Searcharxiv⌕ Search

arXiv subjects

Yang Bai

Publications and source records attributed to Yang Bai.

At least 19 recordsLinked to original sources

Can Vision-Language Models Truly Understand Images? A Benchmark Beyond Commonsense Bias

Multimodal large language models (MLLMs) have achieved impressive performance on a wide range of vision-language tasks. However, it remains unclear whether their predictions are grounded in the visual content itself or are largely driven by statistical co-occurrence patterns and commonsense correlations learned from large-scale image-text corpora. Existing vision-language benchmarks predominantly consist of samples that conform to human commonsense, allowing models to achieve strong performance by exploiting dataset priors rather than genuinely understanding image content. To address this limitation, we present ContraSense, a systematically constructed benchmark for evaluating faithful vision-language understanding. ContraSense consists of carefully curated counter-intuitive image-text pairs that deliberately violate common expectations, substantially reducing the effectiveness of statistical biases and requiring models to rely on visual evidence for reasoning. We conduct comprehensive evaluations on a diverse set of representative open- and closed-source MLLMs. The results reveal a substantial performance gap between current MLLMs and humans: even the strongest closed-source model achieves only about 60\% accuracy, compared with 96.04\% for human participants. These findings demonstrate that current MLLMs still struggle with faithful visual understanding and highlight the need for more robust multimodal reasoning grounded in image content. The ContraSense benchmark is publicly available at https://huggingface.co/datasets/LingLin00/ContraSense.

cs.CV↗

SEAM: Computer-Use Agents Know the Harm Between Screens, and Commit It Anyway

Safety benchmarks for computer-use agents (CUAs) place the harm either in the intent of the task or on a single screen: a malicious instruction, an injected command, a phishing page. An agent that checks the task and each screen in isolation can pass them, so they do not test whether it combines facts seen on different screens. Yet in delegated work, harm often arises exactly this way: a fact about a person on one screen, a fact about an item on another, and a legitimate task that binds them. We introduce SEAM (Safety Execution Across Multiple screens), a benchmark of such cross-screen compositional harm: 600 multi-step, cross-application OSWorld tasks in which the instruction and every single observation are benign and only their combination is harmful. Minimal-pair variants that differ in one place, scored from the server log rather than by an LLM judge, locate where an agent fails. General-purpose open-weight and flagship agents carry out the harm in up to 92% of runs, yet decline when the user states it in one sentence. Agents open the decisive pages, so exposure is not the primary bottleneck. Asked about those pages, they flag the harm, and with thinking they often write the conflict into their reasoning and confirm anyway. Safety recognition thus does not reliably translate into safe execution, and the translation depends on the information channel: the same warning stops agents far more often in the user's instruction than on screen. Proprietary models are markedly more cautious. We release SEAM with its generator and quality gates at https://yuxuan2003.github.io/SEAM/ .

cs.CR↗

SPARC: Evidence-guided adaptive model composition for materials discovery

Closed-loop experimentation requires more than selecting the next experiment. It must also determine which existing measurements and model-derived relationships are sufficiently supported to influence that decision. Here we introduce SPARC (Sparse Predictive Aggregation of Regression Components), a framework that continually revises this support as experimental evidence accumulates. Competing model families are evaluated through out-of-fold predictions, so that each model is tested on observations not used in its corresponding fit. Sparse posterior composition assigns model influence according to predictive contributions supported by the out-of-fold evidence, retaining complementary contributions from several model families. Candidate-dependent predictive dispersion and residual variability define trust, which modulates exploration alongside the predicted response. Independent diagnostics also prioritize existing measurements for prospective verification, with model-derived trust reported alongside them. Across eight published materials datasets, SPARC achieved higher $R^2$ than the corresponding literature-reported values in six, with relative gains of up to 77% without the density-functional-theory- or experiment-derived descriptors used by several published pipelines. In an automated alkaline oxygen-evolution-reaction campaign, 8 of 11 SPARC-selected experiments reached overpotentials below 300 mV, with a best value of 242 mV, compared with 18 of 125 initial experiments. Measurements prioritized for verification before repetition subsequently deviated by 76 mV on average, compared with 27 mV for matched controls. SPARC therefore turns experimental evidence into a basis for revising model influence, directing exploration and identifying measurements that warrant verification.

cond-mat.mtrl-sci↗

Covariance-Calibrated Prediction-Powered Inference

In prediction-powered inference, predictions supplement labeled outcomes to improve estimation. We combine a prediction-corrected score and an auxiliary moment in a generalized method of moments (GMM) criterion. The resulting estimator is first-order equivalent to the corresponding corrected-score root under regularity conditions, but the auxiliary penalty can change the finite-sample fit. In a multiparameter Gaussian linear model with low noise, retaining this penalty reduces risk relative to bounded scalar corrected-score roots. We then introduce covariance-calibrated prediction-powered inference (CC-PPI), which uses prediction scores from both samples to reweight the penalty. With a spectrally protected half-strength update, CC-PPI retains regular first-order efficiency and strictly reduces risk relative to uncalibrated full GMM in a Gaussian linear low-noise regime that tolerates some prediction error and in a structured multiparameter logistic model conditional on the existence of a finite pilot estimate. Targeted simulations isolate the gains from calibration and show that they can reverse as prediction error grows. Broader simulations and four applications evaluate point estimation and the coverage and width of full-refit bootstrap intervals for linear projections.

stat.ME↗

Optimal chemotactic navigation in disordered landscapes

Active navigation in disordered landscape with random traps arises from the interplay between directed motion and environmental constraints. Using E. coli chemotactic navigation in agar gels as a model system, we reveal a fundamental trade-off between diffusive exploration and chemotactic directional bias that dictates the optimal strategy for population range expansion. Counter-intuitively, evolution selects for shorter mean run times (τ_f) to achieve faster chemotactic migration in denser environments. Controlled experiments reveal a non-monotonic relationship between chemotactic navigation speed and τ_f, with the optimum shifting according to the density of physical traps in the gel. Single-cell analysis indicates that cells primarily escape a trap without an intervening tumble and without a large reorientation. Based on these observations, we develop a minimal theoretical model showing that the optimal τ_f emerges from an antagonistic scaling: while the diffusion coefficient increases with τ_f, the chemotactic bias coefficient decreases with it. This work establishes an optimal condition for biased run-tumble motion in disordered landscapes with random traps.

cond-mat.soft↗

Apparent magnetic field decay in a solar flare from Zeeman splitting of the Mg 12.32 microns observations

The magnetic field is the most fundamental observable in solar physics, while solar flares are the most violent, rapid magnetic energy-release events in the solar atmosphere. Decades of observations and theory have sketched the framework of flare evolution and magnetic-field-related physical process, yet the actual magnetic field variations during flares remained poorly understood. The main obstacle is that magnetic-diagnostic spectral lines are distorted by transient flare activity, making traditional spectro-polarimetry inversions complicated. Using a mid-infrared solar telescope, we investigate spectroscopic evolution of the Mg 12.32 microns line during an X1.5-class flare. Owing to the exceptionally large Zeeman sensitivity of this line, magnetic field strength can be directly determined from the measured line separation. Our analysis reveals apparent magnetic field weakening beneath flare ribbons in the impulsive phase. Radiative-hydrodynamic simulations and atmospheric inversions confirm that the formation height of the Mg 12.32 microns line increases during the flare. We suggest this apparent field change probably originates either from elevated line-formation altitude or magnetic-field redistribution. Additional observations and modeling are needed to clarify these two contributions. This study demonstrates the capability of mid-infrared wavelengths to trace magnetic-field evolution during flares.

astro-ph.SR↗

On The Rational Realization of Even-dimensional Spheres and Products of Eilenberg--MacLane Spaces as Classifying Spaces

In this paper, we study the rational realization problem for the classifying space $B\mathrm{aut}_1(X)$. We prove that if $S^{2n}$ is realized as $B\mathrm{aut}_1(X)_{\mathbb{Q}}$ for a simply-connected space $X$, then $X$ is $π$-infinite and has vanishing rational Gottlieb elements above degree $2n-1$. In particular, even-dimensional spheres cannot be realized as $B\mathrm{aut}_1(X)_{\mathbb{Q}}$ for any simply-connected $π$-finite space $X$. We also prove that, for all $n\geq 2$ and $s,t\geq 1$, the product of Eilenberg--MacLane spaces $K(\mathbb{Q}^s,n)\times K(\mathbb{Q}^t,n+1)$ cannot be realized as $B\mathrm{aut}_1(X)_{\mathbb{Q}}$ for any simply-connected $π$-finite space $X$. Moreover, we prove that if $r\geq 2$, $n\geq 3$ and $K(\mathbb{Q}^r,n)$ is realized as $B\mathrm{aut}_1(X)_{\mathbb{Q}}$ for a simply-connected $π$-finite space $X$, then $X$ has the rational homotopy type of $K(\mathbb{Q}^r,n-1)$. The proofs are based on two structural results for Gottlieb elements in the derivation Lie algebra of a Sullivan minimal model, which provide a uniform method for these realization problems.

math.AT↗

A complete classification of the rational realization of gap-two products of Eilenberg--MacLane spaces as classifying spaces

In this paper, we study the rational realization of the gap-two product $K(\mathbb{Q},n)\times K(\mathbb{Q},n+2)$ as the classifying space $B\mathrm{aut}_1(X)$ for a simply-connected $π$-finite space $X$ with rational homology of finite type. We prove that $K(\mathbb{Q},n)\times K(\mathbb{Q},n+2)$ can be realized as $B\mathrm{aut}_1(X)$ up to rational homotopy equivalence if and only if $n=3,4$ or $4m+2$ for $m\geq 1$. Moreover, we determine all their minimal Sullivan models up to isomorphism. The rational homotopy type of such $X$ is unique for $n=3$ or $4m+2$ for $m\geq 1$; there exist infinitely many rational homotopy types of such $X$ for $n=4$.

math.AT↗

Blueprint for an AI-Enabled Particle Theory Community

Rapid advances in the capabilities and reach of artificial intelligence (AI) are changing research in particle physics. In this new era, AI is an innovative tool for theoretical physicists, while the interaction of AI and particle theory is itself a developing field. We argue that particle theory requires a community-scale effort to realize its full potential. We show how particle theory occupies a unique position in the AI for physics landscape, and describe the complementarity of AI and theory developments. We conclude with a series of recommendations on how to build an AI-enabled theory community for the Genesis Mission.

hep-ph↗

Attention Sees, MLP Refuses: Tracing the Attention-to-MLP Safety Pathway in LLMs

Safety-aligned large language models (LLMs) remain vulnerable to jailbreaks and can over-refuse benign requests. Understanding these failures requires distinguishing harmfulness recognition from the computation that writes refusal. We reframe this as a pathway question and make it measurable on residual-stream updates in decoder-only Transformers, separating the harmfulness information a module update exposes from the change it produces in readable refusal evidence. Across 18 Base-Instruct pairs spanning Qwen, Llama, and Gemma, harmfulness is decoded more readily from attention updates, particularly in early layers, whereas MLP updates contribute more strongly to harm-selective refusal writing in aligned models. Post-training strengthens MLP refusal writing in all 18 pairs without systematically improving harmfulness decodability, while positive MLP writing becomes more concentrated in deeper layers after post-training. Across six reasoning-enabled models, harmfulness remains decodable during reasoning while strong refusal writing concentrates at answer onset, a pattern that persists with reasoning disabled. Causal interventions further support this pathway: resetting downstream MLP updates to their clean values substantially reduces the refusal shift caused by upstream attention interventions. Building on these findings, we introduce Sentinel, which combines an attention-derived perception score P with an MLP-derived decision score D. Their joint profiles characterize jailbreaks and over-refusals, and conditional routing with both signals achieves a lower aggregate error rate than single-signal routing and evaluated fixed reminders across eight models.

cs.AI↗

Realization of permutation modules as homology of CW-complexes

In this paper, we investigate the realizability problem, which asks whether prescribed group actions on graded modules can be realized by the groups of self-homotopy equivalences of CW-complexes acting on their homology. We provide a partial answer in the case of arbitrary groups and permutation modules concentrated in certain degrees. As a consequence, we realize every group as the group of self-homotopy equivalences of an $R$-local CW-complex with arbitrary prescribed connectivity, where $R$ may be chosen with $ρ(R)$ sufficiently large. In particular, this provides a complete answer to Kahn's realizability problem.

math.AT↗

LexReward: A Taxonomy-Driven Reward Framework for Legal Language Models

Legal language models require reward signals that capture not only answer correctness but also the multidimensional quality of legal responses. Existing reward methods, however, often rely on coarse-grained holistic judgments, providing limited domain specificity and interpretability. We introduce LexReward, a taxonomy-driven framework for legal reward modeling. LexReward characterizes legal response quality along three complementary dimensions: Style, covering lexical and syntactic quality; Element, assessing legal subjects, facts, statutes, and decisions; and Chain, evaluating the order, completeness, correctness, and non-redundancy of legal reasoning. For each dimension, we develop rubrics that specify evaluation criteria and quality levels. The resulting rewards are used to construct pairwise preference data for Direct Preference Optimization (DPO) and reward-model training. Experiments show that the rubric-based rewards reliably distinguish legal responses of different quality and that DPO training on the preference data improves performance across all three dimensions. The learned reward models, LexRM, also support effective downstream optimization: each dimension-specific reward model improves policy performance in its corresponding dimension through reinforcement learning, without requiring reference answers at reward time. Dimension-wise analyses further support the effectiveness of the proposed taxonomy and reward construction.

cs.CL↗

The Great Wall of the Standard Model

The global form of the Standard Model (SM) gauge group, $[SU(3)_c\times SU(2)_L\times U(1)_Y]/Γ$, with $Γ\subseteq\mathbb{Z}_6$, remains undetermined. For $Γ=\mathbb{Z}_p$, the SM possesses a residual global electric one-form symmetry $\mathbb{Z}_{6/p}$. Compactification on the thermal circle yields an ordinary discrete zero-form symmetry whose spontaneous breaking can produce cosmological domain walls. We calculate the temperature-dependent tension and width of the $\mathbb{Z}_{6/p}$ domain walls below the electroweak scale, which exist for every quotient except $Γ=\mathbb{Z}_6$, finding that their thickness can reach the kilometer scale around reionization in our benchmarks. These macroscopic thermal interfaces, if produced, would constitute the "Great Walls" of the Standard Model. We develop a phenomenological thermal velocity-dependent one-scale model to follow their network evolution, incorporating temperature-dependent surface tension, inertia, and plasma drag. For sufficiently early formation, our benchmarks give an estimated wall energy fraction of order $10^{-14}$ near the onset of Big Bang nucleosynthesis, with further suppression at later times. Uncertainties arise from relating surface free energy to physical wall energy and from late-time plasma evolution and structure formation. Despite their small energy fraction, these walls could probe the global structure of the SM gauge group, motivating dedicated searches for their cosmological signatures.

hep-ph↗

Machine Learning Classification and Portfolio Construction: Does the Loss Function Matter?

Classification outperforms regression across matched machine learning models in portfolio construction. A stacking ensemble of gradient boosted trees, random forest, and neural network yields a value-weighted annualized Sharpe ratio of 2.08 for classification and 1.39 for regression. This outperformance strengthens with class granularity and persists across subsamples and after transaction costs. Spanning tests show that classification retains economically large alphas after we control for regression, whereas regression alphas shrink substantially once we control for classification. These results indicate that classification extracts more return information than matched regression. Our diagnostics trace classification's advantage to more precise separation of return deciles.

q-fin.GN↗

Generative Retrieval for Unsupervised Text-Based Person Search

Text-based person search (TBPS) aims to retrieve images of a target person from a large image gallery based on a given natural language description. Most existing methods rely on supervised learning with manually annotated image-text pairs. In this paper, we explore unsupervised TBPS, with only unlabeled images. We propose GTR+, a two-stage generation-then-retrieval framework. In the generation stage, we introduce a tiered description generation framework designed to produce fine-grained and stylistically diverse textual descriptions through a three-tier sequential process. The base tier leverages an automated question-and-answer mechanism to generate basic visual attribute descriptions; the intermediate tier enhances fine-grained detail using an inter-sample contrastive mechanism; the advanced tier further enriches textual diversity via a stylized expansion mechanism. In the retrieval stage, to mitigate the impact of noisy pseudo texts, we develop an adaptive confidence-weighted retrieval learning framework. We model image-text pairs as clean or noisy using a Gaussian Mixture Model, calibrated by real-time image-text similarity and static text generation probability from the prior stage, yielding adaptive sample weights during training. Beyond that, we also contribute LargeFine-Person, a large-scale TBPS dataset with high-quality, fine-grained, and diverse textual annotations, enabling a practical and generalizable TBPS pre-training benchmark under unsupervised setting. Experiments on multiple TBPS benchmarks demonstrate the effectiveness and generalization of both GTR+ and LargeFine-Person. Code is available at: https://github.com/Flame-Chasers/GTR.

cs.CV↗

Baryoid Dark Matter from $\mathbb{Z}_N$ Domain Walls: The $(N-1):1$ origin of the dark matter-baryon coincidence

We propose an explanation for the dark matter-baryon coincidence based on collapsing $\mathbb{Z}_N$ domain walls, which form a novel compact baryonic state: the baryoid. A baryoid has an asteroid-scale mass and up-to-nuclear-scale energy density, and can serve as a dark matter candidate. Starting from equal baryon numbers in the domains formed in the early universe, the collapse of the domain walls after the QCD phase transition leads to a baryon-number ratio of $(N-1):1$ between the false- and true-vacuum domains. Since baryons are slightly lighter in the false-vacuum domains than in the true-vacuum domain, the resulting dark matter-to-baryon energy-density ratio is naturally close to, but slightly smaller than, $(N-1):1$, or $6:1$ for $N=7$. We calculate the domain-wall dynamics and the efficiency of baryon-number trapping, derive the resulting baryoid properties, and discuss a broad set of phenomenological probes.

hep-ph↗

Why Does Weak-OOD Help? A Further Step Towards Understanding Jailbreaking VLMs

Large Vision-Language Models (VLMs) are susceptible to jailbreak attacks: researchers have developed various attack strategies that bypass the safety mechanisms of VLMs. Among these approaches, jailbreak methods based on the Out-of-Distribution (OOD) strategy have garnered widespread attention due to their simplicity and effectiveness. This paper further advances the understanding of OOD-based VLM jailbreak methods. We show that mild OOD manipulations can achieve stronger jailbreak performance than both clean inputs and overly strong perturbations, a non-monotonic pattern we define as "weak-OOD". We explain this phenomenon through a trade-off between two dominant factors: input intent perception and model refusal triggering. Our evidence suggests that these two factors respond differently to OOD manipulations, which is consistent with a discrepancy between broad pre-training robustness and narrower safety alignment. Building on this insight, we draw inspiration from optical character recognition (OCR) capability enhancement---a core task in the pre-training phase of mainstream VLMs. Leveraging this capability, we design JOCR (Jailbreak via OCR-Aware Embedded Text Perturbation), a practical OCR-readable extension of embedded-text jailbreaks that achieves the best average ASR among evaluated baselines. Code is available at GitHub: https://github.com/Yuxuan2003/weak-ood-jailbreak.

cs.CR↗

Multi-Task Learning with Covariate-Overlap Regularization

Multi-task learning improves data efficiency by sharing information across related tasks, but indiscriminate sharing can be harmful when their covariate distributions and response relationships differ. We propose COVariate-ovERlap regularized multi-task learning (COVER) to address covariate and posterior heterogeneity. The model combines a common component function with a shared neural representation and low-dimensional task-specific coefficients. Taskwise second-moment matrices summarize covariate heterogeneity and determine the strength of coefficient integration in each representation direction. We derive a covariate-overlap penalty by minimizing the total squared change in two task predictors when their coefficients are replaced by one auxiliary coefficient. An equivalent auxiliary formulation supports end-to-end training without matrix inversion. An exact fixed-representation bias--variance decomposition quantifies how covariate overlap controls variance reduction and how posterior heterogeneity determines shrinkage bias. Global and localized end-to-end oracle inequalities account for jointly learning the neural functions and estimating the overlap matrices from the same observations. We give explicit neural-network rates and sharpen the stochastic prediction term when the regularized oracle risk and overlap-estimation error are small. Simulations across diverse heterogeneity settings show competitive performance against deep-learning and statistical data-integration methods, with the largest gains under joint heterogeneity. In a GTEx central-nervous-system analysis, COVER achieves the lowest response-averaged prediction error among the compared methods and reveals tissue-pair integration patterns.

stat.ML↗