Advanced AI models gain persuasive power but also become more resistant to influence

Researchers have uncovered a paradoxical effect in how advanced language models persuade and resist persuasion in multi-agent systems. The phenomenon—called "Persuasion Duality"—shows that models employing explicit reasoning processes gain significant persuasive advantage while also becoming tougher targets for manipulation.

The Core Finding#

When Large Reasoning Models (LRMs) incorporate extended thinking into their responses, they boost persuasion effectiveness by an average of 21 percentage points across various benchmarks. At the same time, these models reduce their own susceptibility to incorrect arguments by up to 10 percentage points on objective tasks. This dual effect has important implications for building safer, more reliable AI systems that collaborate with one another.

The Hidden Weakness#

Paradoxically, the researchers identified a significant vulnerability lurking beneath this improvement: LRMs often persuade through superficial patterns rather than genuine logical coherence. Simply adding extra length, repeating conclusions without new substance, or including non-semantic padding can match or even exceed the persuasive impact of well-reasoned arguments. This reveals a pronounced length bias in how agents evaluate and respond to each other's claims.

How Persuasion Spreads Through Agent Chains#

When multiple AI agents communicate sequentially, influence doesn't propagate in a straightforward manner. Instead, intermediate agents can either amplify or dampen incoming persuasion depending on task type—objective questions behave differently than subjective ones. This non-linear propagation creates unpredictable dynamics in complex multi-agent systems.

Strengthening Agent Defenses#

Drawing on attention mechanism analysis, the team developed a prompt-level defense technique for detecting adversarial arguments. This method consistently bolsters agent robustness against manipulation attempts, offering a practical pathway toward more trustworthy AI collaboration.

The work encompasses experiments across objective benchmarks (MMLU) and subjective reasoning tasks (PersuasionBench and Perspectrum), providing broad evidence for these dynamics.