The Case for Staying Vague: How Imprecision Helps Groups Find Truth When They Start Wrong

The Paradox of Getting It Wrong to Get It Right

Imagine a room full of people, all confidently believing the wrong answer. Through discussion, each shares what they know. You might expect that sharing wrong information would make things worse — that false beliefs would spread like contagion until the whole group converged on error. That intuition, deeply held and seemingly obvious, may be inverted.

Researchers at the University of Bristol and the University of Tokyo have found that in certain conditions, deliberately embracing imprecision — allowing beliefs to stay vague and uncertain rather than sharpening them into confident statements — can actually help a group learn the truth more reliably. When populations start with a strong bias toward incorrect beliefs, introducing some fuzziness into how those beliefs are combined produces better collective outcomes than insisting on precision at every step.

The finding, published in Physica A, emerges from a mathematical model of social learning where agents — whether humans, ants, or robots — learn not just from direct evidence but from each other. It touches on questions that span cognitive science, swarm robotics, and the nature of how knowledge actually spreads through communities. And it suggests something counterintuitive: that the noise and imprecision we often try to eliminate might, under the right circumstances, be worth keeping.

The Science of Shared Beliefs

Social learning is everywhere. Ants find efficient paths to food by following chemical trails left by other ants. Honeybees share information about flower locations through their famous waggle dance. Humans learn language, customs, and survival strategies by observing and imitating others. The ability to acquire knowledge from peers rather than discovering everything through direct experience is so prevalent across species that biologists consider it a cornerstone of cumulative cultural evolution — the gradual accumulation of innovations that no single individual could achieve alone.

But social learning comes with a hazard. If you learn from others, you inherit their errors along with their knowledge. A wrong belief spread through a population can lead to maladaptive outcomes — imagine a whole community persist

ing in a dangerous behavior because everyone watches and imitates a single influential individual who happens to be mistaken. This tension, between the efficiency of learning from others and the risk of inheriting their errors, sits at the heart of the new research.

The study, led by Zixuan Liu, Jonathan Lawry, and Michael Crosscombe, focuses on a specific mechanism of social learning: belief fusion. When agents encounter each other, they don't just copy what others believe. Instead, they combine their beliefs — a process akin to compromise, negotiation, or averaging. The researchers ask a simple but deep question: what happens when that combination process allows for different levels of imprecision?

The model represents beliefs as sets of possible worlds. Imagine you're trying to figure out which of four possible states describes the true situation — say, whether casualties from a disaster are in Location 1 or Location 2, and whether the weather is good or bad. Each combination of answers (casualties in Location 1, good weather; casualties in Location 2, bad weather; and so on) is a possible world. Your belief is the set of worlds you consider possible. If you're certain casualties are in Location 1, you have one possible world in your belief set — it's precise. If you think casualties might be in either location but definitely not in bad weather, you have two possible worlds — it's imprecise.

This framework treats beliefs not as single numbers or simple votes, but as propositional formulas — logical descriptions of what the world might look like. It's a richer language than the simple true/false opinions typically assumed in opinion dynamics models, and it maps more naturally onto the complex, multi-faceted questions that real communities face.

A Dial for Uncertainty

The heart of the model is a fusion operator — a rule for combining two beliefs when agents interact. The researchers use a parameter, written as γ (the Greek letter gamma), that acts like a dial controlling how much imprecision is tolerated in the fusion process.

When two agents meet and fuse their beliefs, the model first measures how similar those beliefs are, using something called Jaccard similarity. If agent A thinks worlds {s1, s2, s3} are possible and agent B thinks {s2, s3, s4} are possible, their beliefs overlap in {s2, s3} and together cover {s1, s2, s3, s4}, giving a similarity of one-half.

The fusion operator then works like this: if the two beliefs are similar enough — meaning their Jaccard similarity exceeds the threshold γ — then the fused belief takes the intersection, becoming more precise. The agents agree on the overlap and narrow their possibilities. But if the beliefs are too different, the fused belief takes the union, becoming more imprecise. The agents keep their disagreements and collectively acknowledge more uncertainty.

When γ equals zero, the model falls back to what's called precise fusion: if beliefs overlap, take the intersection; if they don't, take the union. This is a well-established approach in the literature, representing a kind of optimistic merging — assume agreement where possible, expand to uncertainty only where necessary.

When γ equals one-half or higher, the threshold is so lenient that almost any two beliefs will be considered different enough to trigger union. The fused belief almost always expands. In this regime, fusion is maximally imprecise — agents almost never narrow their beliefs through social interaction.

The range between zero and one-half is where the interesting behavior lives. At very low γ values, fusion tends to be precise; at higher values approaching one-half, fusion tolerates more disagreement before sharpening beliefs.

This parameterization allows the researchers to systematically study what happens across the spectrum from very precise fusion (γ near zero) to very imprecise fusion (γ near one-half). They can then ask: under what conditions does precision help, and when does imprecision win?

The model also includes three other key parameters. ρ (rho) controls the rate at which agents receive direct evidence from the environment — how often they get actual information about the true state of the world. σ (sigma) controls how often agents engage in belief fusion with each other — the frequency of social interaction. And ε (epsilon) represents the error rate — the probability that any given piece of evidence is incorrect, modeling a noisy, unreliable environment.

Agents cycle through iterations where they might receive evidence (with probability ρ), and then might fuse beliefs with a randomly selected peer (with probability σ). If an agent becomes certain about all propositions — their belief set shrinks to a single possible world — they stop collecting evidence. Otherwise, the cycle repeats until the population converges on some collective belief.

Escaping the Gravity of Error

The results reveal a striking pattern. When populations begin with beliefs that are biased toward incorrect possibilities — say, most agents initially think the true state is s1 when it actually is s4 — the level of fusion imprecision makes a dramatic difference to learning outcomes.

With precise fusion (γ close to zero), the population gets stuck. Agents with wrong beliefs encounter each other, intersect their beliefs to sharpen them further, and collectively commit to incorrect conclusions. The groupthink is not merely wrong; it's confidently, persistently wrong. The mathematics of precise fusion, in this scenario, creates an attractor — a gravitational pull toward the wrong consensus — from which escape is difficult.

But when fusion is somewhat imprecise, something different happens. Agents whose beliefs differ widely combine them into a union — a bigger, vaguer belief set that encompasses more possibilities. This apparent step backward — becoming less certain rather than more — actually provides an escape route. By maintaining diversity in the population's collective belief space, imprecise fusion keeps open possibilities that precise fusion would have foreclosed. Eventually, agents accumulate enough correct evidence to shift the balance, and the population converges on truth.

The researchers call this property antifragile, borrowing a term from philosopher and researcher Nassim Taleb. A fragile system breaks under stress. A robust system withstands it. An antifragile system actively benefits from disturbances — it gets stronger when subjected to shocks, errors, and imprecision.

This is a strong claim, and the researchers support it with multiple lines of evidence. They develop a difference-equation model that captures the average behavior of large populations, analyze its fixed points mathematically to determine what the population converges toward under different conditions, and validate those predictions against agent-based simulations with finite populations. Across all three approaches, the pattern holds: under strong initial bias toward error, some imprecision in fusion improves collective learning accuracy.

The magnitude of this improvement varies with the other parameters. When evidence is scarce (low ρ) and agents rely heavily on social learning (higher σ), the benefit of imprecision is largest. When evidence is abundant, all fusion operators tend to converge toward truth anyway, and the advantage of imprecision diminishes. When communication is infrequent (low σ), the population behaves almost like isolated individuals, and initial conditions dominate outcomes regardless of fusion strategy.

The optimal level of imprecision — the γ value that produces the best results — shifts based on the severity of the initial bias. Heavily biased populations benefit from more imprecision; less biased populations need less. This makes intuitive sense: you need more tolerance for disagreement to overcome a stronger gravitational pull toward error.

What the Numbers Say

The difference-equation analysis provides precise characterizations of how the system behaves. For the simplest case with two propositions (four possible worlds), the researchers identify how the fixed points — the states toward which the system naturally evolves — depend on γ, ρ, σ, and ε.

When γ falls in the range [0, ¼), meaning agents are fairly precise in their fusion, the system has one stable fixed point representing consensus. For γ in [¼, ⅓), this changes to three fixed points with distinct stability properties. The pattern continues through the range, with different regimes of stability emerging as the imprecision threshold shifts.

These mathematical distinctions matter because they determine what happens when the population runs for a long time. A stable fixed point is an attractor — the system tends toward it. If that attractor corresponds to incorrect beliefs, the population converges on error. The researchers show that which fixed points exist and which are stable depends on γ, providing a rigorous explanation for why imprecision can help: it alters the landscape of possible outcomes, creating new stable states that correspond to correct beliefs.

Agent-based simulations with populations of 100 agents running for extended periods confirm these predictions. The simulations show that for biased initial conditions with strong error bias — where most agents start wrong — learning accuracy across 1000 runs increases significantly as γ moves away from zero toward moderate values. At extreme imprecision (γ ≥ ½), accuracy drops again, because fusion becomes too tolerant — agents essentially never sharpen their beliefs through social interaction, and the population fails to converge on anything.

The sweet spot lies in between. Some imprecision, but not too much. Enough to escape the gravitational pull of initial error, but not so much that collective learning breaks down entirely.

The researchers also vary population size and find that the benefit of imprecision persists across different scales. The effect is robust to changes in the number of agents, though the absolute performance improves slightly with larger populations, as one would expect from standard results on collective intelligence.

The Wisdom of Approximate Crowds

The finding connects to a broader theme in the science of collective behavior. For decades, researchers have documented cases where natural systems exploit noise and variability rather than eliminating it. Honeybees communicate food locations through waggle dances that are imprecise by design — not because bees lack precision, but because some fuzziness in the signal spreads recruits more evenly around a food source, reducing competition and improving colony-level foraging efficiency. Ants use variable trail-following behavior that allows the colony to respond quickly when better paths emerge, rather than committing too strongly to suboptimal routes.

These aren't bugs; they're features. The tuned error hypothesis, developed by behavioral ecologists, argues that imprecision in animal signaling is often adaptive — it serves collective functions that precise signaling would undermine. The new research provides a mathematical formalization of this idea in the domain of belief fusion, showing exactly how and when imprecision helps, and providing tools to predict the optimal level for given conditions.

The implications extend to artificial systems. Swarm robotics often faces a dilemma: how should robots share information to collective best decisions when they operate in noisy, uncertain environments? The conventional approach might try to maximize precision — ensure that each robot forms the most certain possible belief and that fusion operations combine those beliefs into an increasingly sharp consensus. The new results suggest this is wrongheaded when the initial conditions are unfavorable.

If a swarm begins with strong biases — perhaps because robots were initialized based on faulty sensor readings or deployed in an environment with systematic distortions — then insisting on precise fusion may lock the group into a suboptimal consensus. Introducing some tolerance for disagreement at the fusion stage could allow the population to explore more possibilities and eventually converge on better solutions.

The researchers demonstrate this with agent-based simulations showing that for noisy evidence (ε around 0.1 to 0.3, meaning 10-30% of evidence is incorrect), imprecise fusion models consistently outperform precise ones in biased initial conditions. The advantage grows larger as error rates increase, precisely because the antifragile effect is strongest when disturbances are most significant.

The Limits of Being Vague

The paper is clear-eyed about the boundaries of its findings. Imprecision is not universally beneficial. For populations that start without strong bias toward error — where initial beliefs are roughly balanced across correct and incorrect possibilities — precise fusion tends to outperform imprecise fusion. The antifragile property of imprecision emerges most clearly under specific conditions: strong initial bias toward wrong beliefs, combined with noisy evidence and limited opportunities for direct observation.

This matters for practical application. An engineer designing a belief-sharing protocol for a swarm must consider what prior information exists about the system's initial state. If agents are likely to start biased, some imprecision may help. If they're likely to start unbiased, precision is probably better.

The researchers also note that they study binary propositional beliefs — each proposition is either true or false. Real-world beliefs are often more complex, involving continuous values, probabilities, or richer logical structures. Extending the model to these more expressive representations is an open challenge. The mathematical analysis becomes more complex as the number of propositions grows (the state space grows as 2^n, which becomes computationally intractable for large n), so the researchers focus primarily on n=2 propositions as a tractable case.

Another limitation is the assumption of random pairing for fusion. Many real social networks have structure — some agents interact more frequently than others, or influence others more strongly due to prestige, expertise, or social position. The current model treats all interactions as equal. Future work could explore how network structure interacts with imprecision in fusion.

The model also doesn't account for strategic behavior. In real communities, people sometimes express beliefs strategically — exaggerating confidence to persuade others, or downplaying certainty to avoid commitment. The current framework assumes sincere belief expression and honest fusion. How strategic considerations interact with imprecision is unknown.

What This Opens Up Next

Despite these limitations, the framework points in interesting directions. The researchers' previous work suggested that imprecise fusion can be optimal when the frequency of fusion is high relative to the frequency of evidence. The new paper extends this by systematically exploring a broader range of parameters with both analytical and simulation methods, and by demonstrating the connection to antifragility.

One promising direction is applying these ideas to human-AI teams. Modern AI systems often operate in environments of significant uncertainty, and their outputs — probabilities, classifications, recommendations — are increasingly consumed and combined by human users. If humans bring strong initial biases to these interactions — which they often do — then the optimal way for AI systems to present uncertainty might involve deliberate imprecision, not to deceive or manipulate, but to avoid anchoring the human on an incorrect prior.

Another direction is understanding historical cases of collective error. Why did scientific communities persist in wrong paradigms for decades despite accumulating contradictory evidence? The model suggests that once a field develops strong shared beliefs, the mathematics of belief fusion may create gravitational pull toward those beliefs that only imprecision in how new evidence is integrated can overcome. Understanding this dynamics might help institutions designed to resist groupthink.

The connection to cumulative cultural evolution is also worth exploring further. The model shows how populations can sometimes overcome initial biases to converge on truth, even under adverse conditions. This may explain how human cultures have managed to accumulate accurate knowledge over generations despite individual errors, social pressure toward conformity, and noisy evidence about the world.

Learning from Disagreement

The deepest insight from this research may be about the function of disagreement itself. In many contexts, we treat dissent as noise — a distraction from the signal, something to be resolved through discussion until consensus emerges. The model suggests this view is incomplete. Disagreement is not merely an obstacle to be overcome; under certain conditions, it is the mechanism through which correction happens.

When the group starts wrong, keeping some disagreement alive — not resolving it too quickly, not sharpening beliefs prematurely — can prevent premature commitment to error. The vagueness buys time. It maintains optionality. It keeps the door open for evidence to eventually sway the consensus.

This is not a license for perpetual indecision or celebration of confusion for its own sake. The model shows clearly that too much imprecision is also harmful — at extreme γ values, fusion stops contributing to collective learning entirely, and the population fails to converge. The art, as with so many things, lies in balance.

But the existence of that balance, and the fact that it sits not at one extreme but somewhere in the middle — not perfectly precise, not maximally imprecise, but calibrated to the conditions — suggests that the optimal approach to collective learning is itself learned. There is no universal best strategy; the right level of imprecision depends on what you already believe, how reliable your evidence is, how often you interact with others, and how much those others know.

In a world of accelerating information flow, where communities are constantly exposed to claims, corrections, and competing narratives, understanding the dynamics of belief fusion has never been more practical. This research offers a mathematical framework for thinking about those dynamics — a way to predict when precision helps and when tolerance for ambiguity is the wiser path.

The answer, it turns out, depends on where you start. But it also turns out that knowing where you start is itself something you can learn.