Searcharxiv⌕ Search

arXiv subjects

Zixuan Zhu

Publications and source records attributed to Zixuan Zhu.

18 recordsLinked to original sources
arXiv2026

Harbor Adapters and Harbor-Index: Infrastructure and a Curated Meta-Dataset for Large-Scale Agentic Evaluation

Evaluating agents on the growing number of agentic benchmarks is challenging because they often require complex environments and agent integrations. We introduce Harbor Adapters, a unified evaluation infrastructure for agentic benchmarks. Our work makes three contributions. First, we develop benchmark adapters that port more than 80 benchmarks to evaluate arbitrary agents, and validate them through rigorous code review and parity experiments. Second, we conduct a large-scale evaluation of 8 models spanning capability tiers across 54 benchmarks; every model is run with Terminus-2 and with one of 3 native harnesses. This enables a broader analysis of agent capabilities and failure modes than was previously possible. Third, we introduce Harbor-Index, a curated set of 82 difficult, diverse, and high-quality tasks spanning 29 benchmarks, refined from the adapted suite through difficulty filtering, AI and human audit, and an audit-and-fix loop. Harbor-Index preserves the challenge and breadth of large-scale agentic evaluations while being affordable to run; no evaluated model-harness configuration exceeds 30% pass rate, and the strongest (GPT-5.5 with Codex) reaches 28.0%. We release the adapters, evaluation results, in-depth analysis, and Harbor-Index as open-source artifacts to support more reliable and comprehensive evaluation of language-model agents.

cs.AI↗
arXiv2026

Lapis: Laplacian Spiking Attention via First-Spike Timing and Membrane Leakage

Self-attention has become central to spiking vision transformers, yet its query-key scoring is still largely inherited from dense networks. Existing spiking variants either simplify dot product scoring or replace it with discrete operators, but spike timing, the native variable of a spiking network, does not directly define how tokens are related. We propose Lapis, a spiking attention mechanism that scores each token pair by the L1 distance between its query and key first-spike latency vectors under time-to-first-spike coding, and maps this distance to an affinity through a Laplacian kernel. The kernel's exponential decay matches the impulse response of a leaky integrate-and-fire membrane, so the accumulated latency difference determines the decay of a membrane trace, while row normalization reduces to a bit shift under power-of-two rounding. Scoring therefore needs only subtraction, absolute value, and accumulation, and removes all multiplication between query and key channels. Under a matched backbone and training schedule, Lapis reaches 96.56% top-1 accuracy on CIFAR-10, within 0.53 points of dot-product scoring. On ImageNet-1K, it reduces the estimated arithmetic energy of the attention path by 14.5x relative to dense dot-product attention. The deployed 6-bit model attains 83.25% top-1 accuracy at an estimated arithmetic energy of 3.28mJ per image.

cs.NE↗
arXiv2026

Mitigating Backdoors via Decoy Shortcuts and Knowledge Decoupling

Backdoor attacks pose a serious threat to deep neural networks, especially when training relies on third-party data, allowing adversaries to inject malicious behaviors through data poisoning. In this work, we reveal that backdoor behaviors tend to be absorbed by a simpler parallel branch when jointly trained with the main network. Motivated by this insight, we propose Trapping and Removing (TR), a simple yet effective training-time defense that introduces a lightweight shortcut branch as a "honeypot" to trap backdoor knowledge. After training, backdoors can be removed by discarding the shortcut, without requiring any additional data. To further enhance backdoor isolation while maintaining benign performance, we design a knowledge decoupling strategy with entropy-based weight assignment, encouraging poisoned samples to flow through the honeypot while guiding the main network to focus on benign learning. In addition, we introduce an automatic shortcut generation strategy to improve generalization across model architectures. Extensive experiments on four benchmark datasets and five model architectures demonstrate that our approach effectively mitigates a wide range of backdoor attacks while preserving performance on benign data. Code: https://github.com/Zixuan-Zhu/TR}{github.com/Zixuan-Zhu/TR.

cs.LG↗
arXiv2026

Matterhorn: Masked Time-to-First-Spike Encoding by Reassigning the Silent State for Sparse and Energy-Efficient Spiking Transformers

Spiking neural networks (SNNs) promise energy-efficient inference for large language models (LLMs), yet most reported savings rely on compute-operation counts that overlook data movement. Energy characterization of representative spiking transformers on a commercial 22-nm process shows that accumulation contributes less than 3% of total energy, while spike-triggered inter-core transfers and weight reads dominate the cost. This makes time-to-first-spike (TTFS) encoding a natural choice, as it limits each neuron to at most one spike. However, standard TTFS maps the silent state, an all-zero spike train that transmits no events, to the rarely occurring smallest value, while the most common activations still spike. This raises a simple question: why reserve the only cost-free codeword for a rare value? This choice inverts a basic principle of energy-aware coding, under which the zero-event codeword should represent the most common value, rather than a rare extreme. Thus, we introduce masked time-to-first-spike encoding (M-TTFS), which uses a temporal mask to reassign the silent state to the most common activation value, and a dead-zone extension that trades a controlled amount of information for greater sparsity. Built on M-TTFS with dead-zone radius $k{=}1$, our spiking transformer Matterhorn reaches an overall spike rate of 1.64% on GLUE at an average score of 84.64, exceeding the best prior spiking transformer by 1.42 percentage points while consuming 67% less energy, with consistent gains on spiking LLaMA models from 7B to 70B parameters. Together, these results show that under hardware-faithful accounting, the energy advantage of SNNs is not a given: it is earned by encodings that align spikes with the data distribution.

cs.LG↗
arXiv2026

Unified Context Evolution for LLM Agents

LLM-based agents can solve multi-step interactive tasks by combining reasoning with environment feedback, yet each episode starts from the same fixed context and any useful strategy discovered along the way is lost once the task ends. Existing approaches either limit learning to the current task or pool all experience into a single untyped store, without distinguishing knowledge types, tracking quality through use, or balancing what the library still lacks. We introduce Unified Context Evolution (UCE), a gradient-free framework that externalizes agent experience into an evolving library of typed Evolvable Context Units (ECUs). UCE decomposes experience into four complementary types (Memory, Strategy, Workflow, and Skill), each generated from trajectories under type-specific conditions, retrieved at decision time, scored through repeated usage outcomes, and pruned when no longer valuable. A scheduling module allocates each cycle's generation budget toward the types where the library is weakest. Across two interactive benchmarks, UCE raises ALFWorld success from 75.4% to 96.3% and WebShop task score from 45.1% to 61.3%, and the accumulated library transfers to alternative actor backbones without retraining.

cs.CL↗
arXiv2026

Rank and Independence of Imaginaries in Proper Pairs of ACF

Let $T_P$ be the theory of beautiful pairs of algebraically closed fields of fixed characteristic. It is known that for real tuples in models of $T_P$, SU-rank coincides with Morley rank and can be computed effectively. Building on Pillay's geometric description (2007) of imaginaries in $T_P$, we define an additive rank on imaginaries of $T_P$, called the geometric rank. It takes values in $ω*\mathbb N + \mathbb Z$ and coincides with SU-rank on real tuples. It refines SU-rank and characterizes forking in $T_P^{\mathrm{eq}}$, from which we derive an explicit criterion for determining forking independence.

math.LO↗
arXiv2026

When Simulation Lies: A Sim-to-Real Benchmark and Domain-Randomized RL Recipe for Tool-Use Agents

Tool-use language agents are evaluated on benchmarks that assume clean inputs, unambiguous tool registries, and reliable APIs. Real deployments violate all these assumptions: user typos propagate into hallucinated tool names, a misconfigured request timeout can stall an agent indefinitely, and duplicate tool names across servers can freeze an SDK. We study these failures as a sim-to-real gap in the tool-use partially observable Markov decision process (POMDP), where deployment noise enters through the observation, action space, reward-relevant metadata, or transition dynamics. We introduce RobustBench-TC, a benchmark with 22 perturbation types organized by these four POMDP components, each grounded in a verified GitHub issue or documented tool-calling failure. Across 21 models from 1.5B to 32B parameters (including the closed-source o4-mini), the robustness profile is sharply uneven: observation perturbations reduce accuracy by less than 5%, while reward-relevant and transition perturbations reduce accuracy by roughly 40% and 30%, respectively; scale alone does not close these gaps. We then propose ToolRL-DR, a domain-randomization reinforcement learning (RL) recipe that trains a tool-use agent on perturbation-augmented trajectories spanning the three statically encodable POMDP components. On a 3B backbone, ToolRL-DR-Full retains roughly three-quarters of clean accuracy and reaches an aggregate perturbed accuracy comparable to open-source 14B function-calling baselines while substantially narrowing the gap to o4-mini. It closes approximately 27% of the Transition gap despite never seeing transition perturbations in training, suggesting that RL on adversarial static tool-use inputs induces a more persistent retry policy that transfers to unseen runtime failures. The dataset, code and benchmark leaderboard are publicly available.

cs.AI↗
arXiv2026

Unveiling the Resilience of LLM-Enhanced Search Engines against Black-Hat SEO Manipulation

The emergence of Large Language Model-enhanced Search Engines (LLMSEs) has revolutionized information retrieval by integrating web-scale search capabilities with AI-powered summarization. While these systems demonstrate improved efficiency over traditional search engines, their security implications against well-established black-hat Search Engine Optimization (SEO) attacks remain unexplored. In this paper, we present the first systematic study of SEO attacks targeting LLMSEs. Specifically, we examine ten representative LLMSE products (e.g., ChatGPT, Gemini) and construct SEO-Bench, a benchmark comprising 1,000 real-world black-hat SEO websites, to evaluate both open- and closed-source LLMSEs. Our measurements show that LLMSEs mitigate over 99.78% of traditional SEO attacks, with the phase of retrieval serving as the primary filter, intercepting the vast majority of malicious queries. We further propose and evaluate seven LLMSEO attack strategies, demonstrating that off-the-shelf LLMSEs are vulnerable to LLMSEO attacks, i.e., rewritten-query stuffing and segmented texts double the manipulation rate compared to the baseline. This work offers the first in-depth security analysis of the LLMSE ecosystem, providing practical insights for building more resilient AI-driven search systems. We have responsibly reported the identified issues to major vendors.

cs.CR↗
arXiv2025

Optimal 4-Approximation for the Correlated Pandora's Problem

The Correlated Pandora's Problem posed by Chawla et al. (2020) generalizes the classical Pandora's Problem by allowing the numbers inside the Pandora's boxes to be correlated. It also generalizes the Min Sum Set Cover problem, and is related to the Uniform Decision Tree problem. This paper gives an optimal 4-approximation for the Correlated Pandora's Problem, matching the lower bound of 4 from Min Sum Set Cover.

cs.DS↗
arXiv2025

Advanced LPeg techniques: A dual case study approach

This paper presents advanced optimization techniques for Lua Parsing Expression Grammars (LPeg) through two complementary case studies: a high-performance JSON parser and a sophisticated Glob-to-LPeg pattern converter. We demonstrate how strategic grammar construction can dramatically improve parsing performance without modifying the underlying LPeg library. For the JSON parser, we implement substitution capture and table construction optimization to reduce memory allocation overhead and improve object processing. For the Glob converter, we introduce segment-boundary separation, implement Cox's flattened search strategy, and develop optimized braced condition handling to prevent exponential backtracking. Comprehensive benchmarks demonstrate that our JSON parser achieves processing speeds up to 125 MB/s on complex documents, consistently outperforming dkjson and showing competitive results against rxi_json across most test cases. Our Glob-to-LPeg converter exhibits 14-92% better performance than Bun.Glob and runs 3-14 times faster than Minimatch across diverse pattern matching scenarios. This research provides practical optimization techniques for LPeg-based parsers, contributing valuable strategies to the text processing ecosystem.

cs.PL↗
arXiv2025

CuriousLLM: Elevating Multi-Document Question Answering with LLM-Enhanced Knowledge Graph Reasoning

Large Language Models (LLMs) have achieved significant success in open-domain question answering. However, they continue to face challenges such as hallucinations and knowledge cutoffs. These issues can be mitigated through in-context learning by providing LLMs with relevant context before generating answers. Recent literature proposes Knowledge Graph Prompting (KGP) which integrates knowledge graphs with an LLM-based traversal agent to substantially enhance document retrieval quality. However, KGP requires costly fine-tuning with large datasets and remains prone to hallucination. In this paper, we propose CuriousLLM, an enhancement that integrates a curiosity-driven reasoning mechanism into an LLM agent. This mechanism enables the agent to generate relevant follow-up questions, thereby guiding the information retrieval process more efficiently. Central to our approach is the development of the new Follow-upQA dataset, which includes questions and supporting evidence as input, with follow-up questions serving as ground truths. These follow-up questions either inquire about what is still missing to fully answer the user's query or use special tokens to signify that the retrieved evidence is sufficient. Our experiments show that CuriousLLM significantly boosts LLM performance in multi-document question answering (MD-QA), circumventing the substantial computational costs and latency from the original KGP framework.

cs.CL↗
arXiv2024

MakeupAttack: Feature Space Black-box Backdoor Attack on Face Recognition via Makeup Transfer

Backdoor attacks pose a significant threat to the training process of deep neural networks (DNNs). As a widely-used DNN-based application in real-world scenarios, face recognition systems once implanted into the backdoor, may cause serious consequences. Backdoor research on face recognition is still in its early stages, and the existing backdoor triggers are relatively simple and visible. Furthermore, due to the perceptibility, diversity, and similarity of facial datasets, many state-of-the-art backdoor attacks lose effectiveness on face recognition tasks. In this work, we propose a novel feature space backdoor attack against face recognition via makeup transfer, dubbed MakeupAttack. In contrast to many feature space attacks that demand full access to target models, our method only requires model queries, adhering to black-box attack principles. In our attack, we design an iterative training paradigm to learn the subtle features of the proposed makeup-style trigger. Additionally, MakeupAttack promotes trigger diversity using the adaptive selection method, dispersing the feature distribution of malicious samples to bypass existing defense methods. Extensive experiments were conducted on two widely-used facial datasets targeting multiple models. The results demonstrate that our proposed attack method can bypass existing state-of-the-art defenses while maintaining effectiveness, robustness, naturalness, and stealthiness, without compromising model performance.

cs.CV↗
arXiv2024

Laminar Matroid Secretary: Greedy Strikes Back

We show that a simple greedy algorithm is $4.75$ probability-competitive for the Laminar Matroid Secretary Problem, improving the $3\sqrt{3} \approx 5.196$-competitive algorithm based on the forbidden sets technique (Soto, Turkieltaub, and Verdugo, 2018).

cs.DS↗
arXiv2024

The Victim and The Beneficiary: Exploiting a Poisoned Model to Train a Clean Model on Poisoned Data

Recently, backdoor attacks have posed a serious security threat to the training process of deep neural networks (DNNs). The attacked model behaves normally on benign samples but outputs a specific result when the trigger is present. However, compared with the rocketing progress of backdoor attacks, existing defenses are difficult to deal with these threats effectively or require benign samples to work, which may be unavailable in real scenarios. In this paper, we find that the poisoned samples and benign samples can be distinguished with prediction entropy. This inspires us to propose a novel dual-network training framework: The Victim and The Beneficiary (V&B), which exploits a poisoned model to train a clean model without extra benign samples. Firstly, we sacrifice the Victim network to be a powerful poisoned sample detector by training on suspicious samples. Secondly, we train the Beneficiary network on the credible samples selected by the Victim to inhibit backdoor injection. Thirdly, a semi-supervised suppression strategy is adopted for erasing potential backdoors and improving model performance. Furthermore, to better inhibit missed poisoned samples, we propose a strong data augmentation method, AttentionMix, which works well with our proposed V&B framework. Extensive experiments on two widely used datasets against 6 state-of-the-art attacks demonstrate that our framework is effective in preventing backdoor injection and robust to various attacks while maintaining the performance on benign samples. Our code is available at https://github.com/Zixuan-Zhu/VaB.

cs.CV↗
arXiv2023

Bounding and Filling: A Fast and Flexible Framework for Image Captioning

Most image captioning models following an autoregressive manner suffer from significant inference latency. Several models adopted a non-autoregressive manner to speed up the process. However, the vanilla non-autoregressive manner results in subpar performance, since it generates all words simultaneously, which fails to capture the relationships between words in a description. The semi-autoregressive manner employs a partially parallel method to preserve performance, but it sacrifices inference speed. In this paper, we introduce a fast and flexible framework for image captioning called BoFiCap based on bounding and filling techniques. The BoFiCap model leverages the inherent characteristics of image captioning tasks to pre-define bounding boxes for image regions and their relationships. Subsequently, the BoFiCap model fills corresponding words in each box using two-generation manners. Leveraging the box hints, our filling process allows each word to better perceive other words. Additionally, our model offers flexible image description generation: 1) by employing different generation manners based on speed or performance requirements, 2) producing varied sentences based on user-specified boxes. Experimental evaluations on the MS-COCO benchmark dataset demonstrate that our framework in a non-autoregressive manner achieves the state-of-the-art on task-specific metric CIDEr (125.6) while speeding up 9.22x than the baseline model with an autoregressive manner; in a semi-autoregressive manner, our method reaches 128.4 on CIDEr while a 3.69x speedup. Our code and data is available at https://github.com/ChangxinWang/BoFiCap.

cs.CV↗
arXiv2023

Beyond Generic: Enhancing Image Captioning with Real-World Knowledge using Vision-Language Pre-Training Model

Current captioning approaches tend to generate correct but "generic" descriptions that lack real-world knowledge, e.g., named entities and contextual information. Considering that Vision-Language Pre-Training (VLP) models master massive such knowledge from large-scale web-harvested data, it is promising to utilize the generalizability of VLP models to incorporate knowledge into image descriptions. However, using VLP models faces challenges: zero-shot inference suffers from knowledge hallucination that leads to low-quality descriptions, but the generic bias in downstream task fine-tuning hinders the VLP model from expressing knowledge. To address these concerns, we propose a simple yet effective method called Knowledge-guided Replay (K-Replay), which enables the retention of pre-training knowledge during fine-tuning. Our approach consists of two parts: (1) a knowledge prediction task on automatically collected replay exemplars to continuously awaken the VLP model's memory about knowledge, thus preventing the model from collapsing into the generic pattern; (2) a knowledge distillation constraint to improve the faithfulness of generated descriptions hence alleviating the knowledge hallucination. To evaluate knowledge-enhanced descriptions, we construct a novel captioning benchmark KnowCap, containing knowledge of landmarks, famous brands, special foods and movie characters. Experimental results show that our approach effectively incorporates knowledge into descriptions, outperforming strong VLP baseline by 20.9 points (78.7->99.6) in CIDEr score and 20.5 percentage points (34.0%->54.5%) in knowledge recognition accuracy. Our code and data is available at https://github.com/njucckevin/KnowCap.

cs.CV↗
arXiv2023

A Novel Scholar Embedding Model for Interdisciplinary Collaboration

Interdisciplinary collaboration has become a driving force for scientific breakthroughs, and evaluating scholars' performance in interdisciplinary researches is essential for promoting such collaborations. However, traditional scholar evaluation methods based solely on individual achievements do not consider interdisciplinary cooperation, creating a challenge for interdisciplinary scholar evaluation and recommendation. To address this issue, we propose a scholar embedding model that quantifies and represents scholars based on global semantic information and social influence, enabling real-time tracking of scholars' research trends. Our model incorporates semantic information and social influence for interdisciplinary scholar evaluation, laying the foundation for future interdisciplinary collaboration discovery and recommendation projects. We demonstrate the effectiveness of our model on a sample of scholars from the Beijing University of Posts and Telecommunications.

cs.DL↗
arXiv2019

3D quench modeling based on T-A formulation for high temperature superconductor CORC cables

High temperature superconductor (HTS) (RE)Ba2Cu3Ox (REBCO) conductor on round core cable (CORC) has high current carrying capacity for high field magnet and power applications. In REBCO CORC cables, current redistribution occurs among tapes through terminal contact resistances when a local quench occurs. Therefore, the quench behaviour of CORC cable is different from single tape situation, for it is significantly affected by terminal contact resistances. To better understand the underlying physical process of local quenches in CORC cables, a new 3D multi-physics modelling tool for CORC cables is developed and presented in this paper. In this model, the REBCO tape is treated as a thin shell without thickness, and four models are coupled: T-formulation model, A-formulation model, a heat transfer model and an equivalent circuit model. The current redistribution, temperature and tape voltage of CORC cable during hot spot induced quenches are analysed using this model. The results show that the thermal stability of CORC cable can be considerably improved by reducing terminal contact resistance. The minimum quench energy (MQE) increases rapidly with the reduction of terminal contact resistance when the resistance is in a middle range. When the terminal contact resistance is too low or too high, the MQE shows no obvious variation with terminal contact resistances. With a low terminal contact resistance, a hot spot in one tape may induce an over-current quench on the other tapes without hot spots. This will not happen in a cable with high terminal contact resistance. In this case, the tape with hot spot will quench and burn out before inducing a quench on other tapes. The modelling tool developed can be used to design CORC cables with improved thermal stability.

cond-mat.supr-con↗