it facilitates an important optimization. Namely, if the resulting circuit is later controlled, only f needs to be controlled; the circuits g and g-dagger need not even be controllable. In this paper, we formalize a semantics for reversible and controllable circuits, using a dagger symmetric monoidal category R to interpret reversible circuits, and a new notion we call a controllable category N, which encompasses the control and with-computed operations in Quipper. We extend the language Proto-Quipper with reversing, control and the with-computed operation. Since not all circuits are reversible and/or controllable, we use a type system with modalities to track reversibility and controllability. This generalizes the modality of Fu-Kishida-Ross-Selinger 2023. We give an abstract categorical semantics, and show that the type system and operational semantics are sound with respect to this semantics.

cs.LO↗

arXiv2025

Nonreciprocal quantum phase transition in cavity magnonics

We investigate the nonreciprocal quantum phase transition in a cavity magnonic system driven by a parametric field, where an yttrium iron garnet (YIG) sphere is placed in a spinning microwave resonator. The system exhibits a rich phase diagram due to both magnon Kerr nonlinearity in YIG and parametric drive on the resonator. Especially, Sagnac-Fizeau shift caused by the spinning of the resonator brings about a significant modification in the critical driving strengths for second- and first-order quantum phase transitions, which means that the highly controllable quantum phase can be realized by the spinning speed of the resonator. More importantly, based on the difference in the detunings of the counterclockwise and clockwise modes induced by spinning direction of the resonator, we show that the phase transition in this system is nonreciprocal, that is, the quantum phase transition occurs when the cavity is driven in one direction but not the other. Our work offers an alternative path to engineer and design nonreciprocal magnonic devices.

quant-ph↗

arXiv2025

BeamLoRA: Beam-Constraint Low-Rank Adaptation

Due to the demand for efficient fine-tuning of large language models, Low-Rank Adaptation (LoRA) has been widely adopted as one of the most effective parameter-efficient fine-tuning methods. Nevertheless, while LoRA improves efficiency, there remains room for improvement in accuracy. Herein, we adopt a novel perspective to assess the characteristics of LoRA ranks. The results reveal that different ranks within the LoRA modules not only exhibit varying levels of importance but also evolve dynamically throughout the fine-tuning process, which may limit the performance of LoRA. Based on these findings, we propose BeamLoRA, which conceptualizes each LoRA module as a beam where each rank naturally corresponds to a potential sub-solution, and the fine-tuning process becomes a search for the optimal sub-solution combination. BeamLoRA dynamically eliminates underperforming sub-solutions while expanding the parameter space for promising ones, enhancing performance with a fixed rank. Extensive experiments across three base models and 12 datasets spanning math reasoning, code generation, and commonsense reasoning demonstrate that BeamLoRA consistently enhances the performance of LoRA, surpassing the other baseline methods.

cs.CL↗

arXiv2025

DIVE into MoE: Diversity-Enhanced Reconstruction of Large Language Models from Dense into Mixture-of-Experts

Large language models (LLMs) with the Mixture-of-Experts (MoE) architecture achieve high cost-efficiency by selectively activating a subset of the parameters. Despite the inference efficiency of MoE LLMs, the training of extensive experts from scratch incurs substantial overhead, whereas reconstructing a dense LLM into an MoE LLM significantly reduces the training budget. However, existing reconstruction methods often overlook the diversity among experts, leading to potential redundancy. In this paper, we come up with the observation that a specific LLM exhibits notable diversity after being pruned on different calibration datasets, based on which we present a Diversity-Enhanced reconstruction method named DIVE. The recipe of DIVE includes domain affinity mining, pruning-based expert reconstruction, and efficient retraining. Specifically, the reconstruction includes pruning and reassembly of the feed-forward network (FFN) module. After reconstruction, we efficiently retrain the model on routers, experts and normalization modules. We implement DIVE on Llama-style LLMs with open-source training corpora. Experiments show that DIVE achieves training efficiency with minimal accuracy trade-offs, outperforming existing pruning and MoE reconstruction methods with the same number of activated parameters.

cs.CL↗

arXiv2025

Adapt Once, Thrive with Updates: Transferable Parameter-Efficient Fine-Tuning on Evolving Base Models

Parameter-efficient fine-tuning (PEFT) has become a common method for fine-tuning large language models, where a base model can serve multiple users through PEFT module switching. To enhance user experience, base models require periodic updates. However, once updated, PEFT modules fine-tuned on previous versions often suffer substantial performance degradation on newer versions. Re-tuning these numerous modules to restore performance would incur significant computational costs. Through a comprehensive analysis of the changes that occur during base model updates, we uncover an interesting phenomenon: continual training primarily affects task-specific knowledge stored in Feed-Forward Networks (FFN), while having less impact on the task-specific pattern in the Attention mechanism. Based on these findings, we introduce Trans-PEFT, a novel approach that enhances the PEFT module by focusing on the task-specific pattern while reducing its dependence on certain knowledge in the base model. Further theoretical analysis supports our approach. Extensive experiments across 7 base models and 12 datasets demonstrate that Trans-PEFT trained modules can maintain performance on updated base models without re-tuning, significantly reducing maintenance overhead in real-world applications.

cs.CL↗

arXiv2024

Multimodal Hypothetical Summary for Retrieval-based Multi-image Question Answering

Retrieval-based multi-image question answering (QA) task involves retrieving multiple question-related images and synthesizing these images to generate an answer. Conventional "retrieve-then-answer" pipelines often suffer from cascading errors because the training objective of QA fails to optimize the retrieval stage. To address this issue, we propose a novel method to effectively introduce and reference retrieved information into the QA. Given the image set to be retrieved, we employ a multimodal large language model (visual perspective) and a large language model (textual perspective) to obtain multimodal hypothetical summary in question-form and description-form. By combining visual and textual perspectives, MHyS captures image content more specifically and replaces real images in retrieval, which eliminates the modality gap by transforming into text-to-text retrieval and helps improve retrieval. To more advantageously introduce retrieval with QA, we employ contrastive learning to align queries (questions) with MHyS. Moreover, we propose a coarse-to-fine strategy for calculating both sentence-level and word-level similarity scores, to further enhance retrieval and filter out irrelevant details. Our approach achieves a 3.7% absolute improvement over state-of-the-art methods on RETVQA and a 14.5% improvement over CLIP. Comprehensive experiments and detailed ablation studies demonstrate the superiority of our method.

cs.CV↗

arXiv2024

Are Large Language Models Table-based Fact-Checkers?

Table-based Fact Verification (TFV) aims to extract the entailment relation between statements and structured tables. Existing TFV methods based on small-scaled models suffer from insufficient labeled data and weak zero-shot ability. Recently, the appearance of Large Language Models (LLMs) has gained lots of attraction in research fields. They have shown powerful zero-shot and in-context learning abilities on several NLP tasks, but their potential on TFV is still unknown. In this work, we implement a preliminary study about whether LLMs are table-based fact-checkers. In detail, we design diverse prompts to explore how the in-context learning can help LLMs in TFV, i.e., zero-shot and few-shot TFV capability. Besides, we carefully design and construct TFV instructions to study the performance gain brought by the instruction tuning of LLMs. Experimental results demonstrate that LLMs can achieve acceptable results on zero-shot and few-shot TFV with prompt engineering, while instruction-tuning can stimulate the TFV capability significantly. We also make some valuable findings about the format of zero-shot prompts and the number of in-context examples. Finally, we analyze some possible directions to promote the accuracy of TFV via LLMs, which is beneficial to further research of table reasoning.

cs.CL↗

arXiv2024

Advancing Academic Knowledge Retrieval via LLM-enhanced Representation Similarity Fusion

In an era marked by robust technological growth and swift information renewal, furnishing researchers and the populace with top-tier, avant-garde academic insights spanning various domains has become an urgent necessity. The KDD Cup 2024 AQA Challenge is geared towards advancing retrieval models to identify pertinent academic terminologies from suitable papers for scientific inquiries. This paper introduces the LLM-KnowSimFuser proposed by Robo Space, which wins the 2nd place in the competition. With inspirations drawed from the superior performance of LLMs on multiple tasks, after careful analysis of the provided datasets, we firstly perform fine-tuning and inference using LLM-enhanced pre-trained retrieval models to introduce the tremendous language understanding and open-domain knowledge of LLMs into this task, followed by a weighted fusion based on the similarity matrix derived from the inference results. Finally, experiments conducted on the competition datasets show the superiority of our proposal, which achieved a score of 0.20726 on the final leaderboard.

cs.IR↗

arXiv2024

Think out Loud: Emotion Deducing Explanation in Dialogues

Humans convey emotions through daily dialogues, making emotion understanding a crucial step of affective intelligence. To understand emotions in dialogues, machines are asked to recognize the emotion for an utterance (Emotion Recognition in Dialogues, ERD); based on the emotion, then find causal utterances for the emotion (Emotion Cause Extraction in Dialogues, ECED). The setting of the two tasks requires first ERD and then ECED, ignoring the mutual complement between emotion and cause. To fix this, some new tasks are proposed to extract them simultaneously. Although the current research on these tasks has excellent achievements, simply identifying emotion-related factors by classification modeling lacks realizing the specific thinking process of causes stimulating the emotion in an explainable way. This thinking process especially reflected in the reasoning ability of Large Language Models (LLMs) is under-explored. To this end, we propose a new task "Emotion Deducing Explanation in Dialogues" (EDEN). EDEN recognizes emotion and causes in an explicitly thinking way. That is, models need to generate an explanation text, which first summarizes the causes; analyzes the inner activities of the speakers triggered by the causes using common sense; then guesses the emotion accordingly. To support the study of EDEN, based on the existing resources in ECED, we construct two EDEN datasets by human effort. We further evaluate different models on EDEN and find that LLMs are more competent than conventional PLMs. Besides, EDEN can help LLMs achieve better recognition of emotions and causes, which explores a new research direction of explainable emotion understanding in dialogues.

cs.CL↗

arXiv2024

Light-PEFT: Lightening Parameter-Efficient Fine-Tuning via Early Pruning

Parameter-efficient fine-tuning (PEFT) has emerged as the predominant technique for fine-tuning in the era of large language models. However, existing PEFT methods still have inadequate training efficiency. Firstly, the utilization of large-scale foundation models during the training process is excessively redundant for certain fine-tuning tasks. Secondly, as the model size increases, the growth in trainable parameters of empirically added PEFT modules becomes non-negligible and redundant, leading to inefficiency. To achieve task-specific efficient fine-tuning, we propose the Light-PEFT framework, which includes two methods: Masked Early Pruning of the Foundation Model and Multi-Granularity Early Pruning of PEFT. The Light-PEFT framework allows for the simultaneous estimation of redundant parameters in both the foundation model and PEFT modules during the early stage of training. These parameters can then be pruned for more efficient fine-tuning. We validate our approach on GLUE, SuperGLUE, QA tasks, and various models. With Light-PEFT, parameters of the foundation model can be pruned by up to over 40%, while still controlling trainable parameters to be only 25% of the original PEFT method. Compared to utilizing the PEFT method directly, Light-PEFT achieves training and inference speedup, reduces memory usage, and maintains comparable performance and the plug-and-play feature of PEFT.

cs.CL↗

arXiv2023

Object Attribute Matters in Visual Question Answering

Visual question answering is a multimodal task that requires the joint comprehension of visual and textual information. However, integrating visual and textual semantics solely through attention layers is insufficient to comprehensively understand and align information from both modalities. Intuitively, object attributes can naturally serve as a bridge to unify them, which has been overlooked in previous research. In this paper, we propose a novel VQA approach from the perspective of utilizing object attribute, aiming to achieve better object-level visual-language alignment and multimodal scene understanding. Specifically, we design an attribute fusion module and a contrastive knowledge distillation module. The attribute fusion module constructs a multimodal graph neural network to fuse attributes and visual features through message passing. The enhanced object-level visual features contribute to solving fine-grained problem like counting-question. The better object-level visual-language alignment aids in understanding multimodal scenes, thereby improving the model's robustness. Furthermore, to augment scene understanding and the out-of-distribution performance, the contrastive knowledge distillation module introduces a series of implicit knowledge. We distill knowledge into attributes through contrastive loss, which further strengthens the representation learning of attribute features and facilitates visual-linguistic alignment. Intensive experiments on six datasets, COCO-QA, VQAv2, VQA-CPv2, VQA-CPv1, VQAvs and TDIUC, show the superiority of the proposed method.

cs.CV↗