Search arXiv⌕ Search

arXiv subjects

Yi Cai

Publications and source records attributed to Yi Cai.

At least 37 records · Page 2Linked to original sources
arXiv2026

Truth or Sophistry? LoFa: A Benchmark for LLM Robustness Against Logical Fallacies

Large Language Models (LLMs) exhibit strong semantic capabilities, yet their resilience to manipulative linguistic patterns such as logical fallacies remains underexplored. Prior work has primarily examined whether LLMs can identify or classify fallacies, leaving their robustness against fallacious persuasion insufficiently studied. To address this gap, we introduce LoFa (Logical Fallacy), a comprehensive benchmark for evaluating LLM robustness against fallacies. LoFa is constructed through a multi-agent pipeline that pairs factual questions with fallacious arguments, and is accompanied by a multi-round debate framework for assessing model resilience under sustained adversarial persuasion. To disentangle fallacy robustness from a model's inherent knowledge limitations, we further propose Logical Fallacy Resistance at k (LFR@k), a metric that quantifies resistance to fallacious attacks. Experiments show that LLMs exhibit varying levels of robustness across different fallacy types, revealing distinct vulnerability profiles among models.

cs.CL↗
arXiv2026

Mixture of Debaters: Learn to Debate at Architectural Level in Multi-Agent Reasoning

Existing multi-agent debate frameworks suffer from two critical limitations: they rely on static architectures where agent roles and coordination patterns are fixed at design time, and they require instantiating multiple model copies, incurring substantial computational overhead. We propose Mixture of Debaters (MoD), a unified framework that enables dynamic self-debate within a single model by leveraging the Mixture-of-Experts paradigm. We address three key challenges in adapting MoE for dialectical reasoning: (1) dual-routing that decouples role allocation from process flow, dynamically determining when to debate versus when to synthesize; (2) momentum switching that smooths token-level routing with local context, reducing expert-switch jitter; and (3) unified self-debate that encapsulates diverse debating personas into lightweight expert modules, eliminating inter-agent communication while preserving behavioral diversity. Extensive experiments on multimodal benchmarks demonstrate that MoD outperforms both single-model baselines and conventional multi-agent systems, achieving superior accuracy with 3.7x lower latency and 87% reduction in token consumption.The source code can be accessed at https://github.com/YongLD/MoD.

cs.AI↗
arXiv2026

CDPR: Counterfactual Advantage-based Credit Assignment for Cost-Aware Sequential Medical Diagnosis

Clinical diagnosis is a step-by-step, cost-aware process: a physician orders examinations one at a time, observes the results, and updates the diagnosis before reaching a final conclusion. Most medical language models instead treat diagnosis as a one-pass classification task and ignore the trade-off between a test's value and its cost. We model diagnosis as a cost-aware sequential decision process and train the policy with reinforcement learning. The main difficulty is credit assignment: the only reliable signal comes once at the end of a long trajectory, so it scores a wasteful workup the same as an efficient one. We propose CDPR (Counterfactual Diagnostic Process Reward), which needs no expert labels and no learned critic. CDPR first finds the states where the policy hesitates, using the uncertainty of its action distribution, and then scores the chosen action by its advantage over the alternatives the policy itself would consider, estimated with short rollouts under a utility that balances correctness against test count, cost, and infeasible requests. A rollout cache reuses within-batch trajectories to keep the cost low. We integrate CDPR into GRPO and test it on one in-domain (MIMIC-IV) and two out-of-domain (ClinicalBench and a private hospital dataset) benchmarks. CDPR improves diagnostic accuracy while clearly reducing the number and cost of examinations.

cs.AI↗
arXiv2026

AGE-MIL: Anchor-Guided Evidence Learning for Patient-Level Prediction

Existing computational pathology methods predominantly operate within whole-slide image (WSI)-level multiple instance learning (MIL) paradigms, while patient-level modeling remains underexplored. In routine pathological practice, however, pathologists derive diagnostic and prognostic conclusions by integrating evidence across multiple WSIs rather than relying on any single slide. This discrepancy creates a fundamental misalignment when patient-level supervision is directly imposed on conventional MIL frameworks, often leading to unstable optimization and degraded predictive reliability. To address this issue, we propose Anchor-Guided Evidence MIL (AGE-MIL), a weakly supervised framework for patient-level prediction. AGE-MIL constructs a patient-level anchor from slide representations to capture global pathological context and guide the retrieval and integration of diagnostically relevant local patches, enabling robust patient-level modeling. Patient-level risk is further modeled as an evidence accumulation process, promoting stable optimization under weak supervision. AGE-MIL is evaluated on six clinically relevant patient-level prediction tasks from two independent cohorts. Experimental results show that the proposed framework consistently outperforms eight state-of-the-art MIL methods. Code is available at https://github.com/wodeniua/AGE-MIL.

cs.CV↗
arXiv2026

TacSE3: Equivariant SE(3) Motion Estimation from Low-Texture Visuotactile Images for In-Gripper Tracking and Compensation

Robotic in-hand manipulation requires reliable object-motion tracking under frequent visual occlusion, yet low-texture visuotactile images provide few stable correspondences for conventional image- or geometry-matching methods. This paper presents TacSE3, a tactile motion-estimation pipeline that converts low-texture visuotactile observations into a decoupled three-dimensional force field and estimates incremental rigid-body motion on SE(3). The method derives planar translation from contact-centroid motion and estimates rotation primarily from shear-related tactile responses, yielding a physically interpretable signal for in-gripper tracking and compensation. Experiments with paired DM-Tac fingertip sensors show that dual-sensor sensing reduces translation-rotation ambiguity, supports rotation tracking across axes and object geometries, and provides a lightweight compensation signal that improves disturbance tolerance in downstream manipulation tasks without retraining the base policy.

cs.RO↗
arXiv2026

Strat-Reasoner: Reinforcing Strategic Reasoning of LLMs in Multi-Agent Games

While Large Language Models (LLMs) excel in certain reasoning tasks, they struggle in multi-agent games where the final outcome depends on the joint strategies of all agents. In multi-agent games, the non-stationarity of other agents brings significant challenges on the evaluation of the reasoning process and the credit assignment over multiple reasoning steps. Existing single-agent reinforcement learning (RL) approaches and their multi-agent extensions fail to address these challenges as they do not incorporate other agents in the reasoning process. In this work, we propose Strat-Reasoner, a novel RL-based framework that improves LLMs' strategic reasoning ability in multi-agent games. We introduce a novel recursive reasoning paradigm where an agent's reasoning also integrates other agents' reasoning processes. To provide effective reward signals for the intermediate reasoning sequences, we employ a centralized Chain-of-Thought (CoT) comparison module to evaluate the reasoning quality. Finally, we compute an accurate hybrid advantage and develop a group-relative RL approach to optimize the LLM policy. Experimental results show that Strat-Reasoner substantially improves strategic abilities of underlying LLMs, achieving 22.1\% average performance improvements across various multi-agent games. Code is publicly available at https://github.com/ydhe1012/Strat-Reasoner.

cs.AI↗
arXiv2026

Unsupervised Semi-Parametric Plug-in Likelihood-Ratio Detection for Covert Communications in the Presence of Disco Reconfigurable Intelligent Surfaces

Covert communications, also referred to as low probability of detection (LPD) communications, provide a higher level of privacy protection than cryptography and physical-layer security (PLS) by hiding transmissions in the ambient environment. In this work, we investigate covert communications in the presence of a disco reconfigurable intelligent surface (DRIS) deployed by the warden Willie, which reduces Willie's detection error probability (DEP), i.e., the sum of the false alarm rate (FAR) and the miss detection rate (MDR), and degrades the communication performance between Alice and Bob, without relying on either channel state information (CSI) or additional jamming power. However, the introduction of the DRIS makes it analytically intractable for Willie to construct the Neyman-Pearson (NP) detector, which is the optimal detector for monitoring potential covert transmissions between Alice and Bob. To this end, we develop an unsupervised semi-parametric plug-in likelihood-ratio detector for Willie. The proposed detector retains the parametric Gamma reference model under the silent hypothesis without requiring prior knowledge of noise, and learns from unlabeled data a one-dimensional monotone normalizing flow model for the analytically intractable distribution under the transmission hypothesis. In particular, it exploits the structural prior inherent in covert communications that Willie's observations reduce to noise only when Alice and Bob are silent. The monitoring performance at Willie is evaluated in terms of DEP, while the communication impact on Alice and Bob is quantified by the signal-to-jamming-plus-noise ratio (SJNR). Simulation results verify the analysis and show that the proposed unsupervised plug-in likelihood-ratio detector achieves monitoring performance close to that of its supervised counterpart.

eess.SP↗
arXiv2026

Discrete hypocoercive estimates for discontinuous Galerkin methods: application to the Vlasov-Poisson-Fokker-Planck system

We develop and analyze a class of structure-preserving discontinuous Galerkin schemes for the nonlinear Vlasov-Poisson-Fokker-Planck model, reformulated as a hyperbolic system through a Hermite expansion in the velocity variable. We discretize the Vlasov-Fokker-Planck equation with the discontinuous Galerkin method, while the Poisson equation is approximated with either a discontinuous Galerkin method or a Raviart-Thomas mixed finite element method. We prove the exponential relaxation to equilibrium for suitable initial data, uniformly with respect to the discretization parameters thanks to discrete hypocoercivity arguments. Moreover, we check that the resulting semi-discrete schemes preserve the physical invariants along with the L 2 variational structure of the linearized model. Numerical simulations verify the accuracy and the long-time behavior of the scheme.

math.NA↗
arXiv2026

Reputation as a Solution to Cooperation Collapse in LLM-based MASs

Cooperation has long been a fundamental topic in both human society and AI systems. However, recent studies indicate that the collapse of cooperation may emerge in multi-agent systems (MASs) driven by large language models (LLMs). To address this challenge, we explore reputation systems as a remedy. We propose RepuNet, a dynamic, dual-level reputation framework that models both agent-level reputation dynamics and system-level network evolution. Specifically, driven by direct interactions and indirect gossip, agents form reputations for both themselves and their peers, and decide whether to connect or disconnect other agents for future interactions. Through three distinct scenarios, we show that RepuNet effectively avoids cooperation collapse, promoting and sustaining cooperation in LLM-based MASs. Moreover, we find that reputation systems can give rise to rich emergent behaviors in LLM-based MASs, such as the formation of cooperative clusters, the social isolation of exploitative agents, and the preference for sharing positive gossip rather than negative ones. The GitHub repository for our project can be accessed via the following link: https://github.com/RGB-0000FF/RepuNet.

cs.AI↗
arXiv2026

Intersections of Cantor sets with hyperbolas and continuous images

Given $λ\in (0,1/2)$, let \begin{equation*} C_λ=\set{(1-λ)\sum_{i=1}^\infty d_iλ^{i-1}:d_i\in\set{0,1}} \end{equation*} be the middle Cantor sets with convex hull $[0, 1]$. We are interested in the set $S_t=\set{(x,y)\in C_λ\times C_λ: xy=t}$, where $t\in[0,1]$. Since the cases where $t=0$ or $t=1$ are trivial, we assume that $t\in(0,1)$ in what follows. We show that there exists a $λ_0=0.4302$ such that for all $λ$ satisfying $λ_0 \le λ< 1/2$, the set $S_t$ has the cardinality of the continuum for every $t \in (0,1)$. Besides, we further investigate the continuous image of $C_λ\times C_λ$, that is, for any given $2\le k\in \nn$, we give a sufficient condition for set $\set{x^ky:x,y\in C_λ}$ to be the interval $[0,1]$. Our observations reveal that the behavior exhibited by the image of the function $f_k(x,y)=x^ky$ is complex and depends on the parameters $k$ and $λ$.

math.NT↗
arXiv2026

Phase transitions for unique codings of fat Sierpinski gaskets with multiple digits

Given an integer $M\ge 1$ and $β\in(1, M+1)$, let $S_{β, M}$ be the fat Sierpinski gasket in $\mathbb R^2$ generated by the iterated function system $\left\{f_d(x)=\frac{x+d}β: d\inΩ_M\right\}$, where $Ω_M=\{(i,j)\in\mathbb Z_{\ge 0}^2: i+j\le M\}$. Then each $x\in S_{β, M}$ may be represented as a series $x=\sum_{i=1}^\infty\frac{d_i}{β^i}=:Π_β((d_i))$, and the infinite sequence $(d_i)\inΩ_M^{\mathbb N}$ is called a \emph{coding} of $x$. Since $β β_c(M)$ then $U_{β, M}$ has positive Hausdorff dimension. Our results can also be applied to the intrinsic univoque set $\widetilde{U}_{β, M}$. Moreover, we show that the first critical base $β_G(M)$ is a Perron number, while the second critical base $β_c(M)$ is a transcendental number.

math.DS↗
arXiv2025

ViRectify: A Challenging Benchmark for Video Reasoning Correction with Multimodal Large Language Models

As multimodal large language models (MLLMs) frequently exhibit errors in complex video reasoning scenarios, correcting these errors is critical for uncovering their weaknesses and improving performance. However, existing benchmarks lack systematic evaluation of MLLMs' ability to identify and correct these video reasoning errors. To bridge this gap, we propose ViRectify, a comprehensive benchmark to evaluate their fine-grained correction capability. Through an AI-assisted annotation pipeline with human verification, we construct a dataset of over 30K instances spanning dynamic perception, scientific reasoning, and embodied decision-making domains. In ViRectify, we challenge MLLMs to perform step-wise error identification and generate rationales with key video evidence grounding. In addition, we further propose the trajectory evidence-driven correction framework, comprising step-wise error trajectory and reward modeling on visual evidence-grounded correction. It encourages the model to explicitly concentrate on error propagation and key timestamps for correction. Extensive evaluation across 16 advanced MLLMs demonstrates that our ViRectify serves as a challenging testbed, where GPT-5 achieves only 31.94% correction accuracy. Our framework enables a Qwen2.5-VL-7B to consistently outperform the variants of 72B on ViRectify, showing the effectiveness of our approach. Further analysis uncovers systematic asymmetries in error correction across models, and our dataset is also a valuable data resource to perform reflection learning. We believe ViRectify provides a new direction for comprehensively evaluating the advanced MLLMs in video reasoning.

cs.CV↗
arXiv2025

Code Comments for Quantum Software Development Kits: An Empirical Study on Qiskit

Quantum computing is gaining attention from academia and industry. With the quantum Software Development Kits (SDKs), programmers can develop quantum software to explore the power of quantum computing. However, programmers may face challenges in understanding quantum software due to the non-intuitive quantum mechanics. To facilitate software development and maintenance, code comments offered in quantum SDKs serve as a natural language explanation of program functionalities and logical flows. Despite their importance, scarce research systematically reports their value and provides constructive guidelines for programmers. To address this gap, our paper focuses on Qiskit, one of the most popular quantum SDKs, and presents CC4Q, the first dataset of code comments for quantum computing. CC4Q incorporates 9677 code comment pairs and 21970 sentence-level code comment units, the latter of which involve heavy human annotation. Regarding the annotation, we validate the applicability of the developer-intent taxonomy used in classical programs, and also propose a new taxonomy considering quantum-specific knowledge. We conduct an empirical study comprehensively interpreting code comments from three perspectives: comment structure and coverage, developers' intentions, and associated quantum topics. Our findings uncover key differences in code comments between classical and quantum software, and also outline quantum-specific knowledge relevant to quantum software development.

cs.SE↗
arXiv2025

Hybrid-DMKG: A Hybrid Reasoning Framework over Dynamic Multimodal Knowledge Graphs for Multimodal Multihop QA with Knowledge Editing

Multimodal Knowledge Editing (MKE) extends traditional knowledge editing to settings involving both textual and visual modalities. However, existing MKE benchmarks primarily assess final answer correctness while neglecting the quality of intermediate reasoning and robustness to visually rephrased inputs. To address this limitation, we introduce MMQAKE, the first benchmark for multimodal multihop question answering with knowledge editing. MMQAKE evaluates (1) a model's ability to reason over 2-5-hop factual chains that span both text and images, including performance at each intermediate step, and (2) robustness to visually rephrased inputs in multihop questions. Our evaluation shows that current MKE methods often struggle to consistently update and reason over multimodal reasoning chains after knowledge edits. To overcome these challenges, we propose Hybrid-DMKG, a hybrid reasoning framework built on a dynamic multimodal knowledge graph (DMKG) to enable accurate multihop reasoning over updated multimodal knowledge. Hybrid-DMKG first uses a large language model to decompose multimodal multihop questions into sequential sub-questions, then applies a multimodal retrieval model to locate updated facts by jointly encoding each sub-question with candidate entities and their associated images. For answer inference, a hybrid reasoning module operates over the DMKG via two parallel paths: (1) relation linking prediction, and (2) RAG reasoning with large vision-language models. A decision module aggregates evidence from both paths to select the most credible answer. Experimental results on MMQAKE show that Hybrid-DMKG significantly outperforms existing MKE approaches, achieving higher accuracy and improved robustness to knowledge updates.

cs.AI↗
arXiv2025

Rethinking Explanation Evaluation under the Retraining Scheme

Feature attribution has gained prominence as a tool for explaining model decisions, yet evaluating explanation quality remains challenging due to the absence of ground-truth explanations. To circumvent this, explanation-guided input manipulation has emerged as an indirect evaluation strategy, measuring explanation effectiveness through the impact of input modifications on model outcomes during inference. Despite the widespread use, a major concern with inference-based schemes is the distribution shift caused by such manipulations, which undermines the reliability of their assessments. The retraining-based scheme ROAR overcomes this issue by adapting the model to the altered data distribution. However, its evaluation results often contradict the theoretical foundations of widely accepted explainers. This work investigates this misalignment between empirical observations and theoretical expectations. In particular, we identify the sign issue as a key factor responsible for residual information that ultimately distorts retraining-based evaluation. Based on the analysis, we show that a straightforward reframing of the evaluation process can effectively resolve the identified issue. Building on the existing framework, we further propose novel variants that jointly structure a comprehensive perspective on explanation evaluation. These variants largely improve evaluation efficiency over the standard retraining protocol, thereby enhancing practical applicability for explainer selection and benchmarking. Following our proposed schemes, empirical results across various data scales provide deeper insights into the performance of carefully selected explainers, revealing open challenges and future directions in explainability research.

cs.LG↗
arXiv2025

On Higher-Power Moments of $ Δ_a(x) $ for $-1/2<a<0$

Let $-1/2<a<0$ be a fixed real number and \begin{equation*} Δ_{a}(x)=\sideset{}{'}\sum_{n\leq x} σ_a(n)-ζ(1-a)x-\frac{ζ(1+a)}{1+a}x^{1+a}+\frac{1}{2}ζ(-a). \end{equation*} In this paper, we investigate the higher--power moments of $Δ_a(x)$ and give the corresponding asymptotic formula for the integral $\int_{1}^{T}Δ_a^k(x)\mathrm{d}x$, which constitutes an improvement upon the previous result of Zhai [9] for $k=3,4,5$ and an enlargement of the upper bound of $k$ to $7$.

math.NT↗
arXiv2025

An efficient asymptotic preserving Monte Carlo method for frequency-dependent radiative transfer equations

In this paper, we develop an efficient asymptotic-preserving (AP) Monte Carlo (MC) method for frequency-dependent radiative transfer equations (RTEs), which is based on the AP-MC method proposed for the gray RTEs in \cite{shi2023efficient}. We follow the characteristics-based approach by Zhang et al. \cite{zhang2023asymptotic} to get a reformulated model, which couples a low dimension convection-diffusion-type equation for macroscopic quantities with a high dimension transport equation for the radiative intensity. To recover the correct free streaming limit due to frequency-dependency, we propose a correction to the reformulated macroscopic equation. The macroscopic system is solved using a hybrid method: convective fluxes are handled by a particle-based MC method, while diffusive fluxes are treated implicitly with central difference. To address the nonlinear coupling between radiative intensity and the Planck function across multiple frequency groups, we adopt a Picard iteration with a predictor-corrector procedure, which decouples a global nonlinear system into a linear system restricted to spatial dimension (independent of frequency) with scalar algebraic nonlinear equations. Once the macroscopic update is done, the transport equation, with a known emission source provided by the macroscopic variables, is efficiently solved using an implicit MC method. This approach enables larger time steps independent of the speed of light and also the frequency across a wide range, significantly enhancing computational efficiency, especially for frequency-dependent RTEs. Formal AP analysis in the diffusive scaling is established. Numerical experiments are performed to demonstrate the high efficiency and AP property of the proposed method.

math.NA↗