🧬 biology

Emergence Requires No Miracle: Eligibility Is All You Need

This paper demonstrates that the eligibility trace is the singular, non-negotiable temporal mechanism required for emergent differentiation in reward-modulated spiking networks, as its removal abolishes order while its extension directly governs the degree of differentiation regardless of network scale.

Original authors: Yahua Ruan

Published 2026-08-06

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yahua Ruan

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The Great Mystery of How Order Grows from Chaos

Imagine you are watching a massive crowd of people, each acting completely on their own, shouting, dancing, and bumping into one another. To an outsider, it looks like pure chaos. Yet, if you wait long enough, you might see something amazing happen: the crowd spontaneously organizes into a parade, or a dance formation, or a complex pattern. This phenomenon, where simple local actions create a big, organized system-level order, is called emergence. It's the secret sauce behind everything from how ant colonies build bridges to how your brain learns to recognize a face.

For decades, scientists have wondered how this happens. The usual suspects are complex feedback loops, fierce competition, or carefully timed schedules. But what if the secret is much simpler? What if the only thing needed is a way for a tiny, local action to "stick around" just long enough to be noticed by a global signal? Think of it like leaving a note on a fridge. If you write a note and the fridge disappears instantly, no one sees it. But if the note stays there for a few hours, someone might eventually walk by, read it, and decide to act on it. In the world of artificial brains, this "note" is a memory trace, and the person reading it is a reward signal. This paper dives into that exact question: Is this simple memory bridge the only thing we need to make order emerge from chaos?

The Paper's Big Discovery: The "Eligibility" Magic Trick

In this study, researchers built a giant, simulated artificial brain made of 800 to 80,000 tiny "spiking" neurons (think of them as digital fireflies that flash when they get excited). They taught this brain using a method called Reward-Modulated STDP. In plain English, this means the neurons change their connections based on when they flash, but those changes only stick if a "reward" signal arrives later to say, "Good job!"

The researchers wanted to know: What is the absolute minimum ingredient required for this brain to spontaneously organize itself? To find out, they played a game of "remove and see." They took apart their brain model, removing one piece at a time to see if the magic of organization (called "emergent sparsification," where a few strong connections take over and the rest fade away) would still happen.

The Engine vs. The Brakes
The results were surprisingly clear. The team discovered that the system has three distinct roles:

The Engine: This is the eligibility trace. It's a short-term memory that keeps a record of what a neuron just did for about 20 steps.
The Brakes: These are things like Reward Prediction Error (RPE) and logistic saturation. They stop the system from going crazy or growing too fast.
The Turbocharger: This is multiplicative potentiation, which speeds things up but isn't strictly necessary at small scales.

The "Aha!" Moment
When the researchers removed the eligibility trace (the memory bridge), the magic died instantly. The brain stopped organizing. The connections stayed messy and random, just like they were at the very beginning. The paper states that without this memory, the local actions of the neurons evaporated before the global reward could ever see them. It's like trying to get a high-five from a friend who is standing 100 miles away; if you don't wait for them to get closer, the high-five never happens.

However, when they removed the other parts—the brakes or the turbocharger—the brain didn't just keep working; it often worked even better or faster. Removing the brakes made the organization explode (in a mathematical sense, the "Gini coefficient" jumped from 0.48 to 0.82), but the system never learned to stop or stabilize. Removing the turbocharger just made the process a bit slower at first, but the organization still happened.

The Verdict
The paper concludes that eligibility is all you need to start the engine of emergence. The memory trace is the "sine qua non"—the one thing that cannot be missing. The other components are just helpers or safety mechanisms.

Scaling Up: Bigger Brains Need Bigger Memories
The researchers didn't stop at small brains. They scaled their simulation up to 80,000 neurons (with 128 million connections!). They found that the rule stays the same: if you remove the memory trace, the big brain also fails to organize. But here's a cool twist: as the brain gets bigger, the "brakes" need to be adjusted. In a small brain, a short memory window is fine. But in a huge brain, the noise is different, so the "braking" timescale needs to be longer (scaling with the square root of the number of neurons). When they adjusted this, the big brain actually organized better than the small one, reaching a higher level of order.

How Sure Are They?
The author is very confident in these findings based on their simulations. They ran the experiments multiple times with different random starting points (seeds) and got the same result every time: without the eligibility trace, the effect vanishes. They tested this on networks ranging from 800 to 80,000 neurons and found the rule held true across all sizes. While they suggest this idea might apply to other systems like evolution or standard machine learning (like how AI learns from data), they are careful to say that for those other systems, this is currently a "structural analogy" or a prediction, not a proven fact. For their specific spiking neural network simulations, however, the evidence is solid: the memory bridge is the non-negotiable key to making order out of chaos.

In short, the paper argues that emergence doesn't require a miracle or a complex, mysterious force. It just requires a simple trick: remembering what you just did, long enough for the universe to tell you if it was a good idea.

Technical Summary: "Emergence Requires No Miracle: Eligibility Is All You Need"

Problem Statement

The paper addresses the mechanistic origins of emergent differentiation—the spontaneous appearance of system-level order (specifically weight sparsification) from local interactions in adaptive systems. While phenomena like emergent abilities in large language models or phase transitions in learning are often attributed to complex architectures, positive feedback loops, or carefully designed schedules, the precise minimal mechanism driving these transitions remains elusive. The author challenges the necessity of these complex explanations, proposing that a single, fundamental mechanism—temporal credit assignment—is both necessary and sufficient.

Methodology

The study employs a systematic ablation study within a Spiking Neural Network (SNN) framework to isolate the components responsible for emergent sparsification.

Network Architecture: Leaky Integrate-and-Fire (LIF) neurons with $N$ ranging from 800 to 80,000 neurons (up to 128M synapses). Connectivity is random ( $p=0.02$ ).
Learning Rule: Reward-Modulated Spike-Timing-Dependent Plasticity (R-STDP).
- Local Events: STDP updates ( $\Delta w_{STDP}$ ) based on pre/post spike timing.
- Global Signal: A scalar reward $R(t)$ derived from population activation.
- Binding Mechanism: Updates are accumulated in an eligibility trace ( $e_i$ ) which decays over time ( $\tau_e = 20$ steps). Weight updates occur only when the global reward multiplies this trace: $\Delta w_i = R(t) \cdot \alpha \cdot e_i(t)$ .
Experimental Design: The author tested five conditions across multiple scales ( $N=800, 8,000, 80,000$ $N = 800, 8, 000, 80, 000$ ) and seeds:
1. FULL: Complete system.
2. no_RPE: Removed Reward Prediction Error (replaced with state reward).
3. no_MULT: Removed multiplicative potentiation (replaced with additive LTP).
4. no_ELIG: Removed the eligibility trace (immediate modulation only).
5. no_LOG: Removed logistic saturation (unbounded potentiation).
Metrics:
- Gini Coefficient: Measures weight distribution inequality (sparsification).
- Self-Arrest: Reward signal $R \to 0$ indicating convergence.
- Attractor Convergence: Cosine similarity of weight vectors approaching 1.0.
- Scaling Law: Investigation of the relationship between the braking timescale ( $\tau_{exp}$ ) and network size ( $N$ ).

Key Contributions

Identification of the "Engine": The ablation study demonstrates that the eligibility trace is the sine qua non for emergent differentiation. Removing it reduces sparsification to zero (or reverses it), whereas removing other components (RPE, multiplicative rules, logistic bounds) either preserves or amplifies the effect.
Engine/Brake/Accelerant Framework: The paper categorizes system components into three distinct roles:
- Engine (Necessary): The eligibility trace (temporal credit assignment).
- Brakes (Limiting): RPE and logistic saturation. These prevent runaway differentiation but do not cause it.
- Accelerant (Scale-Dependent): Multiplicative STDP. It is non-essential at small scales but becomes necessary at large scales ( $N=80,000$ ) to overcome noise.
Scaling Law Discovery: The author identifies a scaling law for the braking timescale: $\tau_{exp} \propto \sqrt{N}$ . Without scaling the reward prediction error's memory window, larger networks converge prematurely to lower differentiation levels. Extending $\tau_{exp}$ from 20,000 to 200,000 at $N=80,000$ raised the equilibrium Gini from 0.365 to 0.584.
Cross-Substrate Unification: The author proposes that temporal credit assignment is a universal mechanism structurally isomorphic across substrates:
- Spiking Networks: Eligibility trace ( $e$ ) $\times$ Reward ( $R$ ).
- SGD: Momentum (first-moment EMA) $\times$ Learning Rate.
- Evolution: Genotype (heritability) $\times$ Selection Pressure.

Results

Necessity of Eligibility: At all scales ( $N=800$ to $80,000$), removing the eligibility trace ( $no\_ELIG$ ) abolished emergent sparsification ( $\Delta Gini \approx -0.006$ ), returning the system to a random initialization state.
Role of Brakes: Removing RPE or logistic saturation led to unbounded differentiation (Gini > 0.8) but prevented self-arrest ( $R$ remained high), indicating these components regulate rather than drive the process.
Scale Invariance: The "engine" (eligibility) is scale-invariant; its removal always nullifies the effect. However, the "accelerant" (multiplicative rule) transitions from non-essential (99% effect retained at $N=800$ ) to necessary (only 16–29% effect retained at $N=8,000$ and $80,000$).
Robustness: Five independent seeds at $N=800$ converged to a stable Gini of $0.509 \pm 0.020$ , with $R \to 0.001$ and cosine similarity $\to 1.0$ , confirming the existence of a fixed-point attractor.
$\tau$ -Scaling Verification: At $N=80,000$ , scaling the braking timescale $\tau_{exp}$ by $\sqrt{N}$ successfully prolonged the learning window, allowing the system to surpass the differentiation equilibrium of smaller networks.

Significance and Claims

The paper claims that emergent differentiation requires no "miracle" or complex architectural constraint; it requires only temporal credit assignment.

Minimal Sufficient Mechanism: The ability of local events (synaptic updates) to persist in a decaying memory trace long enough for a global evaluation signal (reward) to bind to them is the minimal sufficient condition for the emergence of order.
Causal Bridging: The eligibility trace acts as a "memory bridge" that solves the temporal binding problem. Without this bridge, local actions have no global consequences, and differentiation cannot accumulate.
Theoretical Unification: The author proposes that this mechanism is not unique to SNNs but is the structural invariant underlying adaptive systems in SGD (via momentum) and evolution (via genotype).
Limitations: The author modestly notes that these findings are based on synthetic spiking networks with random connectivity and scalar rewards. While the structural isomorphism to SGD and evolution is proposed, direct experimental verification in those specific substrates remains a future task. The claim is that the mechanism is universal, even if the specific implementation varies.

In summary, the paper argues that "Eligibility is all you need" for the engine of emergence, while other components serve as necessary regulators (brakes) or scale-dependent enhancers (accelerants).