What happens when LLMs play the Telephone game? ☎️ In this new preprint, we analyse the evolution of texts as they are transmitted between LLM agents 🤖💬🤖💬🤖💬 Do text properties converge to attractors? 🧲 How is this influenced by the task📝 and model⚙️? 1/13🧵

Master internship positions available in the Flowers Lab!! @marko_cvjetko & I will supervise a project on cultural evolution x artificial life x curiosity: what happens when humans follow their curiosity to collectively explore complex systems? Also check out the other positions!

🔥Our paper PhyloLM got accepted at ICLR 2025 !🔥 In this work we show how easy it can be to infer relationship between LLMs by constructing trees and to predict their performances and behavior at a very low cost with @StePalminteri and @pyoudeyer ! Here is a brief recap ⬇️

Check out this work today at 10 am @iclr_conf! Learn what's wrong with evaluating #LLMs after a single interaction, and how cultural attraction theory can help us do better. Poster #288 @Jeremy__Perez @pyoudeyer @MaximeDerex @KovacGrgur @cedcolas @corentinlger @ClemMoulinFrier

New preprint alert with @MaevaLhotellier and @Jeremy__Perez BUT to discover what it is about you have to go and follow us on @bluesky !!! #keepmoving https://t.co/ks1NDtcQE9

📢Open position: Master internship on hybrid human-LLM cultural evolution We're looking for an intern to explore how LLMs can help human groups discover, select and transmit solutions to problems More info and other positions here: https://t.co/q4zjFVj5ig and in thread below

🚨New preprint🚨 When testing LLMs with questions, how can we know they did not see the answer in their training? In this new paper we propose a simple out of the box and fast method to spot contamination on short texts with @StePalminteri and @pyoudeyer ! 1/16

J-6 pour s'inscrire au #hackathon Hack1robo ! 🚀 Choisi ton challenge parmi 8 projets sur les thèmes #IA #robot #sciencescognitives https://t.co/o0xChPtM6Z Encore 25 places ! Inscriptions ➡️ https://t.co/DEksUlY1er @Inria @Neuro_Bordeaux @Bordeaux @aquinum @univbordeaux

🚀 #Hack1robo is approaching!! 🚀 Together with @sinak12345 and @marko_cvjetko, we propose to explore what happens when you plug music into continuous cellular automata: 🎶+ 🦠 = ? If you're in Bordeaux 15th to 17th November, come explore this topic with us! -6 days to register

Excited to share with @flowersINRIA ADTool, an open-source Python toolkit using curiosity-driven AI to explore behavior diversity of complex systems (cellular automata, physics models, ...)!🐍 Serving off-the-shelf strategies like IMGEP https://t.co/UGl6VhhH3A You said IMGEP? 👇

The conference will be basically opened by our own @MaevaLhotellier with a talk entitled "Achieving Scale-Independent Reinforcement Learning Performance With Reward Range Normalization" Friday October 11th (11:15 AM- 12:15 PM)

If you're interested in using LLMs to simulate artificial societies, or for human-AI collective intelligence, you might want to make sure that they can stick to the role/personality that you give them. This project led by @KovacGrgur introduces a leaderboard exactly for that!

🚨 Paper Alert 🚨 AI learns everything from us human culture. But do we also learn from AI? Our paper *delves* into the question of whether humans have now started to speak like ChatGPT (not just write like it). Check out the video and thread below.

Almost time for @ALifeConf where we will present our work "Collective Innovation in Groups of Large Language Models" We study how social structure affects how LLMs collectively play the game Little Alchemy 2 🔥+🌽=🍿 Work done with @Clement_MF_ @pyoudeyer @criticalneuro @risi1979

This was a very collaborative and thus very pleasing project, huge thanks to all co-authors! @corentinlger @KovacGrgur @cedcolas @gaia_molinaro @MaximeDerex @pyoudeyer @Clement_MF_ Paper: https://t.co/ba4fhXgLf7 Code: https://t.co/c9AIfkpJyS 13/13 🧵

This project builds on our previous work https://t.co/XeUjdaLJf7, and follows recent efforts to bridge AI and cultural evo: https://t.co/XEFNFn4121 @acerbialberto @JStubbersfield https://t.co/a04MZkxbka @edwardfhughes @jzl86 https://t.co/Y45Aidh9Ro @LevinBrinkmann @iyadrahwan

For raw data enthusiasts 🥩, the companion website provides a Data Explorer tool allowing to look at all generated texts: https://t.co/EbOyIjj6gr 12/13

Overall, this simple setting highlights that taking into account collective dynamics is crucial when analyzing LLMs' behavior. Promising directions include considering the effect of communication structure, of model and task heterogeneity, and hybrid human-LLM populations. 11/13

Finally, we ask to what extent different chains converge in terms of semantic similarity. For less constrained tasks, we find that chains starting with very different texts eventually become more similar. The degree to which they converge appears to vary accross models. 10/13

We can then compare how task and model influence attraction for different properties. We find that Toxicity displays stronger attractors than Length, and than more constrained tasks (Rephrase) lead to weaker attraction. Model and task sometimes influence attractor positions. 9/13

To quantitatively compare the evolutionary dynamics , we introduce a method for estimating the strength and position of theoretical attractors. By comparing the properties in initial vs final texts, we estimate the position of a theoretical fixed point, as shown below. 8/13