arXiv subjects
Search indexed arXiv papers on artificial intelligence, large language models, computer vision and robotics. Read source abstracts and follow links to arXiv.
In this paper, we propose a geometric proof of the generalized mirror transformation for multi-point virtual structure constants of degree k hypersurfaces in CP^{N-1}.
Policy gradient methods are among the most successful approaches for solving challenging reinforcement learning problems. Despite their empirical success, many state-of-the-art policy gradient algorithms for discounted problems estimate update directions using a sampling distribution that differs from the discounted state distribution required by the policy gradient theorem. In this work, we analyze the impact of this distribution mismatch on policy gradient methods. Specifically, {for tabular direct and softmax parameterizations, we establish global convergence of objective values under biased updates}. {For tabular softmax policies, we further derive the exact leading-order coefficients of the asymptotic objective errors under biased and normalized exact updates. Their ratio provides an explicit criterion for determining when distribution mismatch accelerates or slows asymptotic convergence.} We then extend the analysis to more general policy parameterizations by deriving explicit bounds on both the state distribution mismatch and the induced normalized policy-gradient mismatch in episodic and continuing MDPs. These mismatches vanish at least linearly as the discount factor approaches one. Under an additional smoothness assumption, we further show that the limit inferior of the squared norm of the normalized exact policy-gradient direction along the biased iterates is bounded by $O((1-γ)^2)$ in both settings. Our results provide theoretical insights into the robustness of policy gradient methods to distribution mismatch and shed light on the gap between their theoretical formulations and practical implementations.
Adverse drug reactions linked to the intake of drug combinations are a critical concern in pharmacovigilance, particularly as the controlled environment of clinical trials often lacks the scale and diversity to detect rare events involving multiple medications. While spontaneous reporting systems provide the necessary breadth for post-market surveillance, identifying overrepresented drug cocktails within such large-scale data remains a significant computational challenge. This study introduces a computational framework for the detection of drug cocktails associated with adverse events, leveraging disproportionality analysis on individual case safety reports. By integrating the Anatomical Therapeutic Chemical classification, the framework extends beyond individual drugs to capture hierarchical pharmacological relationships, enabling exploration of the space of drug combinations beyond pairwise analysis. To address biases inherent in existing disproportionality measures, we employ a hypergeometric risk metric, while a Markov Chain Monte Carlo algorithm provides robust empirical p-value estimation for the risk associated with cocktails. A genetic algorithm further facilitates efficient identification of high-risk drug cocktails. A post-treatment step based on penalized logistic regression allows distinguishing true pharmacological interactions from combined individual effects for cocktails of any size. Validation on synthetic and FDA Adverse Event Reporting System data demonstrates the method's efficacy in detecting established drugs and drug combinations associated with myopathy-related adverse events. Implemented as an R package, this framework offers a reproducible, scalable tool for post-market drug safety surveillance.
Democratic politics requires voters to infer how political actions map into outcomes, yet these causal links are often unclear, giving politicians scope to shape voters' causal accounts. This paper studies the supply and returns of causal rhetoric in politics: politicians' language linking political actions to outcomes through blame and credit. We fine-tune a language model to measure its use in 4.2 million congressional tweets spanning 2012--2023. Causal rhetoric is central to politicians' communication: it accounts for more than one third of tweets, is the language of political disagreement, and rises over time. The political returns of causal rhetoric are concentrated in blame. Combining campaign contributions with quasi-exogenous geographic variation in Twitter adoption, we find that exposure to blame has the largest effects on small donations. Using voting intentions from survey data and an event-study design around blame-only and credit-only days, we find that politicians' support rises after blame-only days. Finally, causal rhetoric appears rarely supported by economic knowledge, particularly blame.
Large Language Models (LLMs) can exhibit factual inaccuracies during long-chain reasoning or under misleading prompts, limiting their reliability in historical analysis. We propose Kongzi, a language model for historical understanding and reasoning that combines continued pre-training on historical corpora, supervised fine-tuning, and fact-enhanced reinforcement learning. Its factuality reward compares entities and relations extracted from generated responses with historical references through knowledge-graph similarity. We evaluate 0.5B and 7B models on historical question answering and counterfactual reasoning, using reported ratings from three LLM judges, and on CMMLU-Chinese-History. Kongzi-7B receives the highest historical-accuracy ratings among the compared models under all three judges. Reinforcement learning improves the mean answer scores of both model sizes and their CMMLU-Chinese-History accuracy, although Kongzi-7B remains below the strongest reported baselines on that benchmark. These descriptive results support further study of factuality rewards for historical language models.
Propensity scores near zero or one can make standard confidence intervals undercover. Econometric theory has targeted the Average Treatment Effect (ATE) under such weak overlap using nonstandard estimators or intervals. Practice has instead favored fixed thresholding rules that restore asymptotic normality by discarding extreme-propensity units for whom treatment may matter most. I show that standard t-tests for the ATE remain valid even when overlap is so weak that every regular estimator has infinite asymptotic variance. The key is to threshold the usual doubly robust estimator aggressively enough to control bias among unthresholded observations, but mildly enough to control bias from thresholding. I provide sufficient conditions for such thresholds to exist, and propose an adaptive rule of thumb to achieve coverage without knowing the weak overlap degree. In a canonical application, fixed trimming excludes 18% of the sample; the adaptive procedure estimates the resulting bias at five percent of the effect.
We introduce a new multiplication for the polytope algebra, defined via the intersection of polytopes. After establishing the foundational properties of this intersection product, we investigate finite-dimensional subalgebras that arise naturally from this construction. These subalgebras can be regarded as volumetric analogues of the graded Möbius algebra, which appears in the context of the Dowling--Wilson conjecture. We conjecture that they also satisfy the injective hard Lefschetz property and the Hodge--Riemann relations, and we prove these in degree one.
Knowledge distillation is widely used to transfer behavior from a large teacher model to a smaller student. It is often assumed to be safe when the teacher is clean, because classic backdoor attacks rely on poisoned labels and triggers in supervised training, whereas distillation trains the student to match a teacher's outputs. We show that this assumption can fail when the distillation dataset itself is poisoned. Our attack injects triggered and manipulated images that a clean teacher already predicts as an attacker chosen target label, which causes the student to learn a backdoor even though the teacher remains unaffected. We evaluate this threat across multiple manipulation strategies, including targeted adversarial perturbations and targeted GAN based class transitions, and study how distillation settings influence both accuracy and attack success. We show that the attack remains effective even at a 10% poisoning rates. The results demonstrate that a clean teacher alone is not a sufficient safeguard: poisoned distillation data can produce a strongly backdoored student while maintaining competitive performance on clean images. These findings show that the integrity and provenance of distillation data are part of the security boundary of data intensive KD pipelines, even when the teacher itself is trusted.
While machine learning fairness has made significant progress in recent years, most existing solutions focus on tabular data and are poorly suited for vision-based classification tasks, which rely heavily on deep learning. To bridge this gap, we introduce DetoxAI, an open-source Python library for improving fairness in deep learning vision classifiers through post-hoc debiasing. DetoxAI implements state-of-the-art debiasing algorithms, fairness metrics, and visualization tools. It supports debiasing via interventions in internal representations and includes attribution-based visualization tools and quantitative algorithmic fairness metrics to show how bias is mitigated. This paper presents the motivation, design, and use cases of DetoxAI, demonstrating its tangible value to engineers and researchers.
Although the $λ$I-calculus is a natural fragment of the $λ$-calculus, obtained by forbidding the erasure of arguments, its equational theories did not receive much attention. The reason is that all proper denotational models studied in the literature equate all non-normalizable $λ$I-terms, whence the associated theory is not very informative. The goal of this paper is to introduce a previously unknown theory of the $λ$I-calculus, induced by a notion of evaluation trees that we call "Ohana trees". The Ohana tree of a $λ$I-term is an annotated version of its Böhm tree, remembering all free variables that are hidden within its meaningless subtrees, or pushed into infinity along its infinite branches. We develop the associated theories of program approximation: the first approach -- more classic -- is based on finite trees and continuity, the second adapts Ehrhard and Regnier's Taylor expansion. We then prove a Commutation Theorem stating that the normal form of the Taylor expansion of a $λ$I-term coincides with the Taylor expansion of its Ohana tree. As a corollary, we obtain that the equality induced by Ohana trees is compatible with abstraction and application. Subsequently, we introduce a denotational model designed to capture the equality induced by Ohana trees. Although presented as a non-idempotent type system, our model is based on a suitably modified version of the relational semantics of the $λ$-calculus, which is known to yield proper models of the $λ$I-calculus when restricted to non-empty finite multisets. To track variables occurring in subterms that are hidden or pushed to infinity in the evaluation trees, we generalize the system in two ways: first, we reintroduce annotated versions of the empty multiset indexed by sets of variables; second, (...)
We study heat conduction in a one-dimensional {finite}, unpinned chain of atoms perturbed by stochastic momentum exchange and coupled to Langevin heat baths at {possibly} distinct temperatures placed at the endpoints of the chain. While infinite systems without boundaries are known to exhibit superdiffusive energy transport described by a fractional heat equation with the generator $-|Δ|^{3/4}$, the corresponding boundary conditions induced by heat baths remain less understood. We establish the hydrodynamic limit for a finite chain with $n+1$ atoms connected to thermostats at the endpoints, deriving the macroscopic evolution of the averaged energy profile. The limiting equation is governed by a non-local Lévy-type operator, with boundary terms determined by explicit interaction kernels that encode absorption, reflection, and transmission of long-wavelength phonons at the baths. Our results provide the first rigorous identification of boundary conditions for fractional superdiffusion arising directly from microscopic dynamics with local interactions, highlighting their distinction from both diffusive and pinned-chain settings
Colored lattice models can be used to describe many different types of special functions of interest in both algebraic combinatorics and representation theory, for example Schur polynomials, nonsymmetric Macdonald polynomials, and characters and Whittaker functions for representations of p-adic groups. A notable example is the metaplectic ice model of which there are actually two different variants: a Gamma and a Delta variant. These variants differ in key aspects but surprisingly produce equal partition functions, which are weighted sums over admissible configurations, and this equality is called the Gamma-Delta duality. The duality was used to prove the analytic continuation of certain multiple Dirichlet series and is highly non-trivial, especially since the number of configurations on each side of the equality can differ. In this paper we generalize the Gamma-Delta duality by constructing a new family of solvable, colored lattice models and prove that they are dual to existing lattice models in the literature. Besides the above metaplectic case we also get a new lattice model and Gamma-Delta type duality for (non-metaplectic) Iwahori Whittaker functions together with a crystal limit for Demazure atoms for Cartan type A. The equality of partition functions is shown using Yang-Baxter equations involving both variants of lattice models simultaneously. For the crystal lattice model we show that the duality refines to a weight-respecting bijection of states given by the Schützenberger involution on the associated Gelfand-Tsetlin patterns or semistandard Young tableaux. We also show how the individual steps exchanging two rows in the proof of the duality for the partition functions refines to Berenstein-Kirillov or, equivalently, Bender-Knuth involutions.
Using data from the Galaxy and Mass Assembly (GAMA) survey, we investigate how galaxy properties correlate with the local environment, focusing on the two southern regions of the survey (G02 and G23) that have not previously been examined in this context. We employ two-point and marked correlation functions to quantify the environmental dependence of galaxy color, stellar mass, luminosity across the $u$, $g$, $r$, $J$, and $K$ bands, as well as star formation rate (SFR) and specific star formation rate (sSFR). We also assess the impact of redshift incompleteness and cosmic variance on these clustering measurements. Our results show that $u-r$ and $g-r$ colors are most strongly correlated with local overdensity, followed by stellar mass. The sSFR exhibits a clear inverse relationship with density of the environment, consistent with the trend observed for $u$-band luminosity, which traces young stellar populations. In contrast, galaxies brighter in the $g$, $J$, and $K$ bands preferentially inhabit denser regions. By comparing our measurements from the southern regions with those from the equatorial regions of GAMA, we find that cosmic variance does not significantly influence our conclusions. However, redshift incompleteness affects the clustering measurements, as revealed through comparisons of subsets within the G02 region. The measured correlations provide key constraints for models of galaxy assembly across mass and environment, while the environmental trends in color and near-infrared luminosity offer a means to trace stellar mass growth and quenching with redshift.
The Chowla--Milnor conjecture predicts the linear independence of certain Hurwitz zeta values. In this paper, we prove that for any fixed integer $k \geqslant 2$, the dimension of the $\mathbb{Q}$-linear span of $ζ(k,a/q)-(-1)^{k}ζ(k,1-a/q)$ ($1 \leqslant a < q/2$, $\gcd(a,q)=1$) is at least $(c -o(1)) \cdot \log q$ as the positive integer $q \to +\infty$ for some absolute constant $c>0$. It is well known that $ζ(k,a/q)+(-1)^{k}ζ(k,1-a/q) \in \overline{\mathbb{Q}}π^k$, but much less is known previously for $ζ(k,a/q)-(-1)^{k}ζ(k,1-a/q)$. Our proof is similar to those of Ball--Rivoal (2001) and Zudilin (2002) concerning the linear independence of Riemann zeta values. However, we use a new type of rational functions to construct linear forms.
We consider smooth flows preserving a smooth invariant measure, or, equivalently, locally Hamiltonian flows on compact orientable surfaces, and show that almost every such locally Hamiltonian flow with only simple saddles has singular spectrum. Furthermore, we prove that for almost every pair of such flows, the elements of the pair are spectrally disjoint. More generally, the results from which these statements are deduced are singularity of the spectrum and pairwise spectral disjointness for special flows over full measure sets of interval exchange transformations under a roof with symmetric logarithmic singularities. Spectral singularity is proved using a criterion based on tightness of Birkhoff sums with exponential tail decay. The assumptions of the criterion are verified exploiting the cancellations proved by the last author to prove the absence of mixing in this class of flows, by showing that the latter can be combined with rigidity by exploiting the local product structure of Rauzy-Veech induction. Pairwise spectral disjointness then follows by producing mixing times (for the second flow), using a new mechanism for shearing based on what we call resonant rigidity times.
We introduce ShIOEnv, a Gymnasium-compatible Bash environment for synthesizing execution-annotated shell input traces. Datasets containing such traces are useful for training models to perform execution-free shell simulation, which in turn is useful for deploying honeypots for cyber deception. Existing datasets, such as NL2CMD and other Bash command corpora, do not include complete execution artifacts (stdout, stderr, exit status, and changes to the system's state) associated with a command's execution. ShIOEnv synthesizes shell inputs, executes them in a controlled system context, and records stdout, stderr, exit status, and structured filesystem and environment patches. Unconstrained (random) generation of shell inputs (command and arguments) results in multiple instances of arguments that are mostly redundant or result in execution errors. We introduce grammar-constrained synthesis to generate syntactically valid arguments with high input irreducibility, which we estimate through self-ablative contrasts of execution artifacts. We prepare 1.6M samples using grammar-constrained and unconstrained synthesis and collect their execution artifacts for 86 Linux/Bash utilities. Using this dataset, we train small language models (CodeT5) and evaluate their accuracy in predicting the actual execution behavior of attacker-sourced Bash commands. These models outperform commonly used rule-based simulation (Cowrie) and prompted large language models by up to 14.2 pp in output edit similarity (O-ES). We enhance the NL2CMD dataset by constructing 500K samples executable in a grounded system context along with their complete execution artifacts. We empirically demonstrate that models trained on grammar-constrained samples outperform those trained on NL2CMD samples by 22.5 pp in O-ES. ShIOEnv, command grammars, and 2.1M samples used in our primary evaluations are released for use by other researchers.
Large language model (LLM) agents have shown impressive reasoning capabilities in interactive decision-making tasks. These agents interact with environment through intermediate interfaces, such as predefined action spaces and interaction rules, which mediate the perception and action. However, mismatches often happen between the internal expectations of the agent regarding the influence of its issued actions and the actual state transitions in the environment, a phenomenon referred to as agent-environment misalignment. While prior work has invested substantially in improving agent strategies and environment design, the critical role of the interface still remains underexplored. In this work, we empirically demonstrate that agent-environment misalignment poses a significant bottleneck to agent performance. To mitigate this issue, we propose ALIGN, an Auto-Aligned Interface Generation framework that alleviates the misalignment by enriching the interface. Specifically, the ALIGN-generated interface enhances both the static information of the environment and the step-wise observations returned to the agent. Implemented as a lightweight wrapper, this interface achieves the alignment without modifying either the agent logic or the environment code. Experiments across multiple domains including embodied tasks, web navigation and tool-use, show consistent performance improvements, with up to a 45.67% success rate improvement observed in ALFWorld. Meanwhile, ALIGN-generated interface can generalize across different agent architectures and LLM backbones without interface regeneration. Code and experimental results are available at https://github.com/THUNLP-MT/ALIGN.
The contemporary paradigm of scaling data annotation, crucial for developing machine learning solutions, is to hire ordinary human annotators and instruct them with expert-crafted guidelines to label data. This paradigm is laborious, tedious, and costly, motivating us to study an open problem, auto-annotation with expert-crafted guidelines (dubbed AutoExpert). We develop benchmarks by redesigning the evaluation protocol and re-annotating data with nuScenes and PandaSet, two 3D detection datasets for autonomous driving research that provide expert-crafted annotation guidelines. Their guidelines define 18 and 25 object classes, respectively, using nuanced language descriptions and a few visual examples. Following the guidelines that require using 3D cuboids to label LiDAR data, AutoExpert requires algorithms to learn on few-shot labeled images and texts to perform the task of 3D detection on LiDAR data. Apparently, the challenges of AutoExpert lie in the data-modality and task discrepancy. Nevertheless, public foundation models (FMs) serve as promising tools to tackle these challenges. To address AutoExpert, we adopt a conceptually simple pipeline consisting of three components: (1) 2D object detection and segmentation in RGB images, (2) lifting 2D detections into 3D using known sensor poses, and (3) 3D cuboids generation for the 2D detections. Within this pipeline, we enhance and evaluate a variety of methods such as open-vocabulary detectors, few-shot detectors, and self-supervised learned detectors. We also develop novel techniques, leading to refined components that boost 3D detection mAP from 12.1 to 25.4 on the AutoExpert-nuScenes benchmark.