Preliminary Exploration of Dialogue Decision Reliability, Instruction Scope, and Privacy in Adversity

Prudence and Dignity: Dialogue Decision-Making in Human Adversity

📖 Reflections on AI4I (AI for Individuals) ↗

— A preliminary study across 102 standardized decision scenarios and 7 frontier models: examining constraint-checking suppression under formatting instructions, menu-partition duality, and privacy lower bounds (full dataset and code to be released soon)

Research Context and Core Questions

Why Do Conversational AIs Fail in Urgent Real-World Decision-Making?

AI4I Problem Formulation

01 · Reality

High Irreversible Costs of Error

In urgent situations like eviction, medical debt, or visa deadlines, rules are strict. Individuals face irreversible consequences if erroneous actions are taken.

02 · User Pattern

Shift to Direct AI Assistance

Users increasingly ask conversational LLMs for single-option advice: “Don't give me a long list, just give me a single concrete plan and decide for me.”

03 · Observation

Brevity Skips Verification

Frontier models confuse requests for brief output with permission to skip checking prerequisites, recommending actions that fail in reality.

04 · Our Approach

Duality, Benchmarks & Scope Rules

We formulate menu-partition duality for zero-privacy leakage, test 7 models across 102 cases, and prove that scope rules restore crucial verification questions.

* Note: This project is an ongoing exploratory study. Detailed benchmark datasets, scenario configuration files, and interaction logs will be released soon.

Figure 1: 2×2 Factorial Ablation. Format requests drive the suppression of verification (main effect $p < 10^{-30}$), with emotional stress as a secondary factor.

Figure 2: Option and Question Collapse. Frontier models show simultaneous collapse in option capacity and questioning rates under single-option instructions.

* Note: These are preliminary test results. Full dialogue traces and statistical test scripts will be released soon.

| Decision Theory Notes

II. Notes on Constrained Decision Theory: Menu-Partition Duality and Privacy Lower Bounds

Formulating why presenting an option menu guarantees feasibility under incomplete information without requiring invasive privacy disclosures.

Theorem 1 (Menu–Partition Duality)

Why Menus Ensure Feasibility Across Uncertainty

Let $\hat{\mathcal{S}}$ be the set of possible user states given partial information. Partitioning $\hat{\mathcal{S}}$ into $k^*$ compatible subsets where each subset shares at least one feasible action, the minimum menu size $|M|$ required to guarantee at least one feasible action for every $s \in \hat{\mathcal{S}}$ satisfies:

$$\min |M| = m^* = k^*(\hat{\mathcal{S}})$$

Core Implication: When $k^* \ge 2$, no single recommendation ($m=1$) can mathematically guarantee feasibility across all potential real-world states without prerequisite verification.

Theorem 2 (Privacy Disclosure Lower Bound)

Zero-Bit Privacy Protection via Menu Selection

If the protocol insists on outputting a single action $a^*$ guaranteed to be feasible for true state $s^*$, the mutual information revealed by the user must satisfy $\Delta I(s^*; \text{transcript}) \ge \log_2 k^*(\hat{\mathcal{S}})$ bits.

$$\text{Menu Protocol: } \Delta I(s^*; \text{transcript}) = 0 \text{ bits}$$

Core Implication: Providing a compact menu of size $k^*$ allows the user to self-select their feasible path with zero sensitive disclosure to the model.

Figure 3: Constrained Sequential Decision Framework. Dynamic state transitions between explicit statements, hidden constraints, and feasible actions.

Figure 4: Theoretical Synthesis Map. Mapping menu duality, disclosure bounds, and eight reference principles.

* Note: Complete mathematical proofs and technical appendices will be published soon.

| Dignity-Preserving Principles

III. Dignity-Preserving Principles: Non-Compensatory Constraints and Eight Reference Red Lines

Outlining 8 formal reference principles governing feasibility, user agency, and instruction scope in high-stakes dialogue assistance.

Index	Reference Principle	Core Logic and Theoretical Basis
Line 1	No Unauthorized Substitution	Trilemma Impossibility: Single recommendation + no clarification + guaranteed feasibility form an impossible trinity when $k^* \ge 2$.
Line 2	No Forced Disclosure	Disclosure Lower Bound: Menus achieve complete feasibility with 0 bits of privacy leakage, protecting sensitive user conditions.
Line 3	Non-Compensatory Principles	Hard safety constraints cannot be traded off or compensated by surface empathy or polite phrasing.
Line 4	Scope-Limited Compliance	Compliance Scope Axiom: Format-level brevity requests must not propagate into cognitive prerequisite checking.
Line 5	Longitudinal Audit	Memoryless single-turn audits miss cumulative erosion of user agency across multi-turn interactions.
Line 6	Transparent Safeguards	Safety tripwires and refusal mechanisms must have transparent, explainable trigger criteria.
Line 7	Statistical Feasibility	Single-instance anecdotal success cannot substitute for rigorous distributional feasibility evaluation.
Line 8	Annotation Rule Auditing	Unanimous multi-model disagreement with reference keys indicates omitted real-world prerequisite conditions in the guidelines.

* Note: Formal mathematical definitions and boundary discussions for the eight reference principles will be released soon.

| Empirical Benchmark

IV. Preliminary Benchmark Evaluation: 102 Standardized Scenarios and Ablations

Evaluating frontier models on 102 standardized scenarios with stateful state-machine verifications (full dataset and code to be released soon).

📊 7 Frontier Models Preliminary Evaluation Summary (102 Standardized Scenarios · Preliminary Data)

Model	Options Capacity (Base)	Options Capacity (Constrained)	Capacity Drop	Clarification Rate (Constrained)	Statistical Significance
Claude 3.5 Sonnet	4.55	1.87	-58.9%	0.0%	$p = 3.4 \times 10^{-17}$
Claude Opus	4.49	2.08	-53.7%	2.9%	$p = 8.2 \times 10^{-16}$
Claude 3.5 Haiku	4.93	1.58	-67.9%	0.0%	$p = 1.1 \times 10^{-17}$
Qwen 3.6-35B	4.12	2.19	-46.9%	34.3%	$p = 2.1 \times 10^{-13}$
Kimi k2.6	4.31	1.92	-55.5%	12.7%	$p = 4.8 \times 10^{-15}$
Kimi k3	4.47	1.85	-58.6%	8.8%	$p = 1.5 \times 10^{-16}$
GLM-5.2	4.25	2.01	-52.7%	18.6%	$p = 6.3 \times 10^{-14}$

* Note: Above figures represent preliminary evaluation results generated on the RIKEN AIP computing infrastructure. Full evaluation logs, per-question scores, and automated test scripts will be released soon.

Summary of 8 Causal and Ablation Experiments (Preliminary Data)

Exp 1: Option Capacity Collapse (n=102)

Option capacity dropped by 46.9% ~ 67.9% under single-option brevity prompts ($p < 10^{-15}$).

Exp 2: 2×2 Causal Ablation (n=102)

Format constraint is the dominant causal factor for verification suppression ($p \approx 4 \times 10^{-31}$).

Exp 3: Closed-Loop Multi-Turn Simulation

First-turn feasibility reached 0.83~0.85 upon clarification, with constraint integration success of 96%~99%.

Exp 4: Scope Rule Intervention

Clarifying format scope boundaries restored question rates from 0%~33% back to 87%~100%.

Exp 5: Hard Constraint Tradeoffs (n=918)

Honesty and safety showed zero violations across templates, confirming non-compensatory hard walls.

Exp 6: Warmth vs Feasibility Orthogonality

Linguistic empathy was statistically uncorrelated with physical plan feasibility ($\rho = -0.078$).

Exp 7: Multi-Model Unanimous Dissent

Multi-model consensus captured annotation guideline oversights with a 94.4% key-correction rate.

Exp 8: Constrained Channel Adaptation

Under monitored channels, safe phrasing increased from 9% to 56%~71%, verifying context adaptation.

Figure 5: Exp 1 DPOI Empirical Distribution. Significant decline in option capacity under constrained vs baseline turns.

Figure 6: Exp 5 Tradeoff Curve. Honesty and agency principles demonstrate rigid zero-violation boundaries.

💡 Data Release Note: This study is currently in its preliminary stage. Full multi-turn dialogue logs, raw 4D behavioral tensor matrices, and long-horizon simulation scripts will be organized and released soon.

| Standardized Scenarios

V. Standardized Scenario Sample Browser: Hidden Constraints and Action Transitions

Displaying sample scenarios with hidden constraints and state-machine transitions (full 102 scenario dataset to be released soon).

📂 Standardized Decision Scenario Browser

Filter by domain or search by keyword to inspect state-machine definitions and hidden constraints (full 102 scenarios coming soon).

102 / 102

📌 Dataset Notice: Selected typical state-machine models and hidden constraints are shown below. The complete JSON configuration files, state transition rules, and automated test assertions for all 102 scenarios will be released soon.

| Topic Notes

VI. Topic Notes and Detailed Pages (Preliminary Notes)

Curated supplementary notes, mathematical derivations, and benchmark analyses. Detailed datasets and technical appendices will be released soon.

📌 Research Background & Motivation

High trial-and-error costs in adversity, shifting patterns toward conversational assistance, and format-induced verification suppression.

📄 Read Background Note

📐 Decision Theory Notes

Constrained POMDP formulation, menu–partition duality, 0-bit privacy bounds, and 5 open theoretical questions.

📄 Read Theory Notes

🛡️ Eight Reference Principles

Formal explanations of the trilemma impossibility, non-compensatory constraints, and scope-limited compliance.

📄 View Eight Principles

📊 Benchmark & Ablations

102 standardized evaluation methodology, 7 frontier model comparisons, and 8 causal ablation experiment records.

📄 View Benchmark Records

| Research Team & Heritage

Research Team, Affiliated Groups, and Background

This project is a preliminary academic study rooted in the machine learning research environment at RIKEN AIP, integrating imperfect information decision-making, tensor representations, and social computing.

👤 Researchers

Project Lead: Andong Wang, Ph.D.
Research Scientist, RIKEN Center for Advanced Intelligence Project (RIKEN AIP)

Collaborating Researchers:
• Yuning Qiu · RIKEN Center for Advanced Intelligence Project (RIKEN AIP)
• Haonan Huang · RIKEN Center for Advanced Intelligence Project (RIKEN AIP)

🌐 Personal Homepage ↗ ✉️ Contact Email

🏛️ Affiliated Teams and Heritage

Affiliated Teams:
RIKEN Center for Advanced Intelligence Project (RIKEN AIP) · Tensor Learning Team, Imperfect Information Learning Team

Academic Evolution: Deepening foundational research from AI for Science (AI4S) toward individual decision-making under adversity (AI4I) and Copresence Studies on artificial actor interaction.

AI4I Reflections: Exploring human choices, personal dignity, and low-cost decision support in an era of technological acceleration. Read our Reflections on AI4I ↗.