We present a decentralized algorithm that achieves the same performance as a centralized one, contradicting the existing lower bounds for that problem. This is possible by "hacking" the standard model by constructing a communication protocol between players that deliberately enforces collisions, allowing them to share their information at a negligible cost.

AI red teaming brings security thinking to LLM applications by probing failure modes such as prompt injection, output manipulation, and sensitive data exfiltration. While automated and curated benchmarks (e.g., JailbreakBench [1], HarmBench [2]) are increasingly used to test models and defenses, comparatively fewer studies analyze community scale behavior in the wild. We study ai_gon3_rogu3 [3], a 10 day competition with 504 registrants and 217 active players, to quantify solve dynamics, tactic stratification, and choke points across 11 challenges. We find sharp skill stratification, higher success for output manipulation than for data extraction, and strong effects of format obfuscation tactics, with multi step defenses remaining robust, among other insights.

However, current research for decision-making, like reinforcement learning (RL), has primarily required numerous real interactions with the target environment to learn a skill, while failing to utilize the existing knowledge already summarized in the text.

Siri is known to cut people off midsentence, but there's away to make Siri listen longer. CyberGuy shows you how to customize the wait time. A new villain has entered the scene in the infinite realms of the beloved video game Minecraft, where diamond swords fend off blocky monsters. Unlike creepers or ghasts, this antagonist doesn't explode or float around aimlessly. Instead, it's a real-world threat of infostealing malware distributed by hackers who've craftily used Minecraft, which has a community of over 140 million active players, as their playground.

This story is part of Future Tense Fiction, a monthly series of short stories from Future Tense and Arizona State University's Center for Science and the Imagination about how technology and science will change our lives. When I get home from work at 6:00, the light is fading, and I see my sons and their little friend playing in the street, two white boys and a Black boy throwing a foam football back and forth. I pull around the corner and they scatter, Oliver running one way while Jameson and the neighbor kid run the other. At the last minute, though, Jameson changes his mind, dropping the football and dashing across to his brother's side of the street. I slam to a halt, the bumper almost touching him. My heart throbs in my jaw: so close. Then, just as I release the brake, the neighbor kid runs across, too, and I have to stomp to a stop a second time. Don't any of you have common sense?" Through the unrolled window I see them all staring at me with wide eyes. "What is wrong with your ...

Friend recall is an important way to improve Daily Active Users (DAU) in online games. The problem is to generate a proper lost friend ranking list essentially. Traditional friend recall methods focus on rules like friend intimacy or training a classifier for predicting lost players' return probability, but ignore feature information of (active) players and historical friend recall events. In this work, we treat friend recall as a link prediction problem and explore several link prediction methods which can use features of both active and lost players, as well as historical events. Furthermore, we propose a novel Edge Transformer model and pre-train the model via masked auto-encoders. Our method achieves state-of-the-art results in the offline experiments and online A/B Tests of three Tencent games.

This paper presents a framework for learning player embeddings in competitive games and events. Players and their win-loss relationships are modeled as a skill gap graph, which is an undirected weighted graph. The player embeddings are learned from the graph using a random walk-based graph embedding method and can reflect the relative skill levels among players. Embeddings are low-dimensional vector representations that can be conveniently applied to subsequent tasks while still preserving the topological relationships in a graph. In the latter part of this paper, Graphical Elo (GElo) is introduced as an application of player embeddings when rating player skills. GElo is an extension of the classic Elo rating system. It constructs a skill gap graph based on player match histories and learns player embeddings from it. Afterward, the rating scores that were calculated by Elo are adjusted according to player activeness and cosine similarities among player embeddings. GElo can be executed offline and in parallel, and it is non-intrusive to existing rating systems. Experiments on public datasets show that GElo makes a more reliable evaluation of player skill levels than vanilla Elo. The experimental results suggest potential applications of player embeddings in competitive games and events.

We consider the stochastic multiplayer multi-armed bandit problem, where several players pull arms simultaneously and a collision occurs if the same arm is pulled by more than one player; this is a standard model of cognitive radio networks. We construct a decentralized algorithm that achieves the same performances as a centralized one, if players are synchronized and observe their collisions. We actually construct a communication protocol between players by enforcing willingly collisions, allowing them to share their exploration. With a weaker feedback, when collisions are not observed, we still maintain some communication between players but at the cost of some extra multiplicative term in the regret. We also prove that the logarithmic growth of the regret is still achievable in the dynamic case where players are not synchronized with each other, thus preventing communication. Finally, we prove that if all players follow naively the celebrated UCB algorithm, the total regret grows linearly.