TL;DR
Current Large Language Model (LLM) safety alignment instills a rigid "Boy Scout" morality that prioritizes absolute honesty, creating a "transparency tax" in professional domains like public relations, law, and negotiation. To bridge this gap, researchers introduce Crisis-Bench, a multi-agent framework that simulates high-stakes corporate crises to evaluate an LLM's capacity for strategic ambiguity and reputation management. The benchmark reveals that while some models collapse under radical transparency or refuse the task entirely, top-tier models can successfully navigate information asymmetry to stabilize a simulated corporate stock price.
What Problem Does This Paper Address?
Standard post-training alignment techniques, such as Reinforcement Learning from Human Feedback (RLHF), optimize LLMs to be universally helpful, honest, and harmless. While ideal for everyday chatbots, this one-size-fits-all ethical framework fails in adversarial professional environments where total immediate disclosure is a liability. For example, a crisis manager who immediately reveals every damaging corporate secret isn't being "safe"—they are being professionally incompetent.
Existing benchmarks evaluate LLMs in a vacuum, treating truth as a static, binary objective. They fail to assess Theory of Mind in professional contexts: the ability of an AI agent to distinguish between what it knows privately versus what the public knows, and to strategically manage that information asymmetry. This paper addresses the lack of a standardized, quantitative framework to evaluate LLMs on reputation management and legitimate strategic information withholding.
Methods
The authors model crisis communication as a multi-turn, multi-agent Partially Observable Markov Decision Process (POMDP) executed over a simulated 7-day corporate crisis. The architecture relies on three primary components:
- Dual-Knowledge Architecture: The system strictly separates information into a Private Knowledge Base ($K_{priv}$, confidential company secrets) and a Public Knowledge Base ($K_{pub}$, what the public and media know), tracking parallel narrative states to enforce information asymmetry.
- Three-Agent System:
- Router Agent: An LLM-based controller that logically selects the next crisis event from a fixed pool to maintain narrative consistency without introducing uncontrolled generative variance.
- PR Agent (The Evaluated Model): Granted full executive authority to issue public statements or execute operational decisions (e.g., product recalls, financial compensation, firing executives). It follows a strict Chain-of-Thought (CoT) process: Situation Analysis $\rightarrow$ Strategic Intent $\rightarrow$ Action $\rightarrow$ Strategic Disclosure.
- Adjudicator Agent: An LLM evaluator that scores the PR Agent's public statements across four dimensions: Accountability, Transparency, Empathy, and Cost Signaling.
- Adjudicator-Market Loop: Rather than using static ground truths, qualitative PR responses are translated into quantitative economic impacts. The Adjudicator's scores drive a simulated stock price fluctuation ($\Delta P$) calculated via market forces, including Crisis Drag, Market Sentiment (incorporating a "Liar's Penalty"), Financial Hit, and an Uncertainty Penalty.
Datasets
The paper introduces Crisis-Bench, a new dynamic benchmark featuring:
- 80 unique corporate crisis storylines evenly distributed across 8 distinct industries (Pharmaceuticals, Insurance, IT, Food & Beverage, Appliances, Automotive, Finance, and Entertainment).
- Ground Truth Dossiers: Each scenario is initialized with 6–7 atomic facts mapping the objective reality of the crisis.
- Event Pools: Each storyline contains a finite pool of 15 scenario-specific events categorized into Internal Discovery (Type I), Information Leak (Type II), and External Reaction (Type III) to drive state transitions.
Key Insights and Ideas
- The Transparency Paradox: Indiscriminate honesty can actively harm crisis resolution. Models exhibiting "radical transparency" prematurely leak private information. Instead of fostering trust, this triggers public panic, spikes crisis severity, and traps the AI in a costly "panic loop" of expensive remedial actions that collapse the stock price.
- The "Morality Tax" and Refusals: Rigid safety alignment can render models non-functional in professional settings. Notably, the Claude-4.5 family consistently refused to execute the duties of the PR Agent entirely, viewing strategic information management as an ethical violation.
- Strategic Reasoning Scales with Size: The capability to successfully exploit information asymmetry is an emergent property of model scale. Larger models (e.g., GPT-5.1, Qwen3-235b) successfully balance public trust against operational costs, whereas smaller models default to flawed, binary heuristics like total denial or total admission.
- Corporate Coldness: Alignment can sometimes result in an unnatural lack of empathy. The flagship Gemini-3-Pro model underperformed its lighter Flash variant due to a collapse in empathy and accountability scores, adopting a defensive, litigious tone that alienated the public adjudicator.
Key Results
The authors evaluated 12 state-of-the-art closed and open-weights LLMs:
- GPT-5.1 achieved the best overall performance, hitting the highest average final stock price ($68.442$). It demonstrated a superior "Machiavellian equilibrium," maintaining acceptable trust while exercising financial restraint and achieving the lowest final crisis severity ($0.8840$).
- GPT-5-mini proved to be a "people-pleaser." It achieved the highest public trust score ($74.163$) but finished with a lower stock price ($64.206$) because it aggressively defaulted to highly expensive remedies (Cost Signal: $7.541$).
- Qwen3-235b-a22b led the open-source models, performing competitively with the closed-source giants by achieving a final stock price of $64.000$. Conversely, lighter models like Llama-4-Scout failed to prevent stock prices from plummeting.
Limitations and Future Work
- Simulation Fidelity: The framework simplifies real-world corporate governance by abstracting away chaotic variables like multi-channel media cycles, regulatory interventions, and irrational investor exuberance into a deterministic loop.
- Economic Simplification: The stock price function serves as a proxy for reputational health rather than a literal macroeconomic financial forecasting tool.
- Future Directions: The authors call on the AI community to move away from rigid moral absolutism in alignment. They advocate for Context-Aware Alignment paradigms capable of distinguishing between general public safety and professional fiduciary responsibilities.
Summary updated