arXiv subjects

Jiajie Wang

Publications and source records attributed to Jiajie Wang.

14 recordsLinked to original sources

arXiv2026

Agent Skill Evolution: How Revisions Affect Coding Agents

Agent Skills, the SKILL.md files that tell an LLM coding agent how a project works, are revised like code, yet what a revision does to the agent is unknown. From 2,608 first/last revision pairs of 3,159 Skills, we characterize how Skills evolve and how they change together with the configuration of the agent's harness. We then focus on rule changes, revisions that add or remove a rule we can check automatically, such as "run allium check". We measure their effect on 21 models in single answers and on four agents in a sandbox, and their cost on 20 of these models and the four agents. Most revisions (55%) change a rule or procedure, and commits that revise a Skill change harness files such as CLAUDE.md more often than other commits of the same size. Across 16 open-weight models, an added rule raises compliance in a single answer by +0.41 on average. Across the four agents, the rate at which the agent takes the required action rises by +0.23 on average (+0.16 to +0.36), and for the three agents that blind judges assessed, final correctness rises by +0.10 on average (+0.06 to +0.14). The gain comes mainly from rules that name a command or path the old Skill did not mention. Real tools load a Skill's body only when the agent decides it needs it. In that setting the four agents keep about half of the action gain on average (51%), and the three open models about 38%. A revision adds 18-19% input tokens to a single answer and no detectable cost to an agent episode, while loading a Skill's body raises the tokens of an episode by 50% on average.

cs.SE↗

arXiv2026

CARET: Training-Free Test-Time Scaling for Repository-Level Code Completion

Retrieval-augmented generation (RAG) dominates repository-level code completion: it retrieves cross-file context (R), then decodes one greedy completion (G). Existing work mainly focuses on retrieval and stops there. We argue both stages can be improved together, with generation in particular gaining from test-time scaling. We present CARET, a training-free method. For R, CARET routes among retrieval contexts using the agreement among its own samples, cascading to an alternative context when the samples scatter. For G, it samples candidates over a cached prompt prefix, so the long retrieved context is encoded once rather than once per sample. It then selects the final completion by reverse-context likelihood: a correct completion makes the code after the cursor more probable, so the same model grades its own candidates by reading ahead. Across CrossCodeEval, RepoEval-Line, and RepoEval-API with six code models (1.1B to 7B, four families), CARET improves exact match in all 18 combinations by 10.8 points on average over greedy decoding and 5.3 over self-consistency@10. Token-level compute stays near 1.45 times one generation (measured wall-clock 1.3-2.9 times, growing with generator size). Improving retrieval and generation together yields more accurate code than improving retrieval alone, at a budget that stays close to a single pass.

cs.SE↗

arXiv2026

ExploreAI: Agentic Exploration Knowledge Bases for Reproducible Observable-Regression Testing of Black-Box VR and 3D Applications

Black-box VR and 3D applications are difficult to regression test because observable failures depend on where a tester moves, what objects are visible, and which views are captured. Manual exploratory testing can find such failures, but its evidence is time-consuming to reproduce; systematic sweeps are reproducible, but they lack semantic guidance and spend exploration budget on low-value viewpoints. We observe that an LLM can make the high-level decisions a human tester makes during exploration: interpreting a task, choosing which objects to inspect, grouping related objects, recording what it saw, and deciding when missing evidence should trigger another attempt. Based on this observation, we present ExploreAI, an LLM-driven agentic framework that offloads repeated perception, navigation, multi-view capture execution, and logging to specialized modules while using the LLM for planning, evidence recording, capture-policy decisions, and verification decisions. ExploreAI constructs an Exploration Knowledge Base (EKB): a structured, per-object record of one exploration run. For each object the agent finds, the EKB stores the scan evidence that exposed it, the selected target, the navigation path, the multi-view capture, and the self-verification result. The EKB is a reusable testing artifact that supports reproducible observable-regression checking across versions of a VR or 3D application. Across six indoor and outdoor scenes in Unity, AI2-THOR, and BeamNG, ExploreAI constructs high-completeness EKBs under both complete and target exploration, and an LLM-module ablation shows where semantic planning, capture policy, evidence recording, and self-verification contribute. Reproduction pilots further show that EKB-guided traces help both humans and LLM-based reproducers reproduce exact object-view evidence more effectively than conditions without EKB context.

cs.SE↗

arXiv2025

KVNAND: Efficient On-Device Large Language Model Inference Using DRAM-Free In-Flash Computing

Deploying large language models (LLMs) on edge devices enables personalized agents with strong privacy and low cost. However, with tens to hundreds of billions of parameters, single-batch autoregressive inference suffers from extremely low arithmetic intensity, creating severe weight-loading and bandwidth pressures on resource-constrained platforms. Recent in-flash computing (IFC) solutions alleviate this bottleneck by co-locating weight-related linear computations in the decode phase with flash, yet still rely on DRAM for the key-value (KV) cache. As context length grows, the KV cache can exceed model weights in size, imposing prohibitive DRAM cost and capacity requirements. Attempts to offload KV cache to flash suffer from severe performance penalties. We propose KVNAND, the first DRAM-free, IFC-based architecture that stores both model weights and KV cache entirely in compute-enabled 3D NAND flash. KVNAND addresses the fundamental performance challenges of flash under intensive KV cache access by leveraging IFC for all memory-bound operations to reduce data transfer overhead, introducing head-group parallelism to boost throughput, and employing page-level KV cache mapping to align token access patterns with flash organization. In addition, we propose a design space exploration framework that evaluates discrete and compact KVNAND variants to balance weight and KV placement, automatically identifying the optimal design trade-off. These techniques mitigate latency, energy, and reliability concerns, turning flash into a practical medium for long-context KV storage. Evaluations on MHA 7B and GQA 70B LLMs show that KVNAND achieves 1.98$\times$/1.94$\times$/2.05$\times$ geomean speedup at 128/1K/10K-token contexts compared to DRAM-equipped IFC designs and addresses out-of-memory failures at 100K context length.

cs.AR↗

arXiv2025

Exploring the Role of Artificial Intelligence and Machine Learning in Process Optimization for Chemical Industry

The crucial field of Optical Chemical Structure Recognition (OCSR) aims to transform chemical structure photographs into machine-readable formats so that chemical databases may be efficiently stored and queried. Although a number of OCSR technologies have been created, little is known about how well they work in different picture deterioration scenarios. In this work, a new dataset of chemically structured images that have been systematically harmed graphically by compression, noise, distortion, and black overlays is presented. On these subsets, publicly accessible OCSR tools were thoroughly tested to determine how resilient they were to unfavorable circumstances. The outcomes show notable performance variation, underscoring each tool's advantages and disadvantages. Interestingly, MolScribe performed best under heavy compression (55.8% at 99%) and had the highest identification rate on undamaged photos (94.6%). MolVec performed exceptionally well against noise and black overlay (86.8% at 40%), although it declined under extreme distortion (<70%). With recognition rates below 30%, Decimer demonstrated strong sensitivity to noise and black overlay, but Imago had the lowest baseline accuracy (73.6%). The creative assessment of this study offers important new information about how well the OCSR tool performs when images deteriorate, as well as useful standards for tool development in the future.

cs.LG↗

arXiv2024

Vector Quantized Diffusion Model Based Speech Bandwidth Extension

Recent advancements in neural audio codec (NAC) unlock new potential in audio signal processing. Studies have increasingly explored leveraging the latent features of NAC for various speech signal processing tasks. This paper introduces the first approach to speech bandwidth extension (BWE) that utilizes the discrete features obtained from NAC. By restoring high-frequency details within highly compressed discrete tokens, this approach enhances speech intelligibility and naturalness. Based on Vector Quantized Diffusion, the proposed framework combines the strengths of advanced NAC, diffusion models, and Mamba-2 to reconstruct high-frequency speech components. Extensive experiments demonstrate that this method exhibits superior performance across both log-spectral distance and ViSQOL, significantly improving speech quality.

cs.SD↗

arXiv2024

A Two-Stage Band-Split Mamba-2 Network For Music Separation

Music source separation (MSS) aims to separate mixed music into its distinct tracks, such as vocals, bass, drums, and more. MSS is considered to be a challenging audio separation task due to the complexity of music signals. Although the RNN and Transformer architecture are not perfect, they are commonly used to model the music sequence for MSS. Recently, Mamba-2 has already demonstrated high efficiency in various sequential modeling tasks, but its superiority has not been investigated in MSS. This paper applies Mamba-2 with a two-stage strategy, which introduces residual mapping based on the mask method, effectively compensating for the details absent in the mask and further improving separation performance. Experiments confirm the superiority of bidirectional Mamba-2 and the effectiveness of the two-stage network in MSS. The source code is publicly accessible at https://github.com/baijinglin/TS-BSmamba2.

cs.SD↗

arXiv2024

Robust Predictions with Ambiguous Time Delays: A Bootstrap Strategy

In contemporary data-driven environments, the generation and processing of multivariate time series data is an omnipresent challenge, often complicated by time delays between different time series. These delays, originating from a multitude of sources like varying data transmission dynamics, sensor interferences, and environmental changes, introduce significant complexities. Traditional Time Delay Estimation methods, which typically assume a fixed constant time delay, may not fully capture these variabilities, compromising the precision of predictive models in diverse settings. To address this issue, we introduce the Time Series Model Bootstrap (TSMB), a versatile framework designed to handle potentially varying or even nondeterministic time delays in time series modeling. Contrary to traditional approaches that hinge on the assumption of a single, consistent time delay, TSMB adopts a nonparametric stance, acknowledging and incorporating time delay uncertainties. TSMB significantly bolsters the performance of models that are trained and make predictions using this framework, making it highly suitable for a wide range of dynamic and interconnected data environments.

cs.LG↗

arXiv2021

Collaborative Label Correction via Entropy Thresholding

Deep neural networks (DNNs) have the capacity to fit extremely noisy labels nonetheless they tend to learn data with clean labels first and then memorize those with noisy labels. We examine this behavior in light of the Shannon entropy of the predictions and demonstrate the low entropy predictions determined by a given threshold are much more reliable as the supervision than the original noisy labels. It also shows the advantage in maintaining more training samples than previous methods. Then, we power this entropy criterion with the Collaborative Label Correction (CLC) framework to further avoid undesired local minimums of the single network. A range of experiments have been conducted on multiple benchmarks with both synthetic and real-world settings. Extensive results indicate that our CLC outperforms several state-of-the-art methods.

cs.LG↗

arXiv2019

Minimal Root's embeddings for general starting and target distributions

Most results regarding Skorokhod embedding problems (SEP) so far rely on the assumption that the corresponding stopped process is uniformly integrable, which is equivalent to the convex ordering condition $\mathrm{U}^μ\leq\mathrm{U}^ν$ when the underlying process is a local martingale. In this paper, we study the existence, construction of Root's solutions to SEP, in the absence of this convex ordering condition. We replace the uniform integrability condition by the minimality condition (Monroe,1972), as the criterion of "good" solutions. A sufficient and necessary condition (in terms of local time) for minimality is given. We also discuss the optimality of such minimal solutions. These results extend the generality of the results given by Cox and Wang [2013] and Gassiat et al. [2015]. At last, we extend all the results mentioned above to multi-marginal embedding problems based on the work of Cox et al. [2018].

math.PR↗

arXiv2018

Collaborative Learning for Weakly Supervised Object Detection

Weakly supervised object detection has recently received much attention, since it only requires image-level labels instead of the bounding-box labels consumed in strongly supervised learning. Nevertheless, the save in labeling expense is usually at the cost of model accuracy. In this paper, we propose a simple but effective weakly supervised collaborative learning framework to resolve this problem, which trains a weakly supervised learner and a strongly supervised learner jointly by enforcing partial feature sharing and prediction consistency. For object detection, taking WSDDN-like architecture as weakly supervised detector sub-network and Faster-RCNN-like architecture as strongly supervised detector sub-network, we propose an end-to-end Weakly Supervised Collaborative Detection Network. As there is no strong supervision available to train the Faster-RCNN-like sub-network, a new prediction consistency loss is defined to enforce consistency of predictions between the two sub-networks as well as within the Faster-RCNN-like sub-networks. At the same time, the two detectors are designed to partially share features to further guarantee the model consistency at perceptual level. Extensive experiments on PASCAL VOC 2007 and 2012 data sets have demonstrated the effectiveness of the proposed framework.

cs.CV↗

arXiv2017

Deep Learning from Noisy Image Labels with Quality Embedding

There is an emerging trend to leverage noisy image datasets in many visual recognition tasks. However, the label noise among the datasets severely degenerates the \mbox{performance of deep} learning approaches. Recently, one mainstream is to introduce the latent label to handle label noise, which has shown promising improvement in the network designs. Nevertheless, the mismatch between latent labels and noisy labels still affects the predictions in such methods. To address this issue, we propose a quality embedding model, which explicitly introduces a quality variable to represent the trustworthiness of noisy labels. Our key idea is to identify the mismatch between the latent and noisy labels by embedding the quality variables into different subspaces, which effectively minimizes the noise effect. At the same time, the high-quality labels is still able to be applied for training. To instantiate the model, we further propose a Contrastive-Additive Noise network (CAN), which consists of two important layers: (1) the contrastive layer estimates the quality variable in the embedding space to reduce noise effect; and (2) the additive layer aggregates the prior predictions and noisy labels as the posterior to train the classifier. Moreover, to tackle the optimization difficulty, we deduce an SGD algorithm with the reparameterization tricks, which makes our method scalable to big data. We conduct the experimental evaluation of the proposed method over a range of noisy image datasets. Comprehensive results have demonstrated CAN outperforms the state-of-the-art deep learning approaches.

cs.CV↗

arXiv2013

Optimal robust bounds for variance options

Robust, or model-independent properties of the variance swap are well-known, and date back to Dupire and Neuberger, who showed that, given the price of co-terminal call options, the price of a variance swap was exactly specified under the assumption that the price process is continuous. In Cox and Wang we showed that a lower bound on the price of a variance call could be established using a solution to the Skorokhod embedding problem due to Root. In this paper, we provide a construction, and a proof of optimality of the upper bound, using results of Rost and Chacon, and show how this proof can be used to determine a super-hedging strategy which is model-independent. In addition, we outline how the hedging strategy may be computed numerically. Using these methods, we also show that the Heston-Nandi model is 'asymptotically extreme' in the sense that, for large maturities, the Heston-Nandi model gives prices for variance call options which are approximately the lowest values consistent with the same call price data.

q-fin.PR↗

arXiv2013

Root's barrier: Construction, optimality and applications to variance options

Recent work of Dupire and Carr and Lee has highlighted the importance of understanding the Skorokhod embedding originally proposed by Root for the model-independent hedging of variance options. Root's work shows that there exists a barrier from which one may define a stopping time which solves the Skorokhod embedding problem. This construction has the remarkable property, proved by Rost, that it minimizes the variance of the stopping time among all solutions. In this work, we prove a characterization of Root's barrier in terms of the solution to a variational inequality, and we give an alternative proof of the optimality property which has an important consequence for the construction of subhedging strategies in the financial context.

q-fin.PR↗