For twenty years, scientists believed they understood how a locust swarm works.
The theory was elegant in its simplicity. Each locust, the thinking went, is essentially a self-propelled particle. It looks at the insects around it, calculates the average direction they are moving, and aligns itself with that flow. Multiply this simple rule across millions of individuals and the result is the terrifying, coordinated wall of movement that strips entire fields bare in minutes.
The theory came from physics. It was called the Vicsek model, named after the physicist who developed it in 1995, and it had been the foundation of collective behavior science for decades. It was elegant. It was mathematical. And in February 2025, a team at the University of Konstanz and the Max Planck Institute of Animal Behavior in Germany proved it was wrong.
Their tool for overturning twenty years of settled science was a virtual reality headset — designed not for humans, but for locusts.
Using a combination of field experiments and an innovative virtual reality system in which locust nymphs moved freely, interacting with virtual locusts while immersed in a 3D virtual environment, the researchers found that the insects don’t just follow their neighbors like self-propelled particles but instead rely on internal cognitive decision-making processes to navigate as a collective.
What they discovered inside that virtual world has implications that extend far beyond locust science — touching robotics, artificial intelligence, our understanding of animal intelligence, and ultimately, our ability to predict and prevent the swarm disasters that threaten the food security of hundreds of millions of people every year.
The Old Theory: Why Scientists Thought Locusts Were “Particles”
To understand why the 2025 discovery is significant, you need to understand what the old model said — and why it seemed so convincing for so long.
The Vicsek model, published in 1995, made a specific and testable prediction: when animals move in groups, each individual aligns its direction with the average direction of its nearest neighbors. The more densely packed the group, the more perfectly aligned the movement becomes. Below a critical density, the group moves in disordered, random directions. Above that critical density, there is a spontaneous phase transition — like water turning to ice — and the entire group snaps into coordinated, unified movement.
This model was not just theoretical. Collective motion, which is ubiquitous in nature, has traditionally been explained by self-propelled particle models from theoretical physics. It described starling murmurations, fish schools, and bacterial colonies. And it seemed to describe desert locust swarms perfectly — millions of juvenile insects marching in the same direction, a biological fluid moving across the landscape.
The appeal was the simplicity. No individual locust needed to understand the swarm. No individual needed intelligence or awareness. Each animal only needed to perform one calculation: what is the average direction of my neighbors? Then move that way. From this single dumb rule, the entire magnificent, catastrophic spectacle of a locust swarm could theoretically emerge.
About once per decade, millions of juvenile locusts collect in the deserts of East Africa and begin marching across the continent. The rules of how these insects move in swarms were mostly wrapped up, or so Iain Couzin of the Max Planck Institute of Animal Behavior and his colleagues thought.
Then they actually tested it — properly, rigorously, with a technology that had not existed when the old theories were developed — and the model collapsed.
The Experiment: A Locust in a Holographic World
The core challenge in studying locust swarm behavior has always been the same: you cannot control what a swarm does. In a natural swarm of millions of insects, there are too many variables, too much noise, and no way to isolate the contribution of any single behavioral rule.
Inferring the mechanism of interaction in mobile animal groups is notoriously difficult. Individuals both influence, and are influenced by, the behavior of others in a complex interplay.
The Konstanz team’s solution was to place individual locusts inside a virtual reality environment — a 3D immersive world where the “other locusts” were computer-generated and entirely under the researchers’ control. They could manipulate density, direction, and movement patterns of the virtual swarm with precision impossible in any natural setting.
The precise control of visual information afforded by virtual reality meant that the researchers could establish how sensory input is translated into movement decisions by locusts.
What they did inside that virtual environment was a series of carefully designed tests, each targeting a specific prediction of the Vicsek model.
Test 1 — The Density Test: According to the Vicsek model, locusts should start aligning with neighbors only above a critical density threshold. Below that threshold, they should move randomly. The team varied virtual swarm density from 1 to 64 locusts per square metre and watched what the real locust did.
Result: Sayin et al. found that locusts do not follow the fixed interaction rules assumed by traditional models, such as explicitly aligning with their moving neighbors when the density of the swarm increases. The alignment behavior showed no density dependence. The Vicsek prediction failed.
Test 2 — The Two-Swarm Test: The researchers presented a single real locust with two separate groups of virtual locusts, both marching in the same direction. According to the Vicsek model, the locust should calculate the average of both groups and align with that average direction — since they were both moving the same way, the locust should simply join the flow.
Rather than align with this uniform motion — as predicted in the Vicsek model — the real locust moved orthogonal to the flow until it had immersed itself in one swarm or the other. This non-Vicsek response showed no dependence on the density.
The locust did not average. It chose. It moved sideways — perpendicular to the flow — until it had placed itself inside one specific group rather than the other. It was not computing averages. It was making a decision about where to be.
Test 3 — The Visual Cue Test: The team tested whether locusts respond to the average movement of the entire visual field around them — as the Vicsek model predicts — or whether they respond selectively to specific, salient visual cues.
Alignment occurred in response to coherent visual cues, almost entirely independent of density. “It’s really about the quality of information, not the quantity,” says Sayin.
A locust does not count neighbors and compute averages. A locust watches for the most compelling visual signal in its environment — the most coherent, the most salient, the clearest directional cue — and moves toward it.
What Locusts Actually Do: The Cognitive Model
The experiment demolished the old model. But science does not simply tear things down — it must replace them with something that better explains the evidence.
Instead, locust marching behavior, across scales, can be explained by a minimal cognitive framework, which incorporates individuals’ neural representation of bearings to neighbors and internal consensus dynamics for making directional choices.
This is a profound shift in how scientists think about locust intelligence.
The new model is called a “ring attractor” cognitive framework — borrowed from neuroscience, specifically from the neural circuits that animals use for spatial navigation and orientation. In this model, each locust maintains an internal neural map of the positions and movements of nearby neighbors. This map is not a simple average. It is a weighted, dynamically updated representation of the locust’s visual environment.
Locusts track the single most salient visual cue — sometimes a distant landmark, sometimes a dense knot of peers — and sprint toward it. When salience drops below a cognitive threshold, the barrier collapses and the swarm flips.
When that internal representation reaches a consensus — when the locust’s neural map of its surroundings crosses a threshold of clarity about which direction is most compelling — the locust moves. Not toward the average of everything it sees. Toward the specific thing that its brain has identified as the most salient target.
“Locusts are not behaving like simple particles that align with one another,” says Couzin. “We realized that we need to model them as cognitive agents — processing their surroundings and making decisions about where to move next.”
This distinction — between averaging and choosing, between computing and deciding — is not merely semantic. It changes everything about how swarm behavior should be modeled, predicted, and ultimately disrupted.
The PLOS Paper Connection: What We Already Knew About Locust Collective Motion
The 2025 Science paper did not emerge from nowhere. It built on decades of foundational research — including a comprehensive 2015 review published in PLOS Computational Biology by Gil Ariel of Bar Ilan University and Amir Ayali of Tel Aviv University that examined every major attempt to mathematically model locust collective motion.
That review documented several important biological realities about locust swarming that the old Vicsek-type models struggled to account for:
The Pause-and-Go Pattern: Locust marching is not continuous. Individual locusts alternate between walking periods and standing still — sometimes as few as 10 percent of a band are marching at any given moment. The rest are pausing. This intermittent behavior creates complex dynamics that simple particle models cannot describe.
The Cannibalism Driver: One of the most surprising findings in locust research is that cannibalism plays a direct role in driving collective motion. Locusts in a marching band face a specific threat: the locusts behind them will eat them if they stop moving. The march is, in part, a flight from being consumed by your own swarm-mates. This biological reality is impossible to capture in a model where each locust is simply a particle following alignment rules.
Phase Transformation: A locust in its solitary phase is a shy, green-colored insect that avoids other locusts. A locust in its gregarious phase is brightly colored, strongly attracted to others, and capable of marching hundreds of kilometers. The transformation between these states — triggered by crowding and a surge in serotonin — is one of the most dramatic behavioral changes documented in any animal. Even a 30-minute exposure to a crowd is sufficient for the induced change in behavior to persist for 24 to 72 hours after re-isolation.
The 2025 VR experiments confirmed and extended these insights. The “pause-and-go” behavior that Ariel and Ayali identified as a fundamental feature of locust movement is directly connected to the cognitive decision-making framework that Sayin and Couzin documented. A locust pauses because its internal cognitive map has not yet reached consensus on where to go next. It moves when that consensus is achieved.
| Feature | Vicsek Model (1995) | Sayin et al. (2025) |
|---|---|---|
| Published | Physical Review Letters 1995 | Science Journal Feb 28 2025 |
| Core Idea | Locusts = self-propelled particles | Locusts = cognitive decision-makers |
| Movement Rule | Align with average of neighbors | Chase most salient visual cue |
| Density Effect | Critical threshold triggers alignment | No density dependence found |
| Intelligence | Zero — pure stimulus-response | Internal neural map + decision |
| Brain Involvement | Not considered | Ring attractor neural framework |
| Prediction Accuracy | Failed in VR tests | Explains all 3 VR test results |
| Used Since | 1995 — 30 years | 2025 — new model |
Why This Changes Everything: From Science to Strategy
The scientific community’s response to the 2025 paper has been described as a paradigm shift — a term used carefully in science for findings that genuinely reorganize the field.
The study represents a paradigm shift in swarm research. By providing fundamental new insights into how locust behavior results in devastating swarms, the Konstanz research may provide critical knowledge for improved locust control strategies, such as for effective modeling of swarm movement.
For the millions of farmers in India, East Africa, and the Middle East whose livelihoods are threatened by desert locust swarms, the implications are direct and practical.
Better Swarm Prediction: Current swarm forecasting models are built on the old Vicsek-type assumptions — that swarm direction is determined by density and average neighbor alignment. The 2025 findings show this is wrong. Swarms are actually driven by each locust chasing the most salient visual cue in its environment. A new generation of forecasting models built on the cognitive framework should be able to predict swarm direction and behavior more accurately — giving farmers and governments more precise advance warning.
New Disruption Strategies: If locusts align with average neighbors, disrupting a swarm means reducing density — an essentially impossible task once a swarm is formed. But if locusts track the most salient visual cue in their environment and sprint toward it, a different disruption strategy becomes theoretically possible: creating competing salient cues that confuse or redirect the swarm’s cognitive consensus.
Moreover, the consequences of these findings will likely extend beyond locusts to broader applications in understanding the coordination of motion in other species, as well as robotics, artificial intelligence and the study of collective intelligence. Swarm robotics and autonomous vehicle coordination, for example, may benefit from algorithms inspired by locusts’ highly effective cognitive strategies for collective motion.
The AI Connection: The cognitive framework that explains locust behavior — ring attractor neural dynamics, probabilistic decision-making based on salient sensory inputs — is directly applicable to autonomous vehicle coordination and swarm robotics. Engineers designing drone swarms that can navigate complex environments without central control are facing exactly the same challenge that evolution solved in the desert locust. The 2025 findings are already being cited in AI and robotics research.
What the VR Experiment Tells Us About Animal Intelligence
Perhaps the most philosophically significant aspect of the 2025 findings is what they imply about the nature of insect intelligence.
For decades, the dominant view of insects in general — and locusts specifically — was that they are essentially stimulus-response machines. Stimulus arrives, response occurs, with minimal processing in between. The Vicsek model embodied this view: the locust sees neighbors, computes average direction, moves. No memory. No weighing of options. No decision in any meaningful sense.
The VR experiment shows something different. A locust placed between two groups of virtual locusts moving in the same direction does not join the flow. It moves sideways — actively searching for the group it wants to be in — before choosing one and joining it. This is not a stimulus-response. This is evaluation, comparison, and selection.
Our findings challenge long-held beliefs about how order can emerge from disorder in animal collectives… Sayin et al. conclude that it is time to move beyond the conception of locusts and other organisms as moving particles behaving according to fixed spatiotemporal rules and to consider organisms as probabilistic decision-makers responding dynamically to their sensory environment.
“Probabilistic decision-makers responding dynamically to their sensory environment.” This is language that, until recently, was reserved for mammals and birds. It is now being applied to an insect with a brain roughly the size of a sesame seed.
This matters for how we think about pest management. A creature that makes decisions can potentially be deceived, misdirected, or confused in ways that a stimulus-response machine cannot. An animal that evaluates options before acting can potentially be given worse options — or made to believe that the most salient cue in its environment points away from the crops it is about to destroy.
The Swarm That Threatens One in Ten People
The scientific elegance of the 2025 findings exists alongside a very practical urgency.
Desert locust swarms are estimated to threaten the livelihood of one in ten people worldwide due to their impact on food security. A single swarm can cover 460 square miles. Each square kilometer of swarm contains between 40 and 80 million locusts. Each locust eats its own body weight — approximately two grams — in vegetation every single day.
In India alone, the 2019–2020 locust attack involved 276 swarms and destroyed approximately 3.75 lakh hectares of crops, with losses exceeding ₹100 crore. In East Africa, the 2019–2021 upsurge was the worst in 70 years for some countries, requiring emergency response operations across Ethiopia, Kenya, Somalia, Tanzania, Uganda, Eritrea, Djibouti, Sudan, South Sudan, and Yemen simultaneously.
Climate change is making this crisis worse. Desert locust swarms, which can cover vast areas and contain billions of individuals, serve as a prominent example of collective motion in nature. Warmer Indian Ocean temperatures increase cyclone frequency in the Arabian Sea, creating the wet breeding conditions that trigger population explosions. The regions most vulnerable to locust attack are simultaneously the regions most affected by climate change — a convergence that is tightening with each decade.
Understanding exactly how a swarm forms, moves, and makes collective decisions is therefore not an abstract scientific question. It is a food security question with direct bearing on hunger, poverty, and political stability across three continents.
The Research Gap: What We Still Do Not Know
Scientific honesty requires acknowledging what the 2025 findings did not resolve, alongside what they revealed.
The VR experiments were conducted with individual locusts interacting with virtual swarms. The leap from individual cognitive decision-making to the emergent behavior of a swarm containing billions of insects across hundreds of square kilometers is still not fully bridged. As Ariel and Ayali noted in their PLOS review, there is a persistent gap between individual-scale models and the macroscopic dynamics of real swarms.
“Nobody would expect the Vicsek model to describe such situations accurately,” says Alexandre Solon, who studies active matter at Sorbonne University, France. He says the Vicsek model is based on simplified assumptions and only describes universal properties, but it can predict the transition to collective motion in many cases.
| Country/Region | Year | Scale | Damage |
|---|---|---|---|
| India — Rajasthan/Gujarat | 2019–20 | 276 swarms | 3.75 lakh hectares, ₹100+ crore |
| East Africa | 2019–21 | Worst in 70 years | 10 countries affected simultaneously |
| Middle East — 1915 | 1915 | Historic plague | 536,000 tonnes of food destroyed |
| Africa (typical) | 2003–04 | Regional upsurge | $30M per nation in control costs |
| Global Risk | Ongoing | 60+ countries | 1 in 10 people’s livelihood threatened |
Other researchers have noted that the new cognitive framework, while more biologically accurate at the individual scale, has not yet been demonstrated to generate accurate predictions of large-scale swarm dynamics. Demonstrating that individual cognitive decision-making scales up correctly to million-strong swarms in the field remains the key validation challenge for the new model.
What the 2025 Science paper has definitively established is that the old model was wrong in its fundamental assumption. Locusts do not align with the average of their neighbors. That part is settled. What the right model looks like at scale — and whether it will prove practically useful for swarm prediction and management — will be the work of the next decade.
Frequently Asked Questions
Q: What did the 2025 Science paper actually prove about locust swarms?
The study by Sayin et al., published in Science on February 28, 2025, proved that desert locusts do not follow the “self-propelled particle” model that has been the dominant framework for collective behavior science since 1995. Using virtual reality experiments at the University of Konstanz and Max Planck Institute, the researchers showed that locusts do not align with the average direction of neighboring insects. Instead, they use an internal cognitive process to identify the most salient visual cue in their environment and move toward it — behaving as probabilistic decision-makers rather than as simple physical particles.
Q: What is the Vicsek model and why did scientists believe it for so long?
The Vicsek model, published in 1995, proposes that collective motion emerges when each individual aligns its direction with the average direction of nearby neighbors. Above a critical density, a spontaneous phase transition occurs and the entire group moves in a coordinated direction. The model is mathematically elegant and appeared to describe many natural collective behaviors including fish schools, bird flocks, and bacterial colonies. For locusts specifically, it seemed to fit field observations of swarm behavior. The 2025 VR experiments provided the first direct, controlled test of the model’s core assumption — and found it failed to predict actual locust behavior.
Q: If locusts make cognitive decisions, does that mean they are more intelligent than we thought?
The 2025 findings do not suggest that locusts are “intelligent” in the way humans use that word. What they show is that even an insect with a brain roughly the size of a sesame seed uses a more sophisticated decision-making process than simple stimulus-response alignment. Each locust maintains an internal neural representation of its visual environment, evaluates the salience of different cues, and makes a directional choice when its internal consensus reaches a threshold. This is more complex than averaging — but it operates through neural mechanisms, not conscious deliberation.
Q: How does this change locust control and prevention strategies?
The practical implications are significant but still emerging. Current swarm forecasting models are built on Vicsek-type assumptions about density-driven alignment — the 2025 findings suggest these models need revision. More importantly, if locusts track specific salient visual cues rather than average neighbor directions, disruption strategies that create competing or confusing visual signals become theoretically more promising than density-reduction approaches. Improved cognitive models may also produce more accurate swarm trajectory predictions — giving farmers and governments better advance warning of where a swarm will move next.
Q: What is the connection between the 2015 PLOS paper and the 2025 Science discovery?
The 2015 PLOS Computational Biology review by Ariel and Ayali comprehensively documented the limitations of existing locust swarm models, identifying specific biological behaviors — including the “pause-and-go” movement pattern and the role of cannibalism in driving collective motion — that Vicsek-type models could not account for. This review established that the existing framework was incomplete. The 2025 VR experiments provided the direct experimental evidence of what the correct framework should be, replacing the particle-alignment model with a cognitive decision-making model informed by neurobiology.
