🤖 AI

Attributing Emergence in Million-Agent Systems

This paper introduces a computationally efficient Aumann–Shapley path-integral method for attributing macro-level emergence in million-agent LLM systems, demonstrating through empirical analysis of Bluesky data and a theoretical scaling bias theorem that small-scale studies fundamentally misrepresent the structural drivers of nonlinear social phenomena compared to full-scale analysis.

Original authors: Ling Tang, Jilin Mei, Qian Chen, Qihan Ren, Linfeng Zhang, Quanshi Zhang, Jing Shao, Xia Hu, Dongrui Liu

Published 2026-05-13

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ling Tang, Jilin Mei, Qian Chen, Qihan Ren, Linfeng Zhang, Quanshi Zhang, Jing Shao, Xia Hu, Dongrui Liu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to figure out who is responsible for a massive, chaotic crowd event, like a sudden stampede or a viral trend that sweeps through a city of 1.6 million people.

For a long time, researchers studying these events using AI agents (computer programs that act like humans) have been stuck in a small room. They could only simulate about 1,000 people at a time. They would look at this tiny group, pick out the "loud" or "popular" people, and say, "Aha! These few people caused the whole thing!"

This paper argues that this approach is fundamentally broken. It's like trying to understand the weather of an entire continent by only looking at the temperature in your backyard.

Here is the breakdown of what the authors discovered, using simple analogies:

1. The "Backyard" Problem (Small vs. Big Scale)

The researchers tested this on a real social network called Bluesky, which had 1.67 million active users over two weeks. They looked at five different topics, from sports (The Masters golf tournament) to politics (Trump tariffs) and entertainment (WrestleMania).

The Small Room (N=100): When they simulated a tiny, biased sample of 100 people (which is what most current AI studies do), the math said: "The famous people with the most followers caused everything." The top 1% of users seemed to be the heroes or villains of the story.
The Big City (N=1.6 Million): When they ran the math on the full 1.6 million people, the story flipped completely. The famous people still mattered, but they were actually responsible for a tiny fraction of the action. The real drivers were the "long tail"—the millions of regular, everyday users who, when combined, created the massive wave of activity.

The Analogy: Imagine a stadium full of 1.6 million people cheering.

The Small View: You only look at the VIP box. You see a few celebrities clapping and conclude, "The celebrities are making the noise!"
The Big View: You look at the whole stadium. You realize the celebrities are barely making a sound compared to the roar of the 1.5 million regular fans in the stands. The "noise" comes from the crowd, not the VIPs.

2. The "Magic Scale" That Doesn't Work

You might think, "Okay, maybe the small group just needs to be multiplied by a magic number to fix the mistake."

The authors proved mathematically that this is impossible for complex, real-world situations.

The Analogy: If you are mixing a simple soup (linear), you can taste a spoonful and multiply the salt by 100 to guess the whole pot's flavor.
The Reality: Social phenomena are like a chemical reaction. If you mix a tiny drop of acid and a tiny drop of base, you get a tiny fizz. If you mix a whole bucket of each, you get an explosion. You cannot just "scale up" the tiny fizz to predict the explosion. The math shows that for complex social indicators (like panic, polarization, or viral cascades), there is no single "magic number" that can fix the small-group data to match the big-group reality.

3. The New "Super-Scanner" Tool

The reason researchers were stuck in the "Backyard" was that the old math tools were too slow. Calculating the true contribution of every single person in a group of 1.6 million used to take so long it was practically impossible (like trying to count every grain of sand on a beach one by one).

The authors built a new tool based on a mathematical concept called the Aumann–Shapley path integral.

The Analogy: Old tools were like a person walking the beach counting grains of sand one by one. It took forever and they could only count a small patch.
The New Tool: This is like a satellite taking a photo of the whole beach instantly. It can calculate the contribution of every single person in less than 10 milliseconds. It is 100,000 times faster than previous methods.

4. Why This Matters

The paper concludes that if you want to understand how large-scale social phenomena happen (like market crashes, viral misinformation, or political polarization), you must look at the full scale.

The Takeaway: You cannot learn about the behavior of a million people by studying a hundred. The "small group" answer isn't just a slightly different version of the truth; it is a structurally different lie. The "long tail" of regular people is almost always the engine of these massive events, not just the famous few.

In short: The paper says, "Stop looking at the VIP box to understand the concert. You need to look at the whole crowd, and we finally have a fast enough tool to do it."

Technical Summary: Attributing Emergence in Million-Agent Systems

Problem Statement

Large language model (LLM)-powered multi-agent systems (MAS) are increasingly used to simulate population-scale social phenomena such as polarization, information cascades, and market panics. A critical scientific challenge in these simulations is emergence attribution: determining which individual agents drove a macro-level phenomenon and when it shifted.

Existing attribution methods, primarily based on axiomatic Shapley values, face a fundamental scalability barrier. They require evaluating the system over many agent subsets, resulting in computational costs that scale combinatorially with the number of agents ( $N$ ). Consequently, prior work has been confined to small-scale systems ( $N \lesssim 10^3$ ). However, the social phenomena these systems aim to explain (e.g., online cascades, election dynamics) occur at scales of $N \ge 10^6$ . It remains an open question whether agent-level conclusions drawn from small-scale simulations survive at realistic, million-agent scales. Furthermore, it is unclear if the structural differences observed between small and large scales can be corrected via post-hoc rescaling.

Methodology

The authors propose a framework based on the Aumann–Shapley path-integral attribution adapted for LLM-powered MAS.

1. Theoretical Framework

The method satisfies the four Aumann–Shapley axioms (efficiency, symmetry, dummy, and linearity) for non-atomic games. Instead of the combinatorial subset evaluation required by standard Shapley values, the authors define the attribution of agent $i$ as the integrated marginal effect of the macro indicator $f$ as the agent's feature vector $z_i$ fades in from a baseline $z_0$ to its observed value along a straight path:
$\phi_i = \sum_{d=1}^D (z_{i,d} - z_{0,d}) \int_0^1 \frac{\partial f}{\partial z_{i,d}} \left( z_0 + \tau(z - z_0) \right) d\tau$
For specific analytic macro indicators (linear, multiplicative-saturating, variance, and Gini), the authors derive closed-form expressions that reduce the integration to explicit per-agent calculations, eliminating the need for numerical integration or sampling.

2. Attribution Scaling Bias Theorem

A central theoretical contribution is Theorem 1 (Attribution Scaling Bias), which establishes a structural dichotomy:

Linear Case: If the macro indicator is linear (decomposable additively over agents), small-scale and full-scale attributions differ only by a global rescaling factor.
Nonlinear Case: If the macro indicator is nonlinear (involving cross-agent interactions), no global rescaling factor exists that can reconcile small-scale attribution with full-scale attribution. The disagreement is structural, not merely a matter of magnitude.

3. Empirical Validation

The authors conducted a controlled cross-scale study using 14 days of public Bluesky data ( $N = 1,671,587$ active users).

Full Scale: Attribution was computed on the entire panel.
Small Scale: Attribution was computed on a visibility-biased convenience sample ( $N=102$ ), mimicking the constraints of current small-scale LLM-MAS studies.
Protocols: The study compared visibility-biased sampling against uniform random sampling to isolate the effect of sampling strategy.
Value Functions: Four analytic macro indicators were tested: linear additive, multiplicative-saturating (heat), population variance, and absolute Gini mean difference.

Key Results

1. Structural Flip in Attribution

The study reveals a "flip" in attribution structure when moving from small, biased samples to the full population:

Full Scale: The "long tail" (bottom 90% of users by follower count) and the "middle tier" jointly carry the majority of the attribution mass. For example, under the multiplicative-saturating function ( $f_{heat}$ ), the long tail carried 76.4% of the attribution on the full panel.
Biased Small Scale: The same axiomatic method applied to a visibility-biased sample of 102 users concentrates almost all attribution mass on the top-tier accounts (high-follower users). On the same data, the top 1% carried 11.3% of the attribution (a 3x increase over the full scale), while the long tail collapsed to 52.0%.
Random Control: A uniform random sample of the same size ( $N=102$ ) recovered the population fractions, confirming that the discrepancy is specific to the biased convenience samples used in practice, not small sample sizes in general.

2. Theoretical Impossibility of Correction

The empirical results align with Theorem 1. For nonlinear macro indicators (variance, Gini, heat), the residual error after optimal global rescaling was high ( $\epsilon \approx 0.13$ to $0.97$), proving that no single scalar can map small-scale results to full-scale results. Even for the linear baseline, while a rescaling factor theoretically exists, a significant "flip" (approx. 20 percentage points) still occurred due to the sampling structure, though it was theoretically correctable.

3. Computational Efficiency

The proposed Aumann–Shapley method scales linearly with $N$ .

Speed: It computes attribution for $N=10^6$ agents in under 10 milliseconds on a single CPU node.
Comparison: This is 4–5 orders of magnitude faster than sampled Shapley methods on the same hardware. It extends the feasible scale of axiomatic attribution by over three orders of magnitude (from $N \approx 10^3$ to $N \approx 10^6$ ).

Significance and Claims

The paper claims that full-scale attribution is a theoretical requirement, not merely a methodological choice, for any study involving nonlinear macro indicators in multi-agent systems.

Validity of Small-Scale Studies: The authors argue that conclusions of the form "agent type A drives phenomenon E" drawn from visibility-biased small panels are structurally unreliable when the underlying macro indicator is nonlinear. The "flip" means that small-scale studies may fundamentally misidentify the drivers of emergent phenomena.
Methodological Necessity: The paper does not claim that every study must be run at million-agent scale, but rather that any claim about causal drivers must rest either on analysis at the phenomenon's realistic scale or on evidence of cross-scale consistency.
Theoretical Limit: The Attribution Scaling Bias theorem provides a rigorous proof that post-hoc rescaling cannot fix the structural disagreement between small and large scales for nonlinear systems.

The work bridges the gap between axiomatic game theory and large-scale social simulation, providing the first systematic evidence that agent-level conclusions are not scale-invariant in the presence of nonlinear interactions and biased sampling.