arXiv subjects

Yizhou Lu

Publications and source records attributed to Yizhou Lu.

At least 19 recordsLinked to original sources

arXiv2026

Concept Modulation Models: A Unified Framework for Identifiability and Extrapolation

Reliable generalization in conditional latent variable models requires understanding both identifiability and extrapolation: how observed variation across attributes determines latent structure, and how that structure determines distributions at unseen attributes. However, existing identifiability and extrapolation guarantees are largely model-specific, with separate analyses in nonlinear ICA, causal representation learning, perturbation modeling, and related conditional latent variable models. We introduce concept modulation models (CMMs), an attribute-indexed class of conditional generative models with structure $A\to \Lambda \to C\to X$, where attributes select modulators, modulators induce latent concept laws, and concepts generate observed features. CMMs lift transition-based identifiability to conditional settings by showing that feature agreement on observed attributes induces a latent concept transition constrained by the CMM class. We express these constraints through attribute potentials, log-density ratios between attribute-conditioned concept laws, separating the generic lifting step from model-specific rigidity arguments. The same potentials control extrapolation: agreement at unseen attributes holds exactly when the transported attribute-potential identities extend to those attributes. This yields algebraic extrapolation criteria, identifies the common potential-based proof objects behind several existing identifiability and extrapolation results, and, when combined with the model-specific rigidity arguments in those works, recovers their stated conclusions.

cs.LG↗

arXiv2026

Odysseus: Scaling VLMs to 100+ Turn Decision-Making in Games via Reinforcement Learning

Given the rapidly growing capabilities of vision-language models (VLMs), extending them to interactive decision-making tasks such as video games has emerged as a promising frontier. However, existing approaches either rely on large-scale supervised fine-tuning (SFT) on human trajectories or apply reinforcement learning (RL) only in relatively short-horizon settings (typically around 20--30 turns). In this work, we study RL-based training of VLMs for long-horizon decision-making in Super Mario Land, a visually grounded environment requiring 100+ turns of interaction with coordinated perception, reasoning, and action. We begin with a systematic investigation of key algorithmic components and propose an adapted variant of PPO with a lightweight turn-level critic, which substantially improves training stability and sample efficiency over critic-free methods such as GRPO and Reinforce++. We further show that pretrained VLMs provide strong action priors, significantly improving sample efficiency during RL training and reducing the need for manual design choices such as action engineering, compared to classical deep RL trained from scratch. Building on these insights, we introduce Odysseus, an open training framework for VLM agents, achieving substantial gains across multiple levels of the game and at least 3 times average game progresses than frontier models. Moreover, the trained models exhibit consistent improvements under both in-game and cross-game generalization settings, while maintaining general-domain capabilities. Overall, our results identify key ingredients for making RL stable and effective in long-horizon, multi-modal settings, and provide practical guidance for developing VLMs as embodied agents.

cs.LG↗

arXiv2026

Questioning the Coverage-Length Metric in Conformal Prediction: When Shorter Intervals Are Not Better

Conformal prediction(CP) has become a cornerstone of distribution-free uncertainty quantification, conventionally evaluated by its coverage and interval length. This work critically examines the sufficiency of these standard metrics. We demonstrate that the interval length might be deceptively improved through a counter-intuitive approach termed Prejudicial Trick(PT), while the coverage remains valid. Specifically, for any given test sample, PT probabilistically returns an interval, which is either null or constructed using an adjusted confidence level, thereby preserving marginal coverage. While PT potentially yields a deceptively lower interval length, it introduces practical vulnerabilities: the same input can yield completely different prediction intervals across repeated runs of the algorithm. We formally derive the conditions under which PT achieves these misleading improvements and provide extensive empirical evidence across various regression and classification tasks. Furthermore, we introduce a new metric interval stability which helps detect whether a new CP method implicitly improves the length based on such PT-like techniques. Code is available at https://github.com/benben-cd/PT-Conformal-Prediction.

stat.ML↗

arXiv2025

Smoothing-Based Conformal Prediction for Balancing Efficiency and Interpretability

Conformal Prediction (CP) is a distribution-free framework for constructing statistically rigorous prediction sets. While popular variants such as CD-split improve CP's efficiency, they often yield prediction sets composed of multiple disconnected subintervals, which are difficult to interpret. In this paper, we propose SCD-split, which incorporates smoothing operations into the CP framework. Such smoothing operations potentially help merge the subintervals, thus leading to interpretable prediction sets. Experimental results on both synthetic and real-world datasets demonstrate that SCD-split balances the interval length and the number of disconnected subintervals. Theoretically, under specific conditions, SCD-split provably reduces the number of disconnected subintervals while maintaining comparable coverage guarantees and interval length compared with CD-split.

stat.ML↗

arXiv2025

Deep Feature-specific Imaging

Modern photon-counting sensors are increasingly dominated by Poisson noise, yet conventional feature-specific imaging (FSI), based on principal component analysis (PCA), is optimized for additive Gaussian noise and variance preservation rather than task-specific objectives, leading to suboptimal performance and a loss of its advantages under Poisson noise. To address this, we introduce DeepFSI, what we believe to be a novel end-to-end optical-electronic framework. DeepFSI "unfreezes" PCA-derived masks, enabling a deep neural network to learn globally optimal measurement masks by computing gradients directly under realistic Poisson and additive noise conditions. Simulations and hardware experiments demonstrate that DeepFSI achieves improved classification accuracy and stronger transfer robustness compared to PCAbased FSI across varying photon budgets, particularly in Poisson-noise-dominant environments. DeepFSI also exhibits enhanced robustness to design choices and performs well under additive Gaussian noise, representing a significant advance for noise-robust computational imaging in photon-limited applications.

physics.optics↗

arXiv2025

Loops in inflation with strongly non-geodesic motion

We study loop corrections in the effective field theory of inflation with imaginary speed of sound, which has been shown to provide an effective description of multi-field inflationary models characterized by strongly non-geodesic motion and heavy entropic perturbations. We focus on the one-loop corrections to the scalar and tensor power spectra, taking into account all relevant vertices at leading order in derivatives and in slow-roll. We find a power-law dependence of the scalar two-point function on the scale that defines the range of validity of the effective theory, analogous to the enhancement observed in tree-level correlation functions. Even more dramatic, the relative correction to the tensor spectrum is exponentially enhanced, albeit also suppressed in the slow-roll limit. In spite of these large effects, our results show that this class of models can satisfy the requirement of perturbative control and a consistent loop expansion within a range of parameters of phenomenological interest. On the other hand, models predicting large values of the power spectrum on small scales are found to be under strong tension. As a technical bonus, we carefully explain the prescription for the regularization and manipulation of loop integrals in this set-up, where one has a non-trivial domain of integration for time and momentum integrals owing to the regime of validity of the effective field theory. This procedure is general enough to be of potential applicability in other contexts.

hep-th↗

arXiv2025

On the Gauge Invariance of Secondary Gravitational Waves

Second-order tensor perturbations induced by primordial fluctuations play a crucial role in probing small-scale physics, but gauge dependence of their energy density has remained a fundamental challenge in cosmological perturbation theory. We address this issue by introducing a boundary condition-based filtering method that extracts physical radiation through the Sommerfeld criterion. We demonstrate that after filtering non-physical modes, the energy density of secondary gravitational waves becomes gauge-invariant and exhibits physically consistent behavior in the sub-horizon limit. This approach provides a unified framework for both adiabatic and isocurvature perturbations, enhancing theoretical predictions and observational signatures of early universe physics.

astro-ph.CO↗

arXiv2024

DrugAgent: Automating AI-aided Drug Discovery Programming through LLM Multi-Agent Collaboration

Recent progress in Large Language Models (LLMs) has drawn attention to their potential for accelerating drug discovery. However, a central problem remains: translating theoretical ideas into robust implementations in the highly specialized context of pharmaceutical research. This limitation prevents practitioners from making full use of the latest AI developments in drug discovery. To address this challenge, we introduce DrugAgent, a multi-agent framework that automates machine learning (ML) programming for drug discovery tasks. DrugAgent employs an LLM Planner that formulates high-level ideas and an LLM Instructor that identifies and integrates domain knowledge when implementing those ideas. We present case studies on three representative drug discovery tasks. Our results show that DrugAgent consistently outperforms leading baselines, including a relative improvement of 4.92% in ROC-AUC compared to ReAct for drug-target interaction (DTI). DrugAgent is publicly available at https://anonymous.4open.science/r/drugagent-5C42/.

cs.LG↗

arXiv2024

Seed-ASR: Understanding Diverse Speech and Contexts with LLM-based Speech Recognition

Modern automatic speech recognition (ASR) model is required to accurately transcribe diverse speech signals (from different domains, languages, accents, etc) given the specific contextual information in various application scenarios. Classic end-to-end models fused with extra language models perform well, but mainly in data matching scenarios and are gradually approaching a bottleneck. In this work, we introduce Seed-ASR, a large language model (LLM) based speech recognition model. Seed-ASR is developed based on the framework of audio conditioned LLM (AcLLM), leveraging the capabilities of LLMs by inputting continuous speech representations together with contextual information into the LLM. Through stage-wise large-scale training and the elicitation of context-aware capabilities in LLM, Seed-ASR demonstrates significant improvement over end-to-end models on comprehensive evaluation sets, including multiple domains, accents/dialects and languages. Additionally, Seed-ASR can be further deployed to support specific needs in various scenarios without requiring extra language models. Compared to recently released large ASR models, Seed-ASR achieves 10%-40% reduction in word (or character, for Chinese) error rates on Chinese and English public test sets, further demonstrating its powerful performance.

eess.AS↗

arXiv2024

Cutoff brane vs the Karch-Randall brane: the fluctuating case

Recently, certain holographic Weyl transformed CFT$_2$ is proposed to capture the main features of the AdS$_3$/BCFT$_2$ correspondence \cite{Basu:2022crn,Basu:2023wmv}. In this paper, by adapting the Weyl transformation, we simulate a generalized AdS/BCFT set-up where the fluctuation of the Karch-Randall (KR) brane is considered. In the gravity dual of the Weyl transformed CFT, the so-called cutoff brane induced by the Weyl transformation plays the same role as the KR brane. Unlike the non-fluctuating configuration, in the $2d$ effective theory the additional twist operator is inserted at a different place, compared with the one inserted on the brane. Though this is well-understood in the Weyl transformed CFT set-up, it is confusing in the AdS/BCFT set-up where the effective theory is supposed to locate on the brane. This confusion indicates that the KR brane may be emergent from the boundary CFT$_2$ via the Weyl transformations. We also calculate the balanced partial entanglement (BPE) in the fluctuating brane configurations and find it coincide with the entanglement wedge cross-section (EWCS). This is a non-trivial test for the correspondence between the BPE and the EWCS, and a non-trivial consistency check for the Weyl transformed CFT set-up.

hep-th↗

arXiv2024

Transform then Explore: a Simple and Effective Technique for Exploratory Combinatorial Optimization with Reinforcement Learning

Many complex problems encountered in both production and daily life can be conceptualized as combinatorial optimization problems (COPs) over graphs. Recent years, reinforcement learning (RL) based models have emerged as a promising direction, which treat the COPs solving as a heuristic learning problem. However, current finite-horizon-MDP based RL models have inherent limitations. They are not allowed to explore adquately for improving solutions at test time, which may be necessary given the complexity of NP-hard optimization tasks. Some recent attempts solve this issue by focusing on reward design and state feature engineering, which are tedious and ad-hoc. In this work, we instead propose a much simpler but more effective technique, named gauge transformation (GT). The technique is originated from physics, but is very effective in enabling RL agents to explore to continuously improve the solutions during test. Morever, GT is very simple, which can be implemented with less than 10 lines of Python codes, and can be applied to a vast majority of RL models. Experimentally, we show that traditional RL models with GT technique produce the state-of-the-art performances on the MaxCut problem. Furthermore, since GT is independent of any RL models, it can be seamlessly integrated into various RL frameworks, paving the way of these models for more effective explorations in the solving of general COPs.

cs.LG↗

arXiv2024

Geometrizing the Partial Entanglement Entropy: from PEE Threads to Bit Threads

We give a scheme to geometrize the partial entanglement entropy (PEE) for holographic CFT in the context of AdS/CFT. More explicitly, given a point $\textbf{x}$ we geometrize the two-point PEEs between $\textbf{x}$ and any other points in terms of the bulk geodesics connecting these two points. We refer to these geodesics as the \textit{PEE threads}, which can be naturally regarded as the integral curves of a divergenceless vector field $V_{\textbf{x}}^μ$, which we call \emph{PEE thread flow}. The norm of $V_{\textbf{x}}^μ$ that characterizes the density of the PEE threads can be determined by some physical requirements of the PEE. We show that, for any static interval or spherical region $A$, a unique bit thread configuration can be generated from the PEE thread configuration determined by the state. Hence, the non-intrinsic bit threads are emergent from the intrinsic PEE threads. For static disconnected intervals, the vector fields describing a divergenceless flow is are longer suitable to reproduce the RT formula. We weight the PEE threads with the number of times it intersects with any homologous surface. Instead the RT formula is perfectly reformulated to be the minimization of the summation of the PEE threads with all possible assignment of weights.

hep-th↗

arXiv2024

Weaving the (AdS) spaces with partial entanglement entropy threads

In the context of the AdS/CFT correspondence, we propose a general scheme for reconstructing bulk geometric quantities in a static pure AdS background using the partial entanglement entropy (PEE), a measure of the entanglement structure on the boundary CFT. The PEE between any two points $\mathcal{I}(\vec{x}, \vec{y})$ serves as the fundamental building block of the PEE structure. Any two-point PEE $\mathcal{I}(\vec{x}, \vec{y})$ can be geometrized by the bulk geodesic connecting two boundary points $\vec{x}$ and $\vec{y}$, which we call the PEE thread, with the density of the threads determined by the boundary PEE structure. In the AdS bulk, the set of all the PEE threads forms a continuous ``network'', which we call the PEE network. In this paper, we show that the density of the PEE threads passing through any bulk point is exactly $1/(4G)$. Based on this observation we give a reformulation of the Ryu-Takayanagi (RT) formula for a generic boundary region in general dimensional Poincar\'e AdS space. More explicitly, for any static boundary region $A$, the homologous surface $\Sigma_{A}$ that has the minimal number of intersections with the bulk PEE network is exactly the RT surface of $A$, and the minimal number of intersections reproduces the holographic entanglement entropy. The reconstruction for the area of bulk geometric quantities by counting the number of intersections with the bulk PEE network applies to generic bulk geometric quantities. Interestingly, this reconstruction indicates a pure geometric statement, which is exactly the so-called \emph{Crofton formula} in Poincar\'e AdS.

hep-th↗

arXiv2023

Ownerless island and partial entanglement entropy in island phases

In the context of partial entanglement entropy (PEE), we study the entanglement structure of the island phases realized in several 2-dimensional holographic set-ups. The self-encoding property of the island phase changes the way we evaluate the PEE. With the contributions from islands taken into account, we give a generalized prescription to construct PEE and balanced partial entanglement entropy (BPE). Here the ownerless island region, which lies inside the island $\text{Is}(AB)$ of $A\cup B$ but outside $\text{Is}(A)\cup \text{Is}(B)$, plays a crucial role. Remarkably, we find that under different assignments for the ownerless island, we get different BPEs, which exactly correspond to different saddles of the entanglement wedge cross-section (EWCS) in the entanglement wedge of $A\cup B$. The assignments can be settled by choosing the one that minimizes the BPE. Furthermore, under this assignment we study the PEE and give a geometric picture for the PEE in holography, which is consistent with the geometric picture in the no-island phases.

hep-th↗

arXiv2023

Scalar-Induced Gravitational Waves from Ghost Inflation and Parity Violation

We calculate the scalar-induced gravitational wave energy density in the theory of Ghost Inflation, assuming scale invariance and taking into account both the power spectrum- and trispectrum-induced contributions. For the latter we consider the leading cubic and quartic couplings of the comoving curvature perturbation in addition to two parity-violating quartic operators. In the parity-even case, we find the relative importance of the trispectrum-induced signal to be suppressed by the requirement of perturbativity, strengthening a no-go theorem recently put forth. The parity-odd signal, even though also bound to be small, is non-degenerate with the Gaussian contribution and may in principle be comparable to the parity-even non-Gaussian part, thus potentially serving as a probe of the Ghost Inflation scenario and of parity violating physics during inflation.

gr-qc↗

arXiv2023

An approach to Physics Based Single Photon Vision System Design

The design of the camera and optical measurement is a crucial part of optimizing machine vision systems. However, camera designs are usually optimized to produce human-interpretable images. Moreover, camera optimization typically makes the assumption of additive noise, while modern optical imaging systems are mainly affected by photon noise which is not additive. Previous studies have highlighted the fundamental effect of the noise model on the outcome of the design process in the context of coded single-pixel and compressed sensing cameras. In addition to the exact noise model, the nature of the data encountered and the vision task fundamentally affect the optimal hardware design. This means that a camera or compressed sensing code that is optimal for capturing images is almost certainly not optimal for any downstream vision task. In this work, we used a simple end-to-end model on a single-pixel camera to study the effects of the noise model for different vision tasks. We demonstrate the importance of incorporating non-linear noise features in optimizing even the simplest vision systems.

eess.IV↗

arXiv2023

The Markov gap in the presence of islands

The Markov gap \cite{Hayden:2021gno}, namely the difference between reflected entropy and mutual information, is explicitly computed in the defect extremal surface model, JT gravity, and the generic 2d extremal black holes, in vacuum states. The phases that contain various island contributions are considered, and their existence is carefully checked. Moreover, we show explicitly how the Markov gap originates from the OPE coefficient of the boundary CFT. And, as a generalization of \cite{Hayden:2021gno}, the lower bound of the Markov gap is given by $\frac{c}{3}\log 2$ times the number of EWCS boundaries on minimal surfaces. We propose a boundary way of counting the lower bound for the Markov gap, which states that the lower bound is given by $\frac{c}{3}\log 2$ times the number of gaps between two boundary regions in vacuum states. We discuss the limitation and possible generalization of the boundary counting, and its relation to tripartite entanglement.

hep-th↗

arXiv2023

Primordial black holes and scalar induced gravitational waves from Higgs inflation with non-canonical kinetic term

We resolve the potential-restriction problem in K/G inflation by introducing nonminimal coupling. In this context, Higgs field successfully drives inflation satisfying CMB observations while enhancing curvature perturbations at small scales, which in turn accounts for primordial black holes (PBHs) and scalar induced gravitational waves (SIGWs). We then uncover the effect of the non-canonical kinetic coupling function in more detail and study its the observational constraint. Besides, we also give the gauge invariant expression for the integral kernel of SIGWs, which is related to terms propagating with the speed of light. Finally, the non-Gaussian effect on PBH abundance and SIGWs is studied. We find that non-Gaussianity makes PBHs form more easily, but its effect on the energy density of SIGWs is negligible.

gr-qc↗