TL;DR
This paper demonstrates that when sequential autonomous agents use Large Language Models (LLMs) to flag online hate speech, they inevitably fall into "information cascades" (herding), ignoring their own textual evidence to blindly copy previous agents. To combat this, the authors propose a stochastic control framework based on an optimal stopping-time formulation that strategically delays herding to preserve classification accuracy.
What Problem Does This Paper Address?
When multiple autonomous AI agents sequentially analyze data to make decisions—such as content moderation tools flagging toxic users on a social network—they typically only see the actions (e.g., flags) of previous agents, not their raw private observations (the specific text comments analyzed).
In classical social learning theory, this asymmetry leads to a phenomenon called herding. Once a few early agents take the same action, a powerful "public prior" forms. Subsequent agents will completely ignore their own high-quality private data because the public consensus mathematically outweighs it.
This paper addresses this challenge in the context of modern Language-Driven Decision-Making Agents (LDDAs). If the initial public belief is slightly wrong, the entire chain of LLM-driven agents will cascade into herding and systematically misclassify users, rendering the expensive LLM analysis useless.
Methods
The authors formalize the social learning protocol for LDDAs using a two-part architecture for each agent:
- LLM as a Noisy Sensor: Instead of making the final decision, the LLM (Mixtral-8x7B) parses high-dimensional text comments and maps them to a low-dimensional, discrete feature space (e.g., detecting indicators like violence, bias, or dehumanization).
- Neural Bayesian Engine: A Restricted Boltzmann Machine (RBM) parameterizes the likelihood of these features, and Bayes' rule updates the agent's internal belief to output an action that minimizes expected cost.
Mathematical Contributions
- Theorem 1 (Herding Proof): The authors mathematically prove that under standard sequential protocols, LDDAs will herd in finite time with a probability of 1.
- Theorem 2 (Mitigation via Stochastic Control): To delay herding, they introduce a "socialistic" cumulative cost function. This formulation balances privacy against collective accuracy, determining exactly when an agent should reveal its private observations to correct the public drift. They prove the optimal policy follows a threshold switching curve relative to public belief.
Datasets
The framework was validated using the Measuring Hate Speech Dataset (Sachdeva et al., 2022), which contains 40,000 annotated comments from Twitter, Reddit, and Gab. The dataset utilizes a Rasch measurement scale to score hate speech intensity on a scale from 1 to 5.
Key Insights and Ideas
- LLMs as Compressors, Not Deciders: A key conceptual shift is treating the LLM as a noisy feature extractor feeding into a classical Bayesian decision engine, rather than letting the LLM act as an end-to-end decision-maker.
- The Inevitability of Systemic Blindness: Even if you deploy state-of-the-art LLMs capable of deep textual understanding, if the agents only communicate binary flags, the system will eventually ignore the LLM's insights entirely once a public consensus forms.
- Socialistic Utility Over Opportunism: To prevent catastrophic herding, agents must occasionally act "benevolently" by following a threshold policy that prioritizes updating the public prior accurately over minimizing their own immediate opportunistic cost.
Key Results
The Reality of Herding
The authors simulated a sequence of 100 agents evaluating users of varying hate speech intensities. As shown in Figure 3, when the initial public prior strongly favors the "not a hate speech peddler" state (moving toward 1.0 on the x-axis), the agents completely ignore the true underlying state—even when evaluating users with the highest intensity of hate speech (Intensity 5)—and falsely output an average action of 0.
Delaying the Cascade
By implementing the threshold policy derived from their stochastic control framework, the authors successfully delayed herding. As illustrated in Figure 4, adjusting the threshold parameter ($\gamma$) forces agents to rely longer on their private LLM observations, effectively shrinking the black "herding zone" and keeping classification accurate even under strong opposing priors.
Limitations and Future Work
- Assumption of Cooperation: The mitigation strategy relies on a centralized, cooperative "socialistic" objective where agents are willing to optimize for collective welfare. In decentralized or commercial applications, agents might be purely self-interested.
- Future Directions: The authors note that future work should focus on designing game-theoretic incentive mechanisms that naturally encourage independent LLM agents to delay herding without requiring a forced global cost function.
Summary updated