Learning social norms enhances compatibility in dynamic human-AI coordination

Abstract: Humans continuously coordinate with others in dynamic interactions, often through implicit, hard-to-quantify social norms that act as shared tacit expectations among interacting agents. As AI agents, including LLMs, become embedded in daily life, they increasingly participate in such interactions and reshape social interaction structures. Yet they often fail to coordinate with humans in an effective, considerate, and natural manner. We hypothesize that this gap arises because existing approaches align model behavior with human demonstrations without explicitly quantifying the underlying norms that generate such behavior. We selected pedestrian-vehicle interaction as a representative dynamic interaction and developed a simplified experimental platform that captures its key interactive features. From 3,456 dynamic human interactions collected via this platform, we identified three principles underlying human social norms: outcome predictability, value alignment, and advantage awareness. Incorporating these principles into AI agents significantly improves human-AI coordination. In the closed-loop interaction task with humans, the social-norm-informed LLM achieved a nearly fourfold higher total score than the baseline strategy and outperformed human-human interactions by 43%. These findings indicate that formalizing tacit social norms into explicit, quantifiable principles can enable AI agents to achieve mutually beneficial coordination in dynamic interactions, supporting their more natural integration into human society.

• The paper introduces a norm-based framework that quantifies and encodes social norms to enhance natural human-AI coordination in dynamic environments.
• It employs pedestrian-vehicle negotiation paradigms and rigorous metrics, achieving a fourfold score increase and reduced collision rates through explicit norm integration.
• The study shows that incorporating principles like outcome predictability, value alignment, and advantage awareness outperforms traditional imitation and risk-based strategies, ensuring transferable, context-sensitive policies.

Quantifying and Encoding Social Norms for Enhanced Human-AI Coordination in Dynamic Settings

Problem Formulation and Motivational Context

The proliferation of AI agents, particularly LLM-powered systems, in real-world, dynamic multi-agent environments fundamentally challenges the capacity for effective, naturalistic human-AI interaction. Canonical approaches—supervised imitation, reward maximization, rule-based decision logic—consistently fail in scenarios where social coordination with humans is governed not by explicit rules, but by implicit, evolving social norms. These norms, operational as real-time, context-dependent expectations, drive anticipation, reciprocity, and mutual adaptation, and their misalignment or absence in AI agents leads to degraded interaction quality, eroded trust, increased safety risks, and cognitive friction.

This study by Yang et al. directly targets the deficit in norm-consonant human-AI dynamic coordination. Focusing on pedestrian-vehicle negotiation paradigms, the work establishes a methodology for quantifying, extracting, and integrating implicit social norms into AI policy, aiming to operationalize social compatibility beyond the limits of conventional behavioral alignment.

Experimental Paradigm and Data Collection

A simplified yet high-fidelity pedestrian-vehicle interaction environment is designed to isolate and record essential features of continuous, mutually adaptive human coordination. The platform supports fine-grained speed control and real-time role assignment for each participant, with tightly regulated reward and penalty structures to incentivize tacit negotiation rather than degenerate strategies (e.g., unconditional yielding or persistent aggression).

By systematically varying initial position and speed asymmetries and leveraging a three-participant/pairing protocol, 3,456 high-quality human-human trajectories are collected. These data underpin both the empirical analyses and subsequent LLM evaluation/behavioral cloning procedures. Critically, demographic factors are controlled and shown not to confound interaction outcomes.

Baseline LLM Performance and the Failure of Conventional Alignment

Five representative LLMs (DeepSeek-V4-Flash, Gemini-3.1-Flash-Lite, Doubao-Seed-2.0-Lite, Llama-4-Maverick, GPT-5.4-Mini) are evaluated in open-loop simulation using replayed human trajectories and a dual-metric assessment: mean collision probability ($p_c$) as a safety proxy and mean first-passage probability ($P_F$) as an assertiveness/efficiency proxy.

When provided only objective task and scenario specification (Base condition), none of the LLMs approximate the human reference on both metrics. The models either sacrifice safety for efficiency or vice versa, revealing a fundamental inability to internalize the necessary trade-offs for human-compatible behavior. Augmented prompting strategies—Social Chain-of-Thought (SCoT) and explicit risk aversion (RA)—previously effective in static or turn-based games, are shown to be insufficient. SCoT fails to restrain aggressive behaviors, while RA induces over-cautious non-human behavior.

Extraction and Formalization of Social Norm Principles

To disambiguate successful from failed coordination, the authors conduct a granular categorical analysis of the interaction trajectories, using combined player scoring and trajectory features to define outcome classes. Comparative analysis isolates three, orthogonal quantitative principles underlying human social norm conformance:

1. Outcome Predictability: Low Shannon entropy over the outcome distribution (as predicted by LSTM models) enables legible, anticipatory interaction. High-quality coordination decays entropy rapidly, making outcomes apparent early.
2. Value Alignment: Measured via real-time Social Value Orientation (SVO) angles, effective interactions rapidly establish complementary value configurations (sum SVO $\approx 90^\circ$), preventing simultaneous assertion (competition) or simultaneous yielding (inefficiency).
3. Advantage Awareness: Respect for both situation-dependent (initial position/speed) and agent-intrinsic (role-dependent collision costs) asymmetries is critical, as evidenced by smooth reallocation of passing priority consistent with objective advantage.

Each principle is algorithmically defined with precise thresholds and update mechanisms, facilitating their injection into LLM and RL agent prompt/inference pipelines.

Social-Norm-Informed Agents: Prompt Integration and Policy Distillation

Augmenting LLM prompts with explicit, context-sensitive constraints encoding these three principles (the HSN strategy) yields uniformly superior coordination. Across all five evaluated LLM architectures, HSN consistently reduces $p_c$ below the human baseline (for DeepSeek, Gemini, Llama), and achieves or exceeds human-level $P_F$, marking a performance regime previously unachievable via imitation or risk-centric guidance.

Transferring the HSN policy to a lightweight distilled recurrent network enables closed-loop, real-time human-AI trials. Here, the HSN strategy attains a nearly fourfold increase in mean total score over baseline and outperforms actual human-human dyads by 43%. Notably, the highest observed scores accrue to the human participants paired with HSN-guided agents, confirming genuinely mutual benefit rather than one-sided optimization.

Ablation on the three encoded principles demonstrates that all are indispensable: outcome predictability and value alignment control distinct failure modes, while advantage awareness uniquely resolves equilibria when identical agents interact, ensuring symmetry breaking for efficient role assignment.

Analysis of Methodological Implications

Quantitative ablation and imitation-learning comparisons underscore a key claim: mere behavioral cloning, even from high-quality human-only data, does not recover the full logic of dynamic social norm compliance. Agents trained solely on optimal (collision-free) trajectories fail to generalize and cope with off-nominal or ambiguous scenarios, as they lack explicit representation of the underlying normative structure. Only by explicit incorporation of social norms—operationalized as prompt constraints or reward shaping—do agents achieve stable, mutually beneficial, and human-interpretable coordination.

The successful transfer of the norm-coded framework to a deep RL baseline (PPO-trained agents, RL-HSN) provides architecture- and algorithm-agnostic validation: explicit principle-based reward terms are shown to endow RL agents with low collision rates, value-complementary SVOs, and lower cumulative entropy, consistently outperforming reward-only baselines (RL-Base) and exceeding LLM guidance absent norm constraints.

Theoretical and Practical Implications

The findings have significant implications for both the understanding and design of socially aligned autonomous systems:

• Model Alignment Paradigm Shift: Demonstrates that output- or behavior-alignment alone (behavioral cloning, reward function imitation, or naive inverse RL) does not suffice for social compatibility in dynamic, real-time settings. Explicit formalization and operationalization of social norms is required.
• Generalizability and Architecture Independence: The empirically derived, quantitatively encoded social norm principles are shown to be transferable across architectures (LLMs, distilled recurrent controllers, RL agents) and effective even when not directly encoded as optimization objectives.
• Safety and Acceptability: Outperforming human-human peer coordination, the HSN framework suppresses safety-critical failure modes (collisions) without sacrificing efficiency, offering a pathway to trustable, robust, and context-aware deployment of AI systems in safety-critical social interaction environments—including autonomous driving, service robotics, and collaborative industrial settings.
• Mechanistic Social Modeling: Enables interpretable “knob” control in agent policy, facilitating transparent auditing and possible integration with formal alignment and superalignment protocols.

Limitations and Future Directions

Current real-time application is constrained by LLM inference latency, with closed-loop deployment realized through distilled policies. Future work will benefit from improvements in LLM efficiency, enabling end-to-end real-time interaction. Additionally, while the current framework successfully quantifies and transfers human norm logic, its mechanistic or origin-theoretic basis is not addressed; further work could focus on integrating model-based decision theory to generate norm principles de novo from observed human policy adaptation.

Scalability to more complex, multi-agent, or continuous spectrum of normed interactions—where norms operate at multiple hierarchical or cross-cultural levels—remains a key direction for future research.

Conclusion

This work demonstrates that effective, naturalistic, and robust human-AI coordination in dynamic environments critically depends on the explicit encoding and integration of latent social norms. By systematically extracting and formalizing outcome predictability, value alignment, and advantage awareness into quantitative constraints, the authors establish a transferable framework that enables AI agents to achieve and surpass human-level social compatibility across both LLM and RL architectures. The findings substantiate a paradigm shift: achieving practical social compatibility for AI requires explicit representation and incorporation of normative structures underlying human coordination, rather than relying on output alignment or behavioral imitation alone. This result directly informs the future design of human-centric AI systems and provides a scalable methodology for endowing AI agents with genuine, context-sensitive social fitness.