Pith. sign in

REVIEW 4 major objections 7 minor 1 references

Learning social norms enhances compatibility in dynamic human-AI coordination

T0 review · 4 major / 7 minor · reviewed 2026-07-09 · glm-5.2

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read AI agents beat humans at coordination by learning three social norms

desk verdict Solid empirical work with a real circularity problem in the threshold derivation read the letter →

arxiv 2607.07021 v1 pith:BEE3WXB5 submitted 2026-07-08 cs.AI cs.HC

Yi Yang , Siyuan Liu , Xin Gao , Huamu Sun , Chao Liu , Qing Zhou , Bingbing Nie This is my paper
classification cs.AIcs.HC
keywords dynamicinteractionsnormssocialagentshumaninteractioncoordination
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that the reason AI agents fail to coordinate naturally with humans in dynamic, real-time interactions is not a lack of reasoning ability or behavioral imitation data, but the absence of explicit representations of the implicit social norms that humans use to coordinate. The authors collect 3,456 human-human interactions in a simplified pedestrian-vehicle crossing game, compare successful and failed coordination outcomes, and extract three quantifiable principles that distinguish them: outcome predictability (the interaction outcome should become legible early to both parties), value alignment (the two parties should adopt complementary orientations—one asserting, one yielding—rather than both racing or both holding back), and advantage awareness (decisions should respect asymmetries in position, speed, and collision penalties). When these three principles are encoded into LLM prompts and distilled into deployable policies, the resulting agents coordinate with humans far more effectively than baseline strategies, imitation-learning agents trained on human trajectories, or even humans coordinating with each other—achieving a fourfold score improvement over baseline and a 43% improvement over human-human interaction.

What carries the argument

The three principles serve as the load-bearing mechanism. Outcome predictability is operationalized via game entropy—the Shannon entropy of a neural network's predicted outcome distribution at each timestep, integrated over the interaction. Value alignment is operationalized via social value orientation (SVO) angle, computed from each agent's reward-to-self and reward-to-other at each timestep, where the sum of both agents' SVO angles approaching 90° signals complementary (one assertive, one yielding) rather than conflicting behavior. Advantage awareness is operationalized through situation-dependent asymmetries (initial position and speed advantages) and agent-intrinsic asymmetries (differi

What would settle it

Train imitation-learning agents exclusively on the highest-quality human interactions (categories C1-C2) and test whether they match the performance of the three-principle-informed agent. The paper reports they do not—H1-2 even underperforms agents trained on broader data—supporting the claim that norms cannot be recovered from trajectories alone. A stronger falsification would be to find a coordination task where the three principles are encoded but coordination still fails, suggesting the principle set is incomplete.

Watch

Extended reading notes

Core claim

The central discovery is that the implicit social norms governing human dynamic coordination can be decomposed into three measurable principles—outcome predictability, value alignment, and advantage awareness—and that explicitly encoding these principles into AI agents produces coordination that surpasses human baseline performance. Critically, the paper shows that neither behavioral imitation from high-quality human trajectories nor task-reward optimization alone recovers these norms, demonstrating that the normative structure must be represented explicitly rather than learned implicitly from data.

Load-bearing premise

The claim that these three principles constitute a transferable normative structure for dynamic coordination rests on the assumption that the simplified one-dimensional pedestrian-vehicle game—where two agents approach a single intersection point along fixed paths with no lateral movement—preserves the essential mechanics of real-world dynamic interactions. If this simplification strips away features that are actually essential to coordination (such as continuous spatial re

Editorial extensions

If this is right

  • If the three principles generalize beyond pedestrian-vehicle games, they could be incorporated into autonomous vehicle planners, service robot navigation, and any real-time system where AI must negotiate right-of-way or priority with humans.
  • The finding that imitation learning from human trajectories fails to recover normative structure suggests that current approaches to human-AI alignment that rely on behavioral cloning or demonstration data are missing a fundamental layer of representation.
  • The same principles might apply to multi-agent AI systems coordinating with each other, not just human-AI pairs, since the norms concern complementary role adoption rather than human-specific psychology.
  • The RL validation showing that the principles work as reward terms, not just as LLM prompts, suggests the normative structure is architecture-independent and could be integrated into any learning framework.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the three principles are truly the minimal sufficient set for dynamic coordination, then adding a fourth principle—such as temporal commitment (signaling one's intention through sustained behavior rather than momentary actions)—might further improve coordination in scenarios with richer strategy spaces than a single intersection.
  • The fact that advantage awareness includes agent-intrinsic asymmetry (collision penalties differ by role) suggests that social norms are partly constructed from structural incentives, not purely from convention—raising the question of whether changing the penalty structure would shift the norms humans develop.
  • The 43% improvement over human-human baselines in a controlled game raises the possibility that explicit norm representation could let AI teach humans better coordination strategies, not merely match them—a form of normative enhancement rather than normative compliance.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 7 minor

Summary. This manuscript investigates how implicit social norms governing dynamic human coordination can be formalized into explicit, quantifiable principles and incorporated into AI agents to improve human-AI coordination. The authors develop a simplified pedestrian-vehicle interaction platform, collect 3,456 human-human interactions from 96 participants, and identify three principles—outcome predictability (measured via game entropy), value alignment (measured via social value orientation angle sums), and advantage awareness (measured via initial-state and agent-intrinsic asymmetries). These principles are encoded into LLM prompts (HSN strategy) and evaluated against baseline, social chain-of-thought, and risk-aversion strategies across five LLMs in an open-loop paradigm, then distilled into lightweight policies for closed-loop human-AI evaluation. The HSN strategy achieves a nearly fourfold total score improvement over baseline and outperforms the human-human reference by 43%. Ablation analysis confirms all three principles contribute, and RL validation (Supplementary Note 7) shows the principles transfer beyond prompting to reward shaping. The experimental design is thorough: multiple LLM families tested, open-loop and closed-loop evaluation, ablation, imitation-learning comparison, and RL validation. The central claim—that making implicit norms explicit as computable principles improves coordination beyond what behavior imitation or task optimization achieves—is well-supported by

Significance. The paper addresses a genuine gap: most work on social norms in AI focuses on static or turn-based interactions, while dynamic, continuously coupled coordination remains underexplored. The finding that existing alignment strategies (SCoT, RA) fail to generalize from static to dynamic settings is a valuable negative result. The three principles identified are intuitively grounded in the coordination literature (legibility/predictability, social value orientation, asymmetry awareness) and their operationalization as quantitative metrics is a concrete contribution. The RL validation in Supplementary Note 7, showing that the same principles work as reward terms, strengthens the architecture-independence claim. The closed-loop human-AI evaluation with 24 new participants and the pairwise round-robin tournament across ten policies provide converging evidence. The paper ships falsifiable predictions (specific entropy threshold, SVO target angle) and a reproducible experimental platform. However, the significance is tempered by the fact that all quantitative thresholds are derived from a single simplified task and evaluated against metrics computed by models trained on the same data, which

major comments (4)
  1. §2.3, Fig. 3b: The entropy threshold of 5.27 is defined as 'the mean of the upper quartile of C1–C2 and the lower quartile of C3–C4' from the same 3,456 interactions used to define coordination quality categories. This threshold is then encoded verbatim into the HSN prompt (Supplementary Note 3.3, Fig. S7: 'High-quality interaction requires cumulative entropy below 5.27'). The open-loop evaluation metrics (p̄_C, p̄_F) are estimated by an LSTM trained on the same human dataset (§4.2). The closed-loop distilled HSN policy was trained on LLM decisions that explicitly used this data-derived threshold. The RL validation (Supplementary Note 7, Table S9) reuses E₀=5.27. No sensitivity analysis, cross-validation, or held-out threshold derivation is reported. This creates a circularity concern: the threshold that defines 'high-quality coordination' is extracted from the same data used to evaluate
  2. §2.3, Fig. 3c: The SVO complementary target of 90° is similarly derived from observing that C1–C2 interactions converge to this value, then encoded as a decision rule in the HSN prompt. The SVO computation itself (Eq. 2–3) involves a scaling coefficient α (pedestrian: 2, vehicle: 1) that is justified as compensating for speed-range differences (Supplementary Note 1.2) but whose sensitivity is not analyzed. If α were set differently, the SVO angles and the 90° target would shift. The paper should report how sensitive the HSN performance is to variations in E₀ (e.g., ±20%) and in the SVO target, or alternatively derive these thresholds on a held-out subset of the human data and evaluate on the remainder.
  3. §1, §2.1: The simplified 1D pedestrian-vehicle game removes lateral movement, multi-agent scenarios, and richer goal structures. The paper claims this 'preserves the core mechanics of dynamic coordination while minimizing irrelevant confounding factors,' but does not empirically justify which features are dispensable. The 43% improvement over human-human baselines may partly reflect the constrained task structure (e.g., binary roles, single conflict point, symmetric scoring except for collision penalties). The authors should discuss this limitation more substantively, perhaps by identifying which features of real-world coordination the simplified task does and does not capture, and whether the three principles would need modification in richer settings.
  4. §4.4, Fig. 4a: The closed-loop evaluation uses 24 new participants, each completing 72 trials across four strategies. The statistical comparisons between strategies (e.g., HSN vs. RA, HSN vs. SCoT) use Welch's t-test, but the experimental design appears to be within-subject (each participant experiences all four strategies). If so, paired tests would be more appropriate, and the report should clarify whether the t-tests account for within-subject correlation. Additionally, the 43% improvement over the human-human baseline compares HSN (24 participants, AI opponent) against the full human-human dataset (96 participants, human opponent)—different participant pools and opponent types. This comparison should be qualified as
minor comments (7)
  1. §2.2: The five LLMs are listed as 'DeepSeek-V4-Flash, Gemini-3.1-Flash-Lite, Doubao-Seed-2.0-Lite, Llama-4-Maverick, and GPT-5.4-Mini.' These version numbers appear to be placeholders or future versions. The authors should confirm these are the actual models used and provide access dates or API versions for reproducibility.
  2. Fig. 2: The caption states 'only the upper 30% of the distribution is shown for clarity' for p̄_C. This makes it difficult to assess the full collision probability distribution. Consider showing the full distribution or providing summary statistics (mean, median) in a table.
  3. §4.3, Eq. 2: The SVO reward computation uses exp(α·s(t)/t)·exp(α·v_s(t)−v_mean) for r_s but only exp(α·s(t)/t) for r_o. The asymmetry in formulation (including instantaneous speed deviation for self but not other) is not motivated. Why is the other's reward not also adjusted for speed deviation?
  4. Fig. 3a: The category labels C1–C7 are defined by total score intervals, but the specific boundaries (e.g., '116 < TS < 150' for C1–C2) are only shown in the figure. These should be stated in the main text for clarity.
  5. Supplementary Note 7.2.3, Table S9: The reward coefficients were 'tuned via grid search on a validation set comprising 5% of human interaction data.' It would be useful to report the sensitivity of RL-HSN performance to these coefficients, or at minimum state the grid search range.
  6. §3 (Discussion): The limitation paragraph mentions LLM inference latency but does not discuss the circularity concern raised in Major Comment 1. Adding a discussion of this limitation would strengthen the manuscript's transparency.
  7. Fig. 5b: The Base bar is truncated but the truncation is not visually indicated with a break mark. Consider adding a standard axis-break symbol.

Circularity Check

2 steps flagged · score 2.0 of 10

Thresholds fitted to the same human dataset used for open-loop evaluation, but closed-loop evaluation with new participants provides independent outcome measurement; no definitional circularity.

  1. fitted input called prediction [§2.3, Fig. 3b; Supplementary Note 3.3 (Fig. S7); §4.2]
    "The threshold (5.27), defined as the mean of the upper quartile of C1–C2 and the lower quartile of C3–C4, separates high-quality coordination from other interactions. [...] High-quality interaction requires cumulative entropy below 5.27 [...] We evaluated LLM coordination using two representative metrics: mean collision probability p̄C and mean first-passage probability p̄F [...] both estimated by a long short-term memory (LSTM) network trained on human interaction data"

    The entropy threshold 5.27 is fitted to quality categories (C1–C4) derived from the same 3,456 interactions. This threshold is then encoded verbatim into the HSN prompt. The open-loop evaluation metrics (p̄C, p̄F) are estimated by an LSTM trained on the same 3,456 interactions, and the keyframes (6,924) are extracted from that same dataset. So the open-loop evaluation tests an intervention whose parameters were fitted to the same data used to compute the evaluation metrics. However, this is overfitting risk, not strict circularity: the evaluation metrics (collision probability, passage probability) are different quantities from the fitted parameter (entropy threshold), and the LSTM is a general predictor, not a definitional identity. The closed-loop evaluation with 24 new participants uses

  2. fitted input called prediction [§2.3, Fig. 3c; Supplementary Note 3.3 (Fig. S7); Supplementary Note 7 (Table S9)]
    "it rapidly converges to 90° in C1–C2 [...] sum of both parties' SVO angles equaling roughly 90° (one aggressive, one yielding) [...] E0 Entropy threshold - 5.27"

    The SVO target of 90° is observed in the highest-quality category (C1–C2) of the same 3,456 interactions, then encoded into the HSN prompt and reused as the RL reward parameter (Table S9). The RL validation (Supplementary Note 7) reuses the identical threshold E₀=5.27, so it does not independently test whether this specific value is robust. However, the RL validation tests transfer across architectures (RL vs. LLM prompting), not threshold correctness, and the closed-loop human evaluation (24 new participants) measures actual game scores rather than data-derived metrics. The concern is parameter overfitting to a single task, not definitional circularity.

full rationale

The paper's central claim—that three formalized principles improve human-AI coordination—is not definitionally circular. The thresholds (entropy=5.27, SVO sum=90°) are fitted to the 3,456 human interactions and then encoded into the HSN prompt, and the open-loop evaluation metrics are computed by an LSTM trained on the same dataset. This creates a risk of overfitting to the specific task structure, and the RL validation reuses the same thresholds without independent derivation. However, the closed-loop evaluation with 24 new participants measures actual game scores (not LSTM predictions), providing an independent outcome test. The evaluation metrics (collision probability, passage efficiency, game scores) are distinct quantities from the fitted parameters (entropy threshold, SVO target), so the improvement is not forced by construction. No self-citation chain is load-bearing: the principles are derived from data collected in this study, not from prior work by the same authors. The concern is methodological (potential overfitting of thresholds to a simplified task) rather than circular (result equivalent to input by definition). Score 2 reflects the mild concern that open-loop evaluation and threshold derivation share the same dataset, without the central claim being forced.

Assumptions & free parameters 6 free parameters · 5 assumptions · 3 invented entities

The paper introduces three quantifiable principles with associated metrics, but several key thresholds (entropy = 5.27, SVO = 90°) and scaling parameters (α = 2, 1) are fitted to or derived from the same human interaction dataset used for evaluation. The RL validation provides partial independent support, but the core normative thresholds are data-derived rather than independently predicted.

free parameters (6)
  • Entropy threshold E0 = 5.27
    Defined as the mean of the upper quartile of C1–C2 and the lower quartile of C3–C4 from the same human interaction dataset used for evaluation (§2.3).
  • SVO scaling coefficient α (pedestrian) = 2
    Hand-set to compensate for speed range differences between pedestrian and vehicle (§4.3, Eq. 2). No sensitivity analysis provided.
  • SVO scaling coefficient α (vehicle) = 1
    Hand-set reference value for the vehicle agent (§4.3, Eq. 2).
  • Temperature T for outcome prediction softmax = 4
    Fixed to reduce short-term probability fluctuations in the outcome prediction model (§4.2). No justification for this specific value.
  • SVO complementary target = 90°
    Derived from observation that successful coordination (C1–C2) converges to SVO sum of 90°, then used as a normative target in HSN prompts and RL reward (§2.3, Supplementary Note 7).
  • RL reward coefficients (α_p, α_a, α_s, α_u, α_c, α_w, α_e, α_SVO, β) = See Table S9
    Tuned via grid search on 5% validation set of human interaction data (Supplementary Note 7.2.3).
assumptions (5)
  • domain assumption The simplified 1D pedestrian-vehicle game preserves the core mechanics of dynamic coordination.
    Stated in §1 and §2.1; the entire framework's generalizability depends on this simplification being valid.
  • domain assumption Social value orientation (SVO) is a valid psychometric construct for quantifying value alignment in dynamic interactions.
    Invoked in §2.3 to measure complementary value orientations; the SVO angle computation (Eqs. 2–3) assumes reward-to-self and reward-to-other can be meaningfully computed from kinematic states.
  • standard math Game entropy (Shannon entropy of outcome distribution) is a valid measure of outcome predictability.
    Standard information-theoretic construct applied to the output of an LSTM outcome predictor (§4.3, Eq. 1).
  • domain assumption The LSTM outcome prediction model provides unbiased probability estimates for interaction outcomes.
    The entropy and collision probability metrics depend on this model's calibration; no calibration analysis is provided (§4.2, Supplementary Note 6).
  • ad hoc to paper Three principles are sufficient to capture the normative structure of human dynamic coordination.
    The paper identifies exactly three principles from the data; whether additional principles exist that are not captured by this framework is not addressed.
invented entities (3)
  • Game entropy (cumulative) independent evidence
    purpose: Quantify outcome predictability as the temporal integral of Shannon entropy of the real-time outcome distribution.
    Falsifiable: the threshold of 5.27 separates high-quality from lower-quality coordination and is tested on new participants in closed-loop evaluation.
  • SVO sum as value alignment metric independent evidence
    purpose: Quantify whether two agents have complementary value orientations (target: 90°).
    Falsifiable: the 90° target is tested in ablation (Fig. 5) and RL validation (Supplementary Note 7), where RL-HSN converges to SVO sum near 90°.
  • Advantage awareness heuristic independent evidence
    purpose: Operationalize recognition of agent-intrinsic and situation-dependent asymmetries.
    Falsifiable: ablation (Fig. 5b) shows that self-play without advantage awareness produces lower scores because identical agents cannot adopt complementary roles.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning social norms enhances compatibility in dynamic human-AI coordination." pith.science (2026). https://pith.science/paper/BEE3WXB5

@misc{pith2026260707021,
  author       = {Pith},
  title        = {Pith review of: Learning social norms enhances compatibility in dynamic human-AI coordination},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BEE3WXB5}},
  note         = {Machine review of arXiv:2607.07021}
}
read the original 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 large language models (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.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

1 extracted references · 1 canonical work pages

  1. [1]

    1 Hawkins, R. X. D., Goodman, N. D. & Goldstone, R. L. The emergence of social norms and conventions. Trends Cogn Sci 23, 158-169 (2019). https://doi.org/10.1016/j.tics.2018.11.003 2 Young, H. P. The evolution of social norms. Annu Rev Econ 7, 359 -387 (2015). https://doi.org/10.1146/annurev-economics-080614-115322 3 Rahwan, I. et al. Machine behaviour. N...

    work page Pith review arXiv doi:10.1016/j.tics.2018.11.003 2019

Pith tools

Reviewed July 9, 2026 · model on record in the stance chip above.