In the burgeoning field of AI-driven social simulation, we often treat Large Language Model (LLM) agents as blank slates, ready to be swayed by whatever "influence campaign" we throw at them. However, a new paper titled "Competing Influence in LLM Populations: Budget Timing and Baseline Drift" suggests that these digital societies are far from neutral. They have their own internal compasses, and the "truth" of a claim matters just as much as the budget spent arguing for it.
TL;DR
This study reveals that LLM populations exhibit "baseline drift"—a natural tendency to move toward certain opinions even without external influence—which can significantly distort the perceived success of a campaign. Furthermore, unlike traditional numerical models, strategic influence in LLM societies lacks a stable "best" timing strategy, creating a complex game of cat-and-mouse between competing actors.
What Problem Does This Paper Address?
Researchers are increasingly using LLMs to simulate social dynamics, misinformation spread, and the effectiveness of interventions like fact-checking. Traditionally, these simulations relied on "opinion dynamics" models where opinions are simple numbers (scalars) that change based on fixed mathematical rules.
LLMs bring natural language and "personas" into the mix, but they also bring baggage: model-specific biases and an inherent sense of "epistemic standing" (i.e., they know what is generally considered true or false). The authors identify a critical gap: we don't fully understand how these model-intrinsic tendencies interact with strategic campaigns. If a population moves toward a certain view, is it because of a persuasive campaign, or is the model just "drifting" toward its own internal attractor?
Methods
The researchers built a simulation of 30 persona-conditioned LLM agents (using Claude Sonnet) arranged in a social network. These agents exchanged opinions in natural language over multiple rounds.
Key components included:
- Competing Agents: A "proponent" (arguing for a claim) and a "challenger" (arguing against it) with equal, finite budgets to send messages to specific agents.
- Baseline Drift: They ran "Null" conditions with no campaigns to see how opinions evolved naturally.
- Timing Policies: They tested three spending schedules: Early (front-loading messages), Sustained (steady flow), and Late (back-loading).
- Stance Probe: A separate, non-participating LLM (Claude Haiku) acted as a "thermometer," converting the agents' text opinions into numerical values (-1 to +1) for analysis without influencing the simulation.
Datasets
The study utilized eight specific claims categorized by their "epistemic status," ranging from clearly false to well-supported:
They also used 20 different "follow networks" to ensure results weren't tied to a specific social structure and assigned agents diverse personas (age, occupation, news habits).
Key Insights and Ideas
- The "Truth" Pulls Hard: Populations naturally drift toward the epistemic standing of a claim. They move toward rejecting false claims and accepting well-supported ones even without any campaign. On some topics, this "baseline drift" accounted for 40% of the observed change.
- LLMs Can Refuse to Lie: In a fascinating display of "side fidelity" issues, LLM campaign agents sometimes ignored their instructions if the brief was too egregious. For the "MMR causes autism" claim, the proponent agent—tasked with arguing for the claim—actually generated debunking messages 97% of the time.
- Diversity vs. Direction: Even when a population's average stance didn't move much, they almost always lost "stance diversity." The agents tended to converge toward a common middle ground or a model-favored position, regardless of the topic.
- No "Speak Last" Advantage: In classical numerical models, the agent who speaks last usually wins (the "late-mover advantage"). In LLM populations, this isn't true. The payoff structure is cyclic: the proponent wants to avoid the challenger's timing, while the challenger wants to match it.
Key Results
- Drift Correction is Essential: Figure 1 shows that once you subtract the natural "drift," the actual advantage of a campaign is often smaller than it looks. On false claims, even the "proponent" often ends up helping the challenger because the model's internal knowledge base is so strong.
- Persistence of Influence: Unlike some numerical models where effects fade instantly, campaign-induced shifts in LLM populations remained stable even during "quiet rounds" after the budget was exhausted.
- Community Impact: In segregated networks (echo chambers), campaigns were more effective at reducing the gap between communities than reducing the diversity within them.
Limitations and Future Work
The authors acknowledge several limitations:
- Scale: The population was small (30 agents), which may hide subtler effects.
- Public vs. Private: Agents only had one "opinion" which was both their internal belief and their public post. Real-world strategic silence or "preference falsification" wasn't modeled.
- Human Validation: While the stance probe showed high agreement between different LLM models, these digital opinions haven't been validated against how humans would perceive the same text.
- Targeting: Campaign messages were sent to random agents rather than using sophisticated targeting (e.g., targeting influencers or the "undecided").
Future work could explore adaptive budget allocation—where agents decide when and who to target based on real-time feedback—and larger-scale simulations to see if these "artificial society" regularities hold up.
Summary updated