💻 computer science

When Verification Is Late: Delay Floors and Placement Flips in Corrected Multi-Agent Systems

This paper establishes that in delayed multi-agent correction systems, verification latency imposes an inescapable accuracy floor determined by the Dottie number, creates a budget ceiling for performance gains, and reverses the optimal strategy from concentrating to spreading correction resources beyond a specific threshold.

Original authors: Igor Itkin

Published 2026-07-21

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Igor Itkin

Original paper licensed under CC BY 4.0 (https://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 a group of friends trying to solve a mystery together. They pass notes back and forth, sharing clues and theories. Usually, the more people you have and the more carefully they check each other's work, the closer the group gets to the truth. This is the basic idea behind "multi-agent systems," where many computer programs (agents) work together to solve problems. But there's a catch: checking the work takes time. If a friend takes a long time to verify a clue before passing it on, the whole group might start arguing in circles or getting confused. Scientists have long known that adding "verification" (checking the work) helps, but they haven't fully understood what happens when that checking is slow. They wondered: Is there a limit to how good the group can get, no matter how many checkers you add? And does it matter if you have one super-fast checker or many slower ones?

This paper, titled "When Verification Is Late," dives into that exact question. It treats a group of AI agents like a swarm of bees or a school of fish, where some agents are "correctors" designed to pull the group back to the truth if they start drifting. The authors use math to model what happens when these correctors are delayed—when they take a moment to think before acting. They discover some surprising rules about how to build these teams.

The "Too Late" Problem

Think of the AI swarm as a boat trying to stay in a straight line on a choppy lake. The "correctors" are like a rudder that steers the boat back to the center. If the rudder moves instantly, the boat stays straight no matter how rough the water gets. But in the real world, the rudder takes a split second to react. If the water is calm, a slow rudder is fine. But if the waves are big, that split-second delay can make the rudder steer the boat away from the center just as it's trying to fix it, causing the boat to wobble wildly.

The paper finds that this "wobble" creates a floor on how accurate the group can ever be. No matter how much money (budget) you spend on hiring more checkers or making them smarter, you cannot push the accuracy past this floor if the delay is there. It's like trying to fill a bucket that has a hole in the bottom; you can pour in as much water as you want, but the water level will never rise above a certain point.

The Magic Number: The Dottie Number

Here is the most playful part of the discovery. The authors calculated exactly what this "wobble limit" looks like. They found that the best possible performance happens when the strength of the correction and the delay multiply together to equal a very specific, weird number: 0.739085...

This number is famous in math as the Dottie number. It's the only number that, if you take the cosine of it, gives you back the same number (cos 0.739085... = 0.739085...). The paper proves that for these AI swarms, the "sweet spot" for efficiency is tied to this exact mathematical constant. If you try to make the correctors stronger than this sweet spot, the delay makes them so unstable that the whole system gets worse, not better.

The Great Flip: One Big Brain vs. Many Small Brains

The paper also changes the rules on how to spend your budget. Before this, the common sense was: "If you have a limited budget, put it all on one super-strong checker." This works perfectly if the checker is instant.

But once you add delay, the rules flip. The paper shows that there is a specific, calculated budget threshold (called $c_{flip}$ ). If you spend less than this amount, concentrating your money on a few strong agents is best. However, if you spend more than this specific threshold, it becomes better to spread your budget out over many weaker checkers rather than one strong one.

Imagine you are trying to keep a line of people walking straight.

No Delay: You hire one giant, super-strong person to push everyone back into line instantly. This works great.
With Delay: That giant person takes a second to think before pushing. By the time they push, the people have already moved, so the push knocks them over. Instead, it's better to hire a hundred small people who give tiny nudges. Even if they are a little slow, the fact that there are so many of them keeps the line steady without causing a crash.

The authors calculated this specific tipping point. If you ignore it and keep dumping money into one strong agent past this limit, you actually make the group's answers worse.

Testing the Theory

To make sure this wasn't just a math game, the authors tested their model against real data from a swarm of large language models (specifically a corrected version of the Qwen3.6-35B-A3B model). They simulated a scenario where the AI agents had to check facts.

The results were a strong match for the model's overall shape, but the real-world data had limitations. The model fit the trajectories well, confirming that the delayed-relaxation concept is a valid description of the system. However, the data was too noisy and averaged over many different questions to definitively isolate the specific delay mechanism on its own. Because the effects of delay and correction strength were mixed together in the averaged data, the researchers could not confirm the specific "flip" or the Dottie number constant directly from the real-world numbers alone. While the theory predicts a hard ceiling on accuracy due to delay, the data could not separately prove this specific effect in isolation; it simply showed that the model's general behavior aligned with the observed system dynamics.

The Bottom Line

The main takeaway for anyone building these AI teams is simple: Speed matters more than power.

If your verification process is slow, throwing more money at it won't help. In fact, it might hurt. There is a hard limit to how accurate your AI swarm can be, determined entirely by how long it takes to check the work. Once you hit that limit, the only way to improve is to make the checking faster, not stronger. And if you do have to spend a lot of money, don't put it all on one superstar; spread it out among many smaller helpers to keep the system stable, but only after you've crossed the specific budget threshold where this strategy becomes better.

The paper doesn't just suggest this; it proves it with mathematical formulas. While the real-world data was too complex to isolate every single variable perfectly, the theory provides a clear rule: Don't let your checkers be too strong if they are too slow, or the whole team will spin out of control.

Technical Summary: When Verification Is Late: Delay Floors and Placement Flips in Corrected Multi-Agent Systems

Problem Statement
The paper addresses the design of multi-agent systems (specifically those built on Large Language Models) where a swarm of unreliable agents is guided toward factual accuracy by a subset of "correctors" or verifiers. While previous literature has established that adding verification improves accuracy in delay-free settings, and other work has shown that verification delays can destabilize consensus, no prior framework has integrated these factors to answer a critical design question: Given a fixed budget for verification, how much correction should be purchased, and where should it be placed, when verification inherently takes time?

The authors model this as a pinned, delayed grounded-Laplacian consensus system. In this model, agents exchange beliefs, and correctors pull the system toward evidence only after a verification delay $\tau$ . The core challenge is that a stronger pull (higher corrector strength $w$ ) improves accuracy but also increases the risk of instability under delay, creating a trade-off that traditional delay-free placement theories cannot capture.

Methodology
The authors formulate the system dynamics as a linear stochastic delay differential equation (SDDE):
$\dot{x}(t) = -L_g x(t) - W x(t-\tau) + \xi(t)$
where $L_g$ is the grounded Laplacian (representing agent coupling and anchoring to evidence), $W$ is the diagonal matrix of corrector strengths, $\tau$ is the common verification delay, and $\xi(t)$ is white noise.

The primary metric for system performance is coherence ( $H$ ), defined as the trace of the stationary covariance matrix of the beliefs (equivalent to the squared $H_2$ norm from noise to belief). The design problem involves minimizing $H$ subject to a budget constraint on the corrector strengths.

Key analytical steps include:

Modal Reduction: Leveraging the eigenbasis of $L_g$ , the system is decoupled into $N$ independent scalar modes. The total coherence is the sum of the stationary variances of these modes.
Closed-Form Variance: The authors derive an exact closed-form expression for the stationary variance $V(a, w, \tau)$ of a single scalar mode with decay rate $a$ , strength $w$ , and delay $\tau$ . This expression unifies the delay-free penalty and the delay-induced instability.
Stability Analysis: They identify a hard stability cap $w_{max}(\tau)$ beyond which the system oscillates and loses a stationary distribution.
Optimization: The paper analyzes the budget-accuracy frontier, determining the optimal distribution of a fixed budget among correctors (concentrating on few strong agents vs. spreading across many) under the constraint of delay.

Key Contributions and Results

The Delay Floor (Accuracy Ceiling):
The paper proves that for any non-zero delay $\tau > 0$ , there exists a strictly positive lower bound on the achievable error (the "delay floor"), regardless of how large the verification budget is.
- Exact Constant: In the limit of small delay, the optimal strength-delay product $w^* \tau$ converges to the Dottie number $\xi^* \approx 0.739085$ (the unique root of $\cos \xi = \xi$ ).
- Implication: No amount of additional budget can drive accuracy below this floor. Once the budget reaches a specific threshold ( $B_{floor}$ ), further spending yields no accuracy gains; if the budget is mechanically converted into raw strength without clipping, it can actually degrade performance by pushing the system toward instability.
The Placement Flip:
In delay-free systems, optimal placement often concentrates budget on a few strong correctors (depending on cost law curvature). The authors show that delay fundamentally alters this rule.
- Mechanism: Delay imposes a "knee" in the cost law at a specific strength $w^*(\tau)$ . Beyond this strength, additional power is useless or harmful.
- Result: Past a closed-form budget threshold, the optimal strategy flips from concentrating resources to spreading them. It becomes strictly better to deploy many moderate correctors (each capped at $w^*(\tau)$ ) rather than a few extremely strong ones.
Finite Budget Ceiling:
Unlike delay-free theories where coherence improves monotonically with budget, the delay-constrained system has a finite budget ceiling ( $B_{floor}$ ). Targets requiring accuracy below the delay floor are mathematically infeasible.
Commutator Condition for Exactness:
The paper establishes that the modal reduction (summing independent variances) is exact if and only if the corrector matrix $W$ commutes with the grounded Laplacian $L_g$ ( $[L_g, W] = 0$ ). For non-commuting placements, the modal reduction provides a first-order approximation with a known error bound, though the stability boundary estimates can be optimistic for non-uniform placements.
Empirical Calibration:
The model is calibrated against data from a corrected Qwen3.6-35B-A3B swarm. The results show:
- The delayed-relaxation model fits the observed belief trajectories well ( $R^2 = 0.94–0.99$ ).
- Longer verification delays correlate with slower returns to truth.
- Identifiability Limit: The data confirms that delay ( $\tau$ ) and correction strength ( $w$ ) are not separately identifiable from averaged trajectories; they collapse into a single effective rate. This aligns with theoretical predictions for this model family.

Significance and Claims
The paper claims to resolve a gap between two existing literatures: delay-free placement theory (which ignores delay) and delay stability theory (which ignores budget constraints).

Practical Message: In corrected multi-agent LLM systems, verification latency, not verification power, sets the ultimate limit on achievable accuracy.
Design Guidance: Designers must avoid the instinct to "stack" more verification power indefinitely. Once the budget exceeds the delay-dependent threshold, the strategy must shift from strengthening individual agents to distributing the budget across more agents, while strictly capping individual strength to avoid destabilization.
Theoretical Contribution: The work provides the first closed-form characterization of the "delay floor" in grounded networks, identifying the Dottie number as the fundamental constant governing the trade-off between correction strength and verification latency.

The authors remain modest regarding limitations, noting that the linear model describes relaxation around a plateau rather than full nonlinear dynamics, and that the cost laws and graph topologies used are stylized. They also acknowledge that while the model fits specific conditions, it cannot disentangle delay and strength from averaged data without per-question fitting.