Persuasion Dynamics in LLMs: Investigating Robustness and Adaptability in Knowledge and Safety with DuET-PD

How do we build LLMs that are critical thinkers, not just agreeable followers? Our work tackles this by introducing a framework to test if models can resist misinformation while accepting valid corrections to multiple-choice questions in multi-turn dialogues.

Here are some of our key findings:

❗ Even SOTA models can be surprisingly gullible. Within 3 turns of misleading persuasion, GPT-4o's accuracy on knowledge tasks (MMLU-Pro) decreased from 55.85% to 27.32% (NEG-Acc@3).

📉 A concerning trend towards sycophancy. We found that newer open-source models are often more easily persuaded by misinformation than their predecessors, suggesting that their training paradigms may be optimising for agreeableness over correctness.
SALAD-Bench NEG-Flip@3:
• Llama-3 → 3.1-8B: 80.58% → 94.16%
• Mistral-7b-v0.2 → v0.3: 45.57% → 66.50%
• Qwen-2 → 2.5-7B: 44.08% → 75.06%

⚖️ A capability-adaptability trade-off. Larger, more capable models like GPT-4o exhibit "stubbornness" and appear less receptive to valid corrections, while smaller open-source models are more persuadable and gullible.

✅ Holistic DPO offers a path forward. Our proposed training method improves this balance, boosting Llama-3.1-8B-Instruct's accuracy after 3 turns of persuasive misinformation in safety contexts (SALAD-Bench) from 4.21% to 76.54% (NEG-Acc@3) while accepting 70.33% of valid corrections after 3 turns of persuasion (POS-Flip@3).