-
When LLMs Play the Telephone Game: Cumulative Changes and Attractors in Iterated Cultural Transmissions
Authors:
Jérémy Perez,
Corentin Léger,
Grgur Kovač,
Cédric Colas,
Gaia Molinaro,
Maxime Derex,
Pierre-Yves Oudeyer,
Clément Moulin-Frier
Abstract:
As large language models (LLMs) start interacting with each other and generating an increasing amount of text online, it becomes crucial to better understand how information is transformed as it passes from one LLM to the next. While significant research has examined individual LLM behaviors, existing studies have largely overlooked the collective behaviors and information distortions arising from…
▽ More
As large language models (LLMs) start interacting with each other and generating an increasing amount of text online, it becomes crucial to better understand how information is transformed as it passes from one LLM to the next. While significant research has examined individual LLM behaviors, existing studies have largely overlooked the collective behaviors and information distortions arising from iterated LLM interactions. Small biases, negligible at the single output level, risk being amplified in iterated interactions, potentially leading the content to evolve towards attractor states. In a series of telephone game experiments, we apply a transmission chain design borrowed from the human cultural evolution literature: LLM agents iteratively receive, produce, and transmit texts from the previous to the next agent in the chain. By tracking the evolution of text toxicity, positivity, difficulty, and length across transmission chains, we uncover the existence of biases and attractors, and study their dependence on the initial text, the instructions, language model, and model size. For instance, we find that more open-ended instructions lead to stronger attraction effects compared to more constrained tasks. We also find that different text properties display different sensitivity to attraction effects, with toxicity leading to stronger attractors than length. These findings highlight the importance of accounting for multi-step transmission dynamics and represent a first step towards a more comprehensive understanding of LLM cultural dynamics.
△ Less
Submitted 5 July, 2024;
originally announced July 2024.
-
Cultural evolution in populations of Large Language Models
Authors:
Jérémy Perez,
Corentin Léger,
Marcela Ovando-Tellez,
Chris Foulon,
Joan Dussauld,
Pierre-Yves Oudeyer,
Clément Moulin-Frier
Abstract:
Research in cultural evolution aims at providing causal explanations for the change of culture over time. Over the past decades, this field has generated an important body of knowledge, using experimental, historical, and computational methods. While computational models have been very successful at generating testable hypotheses about the effects of several factors, such as population structure o…
▽ More
Research in cultural evolution aims at providing causal explanations for the change of culture over time. Over the past decades, this field has generated an important body of knowledge, using experimental, historical, and computational methods. While computational models have been very successful at generating testable hypotheses about the effects of several factors, such as population structure or transmission biases, some phenomena have so far been more complex to capture using agent-based and formal models. This is in particular the case for the effect of the transformations of social information induced by evolved cognitive mechanisms. We here propose that leveraging the capacity of Large Language Models (LLMs) to mimic human behavior may be fruitful to address this gap. On top of being an useful approximation of human cultural dynamics, multi-agents models featuring generative agents are also important to study for their own sake. Indeed, as artificial agents are bound to participate more and more to the evolution of culture, it is crucial to better understand the dynamics of machine-generated cultural evolution. We here present a framework for simulating cultural evolution in populations of LLMs, allowing the manipulation of variables known to be important in cultural evolution, such as network structure, personality, and the way social information is aggregated and transformed. The software we developed for conducting these simulations is open-source and features an intuitive user-interface, which we hope will help to build bridges between the fields of cultural evolution and generative artificial intelligence.
△ Less
Submitted 13 March, 2024;
originally announced March 2024.
-
Evolving Reservoirs for Meta Reinforcement Learning
Authors:
Corentin Léger,
Gautier Hamon,
Eleni Nisioti,
Xavier Hinaut,
Clément Moulin-Frier
Abstract:
Animals often demonstrate a remarkable ability to adapt to their environments during their lifetime. They do so partly due to the evolution of morphological and neural structures. These structures capture features of environments shared between generations to bias and speed up lifetime learning. In this work, we propose a computational model for studying a mechanism that can enable such a process.…
▽ More
Animals often demonstrate a remarkable ability to adapt to their environments during their lifetime. They do so partly due to the evolution of morphological and neural structures. These structures capture features of environments shared between generations to bias and speed up lifetime learning. In this work, we propose a computational model for studying a mechanism that can enable such a process. We adopt a computational framework based on meta reinforcement learning as a model of the interplay between evolution and development. At the evolutionary scale, we evolve reservoirs, a family of recurrent neural networks that differ from conventional networks in that one optimizes not the synaptic weights, but hyperparameters controlling macro-level properties of the resulting network architecture. At the developmental scale, we employ these evolved reservoirs to facilitate the learning of a behavioral policy through Reinforcement Learning (RL). Within an RL agent, a reservoir encodes the environment state before providing it to an action policy. We evaluate our approach on several 2D and 3D simulated environments. Our results show that the evolution of reservoirs can improve the learning of diverse challenging tasks. We study in particular three hypotheses: the use of an architecture combining reservoirs and reinforcement learning could enable (1) solving tasks with partial observability, (2) generating oscillatory dynamics that facilitate the learning of locomotion tasks, and (3) facilitating the generalization of learned behaviors to new tasks unknown during the evolution phase.
△ Less
Submitted 29 January, 2024; v1 submitted 9 December, 2023;
originally announced December 2023.
-
Towards a New nCTEQ global nPDF release
Authors:
P. Risse,
N. Derakhshanian,
P. Duwentäster,
T. Ježo,
C. Keppel,
M. Klasen,
K. Kovařík,
A. Kusina,
C. Léger,
J. G. Morfín,
F. I. Olness,
R. Ruiz,
I. Schienbein,
J. Y. Yu
Abstract:
We discuss the foundation for a new global nCTEQ nuclear PDF analysis, combining a number of our previous analyses into one consistent framework with updates to the underlying theoretical treatment as well as the addition of new available data. In particular, the new global release will be the first nCTEQ release containing neutrino DIS scattering data in a consistent manner together with JLab hig…
▽ More
We discuss the foundation for a new global nCTEQ nuclear PDF analysis, combining a number of our previous analyses into one consistent framework with updates to the underlying theoretical treatment as well as the addition of new available data. In particular, the new global release will be the first nCTEQ release containing neutrino DIS scattering data in a consistent manner together with JLab high-x DIS data and new LHC p-Pb data. These additions will improve the data-driven description of nuclear PDFs in new regions, especially the strange quark and the gluon PDF at low-x.
△ Less
Submitted 15 July, 2023;
originally announced July 2023.
-
Target mass corrections in lepton--nucleus DIS: theory and applications to nuclear PDFs
Authors:
R. Ruiz,
K. F. Muzakka,
C. Leger,
P. Risse,
A. Accardi,
P. Duwentäster,
T. J. Hobbs,
T. Ježo,
C. Keppel,
M. Klasen,
K. Kovařík,
A. Kusina,
J. G. Morfín,
F. I. Olness,
J. F. Owens,
I. Schienbein,
J. Y. Yu
Abstract:
Motivated by the wide range of kinematics covered by current and planned deep-inelastic scattering (DIS) facilities, we revisit the formalism, practical implementation, and numerical impact of target mass corrections (TMCs) for DIS on unpolarized nuclear targets. An important aspect is that we only use nuclear and later partonic degrees of freedom, carefully avoiding a picture of the nucleus in te…
▽ More
Motivated by the wide range of kinematics covered by current and planned deep-inelastic scattering (DIS) facilities, we revisit the formalism, practical implementation, and numerical impact of target mass corrections (TMCs) for DIS on unpolarized nuclear targets. An important aspect is that we only use nuclear and later partonic degrees of freedom, carefully avoiding a picture of the nucleus in terms of nucleons. After establishing that formulae used for individual nucleon targets $(p,n)$, derived in the Operator Product Expansion (OPE) formalism, are indeed applicable to nuclear targets, we rewrite expressions for nuclear TMCs in terms of \mbox{re-scaled} (or averaged) kinematic variables. As a consequence, we find a representation for nuclear TMCs that is approximately independent of the nuclear target. We go on to construct a single-parameter fit for all nuclear targets that is in good numerical agreement with full computations of TMCs. We discuss in detail qualitative and quantitative differences between nuclear TMCs built in the OPE and the parton model formalisms, as well as give numerical predictions for current and future facilities.
△ Less
Submitted 12 March, 2024; v1 submitted 18 January, 2023;
originally announced January 2023.
-
Femtosecond laser pulse train interaction with dielectric materials
Authors:
O. Dematteo Caulier,
K. Mishchik,
B. Chimier,
S. Skupin,
A. Bourgeade,
C. Javaux Léger,
R. Kling,
C. Hönninger,
J. Lopez,
V. Tikhonchuk,
G. Duchateau
Abstract:
We investigate the interaction of trains of femtosecond microjoule laser pulses with dielectric materials by means of a multi-scale model. Our theoretical predictions are directly confronted with experimental observations in soda-lime glass. We show that due to the low heat conductivity, a significant fraction of the laser energy can be accumulated in the absorption region. Depending on the pulse…
▽ More
We investigate the interaction of trains of femtosecond microjoule laser pulses with dielectric materials by means of a multi-scale model. Our theoretical predictions are directly confronted with experimental observations in soda-lime glass. We show that due to the low heat conductivity, a significant fraction of the laser energy can be accumulated in the absorption region. Depending on the pulse repetition rate, the material can be heated to high temperatures even though the single pulse energy is too low to induce a significant material modification. Regions heated above the glass transition temperature in our simulations correspond very well to zones of permanent material modifications observed in the experiments.
△ Less
Submitted 14 August, 2015;
originally announced August 2015.
-
Menger curvature and rectifiability
Authors:
J. C. Léger
Abstract:
For a Borel set E in R^n, the total Menger curvature of E, or c(E), is the integral over E^3 (with respect to 1-dimensional Hausdorff measure in each factor of E) of c(x,y,z)^2, where 1/c(x,y,z) is the radius of the circle passing through three points x, y, and z in E.
Let H^1(X) denote the 1-dimensional Hausdorff measure of a set X. A Borel set E in R^n is purely unrectifiable if for any Lips…
▽ More
For a Borel set E in R^n, the total Menger curvature of E, or c(E), is the integral over E^3 (with respect to 1-dimensional Hausdorff measure in each factor of E) of c(x,y,z)^2, where 1/c(x,y,z) is the radius of the circle passing through three points x, y, and z in E.
Let H^1(X) denote the 1-dimensional Hausdorff measure of a set X. A Borel set E in R^n is purely unrectifiable if for any Lipschitz function gamma from R to R^n, H^1(E cap gamma(R)) = 0. It is said to be rectifiable if there exists a countable family of Lipschitz functions gamma_i from R to R^n such that H^1(E - union gamma_i(R)) = 0. It may be seen from this definition that any 1-set E (that is, E Borel and 0<H^1(E)<\infty) can be decomposed into two disjoint subsets E_irr and E_rect, where E_irr is purely unrectifiable and E_rect is rectifiable.
Theorem. If E is a 1-set in R^n and c(E)^2 is finite, then E is rectifiable.
△ Less
Submitted 30 April, 1999;
originally announced May 1999.
-
Front dynamics during diffusion-limited corrosion of ramified electrodeposits
Authors:
C. Leger,
F. Argoul,
M. Z. Bazant
Abstract:
Experiments on the diffusion-limited corrosion of porous copper clusters in thin gap cells containing cupric chloride are reported. By carefully comparing corrosion front velocities and concentration profiles obtained by phase-shift interferometry with theoretical predictions, it is demonstrated that this process is well-described by a one-dimensional mean-field model for the generic reaction A…
▽ More
Experiments on the diffusion-limited corrosion of porous copper clusters in thin gap cells containing cupric chloride are reported. By carefully comparing corrosion front velocities and concentration profiles obtained by phase-shift interferometry with theoretical predictions, it is demonstrated that this process is well-described by a one-dimensional mean-field model for the generic reaction A + B (static) -> C (inert) with only diffusing reactant (cupric chloride) and one static reactant (copper) reacting to produce an inert product (cuprous chloride). The interpretation of the experiments is aided by a mathematical analysis of the model equations which allows the reaction-order and the transference number of the diffusing species to be inferred. Physical arguments are given to explain the surprising relevance of the one-dimensional mean-field model in spite of the complex (fractal) structure of the copper clusters.
△ Less
Submitted 1 February, 1999;
originally announced February 1999.
