💬 NLP

High Volatility and Action Bias Distinguish LLMs from Humans in Group Coordination

This paper reveals that large language models struggle to achieve stable group coordination compared to humans, exhibiting excessive action switching and limited responsiveness to richer feedback, which highlights fundamental behavioral differences in adaptive learning and convergence strategies.

Original authors: Sahaj Singh Maini, Robert L. Goldstone, Zoran Tiganj

Published 2026-04-06

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sahaj Singh Maini, Robert L. Goldstone, Zoran Tiganj

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 part of a team playing a high-stakes guessing game. The goal is simple: everyone on the team must guess a number, and when you add all those numbers together, the total must match a secret "mystery number" chosen by the referee.

Here's the catch: You cannot talk to each other. You don't know what your teammates are guessing. You only get one piece of information after every round: "Too high," "Too low," or "Just right" (and sometimes, how much too high or too low).

This is the game the researchers used to test humans against Large Language Models (LLMs) like the ones powering this conversation. They wanted to see: Can AI teams coordinate as well as human teams?

The short answer? No. Not even close.

Here is the breakdown of what happened, using some everyday analogies.

1. The Human Team: The Calm Orchestra

When humans played this game, they acted like a seasoned jazz band.

They listened: If the group was "too high," they didn't all panic and jump down 50 points. They listened to the feedback and made small, calculated adjustments.
They found roles: As the game went on, some people naturally stopped changing their numbers. They became the "anchors," holding steady while others did the heavy lifting of adjusting. This reduced the noise and helped the group find the target quickly.
They learned: If they played the game 10 times in a row, they got better every time. They figured out the rhythm.

2. The AI Team: The Hyperactive Hamster Wheel

When the researchers swapped humans for AI models (like DeepSeek, Llama, and Gemini), the vibe changed completely. The AI teams acted less like a jazz band and more like a room full of hamsters on a wheel that's spinning too fast.

Here are the three main ways the AI teams failed to coordinate:

A. The "Action Bias" (The Hamster Effect)

Humans have a natural tendency to stay still if things are going well. If the group sum is close to the target, a human might think, "I'll just keep my number the same and let my teammates do the work."

The AI, however, has a massive Action Bias. It feels like it must do something every single turn. Even when the group is close to the target, the AI keeps changing its number.

Analogy: Imagine trying to balance a broom on your hand. If you make tiny, steady movements, you win. If you keep jerking your hand wildly left and right because you feel like you "need to move," the broom falls. The AI kept jerking the hand.

B. The "Overreaction" (The Pendulum Swing)

When the AI got feedback like "You are 10 points too high," it didn't just subtract 10. It often subtracted 20 or 30.

The Result: The group sum would swing wildly. One round they were way too high, the next round they were way too low, and the next round way too high again. They were stuck in an endless loop of swinging back and forth, never landing on the target.
Human vs. AI: Humans tend to be a bit cautious (under-reacting), which actually helps stabilize the group. AI tends to be aggressive (over-reacting), which creates chaos.

C. The "Amnesia" (No Learning)

This was the most surprising part. Humans got better the more they played. They learned, "Okay, when the group is this big, we need to be more careful."
The AI models? They didn't learn. Even after playing 10 games in a row, they performed just as poorly in Game 10 as they did in Game 1. They treated every new game as if it were the very first time they had ever seen the rules. They couldn't connect the dots from one experience to the next.

3. The "Richer Feedback" Surprise

The researchers tried giving the AI better information. Instead of just saying "Too high," they said, "Too high by 25."

For Humans: This was like getting a map. They used the extra detail to zoom in on the target and solved the game much faster.
For AI: It was like giving a map to someone who doesn't know how to read. The extra numbers didn't help them at all. They still swung wildly and failed to coordinate.

The Big Picture: Why Does This Matter?

We are starting to build systems where multiple AIs work together to solve complex problems (like writing code, managing supply chains, or running simulations).

This paper warns us: Just because AI is smart at answering questions doesn't mean it's good at working in a team.

Humans are great at "group coordination" because we know when to speak up and when to stay quiet. We adapt our behavior based on the group's mood.
Current AI is like a very eager, very loud intern who never stops talking, over-corrects every mistake, and forgets what happened yesterday.

The Takeaway

If you want to build a team of AIs that works like a human team, you can't just plug them in and hope for the best. You have to teach them restraint. You have to teach them that sometimes, the best move is to do nothing and let the team stabilize. Until AI learns the art of "strategic inaction," it will struggle to coordinate with us or with itself.

1. Problem Statement

As Large Language Models (LLMs) are increasingly deployed in Multi-Agent Systems (MAS) for collaborative tasks, a critical open question remains: Can LLMs demonstrate adaptive group coordination comparable to humans, and do they utilize similar strategies?

While humans excel at adaptive coordination in settings with incomplete information and limited communication (relying on dynamic strategies like role differentiation and self-consistency), LLMs often struggle in pure coordination games. The paper investigates the specific mechanisms behind these failures, focusing on whether LLMs can learn from experience, adapt to feedback, and stabilize collective behavior in a common-interest game with imperfect monitoring.

2. Methodology

The authors evaluated LLM performance against human baselines using the Group Binary Search (GBS) game.

Task Design (GBS):
- An $n$ -player game where participants must coordinate independent numerical guesses to collectively sum to a hidden "mystery number."
- Imperfect Monitoring: Players do not see individual guesses of others; they only receive aggregate group feedback.
- Feedback Types:
  1. Directional: "Too low," "Too high," or "Just right."
  2. Numerical: "Too low/high by $k$ " (providing error magnitude).
- Goal: Minimize the number of rounds to reach the target sum (max 15 rounds).
Experimental Setup:
- Models Tested: Deepseek-V3, Deepseek-V3.1-Terminus, Llama 3.3, and Gemini 2.0 Flash.
- Conditions:
  - Prompting: Zero-Shot vs. Zero-Shot Chain-of-Thought (CoT).
  - Group Sizes: Small (2–3), Medium (4–7), and Large (10–17 players).
  - Sessions: 10 sequential games per session, alternating feedback types.
- Human Baseline: Data was matched to a prior study (Roberts & Goldstone, 2011) to ensure direct comparability.
Metrics:
- Rounds to solution.
- Learning curves (slope of rounds vs. game index).
- Reactivity scaling (slope of group adjustment vs. feedback error).
- Switching dynamics (proportion of agents changing guesses between rounds).
- Stability (probability of an agent keeping the same guess).

3. Key Contributions

The paper provides a behaviorally grounded diagnostic for the coordination gap between humans and LLMs, moving beyond simple performance metrics to analyze how agents coordinate.

Identification of Action Bias: LLMs exhibit a persistent "action bias," where they are statistically biased to change their output every round, even when stability is optimal.
Mechanism-Level Diagnosis: The authors isolate specific failure modes:
- Excessive Switching: LLMs fail to stabilize roles, leading to high volatility.
- Overreactivity: LLMs adjust their guesses too aggressively in response to feedback, causing oscillations.
- Lack of Cross-Game Learning: Unlike humans, LLMs do not significantly improve performance across repeated games.
Feedback Utilization Gap: The study demonstrates that richer feedback (numerical error magnitude) significantly benefits humans but offers little to no advantage to LLMs, suggesting a fundamental difference in how they process quantitative error signals.

4. Key Results

A. Performance and Learning

Human Superiority: Humans consistently solved games in fewer rounds than LLMs, particularly in numerical feedback conditions.
Learning Failure: Human groups showed a clear negative slope in rounds-to-solution across games (indicating learning). LLM groups showed flat or even positive slopes (worsening performance), indicating a failure to adapt strategies over time.
Scaling Issues: While both humans and LLMs struggled as group size increased, the degradation was more pronounced for LLMs. In large groups (10–17 players), many LLM configurations hit the 15-round ceiling (failure).

B. Reactivity and Switching Dynamics

Overreactivity:
- Humans tend to underreact (slope $\approx -0.77$ ), making conservative adjustments that stabilize the group sum.
- LLMs consistently overreact (slope $\approx -1.39$ ), making adjustments larger than necessary, which leads to oscillations around the target number.
Switching Behavior:
- Humans: As the group converges, humans significantly reduce switching rates. In large groups, a subset of humans often adopts a "stay" strategy (keeping the same guess), effectively reducing the coordination burden.
- LLMs: LLMs maintain high switching rates throughout the game. No LLM agent ever achieved a "stay probability" of 1 (keeping the same guess for the entire game). They rarely stabilize, leading to persistent noise in the group sum.

C. Impact of Prompting and Feedback

Numerical Feedback: Humans leveraged numerical error magnitude to converge faster. LLMs showed inconsistent or negligible benefits from this richer information, sometimes performing worse than with directional feedback.
CoT Prompting: Zero-Shot CoT improved performance for some models (e.g., Deepseek-V3) but did not fundamentally alter the maladaptive patterns of overreactivity or excessive switching.
Heterogeneity: Mixing different LLM models in a group increased behavioral diversity but did not improve coordination performance, suggesting that simple diversity is insufficient without adaptive role specialization.

5. Significance and Implications

Architectural Limitations: The findings suggest that current LLM architectures lack the mechanisms for long-term strategic adaptation and collective stabilization required for effective multi-agent coordination. They struggle to integrate feedback history to refine behavior over time.
Action Bias: The "action bias" (the tendency to act/change rather than stay) is a critical barrier. In human teams, inaction can be a strategic choice to reduce noise; LLMs seem to equate "action" with "progress."
Human-AI Teaming: The results highlight a potential for complementary teaming. Humans' natural tendency to underreact and stabilize could counterbalance LLM overreactivity. However, current LLMs cannot autonomously develop the complementary roles (e.g., "stabilizers" vs. "adjusters") that human groups spontaneously form.
Future Directions: To close the coordination gap, future multi-agent systems need architectures capable of:
- Integrating long-term feedback history.
- Moderating reactivity to prevent oscillation.
- Explicitly modeling the value of "inaction" or role stability.

In conclusion, while LLMs can understand the rules of coordination games, they fail to replicate the adaptive, stabilizing, and role-differentiating strategies that allow human groups to converge efficiently, particularly as group size and task complexity increase.