arXiv subjects

Kuan Wang

Publications and source records attributed to Kuan Wang.

At least 19 recordsLinked to original sources

arXiv2026

MemLeak: Diagnosing Information Leaks in Multimodal Agent Memory

When a multimodal AI agent is asked to forget a fact, current memory systems usually delete the text entry and report success. We find that the fact can remain recoverable from retained user images, including images tagged to entirely different facts, because VLMs use implicit visual cues at inference time. We introduce the Information Provenance Graph (IPG), a taxonomy that classifies memory representations by deletion affordance. The IPG reveals that deletion fails through multiple channels. Our benchmark, MemLeak, measures this across a deletion cascade: direct probing of deletion-capable systems yields <1%, but retained correlated text enables 18.3% recovery, and retained images enable 12.0% recovery (0.0% blind baseline, 0.3% FPR) -- with 47% of image leaks not text-recoverable. Content-aware semantic deletion reduces the image residual to 2.0%. The residual appears across multiple VLMs, a production memory system, and real Unsplash-licensed photographs. Dual-annotator human validation (kappa = 0.88) confirms judge reliability.

cs.LG↗

arXiv2026

MemDelta: Controlled Baselines and Hidden Confounds in Agent Memory Evaluation

Agent memory systems are increasingly evaluated against RAG and full-context baselines, but reported gains often mix changes in the memory method with changes in the language model, embedding model, or retrieval pipeline, making it unclear what is actually being measured. We present MemDelta, a controlled evaluation protocol that varies one component at a time on LongMemEval-S (500 questions, 50+ sessions, three model families). Four findings emerge: (1) verbatim RAG matches full-context GPT-4o-mini (47.2% vs. 49.8%, p = 0.34), but the ranking reverses across models: Gemini gains +14pp from full context, while Sonnet gains +31pp from RAG, partly because it refuses 63% of full-context queries; (2) swapping only the embedding model in an identical pipeline shifts accuracy by +6.2pp at n = 500 (p = 0.004), and Mem0 beats MiniLM-RAG by +11pp but loses to cloud-RAG by 1.2pp, so one variable flips the conclusion; (3) agent self-memory (42%) underperforms basic retrieval (47%); (4) on 2 of 6 question types (n = 88), Mem0 matches cloud RAG (72.7% vs. 73.9%, p = 1.0) at 50x the cost, suggesting narrow rather than general gains. We recommend memory evaluations fix embedding models across comparisons, stratify by model family, and report write-path cost before attributing gains to architecture.

cs.CL↗

arXiv2026

An Improved Fit for Linear Halo Bias at High Redshift

High- to ultrahigh-redshift clustering of halos provides a powerful tool to understand cosmology and galaxy formation. However, theoretical predictions are not firmly established in the first billion years, where current and upcoming surveys are beginning to reach percent-level precision. Here we measure dark matter halo biases at $z=6$ - 19 from simulation data, and find they are $\sim$ 3 - 4$\%$ higher than canonical results calibrated at low $z$. We provide an updated linear-bias fit at these early times, reducing the mean systematic offset to $< 1\%$. These results will enable robust interpretation of early-Universe galaxy clustering from JWST, Roman, and intensity-mapping surveys.

astro-ph.CO↗

arXiv2025

Assisted Refinement Network Based on Channel Information Interaction for Camouflaged and Salient Object Detection

Camouflaged Object Detection (COD) stands as a significant challenge in computer vision, dedicated to identifying and segmenting objects visually highly integrated with their backgrounds. Current mainstream methods have made progress in cross-layer feature fusion, but two critical issues persist during the decoding stage. The first is insufficient cross-channel information interaction within the same-layer features, limiting feature expressiveness. The second is the inability to effectively co-model boundary and region information, making it difficult to accurately reconstruct complete regions and sharp boundaries of objects. To address the first issue, we propose the Channel Information Interaction Module (CIIM), which introduces a horizontal-vertical integration mechanism in the channel dimension. This module performs feature reorganization and interaction across channels to effectively capture complementary cross-channel information. To address the second issue, we construct a collaborative decoding architecture guided by prior knowledge. This architecture generates boundary priors and object localization maps through Boundary Extraction (BE) and Region Extraction (RE) modules, then employs hybrid attention to collaboratively calibrate decoded features, effectively overcoming semantic ambiguity and imprecise boundaries. Additionally, the Multi-scale Enhancement (MSE) module enriches contextual feature representations. Extensive experiments on four COD benchmark datasets validate the effectiveness and state-of-the-art performance of the proposed model. We further transferred our model to the Salient Object Detection (SOD) task and demonstrated its adaptability across downstream tasks, including polyp segmentation, transparent object detection, and industrial and road defect detection. Code and experimental results are publicly available at: https://github.com/akuan1234/ARNet-v2.

cs.CV↗

arXiv2025

Reinforcement Learning for Reasoning in Large Language Models with One Training Example

We show that reinforcement learning with verifiable reward using one training example (1-shot RLVR) is effective in incentivizing the math reasoning capabilities of large language models (LLMs). Applying RLVR to the base model Qwen2.5-Math-1.5B, we identify a single example that elevates model performance on MATH500 from 36.0% to 73.6% (8.6% improvement beyond format correction), and improves the average performance across six common mathematical reasoning benchmarks from 17.6% to 35.7% (7.0% non-format gain). This result matches the performance obtained using the 1.2k DeepScaleR subset (MATH500: 73.6%, average: 35.9%), which contains the aforementioned example. Furthermore, RLVR with only two examples even slightly exceeds these results (MATH500: 74.8%, average: 36.6%). Similar substantial improvements are observed across various models (Qwen2.5-Math-7B, Llama3.2-3B-Instruct, DeepSeek-R1-Distill-Qwen-1.5B), RL algorithms (GRPO and PPO), and different math examples. In addition, we identify some interesting phenomena during 1-shot RLVR, including cross-category generalization, increased frequency of self-reflection, and sustained test performance improvement even after the training accuracy has saturated, a phenomenon we term post-saturation generalization. Moreover, we verify that the effectiveness of 1-shot RLVR primarily arises from the policy gradient loss, distinguishing it from the "grokking" phenomenon. We also show the critical role of promoting exploration (e.g., by incorporating entropy loss with an appropriate coefficient) in 1-shot RLVR training. We also further discuss related observations about format correction, label robustness and prompt modification. These findings can inspire future work on RLVR efficiency and encourage a re-examination of recent progress and the underlying mechanisms in RLVR. All resources are open source at https://github.com/ypwang61/One-Shot-RLVR.

cs.LG↗

arXiv2024

Is Your World Simulator a Good Story Presenter? A Consecutive Events-Based Benchmark for Future Long Video Generation

The current state-of-the-art video generative models can produce commercial-grade videos with highly realistic details. However, they still struggle to coherently present multiple sequential events in the stories specified by the prompts, which is foreseeable an essential capability for future long video generation scenarios. For example, top T2V generative models still fail to generate a video of the short simple story 'how to put an elephant into a refrigerator.' While existing detail-oriented benchmarks primarily focus on fine-grained metrics like aesthetic quality and spatial-temporal consistency, they fall short of evaluating models' abilities to handle event-level story presentation. To address this gap, we introduce StoryEval, a story-oriented benchmark specifically designed to assess text-to-video (T2V) models' story-completion capabilities. StoryEval features 423 prompts spanning 7 classes, each representing short stories composed of 2-4 consecutive events. We employ advanced vision-language models, such as GPT-4V and LLaVA-OV-Chat-72B, to verify the completion of each event in the generated videos, applying a unanimous voting method to enhance reliability. Our methods ensure high alignment with human evaluations, and the evaluation of 11 models reveals its challenge, with none exceeding an average story-completion rate of 50%. StoryEval provides a new benchmark for advancing T2V models and highlights the challenges and opportunities in developing next-generation solutions for coherent story-driven video generation.

cs.CV↗

arXiv2024

Mojito: Motion Trajectory and Intensity Control for Video Generation

Recent advancements in diffusion models have shown great promise in producing high-quality video content. However, efficiently training video diffusion models capable of integrating directional guidance and controllable motion intensity remains a challenging and under-explored area. To tackle these challenges, this paper introduces Mojito, a diffusion model that incorporates both motion trajectory and intensity control for text-to-video generation. Specifically, Mojito features a Directional Motion Control (DMC) module that leverages cross-attention to efficiently direct the generated object's motion without training, alongside a Motion Intensity Modulator (MIM) that uses optical flow maps generated from videos to guide varying levels of motion intensity. Extensive experiments demonstrate Mojito's effectiveness in achieving precise trajectory and intensity control with high computational efficiency, generating motion patterns that closely match specified directions and intensities, providing realistic dynamics that align well with natural motion in real-world scenarios.

cs.CV↗

arXiv2024

The Beyond-Halo Mass Effects of the Cosmic Web Environment on Galaxies

Galaxy properties primarily depend on their host halo mass. Halo mass, in turn, depends on the cosmic web environment. We explore if the effect of the cosmic web on galaxy properties is entirely transitive via host halo mass, or if the cosmic web has an effect independent of mass. The secondary galaxy bias, sometimes referred to as ``galaxy assembly bias'', is the beyond-mass component of the galaxy-halo connection. We investigate the link between the cosmic web environment and the secondary galaxy bias in simulations. We measure the secondary galaxy bias through the following summary statistics: projected two-point correlation function, $\wprp$, and counts-in-cylinders statistics, $\Pncic$. First, we examine the extent to which the secondary galaxy bias can be accounted for with a measure of the environment as a secondary halo property. We find that the total secondary galaxy bias preferentially places galaxies in more strongly clustered haloes. In particular, haloes at fixed mass tend to host more galaxies when they are more strongly associated with nodes or filaments. This tendency accounts for a significant portion, but not the entirety, of the total secondary galaxy bias effect. Second, we quantify how the secondary galaxy bias behaves differently depending on the host halo proximity to nodes and filaments. We find that the total secondary galaxy bias is relatively stronger in haloes more associated with nodes or filaments. We emphasise the importance of removing halo mass effects when considering the cosmic web environment as a factor in the galaxy-halo connection.

astro-ph.GA↗

arXiv2024

RNR: Teaching Large Language Models to Follow Roles and Rules

Instruction fine-tuning (IFT) elicits instruction following capabilities and steers the behavior of large language models (LLMs) via supervised learning. However, existing models trained on open-source IFT datasets only have the ability to follow instructions from users, and often fail to follow complex role and rules specified by developers, a.k.a. system prompts. The ability to follow these roles and rules is essential for deployment, as it ensures that the model safely interacts with users within developer defined guidelines. To improve such role and rule following ability, we propose \model, an automated data generation pipeline that generates diverse roles and rules from existing IFT instructions, along with corresponding responses. This data can then be used to train models that follow complex system prompts. The models are evaluated on our newly created benchmarks for role and rule following ability, as well as standard instruction-following benchmarks and general NLP tasks. Our framework significantly improves role and rule following capability in LLMs, as evidenced by over 25% increase in pass-rate on rule adherence, i.e. following all requirements, in our experiments with the Alpaca and Ultrachat datasets. Moreover, our models achieves this increase without any regression on popular instruction following benchmarks.

cs.CL↗

arXiv2024

Subhalos in Galaxy Clusters: Coherent Accretion and Internal Orbits

Subhalo dynamics in galaxy cluster host halos govern the observed distribution and properties of cluster member galaxies. We use the IllustrisTNG simulation to investigate the accretion and orbits of subhalos found in cluster-size halos. We find that the median change in the major axis direction of cluster-size host halos is approximately $80$ degrees between $a\sim0.1$ and present-day. We identify coherent regions in the angular distribution of subhalo accretion, and $\sim68\%$ of accreted subhalos enter their host halo through $\sim38\%$ of the surface area at the virial radius. The majority of galaxy clusters in the sample have $\sim2$ such coherent regions. We further measure angular orbits of subhalos with respect to the host major axis and use a clustering algorithm to identify distinct orbit modes with varying oscillation timescales. The orbit modes correlate with subhalo accretion conditions. Subhalos in orbit modes with shorter oscillations tend to have lower peak masses and accretion directions somewhat more aligned with the major axis. One orbit mode, exhibiting the least oscillatory behavior, largely consists of subhalos that accrete near the plane perpendicular to the host halo major axis. Our findings are consistent with expectations from inflow from major filament structures and internal dynamical friction: most subhalos accrete through coherent regions, and more massive subhalos experience fewer orbits after accretion. Our work offers a unique quantification of subhalo dynamics that can be connected to how the intracluster medium strips and quenches cluster galaxies.

astro-ph.GA↗

arXiv2024

Subhalos are Anisotropically Distributed and Aligned with the Smooth Matter Distribution of Their Host Halos

We investigate the distributions of subhalos about their hosts in two suites of zoom-in N-body simulations of halo growth -- one suite focused on Milky Way-Mass halos ($\sim 10^{12} \mathrm{M}_{\odot}$) and another focused on cluster-mass halos ($\sim 10^{15} \mathrm{M}_{\odot}$) in the Symphony simulation suite. We find, in agreement with previous work on this subject, that subhalos are distributed anisotropically about their host halos. In particular, the positions of subhalos lie preferentially near the major axes of their host halos, possibly implying that satellite galaxies will exhibit a similar alignment. Furthermore, we show that in two-dimensional projection subhalos are more likely to be observed near the halo center (where the central galaxy presumably resides) when the host halo is projected nearly along its major axis. This projection effect is significant. Within projected radii of a few percent of the virial radius of the host halo, the fraction of mass in subhalos is $\sim 175\%$ larger for Milky Way mass halos and as much as $\sim 195\%$ larger for cluster halos when projected along the major axis as compared to the average from a random projection. This result has consequences for many applications including the interpretation of gravitational lenses. Finally, we find that the orbital angular momentum vector of subhalos is aligned with the angular momentum vector of their host halo, indicating that a significant component of a halo's angular momentum may be carried in its subhalos. This has consequences for galaxy formation models which use host halo angular momentum as a proxy for galaxy momentum.

astro-ph.GA↗

arXiv2024

ARL2: Aligning Retrievers for Black-box Large Language Models via Self-guided Adaptive Relevance Labeling

Retrieval-augmented generation enhances large language models (LLMs) by incorporating relevant information from external knowledge sources. This enables LLMs to adapt to specific domains and mitigate hallucinations in knowledge-intensive tasks. However, existing retrievers are often misaligned with LLMs due to their separate training processes and the black-box nature of LLMs. To address this challenge, we propose ARL2, a retriever learning technique that harnesses LLMs as labelers. ARL2 leverages LLMs to annotate and score relevant evidence, enabling learning the retriever from robust LLM supervision. Furthermore, ARL2 uses an adaptive self-training strategy for curating high-quality and diverse relevance data, which can effectively reduce the annotation cost. Extensive experiments demonstrate the effectiveness of ARL2, achieving accuracy improvements of 5.4% on NQ and 4.6% on MMLU compared to the state-of-the-art methods. Additionally, ARL2 exhibits robust transfer learning capabilities and strong zero-shot generalization abilities. Our code will be published at \url{https://github.com/zhanglingxi-cs/ARL2}.

cs.CL↗

arXiv2024

Adapting LLM Agents with Universal Feedback in Communication

Recent advances in large language models (LLMs) have demonstrated potential for LLM agents. To facilitate the training for these agents with both linguistic feedback and non-linguistic reward signals, we introduce Learning through Communication (LTC). We design a universal buffer to store all the feedback, and an iterative pipeline to enable an LLM agent to explore and update its policy in an given environment. To optimize agent interactions for task-specific learning with our universal buffer and pipeline, we introduce diverse communication patterns tailored for both single-agent and multi-agent environments. We evaluate the efficacy of our LTC approach on four diverse datasets: ALFWorld (single-agent), HotpotQA (multi-agent collaboration), Chameleon (multi-agent competition), and GSM8k (multi-agent teacher-student). On these data sets, LTC outperforms the supervised instruction fine-tuning baselines by 3.6% to 12%. These results highlight the versatility and efficiency of LTC in facilitating online adaptation for LLM agents.

cs.CL↗

arXiv2023

Merger Response of Halo Anisotropy Properties

Anisotropy properties -- halo spin, shape, position offset, velocity offset, and orientation -- are an important family of dark matter halo properties that indicate the level of directional variation of the internal structures of haloes. These properties reflect the dynamical state of haloes, which in turn depends on the mass assembly history. In this work, we study the evolution of anisotropy properties in response to merger activity using the IllustrisTNG simulations. We find that the response trajectories of the anisotropy properties significantly deviate from secular evolution. These trajectories have the same qualitative features and timescales across a wide range of merger and host properties. We propose explanations for the behaviour of these properties and connect their evolution to the relevant stages of merger dynamics. We measure the relevant dynamical timescales. We also explore the dependence of the strength of the response on time of merger, merger ratio, and mass of the main halo. These results provide insight into the physics of halo mergers and their effects on the statistical behaviour of halo properties. This study paves the way towards a physical understanding of scaling relations, particularly to how systematics in their scatter are connected to the mass assembly histories of haloes.

astro-ph.CO↗

arXiv2023

The Influence of Subhaloes on Host Halo Properties

Within the $Λ$CDM cosmology, dark matter haloes are comprised of both a smooth component and a population of smaller, gravitationally bound subhaloes. These components are often treated as a single halo when halo properties, such as density profiles, are extracted from simulations. Recent work has shown that density profiles change substantially when subhalo mass is excluded. In this paper, we expand on this result by analysing the change in three specific host halo properties -- concentration ($c_{\rm{NFW}}$), spin ($λ_{\rm Bullock}$), and shape ($c/a$), -- when calculated only from the smooth component of the halo. This analysis is performed on both Milky Way-mass haloes and cluster-mass haloes in high-resolution, zoom-in, $N$-body simulations. We find that when subhaloes are excluded the median value of (1) $c_{\rm{NFW}}$ is enhanced by $\approx 38 \pm 12\%$ and $\approx 88 \pm 7.7\%$ for Milky Way mass ($10^{12.1}\,\text{M}_\odot$) and cluster mass ($10^{14.8}\,\text{M}_\odot$) haloes respectively, (2) $λ_{\rm Bullock}$ is reduced for Milky Way mass by $\approx 16 \pm 6.8\%$ and cluster mass haloes by $\approx 32 \pm 8.9\%$. Additionally, with the removal of subhaloes, cluster mass haloes tend to become more spherical as the ratio of minor-to-major axis, $c/a$, increases by $\approx 12 \pm 4\%$, whereas Milky Way mass haloes remain approximately the same shape with $c/a$ changed by $\approx 1.2 \pm 5.6\%$. The fractional change of each of these properties depends primarily on the amount of mass that is removed from the halo system and, to a lesser extent, mass accretion history. Our findings demonstrate that the properties of the smooth components of dark matter haloes are biased relative to the total mass of the halo including subhaloes.

astro-ph.GA↗

arXiv2023

The DESI One-Percent Survey: Evidence for Assembly Bias from Low-Redshift Counts-in-Cylinders Measurements

We explore the galaxy-halo connection information that is available in low-redshift samples from the early data release of the Dark Energy Spectroscopic Instrument (DESI). We model the halo occupation distribution (HOD) from z=0.1-0.3 using Survey Validation 3 (SV3; a.k.a., the One-Percent Survey) data of the DESI Bright Galaxy Survey (BGS). In addition to more commonly used metrics, we incorporate counts-in-cylinders (CiC) measurements, which drastically tighten HOD constraints. Our analysis is aided by the Python package, galtab, which enables the rapid, precise prediction of CiC for any HOD model available in halotools. This methodology allows our Markov chains to converge with much fewer trial points, and enables even more drastic speedups due to its GPU portability. Our HOD fits constrain characteristic halo masses tightly and provide statistical evidence for assembly bias, especially at lower luminosity thresholds: the HOD of central galaxies in $z\sim0.15$ samples with limiting absolute magnitude $M_r < -20.0$ and $M_r < -20.5$ samples is positively correlated with halo concentration with a significance of 99.9% and 99.5%, respectively. Our models also favor positive central assembly bias for the brighter $M_r < -21.0$ sample at $z\sim0.25$ (94.8% significance), but there is no significant evidence for assembly bias with the same luminosity threshold at $z\sim0.15$. We provide our constraints for each threshold sample's characteristic halo masses, assembly bias, and other HOD parameters. These constraints are expected to be significantly tightened with future DESI data, which will span an area 100 times larger than that of SV3.

astro-ph.CO↗

arXiv2023

MultiCAM: A multivariable framework for connecting the mass accretion history of haloes with their properties

Models that connect galaxy and halo properties often summarize a halo's mass accretion history (MAH) with a single value, and use this value as the basis for predictions. However, a single-value summary fails to capture the complexity of MAHs and information can be lost in the process. We present MultiCAM, a generalization of traditional abundance matching frameworks, which can simultaneously connect the full MAH of a halo with multiple halo and/or galaxy properties. As a first case study, we apply MultiCAM to the problem of connecting dark matter halo properties to their MAHs in the context of a dark matter-only simulation. While some halo properties, such as concentration, are more strongly correlated to the early-time mass growth of a halo, others, like the virial ratio, have stronger correlations with late-time mass growth. This highlights the necessity of considering the impact of the entire MAH on halo properties. For most of the halo properties we consider, we find that MultiCAM models that use the full MAH achieve higher accuracy than conditional abundance matching models which use a single epoch. We also demonstrate an extension of MultiCAM that captures the covariance between predicted halo properties. This extension provides a baseline model for applications where the covariance between predicted properties is important.

astro-ph.CO↗

arXiv2023

ToolQA: A Dataset for LLM Question Answering with External Tools

Large Language Models (LLMs) have demonstrated impressive performance in various NLP tasks, but they still suffer from challenges such as hallucination and weak numerical reasoning. To overcome these challenges, external tools can be used to enhance LLMs' question-answering abilities. However, current evaluation methods do not distinguish between questions that can be answered using LLMs' internal knowledge and those that require external information through tool use. To address this issue, we introduce a new dataset called ToolQA, which is designed to faithfully evaluate LLMs' ability to use external tools for question answering. Our development of ToolQA involved a scalable, automated process for dataset curation, along with 13 specialized tools designed for interaction with external knowledge in order to answer questions. Importantly, we strive to minimize the overlap between our benchmark data and LLMs' pre-training data, enabling a more precise evaluation of LLMs' tool-use reasoning abilities. We conducted an in-depth diagnosis of existing tool-use LLMs to highlight their strengths, weaknesses, and potential improvements. Our findings set a new benchmark for evaluating LLMs and suggest new directions for future advancements. Our data and code are freely available to the broader scientific community on GitHub.

cs.CL↗