arXiv subjects

Qipeng Wang

Publications and source records attributed to Qipeng Wang.

At least 19 recordsLinked to original sources

arXiv2026

Unlocking Multimodal Protein Language Models at Inference Time

Multimodal protein language models (pLMs) learn joint protein sequence-structure distributions, and their generation performance should also depend critically on inference-time sampling strategies. Yet prior work has focused more on model training than on how inference-time strategies behave. In this paper, we establish a three-stage investigation framework to empirically study the inference design space of multimodal pLMs across three representative pLMs and four fundamental tasks. We evaluate vanilla sampling, task-specific classifier-free guidance, and reward-guided beam search on multimodal pLMs, corresponding to controls over sampling distributions, per-step logits, and parallel trajectories. Throughout the complementary advancements centered on exploration-exploitation trade-off, we (1) reveal the suboptimality of default inference protocols and identify task-oriented sampling preferences; (2) observe substantial quantitative gains across tasks, consistently boosting the upper bound performance of multimodal pLMs without updating model parameters; (3) derive conclusions about base models that differ from prior consensus.

cs.CE↗

arXiv2026

Semantic Steering for Controllable Generation: Tuning-Free Concept Erasure in Multimodal Diffusion Transformers

Multimodal Diffusion Transformers (MM-DiTs) have demonstrated remarkable text-to-image generation performance, surpassing traditional U-Net-based diffusion models. Nevertheless, their powerful generative capabilities also raise significant safety concerns, as they may generate sensitive or inappropriate content. While existing concept erasure methods aim to mitigate such risks, most require modifying model parameters, which are often architecture-specific and impractical for deployed larger models. Several tuning-free approaches face challenges when applied to advanced large-scale MM-DiTs due to their deeply embedded knowledge, broad semantic space, and context-dependent text encoders. To address these challenges, we propose to erase concepts by directly manipulating the model's internal representations. Our key insight, derived from an in-depth analysis of MM-DiT's block-wise generative roles, is that text-conditioned semantic representations are most salient in the middle blocks of MM-DiTs. Based on this, we extract representations of an unwanted concept and a desirable safe one from the middle block, construct a steering vector from their difference, and inject this single vector into consecutive early and middle blocks. By operating exclusively on the sparse text-branch tokens and leveraging the straight sampling trajectory of rectified flow, our method achieves effective concept erasure with negligible overhead and without any training. Extensive experiments across MM-DiT models demonstrate that our method achieves state-of-the-art performance in erasing diverse concepts, enables effective control over the final output, and remains robust to adversarial attacks.

cs.CV↗

arXiv2026

NeuroCogMap Reveals Cognitive Organization of Large Language Models

Understanding how complex cognitive functions are organized within artificial systems is central to interpreting large language models (LLMs) and relating them to biological cognition. Yet although LLMs exhibit broad cognitive-like behaviours, it remains unclear whether their internal representations form reproducible functional systems that explain behaviour, failure and links to human cognition. Here we present NeuroCogMap, a cognitive neuroscience-inspired framework that organizes internal features of LLMs into functional parcels and links them to interpretable functions, cognitive capabilities and a cognitive hierarchy. These parcels form a stable and semantically coherent organization that is partly conserved across models and functionally linked to model outputs. Within this organization, major LLM failures, including hallucination, bias, refusal failure and sycophancy, correspond to distinct disruptions in representational and behavioural-control systems, yielding internal signatures for mechanism-guided detection and targeted intervention. Beyond model behaviour, NeuroCogMap improves prediction of human cortical responses during naturalistic language comprehension, with the strongest correspondence in higher-order association cortex. At the cognitive level, its internal signatures expose latent strategies that guide refinements of classical models of human decision-making. Together, these findings establish NeuroCogMap as a system-level framework for mapping functional organization in artificial systems and for relating this organization to human cortical function and cognitive behaviour.

q-bio.NC↗

arXiv2026

Taming Request Imbalance: SLO-Aware Scheduling for Disaggregated LLM Inference

In production environments, large language model (LLM) serving is required to meet stringent service-level objectives (SLOs) amid highly variable request patterns. In practice, request lengths follow a long-tail distribution, which gives rise to head-of-line blocking on the prefill side and underutilization caused by stragglers on the decode side in disaggregated serving architectures. Current systems, which adopt first-come-first-served (FCFS) scheduling for prefill and continuous batching for decode, lack the ability to adapt to this imbalance, resulting in compromised SLO attainment and reduced throughput. To address these challenges, we propose Kairos, an SLO-aware scheduling system equipped with two complementary mechanisms. On the prefill side, Kairos employs urgency-based priority scheduling: it predicts prefill completion times and dynamically selects requests to maximize the attainment of time-to-first-token (TTFT) SLOs. On the decode side, Kairos introduces slack-guided adaptive batching, which leverages the gap between per-step decode time and the time-per-output-token (TPOT) SLO to greedily pack short requests. This approach maximizes throughput while strictly adhering to SLO requirements. We implement Kairos and conduct evaluations using an online serving dataset and a state-of-the-art LLM. Experimental results demonstrate that, compared with state-of-the-art baselines, Kairos improves TTFT SLO attainment by up to 23.9\%, TPOT SLO attainment by up to 27.1\%, end-to-end SLO attainment by up to 33.8\%, and decode throughput by up to 19.3\%.

cs.DC↗

arXiv2026

Deep Search with Hierarchical Meta-Cognitive Monitoring Inspired by Cognitive Neuroscience

Deep search agents powered by large language models have demonstrated strong capabilities in multi-step retrieval, reasoning, and long-horizon task execution. However, their practical failures often stem from the lack of mechanisms to monitor and regulate reasoning and retrieval states as tasks evolve under uncertainty. Insights from cognitive neuroscience suggest that human metacognition is hierarchically organized, integrating fast anomaly detection with selectively triggered, experience-driven reflection. In this work, we propose Deep Search with Meta-Cognitive Monitoring (DS-MCM), a deep search framework augmented with an explicit hierarchical metacognitive monitoring mechanism. DS-MCM integrates a Fast Consistency Monitor, which performs lightweight checks on the alignment between external evidence and internal reasoning confidence, and a Slow Experience-Driven Monitor, which is selectively activated to guide corrective intervention based on experience memory from historical agent trajectories. By embedding monitoring directly into the reasoning-retrieval loop, DS-MCM determines both when intervention is warranted and how corrective actions should be informed by prior experience. Experiments across multiple deep search benchmarks and backbone models demonstrate that DS-MCM consistently improves performance and robustness.

cs.CL↗

arXiv2025

MedKGI: Iterative Differential Diagnosis with Medical Knowledge Graphs and Information-Guided Inquiring

Recent advancements in Large Language Models (LLMs) have demonstrated significant promise in clinical diagnosis. However, current models struggle to emulate the iterative, diagnostic hypothesis-driven reasoning of real clinical scenarios. Specifically, current LLMs suffer from three critical limitations: (1) generating hallucinated medical content due to weak grounding in verified knowledge, (2) asking redundant or inefficient questions rather than discriminative ones that hinder diagnostic progress, and (3) losing coherence over multi-turn dialogues, leading to contradictory or inconsistent conclusions. To address these challenges, we propose MedKGI, a diagnostic framework grounded in clinical practices. MedKGI integrates a medical knowledge graph (KG) to constrain reasoning to validated medical ontologies, selects questions based on information gain to maximize diagnostic efficiency, and adopts an OSCE-format structured state to maintain consistent evidence tracking across turns. Experiments on clinical benchmarks show that MedKGI outperforms strong LLM baselines in both diagnostic accuracy and inquiry efficiency, improving dialogue efficiency by 30% on average while maintaining state-of-the-art accuracy.

cs.CL↗

arXiv2025

TrajLens: Visual Analysis for Constructing Cell Developmental Trajectories in Cross-Sample Exploration

Constructing cell developmental trajectories is a critical task in single-cell RNA sequencing (scRNA-seq) analysis, enabling the inference of potential cellular progression paths. However, current automated methods are limited to establishing cell developmental trajectories within individual samples, necessitating biologists to manually link cells across samples to construct complete cross-sample evolutionary trajectories that consider cellular spatial dynamics. This process demands substantial human effort due to the complex spatial correspondence between each pair of samples. To address this challenge, we first proposed a GNN-based model to predict cross-sample cell developmental trajectories. We then developed TrajLens, a visual analytics system that supports biologists in exploring and refining the cell developmental trajectories based on predicted links. Specifically, we designed the visualization that integrates features on cell distribution and developmental direction across multiple samples, providing an overview of the spatial evolutionary patterns of cell populations along trajectories. Additionally, we included contour maps superimposed on the original cell distribution data, enabling biologists to explore them intuitively. To demonstrate our system's performance, we conducted quantitative evaluations of our model with two case studies and expert interviews to validate its usefulness and effectiveness.

cs.CG↗

arXiv2025

Detection and Mitigation of Hallucination in Large Reasoning Models: A Mechanistic Perspective

Large Reasoning Models (LRMs) have shown impressive capabilities in multi-step reasoning tasks. However, alongside these successes, a more deceptive form of model error has emerged--Reasoning Hallucination--where logically coherent but factually incorrect reasoning traces lead to persuasive yet faulty conclusions. Unlike traditional hallucinations, these errors are embedded within structured reasoning, making them more difficult to detect and potentially more harmful. In this work, we investigate reasoning hallucinations from a mechanistic perspective. We propose the Reasoning Score, which quantifies the depth of reasoning by measuring the divergence between logits obtained from projecting late layers of LRMs to the vocabulary space, effectively distinguishing shallow pattern-matching from genuine deep reasoning. Using this score, we conduct an in-depth analysis on the ReTruthQA dataset and identify two key reasoning hallucination patterns: early-stage fluctuation in reasoning depth and incorrect backtracking to flawed prior steps. These insights motivate our Reasoning Hallucination Detection (RHD) framework, which achieves state-of-the-art performance across multiple domains. To mitigate reasoning hallucinations, we further introduce GRPO-R, an enhanced reinforcement learning algorithm that incorporates step-level deep reasoning rewards via potential-based shaping. Our theoretical analysis establishes stronger generalization guarantees, and experiments demonstrate improved reasoning quality and reduced hallucination rates.

cs.AI↗

arXiv2025

IntelliCircos: A Data-driven and AI-powered Authoring Tool for Circos Plots

Genomics data is essential in biological and medical domains, and bioinformatics analysts often manually create circos plots to analyze the data and extract valuable insights. However, creating circos plots is complex, as it requires careful design for multiple track attributes and positional relationships between them. Typically, analysts often seek inspiration from existing circos plots, and they have to iteratively adjust and refine the plot to achieve a satisfactory final design, making the process both tedious and time-intensive. To address these challenges, we propose IntelliCircos, an AI-powered interactive authoring tool that streamlines the process from initial visual design to the final implementation of circos plots. Specifically, we build a new dataset containing 4396 circos plots with corresponding annotations and configurations, which are extracted and labeled from published papers. With the dataset, we further identify track combination patterns, and utilize Large Language Model (LLM) to provide domain-specific design recommendations and configuration references to navigate the design of circos plots. We conduct a user study with 8 bioinformatics analysts to evaluate IntelliCircos, and the results demonstrate its usability and effectiveness in authoring circos plots.

cs.HC↗

arXiv2025

Intelligent Reflecting Surface Based Localization of Mixed Near-Field and Far-Field Targets

This paper considers an intelligent reflecting surface (IRS)-assisted bi-static localization architecture for the sixth-generation (6G) integrated sensing and communication (ISAC) network. The system consists of a transmit user, a receive base station (BS), an IRS, and multiple targets in either the far-field or near-field region of the IRS. In particular, we focus on the challenging scenario where the line-of-sight (LOS) paths between targets and the BS are blocked, such that the emitted orthogonal frequency division multiplexing (OFDM) signals from the user reach the BS merely via the user-target-IRS-BS path. Based on the signals received by the BS, our goal is to localize the targets by estimating their relative positions to the IRS, instead of to the BS. We show that subspace-based methods, such as the multiple signal classification (MUSIC) algorithm, can be applied onto the BS's received signals to estimate the relative states from the targets to the IRS. To this end, we create a virtual signal via combining user-target-IRS-BS channels over various time slots. By applying MUSIC on such a virtual signal, we are able to detect the far-field targets and the near-field targets, and estimate the angle-of-arrivals (AOAs) and/or ranges from the targets to the IRS. Furthermore, we theoretically verify that the proposed method can perfectly estimate the relative states from the targets to the IRS in the ideal case with infinite coherence blocks. Numerical results verify the effectiveness of our proposed IRS-assisted localization scheme. Our paper demonstrates the potential of employing passive anchors, i.e., IRSs, to improve the sensing coverage of the active anchors, i.e., BSs.

eess.SP↗

arXiv2025

ReARTeR: Retrieval-Augmented Reasoning with Trustworthy Process Rewarding

Retrieval-Augmented Generation (RAG) systems for Large Language Models (LLMs) hold promise in knowledge-intensive tasks but face limitations in complex multi-step reasoning. While recent methods have integrated RAG with chain-of-thought reasoning or test-time search using Process Reward Models (PRMs), these approaches encounter challenges such as a lack of explanations, bias in PRM training data, early-step bias in PRM scores, and insufficient post-training optimization of reasoning potential. To address these issues, we propose Retrieval-Augmented Reasoning through Trustworthy Process Rewarding (ReARTeR), a framework that enhances RAG systems' reasoning capabilities through post-training and test-time scaling. At test time, ReARTeR introduces Trustworthy Process Rewarding via a Process Reward Model for accurate scalar scoring and a Process Explanation Model (PEM) for generating natural language explanations, enabling step refinement. During post-training, it utilizes Monte Carlo Tree Search guided by Trustworthy Process Rewarding to collect high-quality step-level preference data, optimized through Iterative Preference Optimization. ReARTeR addresses three core challenges: (1) misalignment between PRM and PEM, tackled through off-policy preference learning; (2) bias in PRM training data, mitigated by balanced annotation methods and stronger annotations for challenging examples; and (3) early-step bias in PRM, resolved through a temporal-difference-based look-ahead search strategy. Experimental results on multi-step reasoning benchmarks demonstrate significant improvements, underscoring ReARTeR's potential to advance the reasoning capabilities of RAG systems.

cs.CL↗

arXiv2024

SynthLens: Visual Analytics for Facilitating Multi-step Synthetic Route Design

Designing synthetic routes for novel molecules is pivotal in various fields like medicine and chemistry. In this process, researchers need to explore a set of synthetic reactions to transform starting molecules into intermediates step by step until the target novel molecule is obtained. However, designing synthetic routes presents challenges for researchers. First, researchers need to make decisions among numerous possible synthetic reactions at each step, considering various criteria (e.g., yield, experimental duration, and the count of experimental steps) to construct the synthetic route. Second, they must consider the potential impact of one choice at each step on the overall synthetic route. To address these challenges, we proposed SynthLens, a visual analytics system to facilitate the iterative construction of synthetic routes by exploring multiple possibilities for synthetic reactions at each step of construction. Specifically, we have introduced a tree-form visualization in SynthLens to compare and evaluate all the explored routes at various exploration steps, considering both the exploration step and multiple criteria. Our system empowers researchers to consider their construction process comprehensively, guiding them toward promising exploration directions to complete the synthetic route. We validated the usability and effectiveness of SynthLens through a quantitative evaluation and expert interviews, highlighting its role in facilitating the design process of synthetic routes. Finally, we discussed the insights of SynthLens to inspire other multi-criteria decision-making scenarios with visual analytics.

cs.HC↗

arXiv2024

A Survey of Resource-efficient LLM and Multimodal Foundation Models

Large foundation models, including large language models (LLMs), vision transformers (ViTs), diffusion, and LLM-based multimodal models, are revolutionizing the entire machine learning lifecycle, from training to deployment. However, the substantial advancements in versatility and performance these models offer come at a significant cost in terms of hardware resources. To support the growth of these large models in a scalable and environmentally sustainable way, there has been a considerable focus on developing resource-efficient strategies. This survey delves into the critical importance of such research, examining both algorithmic and systemic aspects. It offers a comprehensive analysis and valuable insights gleaned from existing literature, encompassing a broad array of topics from cutting-edge model architectures and training/serving algorithms to practical system designs and implementations. The goal of this survey is to provide an overarching understanding of how current approaches are tackling the resource challenges posed by large foundation models and to potentially inspire future breakthroughs in this field.

cs.LG↗

arXiv2024

Anatomizing Deep Learning Inference in Web Browsers

Web applications have increasingly adopted Deep Learning (DL) through in-browser inference, wherein DL inference performs directly within Web browsers. The actual performance of in-browser inference and its impacts on the quality of experience (QoE) remain unexplored, and urgently require new QoE measurements beyond traditional ones, e.g., mainly focusing on page load time. To bridge this gap, we make the first comprehensive performance measurement of in-browser inference to date. Our approach proposes new metrics to measure in-browser inference: responsiveness, smoothness, and inference accuracy. Our extensive analysis involves 9 representative DL models across Web browsers of 50 popular PC devices and 20 mobile devices. The results reveal that in-browser inference exhibits a substantial latency gap, averaging 16.9 times slower on CPU and 4.9 times slower on GPU compared to native inference on PC devices. The gap on mobile CPU and mobile GPU is 15.8 times and 7.8 times, respectively. Furthermore, we identify contributing factors to such latency gap, including underutilized hardware instruction sets, inherent overhead in the runtime environment, resource contention within the browser, and inefficiencies in software libraries and GPU abstractions. Additionally, in-browser inference imposes significant memory demands, at times exceeding 334.6 times the size of the DL models themselves, partly attributable to suboptimal memory management. We also observe that in-browser inference leads to a significant 67.2% increase in the time it takes for GUI components to render within Web browsers, significantly affecting the overall user QoE of Web applications reliant on this technology

cs.LG↗

arXiv2023

MUSIC Algorithm for IRS-Assisted AOA Estimation

Based on the signals received across its antennas, a multi-antenna base station (BS) can apply the classic multiple signal classification (MUSIC) algorithm for estimating the angle of arrivals (AOAs) of its incident signals. This method can be leveraged to localize the users if their line-of-sight (LOS) paths to the BS are available. In this paper, we consider a more challenging AOA estimation setup in the intelligent reflecting surface (IRS) assisted integrated sensing and communication (ISAC) system, where LOS paths do not exist between the BS and the users, while the users' signals can be transmitted to the BS merely via their LOS paths to the IRS as well as the LOS path from the IRS to the BS. Specifically, we treat the IRS as the anchor and are interested in estimating the AOAs of the incident signals from the users to the IRS. Note that we have to achieve the above goal based on the signals received by the BS, because the passive IRS cannot process its received signals. However, the signals received across different antennas of the BS only contain AOA information of its incident signals via the LOS path from the IRS to the BS. To tackle this challenge arising from the spatial-domain received signals, we propose an innovative approach to create temporal-domain multi-dimension received signals for estimating the AOAs of the paths from the users to the IRS. Specifically, via a proper design of the user message pattern and the IRS reflecting pattern, we manage to show that our designed temporal-domain multi-dimension signals can be surprisingly expressed as a function of the virtual steering vectors of the IRS towards the users. This amazing result implies that the classic MUSIC algorithm can be applied to our designed temporal-domain multi-dimension signals for accurately estimating the AOAs of the signals from the users to the IRS.

eess.SP↗

arXiv2023

A Heterogeneous 6G Networked Sensing Architecture with Active and Passive Anchors

In the future 6G integrated sensing and communication (ISAC) cellular systems, networked sensing is a promising technique that can leverage the cooperation among the base stations (BSs) to perform high-resolution localization. However, a dense deployment of BSs to fully reap the networked sensing gain is not a cost-efficient solution in practice. Motivated by the advance in the intelligent reflecting surface (IRS) technology for 6G communication, this paper examines the feasibility of deploying the low-cost IRSs to enhance the anchor density for networked sensing. Specifically, we propose a novel heterogeneous networked sensing architecture, which consists of both the active anchors, i.e., the BSs, and the passive anchors, i.e., the IRSs. Under this framework, the BSs emit the orthogonal frequency division multiplexing (OFDM) communication signals in the downlink for localizing the targets based on their echoes reflected via/not via the IRSs. However, there are two challenges for using passive anchors in localization. First, it is impossible to utilize the round-trip signal between a passive IRS and a passive target for estimating their distance. Second, before localizing a target, we do not know which IRS is closest to it and serves as its anchor. In this paper, we show that the distance between a target and its associated IRS can be indirectly estimated based on the length of the BS-target-BS path and the BS-target-IRS-BS path. Moreover, we propose an efficient data association method to match each target to its associated IRS. Numerical results are given to validate the feasibility and effectiveness of our proposed heterogeneous networked sensing architecture with both active and passive anchors.

eess.SP↗

arXiv2022

Mandheling: Mixed-Precision On-Device DNN Training with DSP Offloading

This paper proposes Mandheling, the first system that enables highly resource-efficient on-device training by orchestrating the mixed-precision training with on-chip Digital Signal Processing (DSP) offloading. Mandheling fully explores the advantages of DSP in integer-based numerical calculation by four novel techniques: (1) a CPU-DSP co-scheduling scheme to mitigate the overhead from DSP-unfriendly operators; (2) a self-adaptive rescaling algorithm to reduce the overhead of dynamic rescaling in backward propagation; (3) a batch-splitting algorithm to improve the DSP cache efficiency; (4) a DSP-compute subgraph reusing mechanism to eliminate the preparation overhead on DSP. We have fully implemented Mandheling and demonstrate its effectiveness through extensive experiments. The results show that, compared to the state-of-the-art DNN engines from TFLite and MNN, Mandheling reduces the per-batch training time by 5.5$\times$ and the energy consumption by 8.9$\times$ on average. In end-to-end training tasks, Mandheling reduces up to 10.7$\times$ convergence time and 13.1$\times$ energy consumption, with only 1.9%-2.7% accuracy loss compared to the FP32 precision setting.

cs.NI↗

arXiv2022

Trilateration-Based Device-Free Sensing: Two Base Stations and One Passive IRS Are Sufficient

The classic trilateration technique can localize each target based on its distances to three anchors with known coordinates. Usually, this technique requires all the anchors and targets, e.g., the satellites and the mobile phones in Global Navigation Satellite System (GNSS), to actively transmit/receive radio signals such that the delay of the one-way radio signal propagated between each anchor and each target can be measured. Excitingly, this paper will show that the trilateration technique can be generalized to the scenario where one of the three anchors and all the targets merely reflect the radio signals passively as in radar networks, even if the propagation delay between the passive IRS and the passive targets is difficult to be measured directly, and the data association issue for multi-sensor multi-target tracking arises. Specifically, we consider device-free sensing in a cellular network consisting of two base stations (BSs), one passive intelligent reflecting surface (IRS), and multiple passive targets, to realize integrated sensing and communication (ISAC). The two BSs transmit the orthogonal frequency division multiplexing (OFDM) signals in the downlink and estimate the locations of the targets based on their reflected signals via/not via the IRS. We propose an efficient trilateration-based strategy that can first estimate the distances of each target to the two BSs and the IRS and then localize the targets. Numerical results show that the considered networked sensing architecture with heterogenous anchors can outperform its counterpart with three BSs.

eess.SP↗