arXiv subjects

Yuewei Zhang

Publications and source records attributed to Yuewei Zhang.

At least 19 recordsLinked to original sources

arXiv2026

TruthInsightBench: An Evidence-Grounded Benchmark for Automated Evaluation of Open-Ended Scientific Discovery Agents

Autonomous coding agents are increasingly proposed as AI-scientist systems that conduct analyses and write research reports, but executing a prescribed analysis is not the same as making a discovery. Existing benchmarks are configured for reproduction: tasks, data, and rubrics are built around a hidden target study, and recovery of its result is rewarded. We present TruthInsightBench, a benchmark configured for discovery. Its 40 blind tasks, drawn from 40 peer-reviewed studies across 10 scientific domains, expose only a neutral scientific objective and frozen data; source conclusions, expected values, and analysis paths are withheld, leaving the agent to determine what claim the data support. A fixed LLM-based judge scores the evidentiary maturity of an agent's own claims along six dimensions, operationalized as 29 artifact-grounded items, with automated, deterministic aggregation and no per-instance human grading, so evaluation can be repeated automatically as agents evolve. On one frozen base model, four coding agents form a narrow plateau (58.4-60.3 of 100) with no statistically reliable pairwise separation: they execute and document analyses competently, with comparatively strong evidence auditability and novelty, but largely lack the discriminating acts that establish a trustworthy claim (controls, robustness, falsifiability, and cross-dataset generalization). The bottleneck is scientific judgment rather than coding, and genuine discovery remains out of reach. TruthInsightBench makes this gap a measurable target; data and scoring code are at https://github.com/TruthInsight-stack/TruthInsightBench.

cs.AI↗

arXiv2026

Knowing When Not to Reuse: Conditional Experience Transfer in Autonomous LLM Post-Training

Large language models offer broad capabilities, but adapting them to evolving domains, tools, and requirements often entails repeated post-training. Autonomous systems automate parts of this process by proposing updates, training candidates, and using evaluation feedback to select subsequent proposals. As evidence accumulates, a central problem emerges: which past update evidence remains actionable after subsequent training has changed the parent model? An update's effect depends on its parent, data, and training stage. Treating past success as context-free permission can waste compute. If the resulting child is promoted, it can also degrade the subsequent training trajectory. We formulate this problem as conditional experience transfer and introduce Boundary-Calibrated Intervention Transfer (BCIT), a method that authorizes experience reuse before weight-changing training. BCIT binds an observed effect to its source context, checks applicability conditions, vetoes candidates with named hard conflicts, and obtains current-state evidence through a bounded training trial when needed. Fully trained candidates still face a shared adoption rule, and only observed events extend memory. On one 4B model adapted across finance reasoning, text-to-SQL, and function calling, candidate updates exhibit heterogeneous target and retention effects across the evaluated contexts. Under matched candidates, evidence, and compute, BCIT authorizes fewer harmful updates and attains higher equal-budget final-model quality than the evaluated alternatives. These results support treating experience authorization as a distinct problem in autonomous post-training.

cs.AI↗

arXiv2026

Which Model Is Actually Serving You? IRIS: Budgeted Black-Box Auditing of Model Substitution and Routing Dilution in LLM Gateways

Commercial LLM gateways mediate access to hosted models, but the served backend may not match the advertised one: it may substitute a cheaper model on every request or route only a fraction $\epsilon$ of requests to it. Prior black-box auditors often need a privileged signal (log-probabilities, token ranks, or reference samples) or a target-specific probe, fix the query budget in advance, and return a yes/no verdict. We present $\mathrm{IRIS}$, an audit that needs only the returned text: it asks endpoints to generate random numbers or strings, fingerprints the backend, and is the first to combine, in one text-only audit, detection of whole-stream substitution and fractional dilution, attribution of the served backend, routing-fraction ($\epsilon$) estimation, and a query budget it sizes itself. A cheap pilot fits the exponential query-error decay and freezes that budget before any suspect query is issued. On an intra-family Qwen3 ladder $\mathrm{IRIS}$ verifies the backend at $0.99$ AUROC and sharpens attribution as queries accumulate; across a commercial OpenRouter library it catches $\epsilon{=}0.3$ dilution on margin-qualified pairs at $0.85$ mean power ($0.017$ false-positive rate) and recovers $\epsilon$ to within $0.04$ for enrolled diluents; and a live cross-provider audit flags $14$ of $15$ same-model provider pairs by genuine quantization and kernel deviations, corroborated on third-party MET traces. Against comparable black-box auditors, $\mathrm{IRIS}$ matches or beats detection on shared tasks, and adaptive allocation lifts the matched-budget target-hit rate from $73$% to $87$%. Further experiments cover adversarial gateways, knob identifiability, unseen diluents, and false-positive control.

cs.CR↗

arXiv2026

DVAO: Dynamic Variance-adaptive Advantage Optimization for Multi-reward Reinforcement Learning

Reinforcement Learning has become a standard paradigm for aligning Large Language Models with human intent and task requirements. While Group Relative Policy Optimization offers an efficient, value-model-free alternative to Proximal Policy Optimization, adapting it to real-world multi-reward settings remains challenging. Standard scalarization practices, such as Reward Combination and Advantage Combination, suffer from significant drawbacks: Reward Combination frequently generates advantages with excessively large squared magnitudes that lead to training instability, while Advantage Combination relies on static hyperparameters and ignores cross-objective correlations. To address these limitations, we propose Dynamic Variance-adaptive Advantage Optimization (DVAO), which dynamically adjusts combination weights based on the empirical reward variance of each objective within a rollout group, effectively up-weighting objectives with a stronger learning signal while suppressing noisy ones. We mathematically prove that DVAO maintains bounded advantage magnitudes for stable training and introduces a self-adaptive cross-objective regularization mechanism. Extensive experiments on mathematical reasoning and tool-use benchmarks using Qwen3 and Qwen2.5 models demonstrate that DVAO significantly outperforms baseline methods, achieving a superior multi-objective Pareto frontier and robust training stability.

cs.CL↗

arXiv2026

Beyond Stochastic Exploration: What Makes Training Data Valuable for Agentic Search

Reinforcement learning (RL) has become an effective approach for advancing the reasoning capabilities of large language models (LLMs) through the strategic integration of external search engines. However, current RL-based search agents often rely on a process of stochastic exploration guided by carefully crafted outcome rewards, leading to inefficient reasoning trajectories and unstable training. To address these issues, we propose a novel framework, Hierarchical Experience (HiExp), to enhance the performance and training stability of search agents. Specifically, we extract empirical knowledge through contrastive analysis and a multi-level clustering mechanism, transforming raw reasoning trajectories into hierarchical experience knowledge. By leveraging experience-aligned training, we effectively regularize stochastic exploration, evolving it into a strategic and experience-driven search process. Extensive evaluations on multiple complex agentic search and mathematical reasoning benchmarks demonstrate that our approach not only achieves substantial performance gains but also exhibits strong cross-task and cross-algorithm generalization.

cs.AI↗

arXiv2026

Evo-Retriever: LLM-Guided Curriculum Evolution with Viewpoint-Pathway Collaboration for Multimodal Document Retrieval

Visual-language models (VLMs) excel at data mappings, but real-world document heterogeneity and unstructuredness disrupt the consistency of cross-modal embeddings. Recent late-interaction methods enhance image-text alignment through multi-vector representations, yet traditional training with limited samples and static strategies cannot adapt to the model's dynamic evolution, causing cross-modal retrieval confusion. To overcome this, we introduce Evo-Retriever, a retrieval framework featuring an LLM-guided curriculum evolution built upon a novel Viewpoint-Pathway collaboration. First, we employ multi-view image alignment to enhance fine-grained matching via multi-scale and multi-directional perspectives. Then, a bidirectional contrastive learning strategy generates "hard queries" and establishes complementary learning paths for visual and textual disambiguation to rebalance supervision. Finally, the model-state summary from the above collaboration is fed into an LLM meta-controller, which adaptively adjusts the training curriculum using expert knowledge to promote the model's evolution. On ViDoRe V2 and MMEB (VisDoc), Evo-Retriever achieves state-of-the-art performance, with nDCG@5 scores of 65.2% and 77.1%.

cs.CV↗

arXiv2026

FAQ: Mitigating Quantization Error via Regenerating Calibration Data with Family-Aware Quantization

Although post-training quantization (PTQ) provides an efficient numerical compression scheme for deploying large language models (LLMs) on resource-constrained devices, the representativeness and universality of calibration data remain a core bottleneck in determining the accuracy of quantization parameters. Traditional PTQ methods typically rely on limited samples, making it difficult to capture the activation distribution during the inference phase, leading to biases in quantization parameters. To address this, we propose \textbf{FAQ} (Family-Aware Quantization), a calibration data regeneration framework that leverages prior knowledge from LLMs of the same family to generate high-fidelity calibration samples. Specifically, FAQ first inputs the original calibration samples into a larger LLM from the same family as the target model, regenerating a series of high-fidelity calibration data using a highly consistent knowledge system. Subsequently, this data, carrying Chain-of-Thought reasoning and conforming to the expected activation distribution, undergoes group competition under expert guidance to select the best samples, which are then re-normalized to enhance the effectiveness of standard PTQ. Experiments on multiple model series, including Qwen3-8B, show that FAQ reduces accuracy loss by up to 28.5\% compared to the baseline with original calibration data, demonstrating its powerful potential and contribution.

cs.LG↗

arXiv2025

VCRL: Variance-based Curriculum Reinforcement Learning for Large Language Models

Policy-based reinforcement learning currently plays an important role in improving LLMs on mathematical reasoning tasks. However, existing rollout-based reinforcement learning methods (GRPO, DAPO, GSPO, etc.) fail to explicitly consider LLMs' learning ability for samples of different difficulty levels, which is contrary to the human cognitive process of mathematical reasoning tasks from easy to difficult. Intuitively, we find that the variance of the rollout group's reward in RLVR partly reflects the difficulty of the current sample for LLMs. Samples that are too easy or too difficult have a lower variance, while samples with moderate difficulty have a higher variance. Based on this, we propose VCRL, a curriculum reinforcement learning framework that dynamically controls the difficulty of training samples based on the variance of group rewards. Experiments on five mathematical benchmarks and two models reveal the advantages of VCRL over the current LLM RL baselines.

cs.LG↗

arXiv2025

PVPO: Pre-Estimated Value-Based Policy Optimization for Agentic Reasoning

Critic-free reinforcement learning methods, particularly group policies, have attracted considerable attention for their efficiency in complex tasks. However, these methods rely heavily on multiple sampling and comparisons within the policy to estimate advantage, which may cause the policy to fall into local optimum and increase computational cost. To address these issues, we propose PVPO, an efficient reinforcement learning method enhanced by an advantage reference anchor and data pre-sampling. Specifically, we use the reference model to rollout in advance and employ the calculated reward score as a reference anchor. Our approach effectively corrects the cumulative bias introduced by intra-group comparisons and significantly reduces reliance on the number of rollouts during training. Meanwhile, the reference model can assess sample difficulty during data pre-sampling, enabling effective selection of high-gain data to improve training efficiency. Moreover, PVPO is orthogonal to other advanced critic-free RL algorithms, making it compatible with and complementary to these methods. Experiments conducted on nine datasets across two domains demonstrate that PVPO achieves State-Of-The-Art (SOTA) performance. Our approach not only demonstrates robust generalization across multiple tasks, but also exhibits scalable performance across models of varying scales.

cs.LG↗

arXiv2025

DynaSearcher: Dynamic Knowledge Graph Augmented Search Agent via Multi-Reward Reinforcement Learning

Multi-step agentic retrieval systems based on large language models (LLMs) have demonstrated remarkable performance in complex information search tasks. However, these systems still face significant challenges in practical applications, particularly in generating factually inconsistent intermediate queries and inefficient search trajectories, which can lead to reasoning deviations or redundant computations. To address these issues, we propose DynaSearcher, an innovative search agent enhanced by dynamic knowledge graphs and multi-reward reinforcement learning (RL). Specifically, our system leverages knowledge graphs as external structured knowledge to guide the search process by explicitly modeling entity relationships, thereby ensuring factual consistency in intermediate queries and mitigating biases from irrelevant information. Furthermore, we employ a multi-reward RL framework for fine-grained control over training objectives such as retrieval accuracy, efficiency, and response quality. This framework promotes the generation of high-quality intermediate queries and comprehensive final answers, while discouraging unnecessary exploration and minimizing information omissions or redundancy. Experimental results demonstrate that our approach achieves state-of-the-art answer accuracy on six multi-hop question answering datasets, matching frontier LLMs while using only small-scale models and limited computational resources. Furthermore, our approach demonstrates strong generalization and robustness across diverse retrieval environments and larger-scale models, highlighting its broad applicability.

cs.LG↗

arXiv2025

FlowKV: A Disaggregated Inference Framework with Low-Latency KV Cache Transfer and Load-Aware Scheduling

Disaggregated inference has become an essential framework that separates the prefill (P) and decode (D) stages in large language model inference to improve throughput. However, the KV cache transfer faces significant delays between prefill and decode nodes. The block-wise calling method and discontinuous KV cache memory allocation increase the number of calls to the transmission kernel. Additionally, existing frameworks often fix the roles of P and D nodes, leading to computational imbalances. In this paper, we propose FlowKV, a novel disaggregated inference framework, which reduces the average transmission latency of KV cache by 96%, from 0.944s to 0.053s, almost eliminating the transfer time relative to the total request latency by optimizing the KV cache transfer. FlowKV introduces the Load-Aware Scheduler for balanced request scheduling and flexible PD node allocation. This design maximizes hardware resource utilization, achieving peak system throughput across various scenarios, including normal, computational imbalance, and extreme overload conditions. Experimental results demonstrate that FlowKV significantly accelerates inference by 15.2%-48.9% on LongBench dataset compared to the baseline and supports applications with heterogeneous GPUs.

cs.DC↗

arXiv2025

RASD: Retrieval-Augmented Speculative Decoding

Speculative decoding accelerates inference in large language models (LLMs) by generating draft tokens for target model verification. Current approaches for obtaining draft tokens rely on lightweight draft models or additional model structures to generate draft tokens and retrieve context from databases. Due to the draft model's small size and limited training data, model-based speculative decoding frequently becomes less effective in out-of-domain scenarios. Additionally, the time cost of the drafting phase results in a low upper limit on acceptance length during the verification step, limiting overall efficiency. This paper proposes RASD (Retrieval-Augmented Speculative Decoding), which adopts retrieval methods to enhance model-based speculative decoding. We introduce tree pruning and tree fusion to achieve this. Specifically, we develop a pruning method based on the draft model's probability distribution to construct the optimal retrieval tree. Second, we employ the longest prefix matching algorithm to merge the tree generated by the draft model with the retrieval tree, resulting in a unified tree for verification. Experimental results demonstrate that RASD achieves state-of-the-art inference acceleration across tasks such as DocQA, Summary, Code, and In-Domain QA. Moreover, RASD exhibits strong scalability, seamlessly integrating with various speculative decoding approaches, including both generation-based and retrieval-based methods.

cs.CL↗

arXiv2025

Speech Enhancement with Overlapped-Frame Information Fusion and Causal Self-Attention

For time-frequency (TF) domain speech enhancement (SE) methods, the overlap-and-add operation in the inverse TF transformation inevitably leads to an algorithmic delay equal to the window size. However, typical causal SE systems fail to utilize the future speech information within this inherent delay, thereby limiting SE performance. In this paper, we propose an overlapped-frame information fusion scheme. At each frame index, we construct several pseudo overlapped-frames, fuse them with the original speech frame, and then send the fused results to the SE model. Additionally, we introduce a causal time-frequency-channel attention (TFCA) block to boost the representation capability of the neural network. This block parallelly processes the intermediate feature maps through self-attention-based operations in the time, frequency, and channel dimensions. Experiments demonstrate the superiority of these improvements, and the proposed SE system outperforms the current advanced methods.

eess.AS↗

arXiv2025

AirRAG: Autonomous Strategic Planning and Reasoning Steer Retrieval Augmented Generation

Leveraging the autonomous decision-making capabilities of large language models (LLMs) has demonstrated superior performance in reasoning tasks. However, despite the success of iterative or agentic retrieval-augmented generation (RAG) techniques, these methods are often constrained to a single solution space when confronted with complex problems. In this paper, we propose a novel thinking pattern in RAG that integrates autonomous strategic planning with efficient reasoning actions, significantly activating intrinsic reasoning capabilities and expanding the solution space of specific tasks via Monte Carlo Tree Search (MCTS), which we refer to as AirRAG. Specifically, our approach designs five fundamental reasoning actions, which are expanded to a broad tree-based reasoning space using MCTS. The approach also incorporates self-consistency verification to explore potential reasoning paths and inference scaling law. Additionally, computationally optimal strategies are employed to allocate more inference resources to key actions, thereby enhancing overall performance. Experimental results demonstrate the effectiveness of AirRAG, showing significant performance gains on complex question-answering datasets. Furthermore, AirRAG is flexible and lightweight, making it easy to integrate with other advanced technologies and models.

cs.AI↗

arXiv2024

Mixture-of-LoRAs: An Efficient Multitask Tuning for Large Language Models

Instruction Tuning has the potential to stimulate or enhance specific capabilities of large language models (LLMs). However, achieving the right balance of data is crucial to prevent catastrophic forgetting and interference between tasks. To address these limitations and enhance training flexibility, we propose the Mixture-of-LoRAs (MoA) architecture which is a novel and parameter-efficient tuning method designed for multi-task learning with LLMs. In this paper, we start by individually training multiple domain-specific LoRA modules using corresponding supervised corpus data. These LoRA modules can be aligned with the expert design principles observed in Mixture-of-Experts (MoE). Subsequently, we combine the multiple LoRAs using an explicit routing strategy and introduce domain labels to facilitate multi-task learning, which help prevent interference between tasks and ultimately enhances the performance of each individual task. Furthermore, each LoRA model can be iteratively adapted to a new domain, allowing for quick domain-specific adaptation. Experiments on diverse tasks demonstrate superior and robust performance, which can further promote the wide application of domain-specific LLMs.

cs.CL↗

arXiv2024

A Two-Stage Framework in Cross-Spectrum Domain for Real-Time Speech Enhancement

Two-stage pipeline is popular in speech enhancement tasks due to its superiority over traditional single-stage methods. The current two-stage approaches usually enhance the magnitude spectrum in the first stage, and further modify the complex spectrum to suppress the residual noise and recover the speech phase in the second stage. The above whole process is performed in the short-time Fourier transform (STFT) spectrum domain. In this paper, we re-implement the above second sub-process in the short-time discrete cosine transform (STDCT) spectrum domain. The reason is that we have found STDCT performs greater noise suppression capability than STFT. Additionally, the implicit phase of STDCT ensures simpler and more efficient phase recovery, which is challenging and computationally expensive in the STFT-based methods. Therefore, we propose a novel two-stage framework called the STFT-STDCT spectrum fusion network (FDFNet) for speech enhancement in cross-spectrum domain. Experimental results demonstrate that the proposed FDFNet outperforms the previous two-stage methods and also exhibits superior performance compared to other advanced systems.

eess.AS↗

arXiv2023

VSANet: Real-time Speech Enhancement Based on Voice Activity Detection and Causal Spatial Attention

The deep learning-based speech enhancement (SE) methods always take the clean speech's waveform or time-frequency spectrum feature as the learning target, and train the deep neural network (DNN) by reducing the error loss between the DNN's output and the target. This is a conventional single-task learning paradigm, which has been proven to be effective, but we find that the multi-task learning framework can improve SE performance. Specifically, we design a framework containing a SE module and a voice activity detection (VAD) module, both of which share the same encoder, and the whole network is optimized by the weighted loss of the two modules. Moreover, we design a causal spatial attention (CSA) block to promote the representation capability of DNN. Combining the VAD aided multi-task learning framework and CSA block, our SE network is named VSANet. The experimental results prove the benefits of multi-task learning and the CSA block, which give VSANet an excellent SE performance.

eess.AS↗

arXiv2023

Magnitude-and-phase-aware Speech Enhancement with Parallel Sequence Modeling

In speech enhancement (SE), phase estimation is important for perceptual quality, so many methods take clean speech's complex short-time Fourier transform (STFT) spectrum or the complex ideal ratio mask (cIRM) as the learning target. To predict these complex targets, the common solution is to design a complex neural network, or use a real network to separately predict the real and imaginary parts of the target. But in this paper, we propose to use a real network to estimate the magnitude mask and normalized cIRM, which not only avoids the significant increase of the model complexity caused by complex networks, but also shows better performance than previous phase estimation methods. Meanwhile, we devise a parallel sequence modeling (PSM) block to improve the RNN block in the convolutional recurrent network (CRN)-based SE model. We name our method as magnitude-and-phase-aware and PSM-based CRN (MPCRN). The experimental results illustrate that our MPCRN has superior SE performance.

eess.AS↗