arXiv subjects

Feipeng Zhang

Publications and source records attributed to Feipeng Zhang.

10 recordsLinked to original sources

arXiv2026

Spider-Sense: Intrinsic Risk Sensing for Efficient Agent Defense with Hierarchical Adaptive Screening

As large language models (LLMs) evolve into autonomous agents, their real-world applicability has expanded significantly, accompanied by new security challenges. Most existing agent defense mechanisms adopt a mandatory checking paradigm, in which security validation is forcibly triggered at predefined stages of the agent lifecycle. In this work, we argue that effective agent security should be intrinsic and selective rather than architecturally decoupled and mandatory. We propose Spider-Sense framework, an event-driven defense framework based on Intrinsic Risk Sensing (IRS), which allows agents to maintain latent vigilance and trigger defenses only upon risk perception. Once triggered, the Spider-Sense invokes a hierarchical defence mechanism that trades off efficiency and precision: it resolves known patterns via lightweight similarity matching while escalating ambiguous cases to deep internal reasoning, thereby eliminating reliance on external models. To facilitate rigorous evaluation, we introduce S$^2$Bench, a lifecycle-aware benchmark featuring realistic tool execution and multi-stage attacks. Extensive experiments demonstrate that Spider-Sense achieves competitive or superior defense performance, attaining the lowest Attack Success Rate (ASR) and False Positive Rate (FPR), with only a marginal latency overhead of 8.3\%.

cs.CR↗

arXiv2025

Evolutionary hypergame dynamics: Introspection reasoning and social learning

In the realm of evolutionary game theory, standard frameworks typically presuppose that every player possesses comprehensive knowledge and unrestricted access to the entire strategy space. However, real-world human society inherently harbors diverse levels of knowledge, experience, and background among individuals. Hypergames incorporate this heterogeneity by permitting individuals to differ in their access to the full strategy set, reflecting cognitive or informational constraints and giving rise to asymmetric strategic interactions. Yet, their evolutionary consequences remain underexplored. Our inquiry employs prototype models featuring three available strategies, focusing on social dilemmas involving cooperation, defection, and loner. These strategies manifest cyclic dominance, akin to the well-studied rock-paper-scissors dynamics, a foundational model in game theory. Our study spans both well-mixed and spatial lattice populations, delving into the intricacies of learning and evolution of the strategy set within the evolutionary hypergame dynamics. In stark contrast to traditional evolutionary game dynamics, our findings unveil nuanced and intricate phases, encompassing scenarios of loner dominance, coexistence of multiple strategy sets, combinations of cooperation and loner dominance, and more. Remarkably, we discern that heightened rationality significantly promotes cooperative behaviors.

cs.GT↗

arXiv2025

Evolutionary dynamics under coordinated reciprocity

Using past behaviors to guide future actions is essential for fostering cooperation in repeated social dilemmas. Traditional memory-based strategies that focus on recent interactions have yielded valuable insights into the evolution of cooperative behavior. However, as memory length increases, the complexity of analysis grows exponentially, since these strategies need to map every possible action sequence of a given length to subsequent responses. Due to their inherent reliance on exhaustive mapping and a lack of explicit information processing, it remains unclear how individuals can handle extensive interaction histories to make decisions under cognitive constraints. To fill this gap, we introduce coordinated reciprocity strategies ($CORE$), which incrementally evaluate the entire game history by tallying instances of consistent actions between individuals without storing round-to-round details. Once this consistency index surpasses a threshold, $CORE$ prescribes cooperation. Through equilibrium analysis, we derive an analytical condition under which $CORE$ constitutes an equilibrium. Moreover, our numerical results show that $CORE$ effectively promotes cooperation between variants of itself, and it outperforms a range of existing strategies including memory-$1$, memory-$2$, and those from a documented strategy library in evolutionary dynamics. Our work thus underscores the pivotal role of cumulative action consistency in enhancing cooperation, developing robust strategies, and offering cognitively low-burden information processing mechanisms in repeated social dilemmas.

q-bio.PE↗

arXiv2025

Exploring Imbalanced Annotations for Effective In-Context Learning

Large language models (LLMs) have shown impressive performance on downstream tasks through in-context learning (ICL), which heavily relies on the demonstrations selected from annotated datasets. However, these datasets often exhibit long-tailed class distributions in real-world scenarios, leading to biased demonstration selection. In this work, we show that such class imbalances significantly degrade the ICL performance across various tasks, regardless of selection methods. Moreover, classical rebalancing methods, which focus solely on class weights, yield poor performance due to neglecting condition bias--skewed feature distributions within classes. To address this, we propose Reweighting with Conditional Bias (dubbed RCB), a simple and complementary approach to enhance ICL performance under class imbalance. In particular, RCB estimates conditional bias using a balanced subset and re-weights demonstration scores based on both class weight and conditional bias. In effect, RCB prevents over-selection from dominant classes while preserving the efficacy of current selection methods. Extensive experiments on common benchmarks demonstrate the effectiveness of our method, improving the average accuracy of current selection methods by up to 5.42%.

cs.CL↗

arXiv2024

On the Noise Robustness of In-Context Learning for Text Generation

Large language models (LLMs) have shown impressive performance on downstream tasks by in-context learning (ICL), which heavily relies on the quality of demonstrations selected from a large set of annotated examples. Recent works claim that in-context learning is robust to noisy demonstrations in text classification. In this work, we show that, on text generation tasks, noisy annotations significantly hurt the performance of in-context learning. To circumvent the issue, we propose a simple and effective approach called Local Perplexity Ranking (LPR), which replaces the "noisy" candidates with their nearest neighbors that are more likely to be clean. Our method is motivated by analyzing the perplexity deviation caused by noisy labels and decomposing perplexity into inherent perplexity and matching perplexity. Our key idea behind LPR is thus to decouple the matching perplexity by performing the ranking among the neighbors in semantic space. Our approach can prevent the selected demonstrations from including mismatched input-label pairs while preserving the effectiveness of the original selection methods. Extensive experiments demonstrate the effectiveness of LPR, improving the EM score by up to 18.75 on common benchmarks with noisy annotations. Our code is available at https://github.com/ml-stat-Sustech/Local-Perplexity-Ranking.

cs.CL↗

arXiv2024

Efficient subgroup testing in change-plane models

Considered here is a hypothesis test for the coefficients in the change-plane regression models to detect the existence of a change plane. The test that is considered is from the class of test problems in which some parameters are not identifiable under the null hypothesis. The classic exponential average tests do not work well in practice. To overcome this drawback, a novel test statistic is proposed by taking the weighted average of the squared score test statistic (WAST) over the grouping parameter's space, which has a closed form from the perspective of the conjugate priors when an appropriate weight is chosen. The WAST significantly improves the power in practice, particularly in cases where the number of the grouping variables is large. The asymptotic distributions of the WAST are derived under the null and alternative hypotheses. The approximation of critical value by the bootstrap method is investigated, which is theoretically guaranteed. Furthermore, the proposed test method is naturally extended to the generalized estimating equation (GEE) framework, as well as multiple change planes that can test if there are three or more subgroups. The WAST performs well in simulation studies, and its performance is further validated by applying it to real datasets.

math.ST↗

arXiv2024

Adaptive coordination promotes collective cooperation in repeated social dilemmas

Direct reciprocity based on the repeated prisoner's dilemma has been intensively studied. Most theoretical investigations have concentrated on memory-$1$ strategies, a class of elementary strategies just reacting to the previous-round outcomes. Though the properties of "All-or-None" strategies ($AoN_K$) have been discovered, simulations just confirmed the good performance of $AoN_K$ of very short memory lengths. It remains unclear how $AoN_K$ strategies would fare when players have access to longer rounds of history information. We construct a theoretical model to investigate the performance of the class of $AoN_K$ strategies of varying memory length $K$. We rigorously derive the payoffs and show that $AoN_K$ strategies of intermediate memory length $K$ are most prevalent, while strategies of larger memory lengths are less competent. Larger memory lengths make it hard for $AoN_K$ strategies to coordinate, and thus inhibiting their mutual reciprocity. We then propose the adaptive coordination strategy combining tolerance and $AoN_K$' coordination rule. This strategy behaves like $AoN_K$ strategy when coordination is not sufficient, and tolerates opponents' occasional deviations by still cooperating when coordination is sufficient. We found that the adaptive coordination strategy wins over other classic memory-$1$ strategies in various typical competition environments, and stabilizes the population at high levels of cooperation, suggesting the effectiveness of high level adaptability in resolving social dilemmas. Our work may offer a theoretical framework for exploring complex strategies using history information, which are different from traditional memory-$n$ strategies.

cs.GT↗

arXiv2018

Segmented correspondence curve regression model for quantifying reproducibility of high-throughput experiments

The reliability of a high-throughput biological experiment relies highly on the settings of the operational factors in its experimental and data-analytic procedures. Understanding how operational factors influence the reproducibility of the experimental outcome is critical for constructing robust workflows and obtaining reliable results. One challenge in this area is that candidates at different levels of significance may respond to the operational factors differently. To model this heterogeneity, we develop a novel segmented regression model, based on the rank concordance between candidates from different replicate samples, to characterize the varying effects of operational factors for candidates at different levels of significance. A grid search method is developed to identify the change point in response to the operational factors and estimate the covariate effects accounting for the change. A sup-likelihood-ratio-type test is proposed to test the existence of a change point. Simulation studies show that our method yields a well-calibrated type I error, is powerful in detecting the difference in reproducibility, and achieves a better model fitting than the existing method. An application on a ChIP-seq dataset reveals interesting insights on how sequencing depth affects the reproducibility of experimental results, demonstrating the usefulness of our method in designing cost-effective and reliable high-throughput workflows.

stat.ME↗

arXiv2016

A Continuous Threshold Expectile Model

Expectile regression is a useful tool for exploring the relation between the response and the explanatory variables beyond the conditional mean. This article develops a continuous threshold expectile regression for modeling data in which the effect of a covariate on the response variable is linear but varies below and above an unknown threshold in a continuous way. Based on a grid search approach, we obtain estimators for the threshold and the regression coefficients via an asymmetric least squares regression method. We derive the asymptotic properties for all the estimators and show that the estimator for the threshold achieves root-n consistency. We also develop a weighted CUSUM type test statistic for the existence of a threshold in a given expectile, and derive its asymptotic properties under both the null and the local alternative models. This test only requires fitting the model under the null hypothesis in the absence of a threshold, thus it is computationally more efficient than the likelihood-ratio type tests. Simulation studies show desirable finite sample performance in both homoscedastic and heteroscedastic cases. The application of our methods on a Dutch growth data and a baseball pitcher salary data reveals interesting insights.

stat.ME↗

arXiv2016

Robust bent line regression

We introduce a rank-based bent linear regression with an unknown change point. Using a linear reparameterization technique, we propose a rank-based estimate that can make simultaneous inference on all model parameters, including the location of the change point, in a computationally efficient manner. We also develop a score-like test for the existence of a change point, based on a weighted CUSUM process. This test only requires fitting the model under the null hypothesis in absence of a change point, thus it is computationally more efficient than likelihood-ratio type tests. The asymptotic properties of the test are derived under both the null and the local alternative models. Simulation studies and two real data examples show that the proposed methods are robust against outliers and heavy-tailed errors in both parameter estimation and hypothesis testing.

stat.ME↗