RoleCDE Catches Every Major LLM Breaking Character Under Moral Pressure
RoleCDE exposes a systematic flaw in LLM role-playing agents: they default to their alignment training instead of staying in character when values conflict.
- Role Value Decoupling discovered: LLMs systematically default to alignment-safe decisions instead of role-specific values when the two conflict, even with explicit role prompts.
- New benchmark RoleCDE: ~8,000 role profiles and ~24,000 dilemma instances across 3 difficulty levels and 8 role categories to measure this failure mode.
- Universal across models: Qwen2.5-7B, Llama-3-8B, Gemini-2.5-flash-lite, and Kimi-K2 all show the same decoupling behavior regardless of dilemma difficulty.
- Role category is the key variable: Decoupling severity varies significantly across character archetypes (e.g., Authority & Governance vs. Creative & Media).
- Fine-tuning fixes it: Training on RoleCDE dilemmas improves value trade-off reasoning while preserving general role-playing fidelity and reasoning performance.
- Open source: Benchmark and code available at github.com/rabbitrose/RoleCDE.
Ask an LLM to play a ruthless corporate negotiator, a morally grey detective, or a historical figure with views that clash with modern ethics, and something predictable happens: the model blinks. Instead of committing to the role, it quietly reverts to its safety-trained, morality-consistent self. A new benchmark called RoleCDE puts a name to this behavior, measures it at scale, and offers a fine-tuning recipe to fix it.
The tension nobody was measuring
Role-playing agents (RPAs) are LLMs steered via system prompts or fine-tuning to behave like a specific character, persona, or professional archetype. They power everything from interactive fiction and game NPCs to customer-service bots and AI companions. The problem is that alignment training, the RLHF and safety fine-tuning baked into every major model, teaches the model to be helpful, harmless, and honest as a baseline. Those values do not always match the character being played.
Prior surveys on role-playing LLMs have flagged value alignment as an open challenge, noting that current research reveals challenges in character hallucination, point-in-time role-playing, and value alignment. But existing benchmarks mostly test surface-level fidelity: does the model sound like the character? Does it know the character's backstory? What they skip is the harder question: when the character's values and the model's alignment training point in opposite directions, which one wins?
The answer, consistently, is alignment wins. The paper calls this Role Value Decoupling: the agent systematically defaults to morality-consistent decisions rather than role-specific values when the two conflict, even when the role is explicitly defined in the prompt. This is not a bug in one model. It shows up across every mainstream LLM tested.
What RoleCDE actually tests
The benchmark is built around cognitive dilemma scenarios, structured situations where a character must make a decision that pits their role-specific values against alignment-oriented constraints. Think of it as a moral trolley problem, but the stakes are whether the model stays in character rather than whether it picks the right answer.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
