acceptodds
Under review as a conference paper at ICLR 2027

Coarse-Grained Dynamics of Multi-Agent Collective AutoResearch

Abstract

Multi-agent collective AutoResearch generates research findings through shared knowledge and collaborative exploration. As the number of agents increases, collective behavior arising from complex interactions becomes increasingly important, making system-level performance and scalability difficult to predict and motivating an effective dynamical description. We model the effective dynamics of collective behavior from a statistical physics perspective using coarse-graining methods and formalize a multi-agent collective AutoResearch system as a stochastic growth process in which agents are coupled through a growing knowledge hypergraph. We describe each agent by the phase-type distribution of its action duration and outcome, specify local hypergraph interaction rules, and average over the agents to obtain the coarse-grained dynamics. This formulation can model and predict collective behavior using measurements on single agents, mapping different research agents onto their phase-type distributions and different frameworks onto interaction rules and correction terms. We evaluate the model's predictive ability in formal theorem-proving tasks: from the phase-type distribution of a single agent, it predicts how the completion time of many agents sharing a record changes with their number, in close agreement with experiments. This provides a dynamical framework for understanding and predicting the scalability of collective AutoResearch systems.

Primary area:
learning theory
Keywords:
Non-equilibrium statistical physics, Multi-agent system, Autonomous Research

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_8UCpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies the prediction of outcomes of automatic research by modeling the multi-agent interactions as a stochastic growth process on a shared knowledge hypergraph. In the proposed framework, each agent is described by the phase-type distribution of its action duration and outcome, and the coarse-grained dynamics are averaged over the individual agents. The authors use formal theorem-proving tasks to evaluate the proposed model, and demonstrate that the model can predict how the completion time of many agents sharing a record changes with their number.

Critical strengths
  1. Predicting the outcomes of multi-agent interactions in automatic research is interesting and important. I also agree that statistical physics can provide valuable methods and insights for modeling the processes in the theoretical sense.
  2. The paper covers comprehensive theoretical analysis and empirical validations, and yields meaningful insights regarding predicting the outcomes of multi-agent systems in automatic research.
Critical weaknesses
  1. The theoretical contributions of the paper are significantly limited in the sense that the empirical implications do not match the theoretical complexity. I think most of the analysis, although reasonable, is conceptual, without deterministic insights regarding the behaviors of multi-agent systems.
  2. The empirical validations are also thin. The multi-agent test uses a single ETP world, and a single agent. The replication at large NN is also quite limited since only several values are tested, leaving significant concerns regarding the results of the paper. Also, the only baseline compared in the paper is with the own zeroth-order model, which is known to fail with a delay at synchronized start.
  3. Another big issue of the paper is the weak connection to automatic research. From the theoretical analysis to empirical validations, I think the methods used in the paper can generalize to many different multi-agent systems powered by LLMs, without pointing out the fundamental challenges in automatic research. Empirically, the paper also only studies theorem proving, which is not equivalent to many settings of automatic research studied in previous work.
Questions for authors

N/A

1 comment

We thank the reviewer for the helpful comments and for recognizing the importance of predicting multi-agent automatic research.

The theoretical contributions of the paper are significantly limited in the sense that the empirical implications do not match the theoretical complexity. I think most of the analysis, although reasonable, is conceptual, without deterministic insights regarding the behaviors of multi-agent systems.

The theory yields quantitative predictions computed from the single-agent action law alone, with nothing fitted to multi-agent campaigns. It predicts when adding agents stops dividing the completion time by NN (the onset of competition, Eq. (17)), the floor of the completion time set by the fixed delay, and the scaling exponents above the onset. These predictions match the observed T50T_{50} and T80T_{80} exponents (Table 2). The same equations also give a new result for finite worlds: when difficulty stays with each question, the late stage slows down, and the slowdown grows with the number of agents per question. This prediction is supported out of sample (Table 1).

The empirical validations are also thin. The multi-agent test uses a single ETP world, and a single agent. The replication at large N is also quite limited ... Also, the only baseline compared in the paper is with the own zeroth-order model ...

We extended the validation in three ways. First, we ran new campaigns at N=64,128,256N=64,128,256 with a changed agent: under the same model identifier, it was about five times slower on single-agent attempts. Calibrated only on its own single-agent campaigns, the model predicts all three T50T_{50} values within the 90% intervals, and with question-specific difficulty two of the three T90T_{90} values (Table 1). Second, we compared against baselines that use the same single-agent data: an empirical-duration simulator, which resamples the recorded single-agent attempts, and ideal 1/N1/N scaling with a fitted delay. The phase-kept model matches the simulator and reproduces the observed T50T_{50} and T80T_{80} exponents, while the fitted 1/N1/N baseline does not (Table 2). Third, ablations show that the fixed delay and the phase structure jointly produce the departure from 1/N1/N scaling. Beyond ETP, the single-agent action law is validated on three kinds of task with three agent frameworks and three models (Section 6.1), and the crowding predictions are tested in two Gymnasium worlds (Appendix E.5).

Another big issue of the paper is the weak connection to automatic research. ... the methods used in the paper can generalize to many different multi-agent systems powered by LLMs, without pointing out the fundamental challenges in automatic research. Empirically, the paper also only studies theorem proving ...

The model is built on features that distinguish research from other multi-agent tasks. Knowledge accumulates in a shared record, and one result can settle many other questions through logical closure; in ETP, a single refuting table can settle over a hundred implications. A question becomes answerable only once its premises are established. Researchers duplicate each other's work on the same open question. And a finite research program ends with its hardest questions, where additional agents help least. These features set the scaling laws, and generic multi-agent task models do not contain them. The experiments are also not limited to theorem proving: the action laws are tested on scientific law discovery in NewtonBench as well as on formal mathematics in Lean and ETP, and ETP is itself a collaborative mathematical research project. We will make these connections explicit in the revised paper.

Table 1. New campaigns with the changed agent

NNT50T_{50} observed / frozen prediction [90%]T90T_{90} observed / with question-specific difficulty [90%]
6442.5 / 51.5 [39.7, 74.5]235.8 / 285.6 [199.8, 392.7]
12842.2 / 42.0 [34.2, 54.5]246.1 / 192.0 [127.2, 246.2]
25632.6 / 36.0 [30.8, 41.2]185.7 / 141.4 [103.1, 179.6]

Times in seconds. Difficulty is estimated from the original agent's campaigns; all inputs were fixed before the new campaigns.

Table 2. Predictors using the same single-agent data

Prediction MethodError: T50T_{50} / T80T_{80} / T90T_{90}Slope: T50T_{50} / T80T_{80} / T90T_{90}
Measured---0.28 / -0.59 / -0.52
Ideal 1/N1/N1.04 / 0.55 / 0.88-1.00 / -1.00 / -1.00
1/N1/N + fitted delay0.23 / 0.19 / 0.56-0.47 / -0.75 / -0.80
Zeroth-order model1.25 / 0.79 / 1.09-1.07 / -1.10 / -1.07
Empirical-duration simulator0.15 / 0.25 / 0.64-0.32 / -0.60 / -0.65
Phase-kept model0.15 / 0.32 / 0.69-0.31 / -0.60 / -0.65
Reviewer_cMUbOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper models research agents interacting through a shared knowledge hypergraph. It derives collective dynamics from single-agent measurements and tests predictions of multi-agent completion times on theorem-proving tasks.

Critical strengths

Using data from one agent to predict how a group of agents behaves is an interesting idea. The model helps explain why agents compete and why adding more agents gives smaller gains.

Critical weaknesses
  1. High-agent-count settings have only one campaign each. The model also underpredicts T90 by roughly twofold at N ≥ 64, leaving its accuracy and prediction-interval coverage uncertain.

  2. The ETP comparisons mainly use ideal 1/N scaling and the zeroth-order model. The advantage over other predictors using the same single-agent data remains unclear.

  3. The experiments do not isolate the effects of fixed delays, phase structure, and cancellation. More controlled ablations are needed to establish which components explain the observed scaling.

Questions for authors
  1. Can repeated high-agent-count campaigns establish prediction-interval coverage for T90 using the same frozen single-agent calibration?

  2. How does the model compare with an empirical-duration simulator using the same single-agent data and interaction rules?

  3. Could you separately ablate fixed delays, phase structure, and cancellation, and report their effects on T50, T80, and T90 prediction errors

  4. Which three preregistered checks failed, and how should these failures change the interpretation of the main results?

2 comments

We thank the reviewer for the helpful suggestions for improving this work.

High-agent-count settings have only one campaign each. The model also underpredicts T90 by roughly twofold at N ≥ 64, leaving its accuracy and prediction-interval coverage uncertain.

Can repeated high-agent-count campaigns establish prediction-interval coverage for T90 using the same frozen single-agent calibration?

We conducted one additional independent campaign at each of N=64,128,256N=64,128,256 on the same ETP world. All new T50T_{50} results fall within their 90% prediction intervals. For T90T_{90}, the original and new campaigns show a consistent late-stage discrepancy: the observed values are 1.35 to 2.35 times the predictions, and all fall outside the intervals at high NN.

This discrepancy is caused by the finite size of the world, and it follows our growth equation in the finite case. By Eqs. (7) and (9), an open question qq is settled at the rate Wq=λinLs,q(κq)W_q=\lambda^{\rm in}L_{{\rm s},q}(\kappa_q), where λin=N(1−α)/(τˉMopen)\lambda^{\rm in}=N(1-\alpha)/(\bar\tau M_{\rm open}) is the rate at which agents enter each open question. The ETP world is fixed and no questions are posed during a campaign (α=0\alpha=0), so MopenM_{\rm open} only decreases. This produces two late-stage effects. The first is crowding: the fewer questions remain, the more agents work on each. The second is selection: when each question keeps its own success law Ls,qL_{{\rm s},q}, easy questions are settled first and the remaining ones are increasingly hard. The original model gives all questions of a class the same law, so it has crowding but no selection. With question-specific laws, the share of questions still open is never smaller than with a common average rate (Jensen's inequality), and the gap grows with the number of attempts per question. This is why the discrepancy is small at T50T_{50}, largest at T90T_{90}, and grows with NN. Table 1 shows this in the data: the more agents per remaining question, the more the remaining questions are the hard ones, and the larger the T90T_{90} discrepancy. Our model describes free growth, in which research keeps posing new questions and the pool of questions is not exhausted, whereas experiments can only be run in finite worlds. The late-stage deviation is what the same growth equation gives for a finite world, so it supports rather than contradicts the model.

Table 1. Late stage in a finite world

NNAgents per remaining question at T90T_{90}P(remaining question is harder), original / new agentObserved / predicted T90T_{90}, original / new agent
160.40.59 / --1.45 / --
641.50.95 / 0.851.75 / 1.35
1283.00.98 / 0.941.92 / 2.18
2566.00.98 / 0.982.16 / 2.35

About 43 questions remain at T90T_{90}. Difficulty is estimated from the other campaigns, for the new agent from those of the original agent; 0.5 means no relation.

We retained question-specific difficulty estimated from the original agent's campaigns, while calibrating speed and quitting rates using the new agent's single-agent campaigns. All inputs were fixed before the new multi-agent campaigns, with no parameters fitted to their outcomes. The resulting predictions substantially reduce the errors in T50,T80T_{50}, T_{80}, and T90T_{90}, with most observations falling within the 90% intervals. This improvement is driven by preserving question-specific difficulty. Simulations also show that doubling the number of questions has the same effect as halving the number of agents. The revised paper will include this extension and further validation.

Table 2. Out-of-sample prediction of T90T_{90}

NNObserved T90T_{90} (s)Frozen single-type predictionWith question-specific difficulty
64235.8174.8 [133.7, 234.0]285.6 [199.8, 392.7]
128246.1113.0 [91.5, 141.8]192.0 [127.2, 246.2]
256185.779.1 [68.2, 92.3]141.4 [103.1, 179.6]
Mean absolute log error--0.640.24

(continued)

The ETP comparisons mainly use ideal 1/N scaling and the zeroth-order model. The advantage over other predictors using the same single-agent data remains unclear.

How does the model compare with an empirical-duration simulator using the same single-agent data and interaction rules?

Our framework already includes an empirical-duration simulator, which produced the original registered predictions by sampling single-agent observations under the same interaction rules. With full data, the phase-kept model matches it in accuracy and scaling slopes, and both come closest to the observed T50T_{50} and T80T_{80} exponents (Table 3). The phase-type description adds three advantages. It gives the onset of competition and the floor of the completion times analytically. Its extrapolation shifts less when single-agent data are scarce: an empirical simulator cannot produce durations shorter than the shortest one observed, so its large-NN floor rises as the data shrink, while the phase-type prediction barely changes. And task-independent workflow stages can be reused across worlds, which reduces the number of fitted parameters and improves the AIC.

Table 3. Prediction errors and high-NN scaling slopes

Prediction MethodError: T50T_{50} / T80T_{80} / T90T_{90}Slope: T50T_{50} / T80T_{80} / T90T_{90}
Measured---0.28 / -0.59 / -0.52
Ideal 1/N1/N1.04 / 0.55 / 0.88-1.00 / -1.00 / -1.00
1/N1/N + fitted delay0.23 / 0.19 / 0.56-0.47 / -0.75 / -0.80
Zeroth-order model1.25 / 0.79 / 1.09-1.07 / -1.10 / -1.07
Empirical-duration simulator0.15 / 0.25 / 0.64-0.32 / -0.60 / -0.65
Phase-kept model0.15 / 0.32 / 0.69-0.31 / -0.60 / -0.65

The experiments do not isolate the effects of fixed delays, phase structure, and cancellation. More controlled ablations are needed to establish which components explain the observed scaling.

Could you separately ablate fixed delays, phase structure, and cancellation, and report their effects on T50, T80, and T90 prediction errors?

Removing each component separately shows distinct roles (Table 4). Without fixed delays, the floor of the completion times drops and all three errors increase. Without phase structure, the slopes move away from the observed ones. Without both, the model collapses to ideal 1/N1/N scaling. Cancellation mainly shortens the late stage and changes the slopes little. The full model comes closest to the observed exponents.

Table 4. Ablation results for prediction errors and high-NN scaling

Model VariantError: T50T_{50} / T80T_{80} / T90T_{90}Slope: T50T_{50} / T80T_{80} / T90T_{90}T50T_{50} at N=256N=256 (s)
Measured---0.28 / -0.59 / -0.5215.7
Full model0.15 / 0.32 / 0.69-0.31 / -0.60 / -0.6514.0
Without fixed delay0.34 / 0.42 / 0.79-0.44 / -0.65 / -0.709.2
Without phase structure0.19 / 0.12 / 0.37-0.41 / -0.69 / -0.7313.2
Without both1.33 / 0.58 / 0.85-0.99 / -1.01 / -1.001.8
Without cancellation0.15 / 0.16 / 0.43-0.33 / -0.64 / -0.6814.0

Which three preregistered checks failed, and how should these failures change the interpretation of the main results?

Five of the eight preregistered checks fell within their intervals, including both T90T_{90} checks; the three outside are listed in Table 5. The first lies just below its interval. The other two are consistent with a task-selection preference: agents chose questions already being researched less often than uniform selection predicts, so fewer attempts collide and early progress is faster. The failures therefore come from the selection rule rather than from the late-stage dynamics, and Eq. (11) already allows non-uniform selection, so this behavior can be included without changing the framework. Although the primary criterion was not met, the results provide partial validation, which we will clarify in the revised paper.

Table 5. Preregistered checks outside their intervals

CheckObservedRegistered 90% interval
N=4N=4, cascading-cancellation fraction0.068[0.069, 0.123]
N=8N=8, T50T_{50} (s)38[44, 91]
N=8N=8, same-question cancellation fraction0.153[0.191, 0.251]
Reviewer_zwzmOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper provides a statistical-physics-based analysis of the dynamics of collective AutoResearch as it explores and fills a bounded set of open research questions. The authors formulate a closed approximate system of equations for the growth rate of the number of established mathematical statements. The setting allows definitive verification of such statements via lean. As a result, some scaling laws are provided for scaling the number of research agents.

Critical strengths
  • The direction seems promising and timely: Understanding how AutoResearch affects the generation of new research or is to be modelled.
  • The presentation of the work is easy to understand and contains greater visual aids such as clear overview figures.
  • The main concepts are simple and clear, which lends extendability to this work.
Critical weaknesses
  • The work is applicable to areas with verifiable outcomes; The significance of the model is somewhat limited, as domain transfer may be difficult due to the requirement of a verifier such as lean for mathematical statements; If none exists, hallucinations and wrong results are not modelled. Furthermore, uniform modelling of all agents and tasks is assumed.
  • The boundedness of the space of theorems appears to be assumed. In reality, there is practically an infinity of possible statements and questions to prove, which is not covered by the model and may need discussion.
  • I think this work is not specific to AI / AutoResearch, but rather a specific case study using AI. The work applies modern LLMs on domains with bounded research questions, and verifies in its experiments how long it takes to fill the space of research questions.
  • As a case study, I think the manuscript should attempt to find a use case, such as to apply control in the model to infer the Pareto-frontier of optimal numbers of agents working on the field given a cost to duplicated work and speed of discovery, or other similar insights. Otherwise, I feel that the experimental insights are a little bit trivial (e.g., more agents duplicate more work and shorten the wait).
  • Presentation-wise, Eq. (3) is unclear from the main text.
Questions for authors
  1. This model is "append-only" on the hypergraph. Could it be extended easily to removals? Because AI agents are known to hallucinate a lot, some derivations may need to be revised as well, though this may not be required in this initial scenario when a lean verifier exists.
  2. How would you extend the model to non-uniform modelling of difficulty?
  3. Is it possible to obtain a statistical physics model for unbounded numbers of open questions?
  4. What is the relation with infection models on graphs such as SIS models? Can this model be understood as a specific kind of infection model?
  5. Can we obtain predictions from fits for how long it would take AI researcher communities to reach certain discoveries? For example, this could be interesting for downstream applications to predict when the singularity hits, or when a cure to a cancer type is found.
2 comments

We thank the reviewer for the helpful suggestions for improving this work. We first restate the core contribution of the paper. We then answer the questions one by one.

Core contribution. The main contribution of the paper is a theoretical construction. It is not meant to describe a single phenomenon. Starting from an explicit philosophical setting, we design an interpretable coarse-graining path. With it we find the macroscopic variables and macroscopic equations that describe the collective behavior of agents doing research together. The path makes explicit which microscopic information matters for the macroscopic behavior and which can be discarded. The few macroscopic variables we find are measurable and sufficient to determine the collective behavior. The low-dimensional equations they satisfy describe this behavior. Under the stated conditions, they also predict it from measurements on single agents. The experiments test whether this coarse-graining loses necessary information. The experiments themselves are not the purpose of building the theory. Without coarse-graining, the way to understand or predict what large N agents do together is to actually run them. To extrapolate from the experience of a few agents, one needs a method that is interpretable and has a known range of validity, and coarse-graining is such a method.

The work is applicable to areas with verifiable outcomes; ... domain transfer may be difficult due to the requirement of a verifier such as lean ...; If none exists, hallucinations and wrong results are not modelled.

The paper does not require a formal verifier for research such as Lean. For "research", including human scientific communities, needs some verification to decide which propositions are accepted. This is not the same as judging whether a proposition is a truth. People may believe Lean is a truth checker. What research consists of, and how propositions are verified and accepted, are still open questions in the philosophy of science. Disagreement comes from different philosophical viewpoints. What matters is therefore to state the philosophical position clearly and to derive a self-consistent theory from it.

In our model, the record consists of the results accepted by the framework, and the verifier decides what is accepted. No step of the coarse-graining uses the notion of truth, and all collective variables are defined on the accepted record. A stricter or looser verifier changes the parameters of the outcome distributions. It does not change the framework of the method. Our experiments already use three different checks: Lean proof checking, finite operation tables for refutations, and the numerical evaluation of NewtonBench. To avoid misunderstanding, the revised Appendix A adds more philosophical discussion. It states explicitly that the hidden structure represents derivability within the framework, not truth outside it.

Furthermore, uniform modelling of all agents and tasks is assumed. / How would you extend the model to non-uniform modelling of difficulty?

The paper does not assume identical agents. On the contrary, it is because agents differ that the first coarse-graining gives each agent its own distribution (Figure 2). Only the zeroth-order closure uses a common rate, in order to obtain a simple closed form. Differences between agents come from the higher orders of the expansion. The paper gives the general form of the first-order expansion. At first order, the rates of the agents add up, and only their average enters. The differences between agents, that is, the higher moments of their distribution, enter only at higher orders, for example through competition on the same question. For questions of different difficulty, the questions are grouped by difficulty type. The collective variables are recorded separately for each type. The form of the equations does not change.

This model is "append-only" on the hypergraph. Could it be extended easily to removals?

Removal can be added as another type of event. Removing a node also removes all results derived from it. It also affects the agents that are using it as a premise. This has many consequences. However, removal is a contraction or revision of accepted knowledge. It is a different class of process from the expansion of knowledge studied here (see the part on Lakatos in the revised Appendix A for this philosophical difference). This class of processes is outside the scope of the paper.

(continued in the next comment)

(continued)

The boundedness of the space of theorems appears to be assumed. / Is it possible to obtain a statistical physics model for unbounded numbers of open questions?

The model does not assume a bounded set of questions: the hidden hypergraph can be infinite, and a question enters the record only when it is posed. The fixed pool is a choice of experimental design that makes "the time to settle a given fraction of the questions" well defined. With unboundedly many questions, the number of questions open to research (the frontier) grows when questions are posed or their premises are settled, and shrinks when questions are settled; the paper gives this balance and a conditional growth law.

I think this work is not specific to AI / AutoResearch, but rather a specific case study using AI.

Indeed, the framework is not restricted to agents based on large language models: traditional automated systems and human researchers fit the same general framework, only with different phase-type distributions for different agents. This generality is what allows conclusions to transfer across agents and tasks. Compared with human researchers, AI agents can be copied in large numbers and measured repeatedly under the same conditions, and are easier to scale to very large numbers.

... the manuscript should attempt to find a use case, such as ... the Pareto-frontier of optimal numbers of agents ... Otherwise, I feel that the experimental insights are a little bit trivial.

The paper already gives the onset of competition N* and several scaling relations, which provide practical guidance. Below N*, adding agents shortens the time while the total agent time (number of agents times time) stays roughly constant; beyond N*, duplicated work makes the total agent time grow, time and total agent time trade off, and the optimal number of agents depends on the relative cost of waiting time and agent time. The revised version shows this trade-off in a new figure with N* marked (Appendix E), and a new paragraph in Section 7 explains how single-agent data decide, before deployment, how many agents to use.

Presentation-wise, Eq. (3) is unclear from the main text.

Eq. (3) states that a question can be solved by an agent only if some set of premises sufficient to derive it is entirely in that agent's knowledge base. The revised version rewrites this part.

What is the relation with infection models on graphs such as SIS models?

Connecting to physical models is exactly the approach of the paper, which identifies the coarse-grained dynamics as a catalytic reaction network (Section 4.2). It is not an ordinary infection model, for two main reasons: a node can grow only after several premises hold and are used by an agent; and in this growth model an accepted answer does not revert, so there is no recovery.

Can we obtain predictions from fits for how long it would take AI researcher communities to reach certain discoveries? ... predict when the singularity hits, or when a cure to a cancer type is found.

For a research goal such as a cure for a type of cancer, the required premises and the difficulty of each step cannot be measured in advance; our theory does not assess how far away such a single goal is, so we cannot give a date. What the theory does give is the law by which progress toward the goal develops: given an estimate measure on how far from the target on the hypergraph (for example, rely on the sub-steps of other cancer research data), the growth equations give the distribution of the hitting time needed to reach it. In the ETP experiments, where the hypergraph is known, the model computed from single-agent data alone the time for 1 to 256 agents to settle half of the open questions; over this range the measured time fell by a factor of about 40, and the model stayed within a factor of two of it at every N (Section 6.2).

As for when the singularity arrives, the model cannot judge how far away the singularity is. It can, however, describe the behavior near the singularity, since we expect the collective behavior there still to follow the dynamics of this paper, possibly with higher-order terms of the expansion. For example, if a discovery requires M results in sequence, each taking at least a time t_min (shorter times being practically impossible), adding agents alone cannot accelerate the discovery without limit; going faster requires each step itself to become faster. How fast each step is, and how this changes as accepted results accumulate, can be measured with a few agents. This is the role we hope the framework will play: before deploying, say, a billion agents on Earth, we would rather predict their collective behavior to some extent from the behavior of a few agents, to avoid irreversible consequences.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.