Searcharxiv⌕ Search

arXiv subjects

Zhiqi Huang

Publications and source records attributed to Zhiqi Huang.

At least 19 recordsLinked to original sources

Benchmarking Answer-Grounding Consistency in Remote Sensing Visual Question Answering

In grounded remote sensing visual question answering, a model produces a textual answer and identifies relevant objects in the image as grounding evidence. Existing evaluations measure answer correctness and localization quality, but do not directly assess answer-grounding consistency. We propose the Answer-Evidence Support Protocol (AESP) to assess this consistency through evidence support: whether the grounding evidence supports the model's own answer. AESP applies task-specific rules to existence, counting, and comparison questions. We further construct the Grounded Remote Sensing Visual Question Answering (G-RSVQA) benchmark, containing 15,027 samples with questions, reference answers, and all related object-level annotations. Experiments on seven grounded vision--language models reveal failures of evidence support across all three question types. Up to 50.82% of correct answers in the evidence evaluation subset are not supported by the provided grounding evidence.

cs.CV↗

SETA: Scaling Environments for Terminal Agents

Large language models (LLMs) are rapidly shifting toward agents that solve tasks through diverse interfaces, including web and graphical user interfaces (GUIs). Among these, the terminal command line provides a text-based, general-purpose interface, covering tasks from system operations to data science and machine learning. However, scaling terminal-agent training remains challenging, as it requires diverse and coherent task instructions, executable environments, and reliable verification, while lacking naturally grounded supervision data. In this work, we propose SETA, a scalable framework for generating verifiable terminal environments for reinforcement learning (RL). The framework consists of two pipelines sharing a unified verification mechanism: SETA-Synth converts diverse sources into standardized RL environments, and SETA-Evol further expands from existing environments with adaptive control of difficulty and diversity. Together, we construct and release SETA-Env, the largest open-source verifiable terminal RL dataset to date, containing over 4,500 environments. We evaluate our dataset by training Qwen3-8B with GRPO on SETA-Env, achieving 12% pass rate on Terminal-Bench 2.0, the best reported result for an RL-trained model at the 8B scale. We further observe gains on DeepSeek-V4-Flash under the same terminal agent harness, with pass@1 on Terminal-Bench 2.0 improving from 40% to 43% and pass@5 improving from 54% to 58%. These results demonstrate that SETA- Env provides high-quality training environments for terminal agents and serves as a valuable resource for advancing research on terminal-based agent learning.

cs.AI↗

When Noise Meets Long-Tail: Feature-Threshold Dual Calibration for Robust Pseudo-Labeling

Pseudo-labeling has become a cornerstone of learning from unlabeled data in semantic segmentation. Yet its effectiveness drops sharply in real-world scenarios where strong imaging noise and long-tailed class distributions occur together. We trace this failure to a vicious cycle of pseudo-label degradation. Imaging noise entangles foreground and background features, lowering prediction confidence across all classes, while long-tailed distributions leave tail classes with far fewer training samples and inherently lower confidence. Under fixed high-threshold filtering, these tail-class predictions are systematically filtered out, so they receive no supervision from unlabeled data and thus features keep degrading in subsequent iterations. Critically, noise and long-tail are not independent obstacles but mutually amplifying ones, and addressing either alone is insufficient. To break this cycle, we propose FTC-Seg, a Feature-Threshold dual-Calibration framework built on a standard teacher-student framework. At the feature level, Orthogonal Prototype Reconstruction (OPR) uses a set of learnable orthogonal prototypes to residually purify pixel-wise features, widening the margin between weak foreground targets and noisy backgrounds. At the threshold level, Adaptive Threshold Calibration (ATC) dynamically adjusts class-specific thresholds based on learning difficulty and prediction-distribution bias, rescuing low-confidence pseudo-labels of tail classes from systematic exclusion. Extensive experiments on four public benchmarks spanning three distinct noise modalities show that FTC-Seg achieves strong performance against state-of-the-art methods, with particularly substantial gains on tail classes. Our results establish that jointly calibrating features and thresholds is essential for robust pseudo-labeling under compounded noise and class imbalance.

cs.CV↗

Falsifiable anthropic model from the string landscape

We combine the axiverse with a negative cosmological constant $Λ$, both originating from the string landscape, and apply an observation-time-weighted anthropic argument. Adopting a uniform prior on $Λ$ and a typical string-motivated axion decay constant comparable to the reduced Planck scale, this framework predicts a $\sim 40\%$ probability of observing $0.1 < Ω_m < 0.9$, thereby naturally resolving the long-standing fine-tuning and coincidence problems of dark energy. Unlike many scalar-field dark energy models, the anthropic Lambda-axion universe (ALAverse) statistically disfavors slow-roll dynamics, as slow-roll requires fine-tuning of the initial field displacement. Moreover, the negative cosmological constant renders fast-roll scenarios anthropically unfavorable, since they typically yield only a very brief observational window with positive dark energy density. Having ruled out both extremes, the ALAverse characteristically predicts a moderate-roll dynamics. We derive a two-parameter parametrization in terms of $δ_Ω$ and $|\varepsilon_s|$ that covers ALAverse solutions as well as a broad class of canonical and phantom field models. Current observational data yield $δ_Ω= -0.0498 \pm 0.0186$, corresponding to a $2.7σ$ rejection of $Λ$CDM ($δ_Ω= 0$) and single-field thawing phantom models ($δ_Ω> 0$). The data also show a mild preference for the ALAverse over slow-roll quintessence, a trend that can be conclusively tested with future high-precision measurements of the Hubble diagram.

astro-ph.CO↗

Many Paths to the Same Mistake: Generating Diverse Reasoning Paths Behind the Same Student Mistake

Realistic evaluation of AI systems requires user simulators that capture not only cooperative behavior, but also the errors and variability that arise in real interactions. In educational technology, simulating step-by-step flawed student reasoning is essential, yet datasets mapping these pathways are scarce, and Large Language Models (LLMs) often auto-correct rather than sustain simulated mistakes. To address this, we present a generate-and-filter pipeline that produces diverse and realistic student error traces. Our approach utilizes forward consistency, a metric verifying if a flawed reasoning trace can be used to recover the student's actual incorrect answer, paired with an LLM-as-a-judge to ensure realism. Our evaluation suggests that while many LLMs struggle to maintain incorrect reasoning mid-trace, smaller supervised fine-tuned models successfully capture these divergent paths. This work provides a framework for high-fidelity error trace generation, advancing realistic, non-cooperative user simulation.

cs.CL↗

VerTox: Verifiable Reward-Guided Corpus Poisoning Against Neural Ranking Models

Neural ranking models have become core components of modern information retrieval systems and important building blocks of AI systems such as retrieval-augmented generation (RAG) pipelines. However, their robustness remains insufficiently understood in the presence of large language models (LLMs), which can generate fluent and deceptive content at scale. This work investigates the vulnerability of neural ranking models to corpus poisoning attacks, in which an adversary injects a small number of maliciously crafted documents into the corpus to distort ranking behavior. We propose VerTox, the first framework to formulate corpus poisoning as a verifiable reward-guided reinforcement learning (RLVR) problem. By explicitly coupling ranking distortion with factual corruption through specialized reward shaping, we fine-tune compact LLMs into adversarial generators. Experiments demonstrate that our method achieves near-perfect attack success rates, producing adversarial documents that frequently rank higher than target documents across major neural ranking architectures, as well as a proprietary commercial embedding model. The generated adversarial documents are fluent and exhibit low perplexity, making them difficult to detect. Furthermore, by explicitly encouraging factual corruption, our adversarial documents significantly degrade the performance of a downstream RAG application.

cs.CL↗

Modelling the nonlinear matter power spectrum in Hu-Sawicki $f(R)$ gravity with the Web-Halo Model

We develop a semi-analytic extension of the Web-Halo Model (WHM) to Hu-Sawicki $f(R)$ gravity, with the aim of linking nonlinear matter clustering to the successive stages of cosmic-web collapse. The cylindrical sheet and filament windows of the original WHM are replaced by axisymmetric ellipsoidal top-hat windows, yielding a modest improvement around the perturbative-to-nonlinear transition without introducing additional fitting parameters. Modified gravity is incorporated through environment-dependent chameleon spherical collapse, from which the collapse threshold and virial overdensity are obtained and propagated through the sheet, filament, and halo contributions. For $|f_{R0}|=10^{-5}$ and $10^{-6}$, the predicted nonlinear enhancement is broadly consistent with the scale-, redshift-, and field-strength dependence seen in the e-MANTIS emulator, with closer agreement for the weaker field and at higher redshift. The component responses show that sheets and filaments contribute substantially around the transition regime, while the halo response becomes increasingly important at smaller scales. As complementary diagnostics, no displacement of the BAO peak is resolved, whereas tomographic weak-lensing spectra retain a percent-level response to the modified matter power spectrum. The extended WHM therefore provides a physically interpretable framework for tracing screened modified-gravity effects across the cosmic-web collapse hierarchy.

astro-ph.CO↗

Kimi K2.5: Visual Agentic Intelligence

We introduce Kimi K2.5, an open-source multimodal agentic model designed to advance general agentic intelligence. K2.5 emphasizes the joint optimization of text and vision so that two modalities enhance each other. This includes a series of techniques such as joint text-vision pre-training, zero-vision SFT, and joint text-vision reinforcement learning. Building on this multimodal foundation, K2.5 introduces Agent Swarm, a self-directed parallel agent orchestration framework that dynamically decomposes complex tasks into heterogeneous sub-problems and executes them concurrently. Extensive evaluations show that Kimi K2.5 achieves state-of-the-art results across various domains including coding, vision, reasoning, and agentic tasks. Agent Swarm also reduces latency by up to $4.5\times$ over single-agent baselines. We release the post-trained Kimi K2.5 model checkpoint to facilitate future research and real-world applications of agentic intelligence.

cs.CL↗

PerceptionBench: Evaluating Atomic Visual Perception in Multimodal Large Language Models

We introduce PerceptionBench, a benchmark specifically designed to evaluate the atomic visual perception capabilities of Multimodal Large Language Models (MLLMs). Existing benchmarks often fail to isolate perception: holistic evaluations conflate perceptual errors with failures in reasoning or domain knowledge, while application-driven benchmarks only cover narrow, fragmented domains shaped by heuristic designs. To address these limitations, PerceptionBench adopts a bottom-up approach: by diagnosing the earliest failure points in the responses of frontier MLLMs across 42 existing benchmarks, we construct an error taxonomy whose perception branch defines ten atomic perceptual capabilities. Guided by this taxonomy, we construct 3,000 verified questions with short, unambiguous answers, each isolating a single capability, with difficulty stemming from perception rather than reasoning or knowledge. Benchmark results across sixteen frontier MLLMs reveal that atomic perception remains largely unsolved---no model reaches 60\% accuracy, perception-related hallucination is the weakest capability on average, and similar overall scores conceal sharply divergent capability profiles. PerceptionBench thus provides a capability-level standard for measuring and diagnosing the visual perception boundaries of MLLMs.

cs.CV↗

BabyVision: Visual Reasoning Beyond Language

While humans develop core visual skills long before acquiring language, contemporary Multimodal LLMs (MLLMs) still rely heavily on linguistic priors to compensate for their fragile visual understanding. We uncovered a crucial fact: state-of-the-art MLLMs consistently fail on basic visual tasks that humans, even 3-year-olds, can solve effortlessly. To systematically investigate this gap, we introduce BabyVision, a benchmark designed to assess core visual abilities independent of linguistic knowledge for MLLMs. BabyVision spans a wide range of tasks, with 388 items divided into 22 subclasses across four key categories. Empirical results and human evaluation reveal that leading MLLMs perform significantly below human baselines. Gemini3-Pro-Preview scores 49.7, lagging behind 6-year-old humans and falling well behind the average adult score of 94.1. These results show despite excelling in knowledge-heavy evaluations, current MLLMs still lack fundamental visual primitives. Progress in BabyVision represents a step toward human-level visual perception and reasoning capabilities. We also explore solving visual reasoning with generation models by proposing BabyVision-Gen and automatic evaluation toolkit. Our code and benchmark data are released at https://github.com/UniPat-AI/BabyVision for reproduction.

cs.CV↗

Physical Implications and Updated Observational Constraints of the PAge-like Unified Dark Fluid Model

The standard paradigm of cosmology assumes two distinct dark components, namely dark matter and dark energy. However, the necessity of splitting the dark-side world into two sectors has not been experimentally or theoretically proven. Unified dark fluid models provide an alternative in which a single fluid accounts for both phenomena. It is shown in Wang et al. 2024 that a PAge-like unified dark fluid (PUDF) can explain both the cosmic microwave background (CMB) and late-universe data, with the fitting quality not much worse than the standard Lambda cold dark matter ($Λ$CDM) model. Using the Planck 2018 CMB, baryon acoustic oscillations measurement from the dark energy spectroscopic instrument (DESI) data release 2, dark energy survey 5-year supernova data, and cosmic-chronometer data, we update the constraints on PUDF and clarify its physical implications. We show that PUDF can reproduce the primary CMB anisotropies, the background expansion history, and linear growth that are very close to the $Λ$CDM prediction. Nevertheless, the combined datasets still favor $Λ$CDM, largely due to the significant tension between CMB and DESI + SNe data, which exceeds the $4σ$ level in PUDF and remains non-negligible in the $w$CDM framework. Using mock data generated from the Planck best-fit $Λ$CDM model, we find that PUDF and $Λ$CDM cannot be statistically distinguished, indicating that the precision of current data is insufficient to separate the two models. Overall, the apparent preference for $Λ$CDM may be driven by dataset inconsistencies rather than a genuine physical difference, leaving unified dark fluid models as viable alternatives within current observational limits.

astro-ph.CO↗

TCDA: Thread-Constrained Discourse-Aware Modeling for Conversational Sentiment Quadruple Analysis

Conversational Aspect-based Sentiment Quadruple Analysis (DiaASQ) needs to capture the complex interrelationships in multiple rounds of dialogues. Existing methods usually employ simple Graph Convolutional Networks (GCN), which introduce structural noise and fail to consider the temporal sequence of the dialogues, or use standard RoPE, which implicitly captures relative distances in a flat sequence but cannot clearly separate the token-level syntactic order from the utterance-level progression, and may suffer from the Distance Dilution problem. To address these issues, we propose a new framework that combines Thread-Constrained Directed Acyclic Graph (TC-DAG) and Discourse-Aware Rotary Position Embedding (D-RoPE). Specifically, TC-DAG filters out cross-thread noise based on thread constraints, maintains global connectivity through root anchoring, and incorporates the temporal sequence of the dialogues. D-RoPE aligns multi-layer semantics using dual-stream projection and multi-scale frequency signals, captures thread dependencies using tree-like distances, and alleviates the token-level Distance Dilution problem by incorporating utterance-level progressions. Experimental results on two benchmark datasets demonstrate that our framework achieves state-of-the-art performance.

cs.CL↗

MUltiplexed Survey Telescope (MUST) Science White Paper I: Overview of Large-Scale Structure Cosmology in the Era of Stage-V Spectroscopic Surveys

The MUltiplexed Survey Telescope (MUST) is a 6.5-meter telescope under development. Dedicated to highly-multiplexed, wide-field spectroscopic surveys, MUST observes over 20,000 targets simultaneously using 6.2-mm pitch positioning robots within a ~5 deg$^2$ field of view. MUST aims to conduct the first Stage-V spectroscopic survey in the 2030s, mapping the 3D Universe with over 100 million galaxies and quasars, spanning from the nearby Universe to a redshift of z ~ 5.5, corresponding to approximately 1 billion years after the Big Bang. To cover this extensive redshift range, we present an initial conceptual target selection algorithm for different types of galaxies, ranging from local bright galaxies and luminous red galaxies to emission-line galaxies, and high-redshift (2 < z < 5.5) Lyman-break galaxies. Using Fisher forecasts, we demonstrate that MUST can address fundamental questions in cosmology, including the nature of dark energy, tests of gravity theories, and investigations into primordial physics. This is the first paper in the series of science white papers for MUST, with subsequent developments focusing on additional scientific cases such as galaxy and quasar evolution, Milky Way physics, and dynamic phenomena in the time-domain Universe.

astro-ph.CO↗

Learning to Route Queries to Heads for Attention-based Re-ranking with Large Language Models

Large Language Models (LLMs) have recently been explored as fine-grained zero-shot re-rankers by leveraging attention signals to estimate document relevance. However, existing methods either aggregate attention signals across all heads or rely on a statically selected subset identified by heuristic rules. This solution can be suboptimal because the informative heads can vary across queries or domains. Moreover, naively combining multiple heads can degrade performance due to redundancy or conflicting ranking signals. In this paper, we propose a query-dependent head selection method, RouteHead, for attention-based re-ranking with LLMs. Specifically, we learn a lightweight router that can map each query to an optimal head set, and relevance scores are computed by aggregating attention signals only from these heads. Since query-to-head optimal labels are unavailable, we first construct pseudo labels via an offline search. The router represents each head with a learnable embedding and represents each query using an embedding extracted from the hidden states of the frozen LLM. Then it is trained on the pseudo labels with a sparsity regularizer. Experiments on diverse benchmarks and multiple LLM backbones show that the proposed method consistently outperforms strong baselines.

cs.IR↗

SIE3D: Single-Image Expressive 3D Avatar Generation via Semantic Embedding and Perceptual Expression Loss

Generating high-fidelity 3D head avatars from a single image is challenging, as current methods lack fine-grained, intuitive control over expressions via text. This paper proposes SIE3D, a framework that generates expressive 3D avatars from a single image and descriptive text. SIE3D fuses identity features from the image with semantic embedding from text through a novel conditioning scheme, enabling detailed control. To ensure generated expressions accurately match the text, it introduces an innovative perceptual expression loss function. This loss uses a pre-trained expression classifier to regularize the generation process, guaranteeing expression accuracy. Extensive experiments show SIE3D significantly improves controllability and realism, outperforming competitive methods in identity preservation and expression fidelity on a single consumer-grade GPU. Project page: https://huang-zhiqi.github.io/SIE3D/

cs.CV↗

MMErroR: A Benchmark for Erroneous Reasoning in Vision-Language Models

Recent advances in Vision-Language Models (VLMs) have improved performance in multi-modal learning, raising the question of whether these models truly understand the content they process. Crucially, can VLMs detect when a reasoning process is wrong and identify its error type? To answer this, we present MMErroR, a multi-modal benchmark of 1997 samples, each embedding a single coherent reasoning error. These samples span 24 subdomains across six top-level domains, ensuring broad coverage and taxonomic richness. Unlike existing benchmarks that focus on answer correctness, MMErroR targets a process-level, error-centric evaluation that requires models to detect incorrect reasoning and classify the error type within both visual and linguistic contexts. We evaluate 12 representative VLMs, and even the best model, Gemini-3-Pro-Preview, classifies the error correctly in only 66.65\% of cases, underscoring the challenge of identifying erroneous reasoning. Furthermore, the ability to accurately identify errors offers valuable insights into the capabilities of multi-modal models. Project Page: https://mmerror-benchmark.github.io

cs.CV↗

Probing power spectrum enhancement at small scales with the SKA

The reionization process is driven by ionizing photons from dwarf galaxies in halos with virial temperature $T_{\rm vir} \gtrsim 10^4$ K, while minihalos whose $T_{\rm vir}\lesssim 10^4$ K consume ionizing photons and have negative contributions to reionization. Since ionizing sources and minihalos have different clustering characteristics, not only the reionization history, but also the morphology of the ionization field, is sensitive to the small-scale power spectrum. If the power spectrum at small scales is enhanced compared with the standard six-parameter $Λ$CDM model, then both the sources and sinks of ionizing photons would be boosted and the net impact depends on the competition between them. Therefore, the 21 cm signal that can probe the morphology of the ionization field will be a useful tool for detecting the small-scale power spectrum. Using the power spectrum proposed by Cielo et al. (2025) (C25) as a demonstration, we investigate the influence of small-scale power spectrum enhancement on the ionization field and the 21 cm signal. We find that for the C25 model, even under the constraints of observed UV luminosity functions for high-$z$ galaxies and reionization history, the 21~cm power spectrum and the bubble size distribution could be still significantly different from the regular $Λ$CDM model. The upcoming SKA-low AA* telescope, and a further imaging telescope, have the potential to detect the small-scale power spectrum more deeply.

astro-ph.CO↗

A Survey of Model Architectures in Information Retrieval

The period from 2019 to the present marks one of the most significant paradigm shifts in information retrieval (IR) and natural language processing (NLP), culminating in the emergence of powerful large language models (LLMs) from 2022 onward. Methods based on pretrained encoder-only architectures (e.g., BERT) as well as decoder-only generative LLMs have outperformed many earlier approaches, demonstrating particularly strong performance in zero-shot scenarios and complex reasoning tasks. This survey examines the evolution of model architectures in IR, with a focus on two key aspects: backbone models for feature extraction and end-to-end system architectures for relevance estimation. To maintain analytical clarity, we deliberately separate architectural design from training methodologies, enabling a focused examination of structural innovations in IR systems. We trace the progression from traditional term-based retrieval models to modern neural approaches, highlighting the transformative impact of transformer-based architectures and subsequent LLM developments. The survey concludes with a forward-looking discussion of open challenges and emerging research directions, including architectural optimization for efficiency and scalability, robust handling of multimodal and multilingual data, and adaptation to novel application domains such as autonomous search agents, which may represent the next paradigm in IR.

cs.IR↗