←
António Góis79 posts
António Góis avatar

António Góis

@AntGois

PhD student @Mila_Quebec; Interested in performative prediction, causality, evolution of cooperation,NLProc; prev @deep_spin @Unbabel HotClubeJazz @istecnico

◷ Joined June 2013▧ 8 Photos and videos
António Góis avatar
António Góis@AntGois·May 7

.@eurowings Why is your support telling passengers "free seats don't exist" to force paid check-ins? An agent told me I HAD to pay to finish check-in. I waited 20mins, tried the app again, and got a free seat easily. Stop using misleading sales pressure. #Eurowings #Travel

Jose Gallego-Posada avatar
⟳ António Góis retweeted
Jose Gallego-Posada@JGallegoPosada·2 Jul 2025

Meet Cooper — a PyTorch library for constrained optimization in deep learning. 🚀 Cooper makes constraints as easy to use as losses! @juan43ramirez @meraj_hashemi_ @SimonLacosteJ @Mila_Quebec 💻github.com/cooper-org/cooper 🧵👇

ELLIS Unit Lisbon avatar
⟳ António Góis retweeted
ELLIS Unit Lisbon@Lisbon_ELLIS·28 Apr 2025

🚨Today is the last day to apply to the Lisbon Machine Learning School! Hear from top speakers in the field and get to meet new people during a fantastic week in Lisbon. ‼️Apply today: tinyurl.com/mre5as8f

📢 Applications to the 15th Lisbon Machine Learning School 2025 are open! Covering #ML topics such as neural networks, transformers, causality, vision & language. 📅 Dates: 19th - 25th of July 🔗 more info: lxmls.it.pt/2025/index.html 🔗 apply here: tinyurl.com/mre5as8f

Post attachment
Milad Aghajohari avatar
⟳ António Góis retweeted
Milad Aghajohari@MAghajohari·25 Apr 2025

Multi-Agent RL fails in real life. Agents cooperating to solve tasks remains a utopia. -No scalable algorithms for general-sum games. -In a simple apple-harvesting game, PPO agents overharvest and ruin bushes. Advantage Alignment (ICLR 2025 Oral📢) is a huge step forward. 1/n

Post attachment
António Góis avatar
António Góis@AntGois·12 Apr 2025

This connects to open questions in 𝗽𝗵𝗶𝗹𝗼𝘀𝗼𝗽𝗵𝘆 with practical relevance — how to evaluate performative models? Aiming for prosocial outcomes (appraisal view) undesirably injects moral values, endogenizing causes of adaptation (mitigation view) has side effects 8/9🧵

António Góis avatar
António Góis@AntGois·12 Apr 2025

We explore graphical criteria and game parameters that determine tension between accuracy and welfare in our setting. We also show how repeatedly maximizing accuracy can minimize welfare with high probability. 7/9🧵

Post attachment
António Góis avatar
António Góis@AntGois·12 Apr 2025

Agents use a Bayesian update to compute their belief in external predictions of collective behaviour. Depending on past accuracy, 𝘁𝗿𝘂𝘀𝘁 𝗶𝗻 𝗽𝗿𝗲𝗱𝗶𝗰𝘁𝗶𝗼𝗻𝘀 may increase or decrease. We show that welfare-maximizing predictions can induce cycles in trust 6/9🧵

Post attachment
António Góis avatar
António Góis@AntGois·12 Apr 2025

This provides a testbed for different predictive systems, regarding accuracy of predictions and social welfare after adapting to predictions 5/9🧵

António Góis avatar
António Góis@AntGois·12 Apr 2025

We focus on "collective risk dilemma", a game where a critical mass of cooperators is required to avoid a disaster, and all agents simultaneously choose whether to cooperate at a cost. 4/9🧵

Post attachment
António Góis avatar
António Góis@AntGois·12 Apr 2025

We propose a novel framework where: a) predictions influence reality through interdependence among the predicted population b) social welfare arises naturally, since agents' goals are affected by predictions c) agents' trust in predictions is computed as a Bayesian update 3/9🧵