arXiv subjects

Chunyi Peng

Publications and source records attributed to Chunyi Peng.

14 recordsLinked to original sources

arXiv2026

ConceptFormer: Learning Adaptive Latent Concepts for Query-Document Alignment in Visual Document Retrieval

Visual document retrieval is a critical component of multimodal retrieval-augmented generation, aiming to identify query-relevant pages from document collections where evidence is distributed across text, layout, charts, and visual structures. Recent efforts toward finer-grained supervision primarily rely on textual descriptions or localized visual regions as evidence proxies. However, such supervision signals may either overlook complex visual structures or provide incomplete and inaccurate representations of the underlying evidence. To address these limitations, we propose ConceptFormer, a latent concept representation learning framework for visual document retrieval. ConceptFormer models query-relevant evidence as continuous, query-conditioned latent concepts that explicitly bridge localized visual evidence and semantic relevance, without requiring either textual intermediate representations or direct reliance on raw visual annotations. During training, ConceptFormer employs a strong vision-language model to dynamically determine the number of latent concept tokens and uses these concepts as an intermediate representation to bridge the semantic gap between queries and documents, thereby guiding the learning of the embedding space. Experiments on diverse visual document retrieval benchmarks demonstrate that ConceptFormer achieves 16.7\% and 22.1\% relative improvements in average NDCG@10 over the strongest visual retrieval baseline and the strongest OCR-based text retrieval baseline, respectively. Further analysis reveals that latent concepts effectively connect localized visual evidence with semantic relevance, enabling the retriever to capture both fine-grained textual cues and complex document-level visual structures while preserving strong retrieval alignment. Codes and data are available at https://github.com/Neuir/ConceptFormer.

cs.CV↗

arXiv2026

HiEviDR-Bench: A Benchmark for Hierarchical Evidence Aggregation in Deep Research

Deep research requires models to retrieve, connect, and synthesize evidence from large-scale heterogeneous sources to answer complex queries and produce analytical reports. Existing benchmarks mainly evaluate final outcomes, such as answer correctness, report quality, or citation alignment, while providing limited visibility into whether evidence is correctly selected, linked, and aggregated into supported claims and conclusions. To address this gap, we introduce HiEviDR-Bench, a benchmark for evaluating Hierarchical Evidence Aggregation in Deep Research. HiEviDR-Bench covers open-domain and academic-domain settings under both text-only and multimodal conditions, and represents each instance with an explicit evidence graph that captures evidence selection, cross-source linking, and aggregation from evidence to intermediate claims and final conclusions. Based on this formulation, we develop a traceability-oriented evaluation framework with five dimensions: report quality, evidence traceability, citation accuracy, claim verification, and answer correctness, together with a progressive gating mechanism for fine-grained error localization. HiEviDR-Bench contains 2,000 human-validated questions with evidence graphs across multiple difficulty levels. Experiments on 16 representative multimodal large language models show that, although many systems achieve strong report quality, their performance drops markedly on citation accuracy, claim construction, and answer correctness. Further analysis shows that the main bottlenecks lie in evidence identification and intermediate claim construction, revealing that strong surface-level report quality does not necessarily imply grounded multi-stage reasoning on our benchmark.

cs.IR↗

arXiv2026

Memory Shot for Long-Term Dialogue

Large Language Models (LLMs) have demonstrated strong capabilities in general conversation, instruction following, and complex reasoning. However, in long-term dialogue settings, they often struggle to locate and utilize historical information most relevant to the current query. Existing approaches address this issue by constructing structured text-centered memory units through compressing and reorganizing user interaction history. However, these systems often rely on brute-force extraction of crucial evidence to associate episodes across dialogue sessions, causing substantial computational overhead and weakening structural cues such as speaker transitions, turn boundaries, and local contextual relationships. To avoid fragile text-based memory representations, we propose MemShot, which leverages dialogue structuring for long-term dialogue modeling and relies on the model's internal visual reasoning capabilities to associate key episodes. Specifically, MemShot renders local contiguous dialogue spans into structured visual memory units, preserving meta-information and chronological dialogue turns while avoiding heavy-weight textual memory construction. Experimental results show that MemShot achieves stable and competitive performance on both LoCoMo and LongMemEval, while substantially shortening the memory construction pipeline and delivering 70$\times$ speedup. Further analysis reveals that MemShot enhances the localization and utilization of historical evidence by directing memory processing toward structured local dialogue cues rather than surface-level lexical matching in a flat text stream. All codes are released on https://github.com/NEUIR/MemShot.

cs.IR↗

arXiv2026

CC-OCR V2: Fine-Grained Attribution of LMM Failures in Real-World Visual Document Understanding

Recent Large Multimodal Models (LMMs) have achieved remarkable progress on OCR-centric document understanding and processing tasks. Existing benchmarks primarily evaluate LMMs across diverse tasks to reflect practical document-processing workflows or analyze how document characteristics influence model performance. However, they provide limited insight into the reliability of LMMs under real-world document acquisition conditions, where factors such as lighting, screen displays, imaging quality, and capture methods can substantially affect performance. To bridge this gap, we present CC-OCR v2, a comprehensive benchmark for attributing LMM failures in real-world document processing. CC-OCR v2 provides a unified evaluation framework covering five core document-processing capabilities through 16 subtasks, 74 application scenarios, and 7,093 samples. The benchmark spans five evaluation tracks, ten document categories, and 32 languages in the recognition suite. Beyond task-level evaluation, each sample is annotated with ten fine-grained document factors, enabling systematic attribution of model failures across document types and acquisition conditions. Extensive experiments on 17 representative LMMs reveal substantial performance variation across tasks, document categories, and real-world conditions. Moreover, models with comparable overall accuracy often exhibit fundamentally different failure patterns under specific document factors, highlighting a significant gap between benchmark-level performance and reliable deployment in practical applications. The dataset and evaluation toolkit are publicly available at https://github.com/eioss/CC-OCR-V2.

cs.CL↗

arXiv2026

Lang2Act: Fine-Grained Visual Reasoning through Self-Emergent Linguistic Toolchains

Visual Retrieval-Augmented Generation (VRAG) enhances Vision-Language Models (VLMs) by incorporating external visual documents to address a given query. Existing VRAG frameworks usually depend on rigid, pre-defined external tools to extend the perceptual capabilities of VLMs, typically by explicitly separating visual perception from subsequent reasoning processes. However, this decoupled design can lead to unnecessary loss of visual information, particularly when image-based operations such as cropping are applied. In this paper, we propose Lang2Act, which enables fine-grained visual perception and reasoning through self-emergent linguistic toolchains. Rather than invoking fixed external engines, Lang2Act collects self-emergent actions as linguistic tools and leverages them to enhance the visual perception capabilities of VLMs. To support this mechanism, we design a two-stage Reinforcement Learning (RL)-based training framework. Specifically, the first stage optimizes VLMs to self-explore high-quality actions for constructing a reusable linguistic toolbox, and the second stage further optimizes VLMs to exploit these linguistic tools for downstream reasoning effectively. Experimental results demonstrate the effectiveness of Lang2Act in substantially enhancing the visual perception capabilities of VLMs, achieving performance improvements of over 4%. All code and data are available at https://github.com/NEUIR/Lang2Act.

cs.AI↗

arXiv2025

VisRAG2.0: Mitigating Visual Hallucinations via Evidence-Guided Multi-Image Reasoning in Visual Retrieval-Augmented Generation

Visual Retrieval-Augmented Generation (VRAG) has emerged as a promising paradigm for equipping Vision-Language Models (VLMs) with external visual evidence, enabling them to go beyond parametric knowledge when answering visually grounded questions. However, in such multi-image settings, VLMs still often suffer from visual hallucinations and struggle to accurately identify the question-relevant evidence needed for reliable reasoning. Existing methods usually lack an explicit cross-image evidence collection process, and also provide limited credit assignment when jointly optimizing perception and reasoning. To address this issue, we propose EVisRAG, an evidence-guided visual retrieval-augmented framework for multi-image reasoning. EVisRAG first observes the retrieved images, records question-relevant visual evidence from each image, and then performs reasoning and answer generation based on the aggregated evidence. We further introduce RS-GRPO, which aligns reward signals with token spans from different stages, improving training stability and strengthening the joint optimization of evidence localization and reasoning. Experiments on multiple visual question answering benchmarks show that EVisRAG consistently outperforms the backbone VLM by an average of about 19\%, while substantially reducing visual hallucinations. These results demonstrate that explicit evidence collection and scoped reward design are effective for improving visual grounding and reasoning reliability in multi-image settings. Codes and data are available at https://github.com/OpenBMB/VisRAG

cs.CL↗

arXiv2025

ReCUT: Balancing Reasoning Length and Accuracy in LLMs via Stepwise Trails and Preference Optimization

Recent advances in Chain-of-Thought (CoT) prompting have substantially improved the reasoning capabilities of Large Language Models (LLMs). However, these methods often suffer from overthinking, leading to unnecessarily lengthy or redundant reasoning traces. Existing approaches attempt to mitigate this issue through curating multiple reasoning chains for training LLMs, but their effectiveness is often constrained by the quality of the generated data and prone to overfitting. To address the challenge, we propose Reasoning Compression ThroUgh Stepwise Trials (ReCUT), a novel method aimed at balancing the accuracy and length of reasoning trajectory. Specifically, ReCUT employs a stepwise exploration mechanism and a long-short switched sampling strategy, enabling LLMs to incrementally generate diverse reasoning paths. These paths are evaluated and used to construct preference pairs to train two specialized models (Gemini LLMs)-one optimized for reasoning accuracy, the other for shorter reasoning. A final integrated model is obtained by interpolating the parameters of these two models. Experimental results across multiple math reasoning datasets and backbone models demonstrate that ReCUT significantly reduces reasoning lengths by approximately 30-50%, while maintaining or improving reasoning accuracy compared to various baselines. All codes and data will be released via https://github.com/NEUIR/ReCUT.

cs.CL↗

arXiv2025

Mixture-of-Retrieval Experts for Reasoning-Guided Multimodal Knowledge Exploitation

Multimodal Retrieval-Augmented Generation (MRAG) has shown promise in mitigating hallucinations in Multimodal Large Language Models (MLLMs) by incorporating external knowledge. However, existing methods typically adhere to rigid retrieval paradigms by mimicking fixed retrieval trajectories and thus fail to fully exploit the knowledge of different retrieval experts through dynamic interaction based on the model's knowledge needs or evolving reasoning states. To overcome this limitation, we introduce Mixture-of-Retrieval Experts (MoRE), a novel framework that enables MLLMs to collaboratively interact with diverse retrieval experts for more effective knowledge exploitation. Specifically, MoRE learns to dynamically determine which expert to engage with, conditioned on the evolving reasoning state. To effectively train this capability, we propose Stepwise Group Relative Policy Optimization (Step-GRPO), which goes beyond sparse outcome-based supervision by encouraging MLLMs to interact with multiple retrieval experts and synthesize fine-grained rewards, thereby teaching the MLLM to fully coordinate all experts when answering a given query. Experimental results on diverse open-domain QA benchmarks demonstrate the effectiveness of MoRE, achieving average performance gains of over 7% compared to competitive baselines. Notably, MoRE exhibits strong adaptability by dynamically coordinating heterogeneous experts to precisely locate relevant information, validating its capability for robust, reasoning-driven expert collaboration. All codes and data are released on https://github.com/OpenBMB/MoRE.

cs.CL↗

arXiv2021

Towards Live Video Analytics with On-Drone Deeper-yet-Compatible Compression

In this work, we present DCC(Deeper-yet-Compatible Compression), one enabling technique for real-time drone-sourced edge-assisted video analytics built on top of the existing codec. DCC tackles an important technical problem to compress streamed video from the drone to the edge without scarifying accuracy and timeliness of video analytical tasks performed at the edge. DCC is inspired by the fact that not every bit in streamed video is equally valuable to video analytics, which opens new compression room over the conventional analytics-oblivious video codec technology. We exploit drone-specific context and intermediate hints from object detection to pursue adaptive fidelity needed to retain analytical quality. We have prototyped DCC in one showcase application of vehicle detection and validated its efficiency in representative scenarios. DCC has reduced transmission volume by 9.5-fold over the baseline approach and 19-683% over the state-of-the-art with comparable detection accuracy.

cs.NI↗

arXiv2020

I-BOT: Interference-Based Orchestration of Tasks for Dynamic Unmanaged Edge Computing

In recent years, edge computing has become a popular choice for latency-sensitive applications like facial recognition and augmented reality because it is closer to the end users compared to the cloud. Although infrastructure providers are working toward creating managed edge networks, personal devices such as laptops and tablets, which are widely available and are underutilized, can also be used as potential edge devices. We call such devices Unmanaged Edge Devices (UEDs). Scheduling application tasks on such an unmanaged edge system is not straightforward because of three fundamental reasons-heterogeneity in the computational capacity of the UEDs, uncertainty in the availability of the UEDs (due to devices leaving the system), and interference among multiple tasks sharing a UED. In this paper, we present I-BOT, an interference-based orchestration scheme for latency-sensitive tasks on an Unmanaged Edge Platform (UEP). It minimizes the completion time of applications and is bandwidth efficient. I-BOT brings forth three innovations. First, it profiles and predicts the interference patterns of the tasks to make scheduling decisions. Second, it uses a feedback mechanism to adjust for changes in the computational capacity of the UEDs and a prediction mechanism to handle their sporadic exits. Third, it accounts for input dependence of tasks in its scheduling decision (such as, two tasks requiring the same input data). To evaluate I-BOT, we run end-to-end simulations with applications representing autonomous driving, composed of multiple tasks. We compare to two basic baselines (random and round-robin) and two state-of-the-arts, Lavea [SEC-2017] and Petrel [MSN-2018]. Compared to these baselines, I-BOT significantly reduces the average service time of application tasks. This reduction is more pronounced in dynamic heterogeneous environments, which would be the case in a UEP.

cs.DC↗

arXiv2019

Resilient Cyberphysical Systems and their Application Drivers: A Technology Roadmap

Cyberphysical systems (CPS) are ubiquitous in our personal and professional lives, and they promise to dramatically improve micro-communities (e.g., urban farms, hospitals), macro-communities (e.g., cities and metropolises), urban structures (e.g., smart homes and cars), and living structures (e.g., human bodies, synthetic genomes). The question that we address in this article pertains to designing these CPS systems to be resilient-from-the-ground-up, and through progressive learning, resilient-by-reaction. An optimally designed system is resilient to both unique attacks and recurrent attacks, the latter with a lower overhead. Overall, the notion of resilience can be thought of in the light of three main sources of lack of resilience, as follows: exogenous factors, such as natural variations and attack scenarios; mismatch between engineered designs and exogenous factors ranging from DDoS (distributed denial-of-service) attacks or other cybersecurity nightmares, so called "black swan" events, disabling critical services of the municipal electrical grids and other connected infrastructures, data breaches, and network failures; and the fragility of engineered designs themselves encompassing bugs, human-computer interactions (HCI), and the overall complexity of real-world systems. In the paper, our focus is on design and deployment innovations that are broadly applicable across a range of CPS application areas.

cs.CY↗

arXiv2018

The Untold Secrets of Operational Wi-Fi Calling Services: Vulnerabilities, Attacks, and Countermeasures

Since 2016, all of four major U.S. operators have rolled out nationwide Wi-Fi calling services. They are projected to surpass VoLTE (Voice over LTE) and other VoIP services in terms of mobile IP voice usage minutes in 2018. They enable mobile users to place cellular calls over Wi-Fi networks based on the 3GPP IMS (IP Multimedia Subsystem) technology. Compared with conventional cellular voice solutions, the major difference lies in that their traffic traverses untrustful Wi-Fi networks and the Internet. This exposure to insecure networks may cause the Wi-Fi calling users to suffer from security threats. Its security mechanisms are similar to the VoLTE, because both of them are supported by the IMS. They include SIM-based security, 3GPP AKA (Authentication and Key Agreement), IPSec (Internet Protocol Security), etc. However, are they sufficient to secure Wi-Fi calling services? Unfortunately, our study yields a negative answer. We conduct the first study of exploring security issues of the operational Wi-Fi calling services in three major U.S. operators' networks using commodity devices. We disclose that current Wi-Fi calling security is not bullet-proof and uncover four vulnerabilities which stem from improper standard designs, device implementation issues and network operation slips. By exploiting the vulnerabilities, together with several state-of-the-art computer visual recognition technologies, we devise two proof-of-concept attacks: user privacy leakage and telephony harassment or denial of voice service (THDoS); both of them can bypass the security defenses deployed on mobile devices and the network infrastructure. We have confirmed their feasibility and simplicity using real-world experiments, as well as assessed their potential damages and proposed recommended solutions.

cs.CR↗

arXiv2015

New Threats to SMS-Assisted Mobile Internet Services from 4G LTE: Lessons Learnt from Distributed Mobile-Initiated Attacks towards Facebook and Other Services

Mobile Internet is becoming the norm. With more personalized mobile devices in hand, many services choose to offer alternative, usually more convenient, approaches to authenticating and delivering the content between mobile users and service providers. One main option is to use SMS (i.e., short messaging service). Such carrier-grade text service has been widely used to assist versatile mobile services, including social networking, banking, to name a few. Though the text service can be spoofed via certain Internet text service providers which cooperated with carriers, such attacks haven well studied and defended by industry due to the efforts of research community. However, as cellular network technology advances to the latest IP-based 4G LTE, we find that these mobile services are somehow exposed to new threats raised by this change, particularly on 4G LTE Text service (via brand-new distributed Mobile-Initiated Spoofed SMS attack which is not available in legacy 2G/3G systems). The reason is that messaging service over LTE shifts from the circuit-switched (CS) design to the packet-switched (PS) paradigm as 4G LTE supports PS only. Due to this change, 4G LTE Text Service becomes open to access. However, its shields to messaging integrity and user authentication are not in place. As a consequence, such weaknesses can be exploited to launch attacks (e.g., hijack Facebook accounts) against a targeted individual, a large scale of mobile users and even service providers, from mobile devices. Current defenses for Internet-Initiated Spoofed SMS attacks cannot defend the unprecedented attack. Our study shows that 53 of 64 mobile services over 27 industries are vulnerable to at least one threat. We validate these proof-of-concept attacks in one major US carrier which supports more than 100 million users. We finally propose quick fixes and discuss security insights and lessons we have learnt.

cs.CR↗

arXiv2015

iCellular: Define Your Own Cellular Network Access on Commodity Smartphones

Leveraging multi-carrier access offers a promising approach to boosting access quality in mobile networks. However, our experiments show that the potential benefits are hard to fulfill due to fundamental limitations in the network-controlled design. To overcome these limitations, we propose iCellular, which allows users to define and intelligently select their own cellular network access from multiple carriers. iCellular reuses the existing device-side mechanisms and the standard cellular network procedure, but leverages the end device's intelligence to be proactive and adaptive in multi-carrier selection. It performs adaptive monitoring to ensure responsive selection and minimal service disruption, and enhances carrier selection with online learning and runtime decision fault prevention. It is deployable on commodity phones without any infrastructure/hardware change. We implement iCellular on commodity Nexus 6 phones and leverage Google Project-Fi's efforts to test multi-carrier access among two top US carriers: T-Mobile and Sprint. Our experiments confirm that iCellular helps users with up to 3.74x throughput improvement (7x suspension and 1.9x latency reduction) over the state-of-art selection. Moreover, iCellular locates the best-quality carrier in most cases, with negligible overhead on CPU, memory and energy consumption.

cs.NI↗