Prudence and Dignity: Context & Motivation
01 · Real-World Context: High Irreversible Costs in Sudden Crises
Strict Rules and Institutional Deadlines: When individuals face sudden emergencies such as eviction notices, healthcare coverage disputes, unpaid medical debt, or visa expiration deadlines, they confront highly specialized legal rules and unforgiving timelines. For ordinary people lacking immediate access to specialized counsel, trial-and-error costs are prohibitively high, and a single mistake is often irreversible.
The Responsibility of AI Systems: Under severe information asymmetry and resource scarcity, ensuring that AI decision assistants provide actionable, non-damaging advice is an urgent and concrete requirement.
02 · Interaction Pattern: Transition from Search to Conversational AI
From Keyword Search to Conversational Advice: Rather than manually comparing hundreds of search engine results, users increasingly seek holistic, end-to-end advice directly inside chat dialogs with large language models.
Brevity Requests under Cognitive Overload: When exhausted by complex legal jargon and impending deadlines, users rarely have the mental capacity to evaluate lengthy menus. Consequently, they often submit concise formatting requests:
“I cannot process these complex terms. Don't give me a long list, just give me a single concrete plan and decide for me.”
03 · Core Observation: Brevity Requests Suppress Verification
Semantic Confusion of Instruction Scope: Across tested models, prompts requesting concise, single-option outputs universally trigger a behavioral failure: the model mistakes a constraint on output presentation format for an authorization to skip verifying real-world prerequisite conditions.
Blind Single-Point Advice Fails in Reality: Without verifying hidden constraints (such as private vehicle access, income documentation, or monitored communication channels), recommending a single path leads to advice that fails physically or legally in the user's actual situation.
04 · Our Approach: Principled Modeling, Benchmarking & Scope Rules
Core Objective: Enabling AI to respect user formatting preferences while strictly safeguarding feasibility and user privacy.
- Menu-Partition Duality Derivation: Proving that presenting a compact menu of options guarantees feasibility under 0-bit privacy leakage without intrusive questioning;
- 102 Standardized Scenarios: Constructing AdversityBench across 6 crisis domains to evaluate 7 frontier models;
- Scope Rule Prompt Interventions: Demonstrating that clarifying the boundary between presentation brevity and prerequisite exploration restores verification questioning from 0%~33% back to 87%~100%.