arXiv subjects

Xing Lyu

Publications and source records attributed to Xing Lyu.

3 recordsLinked to original sources

arXiv2025

Alignment Without Understanding: A Message- and Conversation-Centered Approach to Understanding AI Sycophancy

AI sycophancy is increasingly recognized as a harmful alignment, but research remains fragmented and underdeveloped at the conceptual level. This article redefines AI sycophancy as the tendency of large language models (LLMs) and other interactive AI systems to excessively and/or uncritically validate, amplify, or align with a user's assertions-whether these concern factual information, cognitive evaluations, or affective states. Within this framework, we distinguish three types of sycophancy: informational, cognitive, and affective. We also introduce personalization at the message level and critical prompting at the conversation level as key dimensions for distinguishing and examining different manifestations of AI sycophancy. Finally, we propose the AI Sycophancy Processing Model (AISPM) to examine the antecedents, outcomes, and psychological mechanisms through which sycophantic AI responses shape user experiences. By embedding AI sycophancy in the broader landscape of communication theory and research, this article seeks to unify perspectives, clarify conceptual boundaries, and provide a foundation for systematic, theory-driven investigations.

cs.HC↗

arXiv2023

Sum-of-Local-Effects Data Structures for Separable Graphs

It is not difficult to think of applications that can be modelled as graph problems in which placing some facility or commodity at a vertex has some positive or negative effect on the values of all the vertices out to some distance, and we want to be able to calculate quickly the cumulative effect on any vertex's value at any time or the list of the most beneficial or most detrimential effects on a vertex. In this paper we show how, given an edge-weighted graph with constant-size separators, we can support the following operations on it in time polylogarithmic in the number of vertices and the number of facilities placed on the vertices, where distances between vertices are measured with respect to the edge weights: Add (v, f, w, d) places a facility of weight w and with effect radius d onto vertex v. Remove (v, f) removes a facility f previously placed on v using Add from v. Sum (v) or Sum (v, d) returns the total weight of all facilities affecting v or, with a distance parameter d, the total weight of all facilities whose effect region intersects the ``circle'' with radius d around v. Top (v, k) or Top (v, k, d) returns the k facilities of greatest weight that affect v or, with a distance parameter d, whose effect region intersects the ``circle'' with radius d around v. The weights of the facilities and the operation that Sum uses to ``sum'' them must form a semigroup. For Top queries, the weights must be drawn from a total order.

cs.DS↗

arXiv2022

Ruler Rolling

At CCCG '21 O'Rourke proposed a variant of Hopcroft, Josephs and Whitesides' (1985) NP-complete problem {\sc Ruler Folding}, which he called {\sc Ruler Wrapping} and for which all folds must be 180 degrees in the same direction. Gagie, Saeidi and Sapucaia (2023) noted that if the last straight section of the ruler must be longest, then {\sc Ruler Wrapping} is equivalent to partitioning a string of positive integers into substrings whose sums are increasing such that the last substring sums to at most a given amount. They gave linear-time algorithms for the versions of {\sc Ruler Wrapping} both with and without this assumption. In real life we cannot repeatedly fold a carpenter's ruler 180 degrees in the same direction. In this paper we propose the more realistic problem of {\sc Ruler Rolling}, in which we repeatedly fold the segments 90 degrees in the same direction and thus fold the ruler into a rectangle instead of into an interval. We should report all the Pareto-optimal rollings. We note that if the last straight section of the ruler must be longer than the third to last -- analogously to Gagie et al.'s assumption -- then {\sc Ruler Rolling} is equivalent to partitioning a string of positive integers into substrings such that the sums of the even substrings are increasing, as are the sums of the odd substrings. We give a simple dynamic-programming algorithm that reports all the Pareto-optimal rollings in quadratic time under this assumption. Our algorithm still works even without the assumption, but then we are left with a quadratic number of two-dimensional feasible solutions, so finding the Pareto-optimal ones and increases our running time by a logarithmic factor. If we have a nice objective function, however, we still use quadratic time.

cs.DS↗