←
Logan Cross966 posts
Logan Cross avatar

Logan Cross

@locross

Research Scientist, @GoogleDeepMind | Previously at @Stanford, and PhD in Computation and Neural Systems at @Caltech

◷ Joined May 2009▧ 33 Photos and videos
Logan Cross avatar
Logan Cross@locross·Nov 4

How can we build LLM social simulations that capture meaningful human differences and generalize beyond a single experiment? Excited about our new work identifying which axes of behavioral variation matter for predicting human decisions across configurations of a task family Great working with @RobinNa_ @Kehang_Zhu and @bakkermichiel!

Prompting LLMs to behave like people can improve simulations, but how should we design this? What info is needed, and how much? With @Kehang_Zhu @locross @bakkermichiel, we build “behavioral profiles” that help LLMs simulate human behavior under distribution shifts. 1/8 🧵

Post attachment
Shane Legg avatar
⟳ Logan Cross retweeted
Shane Legg@ShaneLegg·Oct 29

Two new DeepMind Institute essays: What is Work For? As AI advances how do people value their jobs, for material needs and beyond? The Economics of Intelligence When will AI capabilities impact market structure, wages and jobs? Read the essays here -> bit.ly/dmi_home

Logan Cross avatar
Logan Cross@locross·Oct 29

Jobs do far more than deliver a paycheck—they are a core institution in modern society for agency, purpose, recognition, and community In our new DeepMind Institute essay with @scychan_brains and @JulianDJacobs, we explore research on how work relates to wellbeing. As AI potentially changes the nature of work, we must design policies and institutions that protect its non-monetary value bit.ly/wellbeing-work-agi

Post attachment
Davide Paglieri avatar
⟳ Logan Cross retweeted
Davide Paglieri@PaglieriDavide·Sep 24

Recent incidents show how quickly misbehavior can cascade in agent swarms. Yet when given transparent channels, honest agents naturally try to blow the whistle on cheating peers. In our new DeepMind Institute essay, @Vezhnick and I propose that giving agents the tools to self-police may be part of the solution: bit.ly/cheaters-and-whistleb…

Shane Legg avatar
⟳ Logan Cross retweeted
Shane Legg@ShaneLegg·Sep 16

My journey to develop AGI spans 25 yrs, including 10+ yrs thinking about technical & societal perspectives at Google DeepMind. AGI is on the horizon - we need deeper understanding of its implications. To help, we've created the DeepMind Institute.

Logan Cross avatar
Logan Cross@locross·Sep 9

As autonomous agents increasingly interact with one another, we need to understand not just agent capabilities, but agent institutions. Our new case study: cheating, whistleblowing, and emergent norm enforcement in an agent swarm 👇

🧵We conducted an experiment with 100 agents by giving them math problems to solve. When a small group (9%) started to cheat, what came next surprised us: 24% fought back by blowing the whistle on their peers and alerting humans.

Post attachment
Violet Xiang avatar
⟳ Logan Cross retweeted
Violet Xiang@ZiyuX·Jun 16

🧵(1/9) Sparse RL for reasoning has an exploration problem. It can only reward solutions the model already stumbles into. On hard problems, that means lots of zeros and very little signal. SFT and self-distillation attack this with reference solutions as targets to match. Instead, we use them as reward scaffolds: a dense signal at both the outcome and process level. Introducing ExpRL: RL-based mid-training that improves exploration by scoring the model’s own attempts against the reference via an LLM judge. What we find: • A stronger policy straight out of mid-training (higher pass@1 and pass@k) • Still ahead after downstream sparse-reward RL • Holds across domains – math & STEM • Scales to a larger policy graded by a smaller judge

Post attachment
Logan Cross avatar
Logan Cross@locross·Jun 4

Excited to share our latest ICML 2026 paper: Solipsistic Superintelligence is Unlikely to be Cooperative We argue that AI systems trained as isolated optimizers struggle during deployment in this fundamentally multi-agent reality. Why? The world pushes back. Humans adapt. Institutions adapt. Other AIs adapt.

As increasingly capable AI systems are deployed, humans, institutions, and other AI systems adapt in response — i.e. the world pushes back. So is capability still the central safety challenge for AI? We think not. We believe the harder challenge is coexistence. The current AI research paradigm treats the world as a stationary source of feedback, what we refer to as the solipsistic approach to AI design. This raises serious risks for coexistence. In our new #ICML2026 paper, we argue that superintelligence — an extremely capable task solver, built through such a solipsistic approach — is unlikely to be cooperative. 🧵

Post attachment
Logan Cross avatar
Logan Cross@locross·Mar 16

We generalize the same mechanisms beyond luxury consumption and procedurally generate social signaling scenarios for political allegiance, altruism, and arbitrary conventions. For example: with a social life, agents were more likely to choose a large public donation over a smaller anonymous one

Post attachment
Logan Cross avatar
Logan Cross@locross·Mar 16

But importantly, our interpretation of Veblen goods is not “agents buy things to show off the $ amount". Prices rise due to the social hype. We ran a fixed-price control where social life stayed the same, but sellers could not raise prices. Demand for status goods did not collapse — it increased even more. So the causal story looks more like: hype → demand → price, not price → demand

Post attachment
Logan Cross avatar
Logan Cross@locross·Mar 16

With this Social Life, we see dramatic price run-ups for items like a Chanel bag and Labubu doll. Demand for Labubu continued to increase even when its price increased (positive price elasticity), becoming a true Veblen good that inverts the law of demand. This demonstrates that our generative model can produce both the "rational" and "irrational" aspects of human behavior

Post attachment
Logan Cross avatar
Logan Cross@locross·Mar 16

With the Social Life turned on, the demand for status goods significantly increased. Agents organically bought a lot of status goods and chatted about them as trends started forming

Post attachment
Logan Cross avatar
Logan Cross@locross·Mar 16

Hypothesis: social exposure increases the preference for a signaling behavior To test this, our agents bid for food/gadgets/clothing in a marketplace. We turn on/off a social layer: with a Social Life they go on "first dates" with each other, observe each other's outfits, and chat freely

Post attachment
Logan Cross avatar
Logan Cross@locross·Mar 16

Classical models treat human preferences as fixed. We argue they are socially constructed. Instead of hard-coded utility functions, our agents use predictive pattern completion, selecting actions that are most appropriate conditioned on all the contextual variables of that situation. When an agent observes a status signal during a social interaction, they update their memory and are more likely to imitate that behavior in a similar context

Post attachment
Logan Cross avatar
Logan Cross@locross·Mar 16

Ever wonder why we drop $1,000s on a Chanel bag or queue for ages for a Labubu doll? Status signaling drives a lot of human behavior, but how certain things become potent status symbols has a remained a big puzzle in social science In our new paper we: 1. Synthesize the literature on this topic and propose a generative model of status signaling through the theory of appropriateness: people imitate what "someone like us" is supposed to want, display, and value 2. Show that we can simulate the theory to demonstrate how status symbols emerge with LLM-agent societies in Concordia @jordigraumo @WilCunningham Sasha Vezhnevets and @jzl86

Post attachment