Source-linked AI summary
Runtime Action Interference for AI Control of AlphaStar in StarCraft II
Jaymari Chua, Chen Wang, Liming Zhu, Lina Yao
TL;DR
Interactive systems can alter which actions a learned policy executes, raising questions about post-inference control and how capability disclosure affects user perceptions. This paper introduces runtime action interference and evaluates identical controlled opponents with capability claims withheld or disclosed, finding that disclosure changed perceived fairness, trust, and toxicity.
Problem
The paper examines how to regulate reinforcement-learning actions after inference and how capability disclosure affects perceptions of an otherwise identical controlled opponent.
Method
The study inserts deterministic runtime action interference between policy inference and dispatch, combining cooldown-based pacing with targeted content filtering, then compares two opponent presentations.
Results
Capability disclosure corresponded with lower perceived fairness and higher perceived toxicity across expertise groups, while trust varied by expertise despite unchanged opponent control.
Takeaways & Limitations
Human-computer evaluations should distinguish execution-level control from capability disclosure and assess fairness, trust, and toxicity separately.
Takeaways & Limitations
Because expertise cells were small, the analysis is descriptive and does not support causal claims from condition contrasts.
Abstract
from arXiv · showhide
A trained reinforcement learning policy does not determine the complete behavior that users encounter: deployment code still schedules, admits, suppresses, or replaces its proposed actions. We contribute \emph{runtime action interference} (RAI), an AI control mechanism that preserves policy parameters while regulating action pacing and filtering configured action patterns after inference. RAI releases a proposed action only when its cooldown condition is satisfied and its content detector does not flag the action; otherwise, it dispatches a no-op. The detector covers specified toxic behaviors, including worker-unit harassment, while the cooldown controls action rate. We implement RAI in a replication of AlphaStar actor.py and make the implementation and reproducibility materials available through an open source code repository. We deployed RAI in a \textit{StarCraft~II} human participant study that compared two presentations of the same opponent with high capability and rate limited actions; we withheld its capability claim in one presentation and disclosed it in the other. On response scales from 1 to 5, we observed pooled fairness, trust, and toxicity means of 3.90, 3.50, and 2.00 under claim withholding, compared with 2.62, 4.31, and 2.85 under disclosure. Disclosure corresponded with lower perceived fairness and higher perceived toxicity across every expertise group, whereas trust increased among novices and experts but decreased among intermediate participants. Our human evaluation therefore shows that perceptions of an opponent controlled through RAI can vary substantially with the capability information presented to users, even when the configured control remains constant. We conclude that human-computer evaluations must separate control within the execution stack from capability disclosure and assess fairness, trust, and toxicity as distinct dimensions of human experience.
1 Introduction
The paper introduces runtime action interference (RAI), a deterministic post-inference mechanism that admits or suppresses policy proposals without changing policy parameters. It studies how withholding or disclosing the same opponent’s capability claim shapes players’ perceptions in StarCraft II.
- Runtime action interference: RAI admits a proposed action only when it satisfies a safety condition and passes an action-content detector; otherwise, it dispatches a no-op.The mechanism operates between policy inference and environment dispatch, leaving policy parameters fixed and avoiding retraining.
- Human-participant study: The study retained 32 post-match responses across three conditions, including 23 focal responses involving the same documented high-capability, rate-controlled opponent.Capability was withheld for 10 responses and disclosed for 13; the remaining responses served as baseline.
- Research questions: The research questions address auditable post-inference regulation and descriptive differences in fairness, trust, toxicity, and narrative interpretation when capability claims are withheld or disclosed.Both questions concern users encountering the same documented RAI configuration under different capability-information presentations.
- Contributions: The paper contributes an action-admission model, an auditable StarCraft II implementation linking proposals, controller decisions, and released actions, and a human evaluation.The audit contract connects runtime decisions to the actions ultimately executed.
2 Related Work
Related work places RAI at the deployment stage, where a system can review, withhold, defer, or replace policy actions after inference rather than constraining policy learning. Prior research also shows that execution resources, verification mechanisms, and capability disclosures shape competitive-agent evaluation and human judgments.
- Deployment-stage control: RAI differs from policy-learning constraints by filtering or replacing proposed actions after inference while preserving the policy’s learned parameters.Constrained policy optimization incorporates restrictions into the learning objective, whereas runtime shielding blocks or replaces specification-violating actions.
- Deployment-stage control: RAI applies self-repair to runtime deployment by treating policy proposals as reviewable candidates that may be withheld, deferred, or replaced with specification-conforming actions.This relocates corrective logic from program modification to post-inference system control.
- Competitive evaluation: Competitive-agent evaluation already incorporates execution constraints, including AlphaStar’s action-rate and camera-access limits and efforts to make agents resemble human play.Benchmark fairness also depends on whether people and artificial agents receive comparable information and action resources.
- Runtime verification: Runtime verification traditions treat inferred actions as candidates for review, with configurable components able to intervene before execution.AgentSpec evaluates rules while an agent runs, while Reinforced Agent delegates tool-call review to an independently configurable component.
- Transparency and disclosure: Transparency does not yield uniform benefits: disclosures and explanations can increase acceptance, reduce excessive reliance, or reduce perceived legitimacy depending on context and presentation.Prior studies also find that transparency can alter judgments of human and AI contribution.
3 Method
RAI operates between policy inference and environment dispatch, regulating proposed actions through cooldown-based admission and configured content detection while keeping policy parameters fixed. It therefore changes the executed action process without changing the policy that generated each proposal.
- Implementation: RAI is implemented in actor.py after policy inference and before dispatch to the environment.
- Action admission: A stateful cooldown registry and timeframe-eligibility predicate determine whether a proposed action may execute.
- Content filtering: A configured detector identifies prohibited patterns, including worker-unit harassment, and blocks flagged proposals.
- Policy preservation: The policy parameters θ remain fixed, while controller parameters ϕ govern admission and state evolution and detector configuration ψ specifies prohibited patterns.
- Intervention: If either admission predicate fails, RAI vetoes the proposal and substitutes no_act.
2. Content Detector (Tψ(aπ
RAI’s content detector evaluates proposed actions against historical context and configured prohibitions, including toxic patterns such as worker-unit harassment. In the AlphaStar actor path, it works with cooldown control and no-op substitution while leaving policy parameters unchanged.
- Content detector: The content detector scrutinizes each proposed action against historical context using configuration ψ and flags prohibited behavioral paradigms.
- Runtime boundary: RAI intervenes only at the terminal output layer, while policy parameters θ remain immutable.
- Runtime decision rule: RAI combines content detection, target APM cooldown control, and no-op substitution, distinguishing temporal interventions from content interventions in its audit records.Proposals execute only when they satisfy cooldown and pass the configured detector; otherwise, RAI substitutes a no-op.
- Human study design: The focal human study compared withheld versus disclosed capability claims for the same high-capability opponent with RAI enabled, varying participant information while holding the configured opponent constant.The study involved three conditions and 32 responses.
- Human evaluation measures: Fairness, trust, and toxicity were measured as distinct outcomes on scales from 1 to 5.The measures addressed community-consistent behavior, willingness to play or rely on the opponent, and aversive interaction respectively.
4 Results
The results are exploratory because expertise-by-presentation cells contain only two to five responses. The paper therefore reports descriptive means, standard deviations, and arithmetic contrasts rather than statistical significance claims.
- Results: Expertise-by-presentation cells contained between two and five participant responses, limiting stable hypothesis tests and precise interval estimates.Participant-level records were released, but the small cells constrained inferential analysis.
- Results: Disclosed-minus-withheld contrasts show direction and magnitude across separate participant groups, not within-participant change.The reported contrasts are arithmetic and descriptive rather than claims of statistical significance.
- Results: Table 2 reports every focal cell as M (descriptive SD) [n] on 1-to-5 response scales, with pooled rows using sample-size-weighted means.Pooled rows omit standard deviations.
A. RAI Effect Baseline →RAI, Claim withheld · B. RAI w/ Disclosure Effect Claim withheld →Disclosed
Under the same configured RAI controller, disclosing capability was associated with lower fairness and higher toxicity, while trust varied by expertise group. The baseline-to-RAI comparison was descriptive, whereas the disclosure contrasts were composition-standardized.
- A. RAI Effect Baseline →RAI, Claim withheld: Panel A reports descriptive pooled differences between the separately reported baseline and RAI with the capability claim withheld, without identifying a causal RAI effect.The baseline implementation was not recovered.
- B. RAI w/ Disclosure Effect Claim withheld →Disclosed: Fairness was lower under disclosure in every expertise stratum, with disclosed-minus-withheld contrasts of −1.30 for novices, −1.58 for intermediates, and −0.75 for experts.The pooled means were 3.90 under claim withholding and 2.62 under disclosure.
- B. RAI w/ Disclosure Effect Claim withheld →Disclosed: Composition standardization yielded ∆std = −1.17 for fairness, while equal weighting of expertise tiers yielded −1.21.Both standardized contrasts preserve lower fairness under disclosure.
- B. RAI w/ Disclosure Effect Claim withheld →Disclosed: Trust increased under disclosure among novices (∆= +1.30) and experts (∆= +1.35), but decreased among intermediates (∆= −0.75).The pooled means were 3.50 under claim withholding and 4.31 under disclosure.
- B. RAI w/ Disclosure Effect Claim withheld →Disclosed: Trust’s pooled contrast was positive, with ∆std = +0.70 under composition standardization and +0.63 under equal weighting, despite opposing expertise-specific directions.The positive aggregate contrast coexists with an expertise-specific reversal.
- B. RAI w/ Disclosure Effect Claim withheld →Disclosed: Toxicity was higher under disclosure in every expertise stratum, with contrasts of +2.00 for novices, +0.50 for intermediates, and +0.60 for experts.Pooled means were 2.00 under claim withholding and 2.85 under disclosure; standardization yielded ∆std = +1.00 and equal weighting yielded +1.03.
5 Discussion
The discussion distinguishes preserved policy parameters from preserved strategy or capability, because RAI suppression and no-op substitution alter the deployed closed-loop behavior. It also frames the human-study contrast as exploratory evidence about capability framing and calls for execution logs linking controller decisions to outcomes.
- Policy and capability: Suppressing or replacing proposed actions with no-ops creates a different closed-loop policy from the one represented by the original weights alone.Substituted actions change subsequent game states and the histories that condition later proposals.
- Deployment interpretation: RAI’s content detector suppressed configured patterns such as worker-unit harassment, while permissible action rates controlled release of otherwise admissible actions.The deployment confirms that participants encountered the RAI execution path, but not how often intervention occurred without proposal and execution logs.
- Human-study limitations: The human study provides exploratory evidence about capability framing, not evidence that RAI improved fairness, increased trust, or reduced perceived toxicity.Both focal presentations used the same RAI controller, so their contrast cannot establish causal benefits from RAI itself.
- Human-study limitations: User perceptions can vary substantially with capability disclosure even when deterministic guardrails remain active.The discussion characterizes perception as highly volatile and anchored by capability information.
- Implementation and reporting: Deployment claims should describe encountered behavior and connect proposals, controller decisions, releases, timestamps, and states to replays and participant outcomes.Figure 2’s event record supports calculating rate interventions, content interventions, and executed-action timing.
6 Conclusion
RAI regulates post-inference action execution through cooldown admission, toxic-action detection, and no-op substitution without retraining the policy. The conclusion formalizes this control point and its audit record, then reports participant perceptions of a StarCraft II opponent using RAI.
- Runtime action interference: RAI regulates which policy proposals reach the environment through cooldown admission, configured toxic-action detection, and no-op substitution without retraining.The mechanism operates after policy inference.
- Runtime action interference: The study formalizes RAI and the event record required to audit its operation.The reported evaluation concerned a StarCraft II opponent documented as using the shared RAI mechanism.
- Human evaluation: Under claim withholding, pooled fairness, trust, and toxicity means were 3.90, 3.50, and 2.00.These perceptions were measured for a StarCraft II opponent using RAI.
A Aggregate Results · A.1 Condition Mapping
The study reports aggregate condition-by-expertise results for two focal presentations that used the same RAI configuration while differing in whether the opponent’s capability claim was disclosed. Aggregate summaries retain descriptive statistics by presentation and expertise, alongside weighted disclosed-minus-withheld outcome differences.
- A Aggregate Results: The focal presentations used the same RAI configuration, including temporal release and configured content-detection rules.The nondisclosure presentation withheld the capability claim.
- A Aggregate Results: Means, descriptive standard deviations, variances, and cell sizes were retained for every presentation-by-expertise combination.Individual ratings and several study records, including the assignment and exclusion ledger and verbatim questionnaire, were not retained.
- A Aggregate Results: −1.17 for fairness, +0.70 for trust, and +1.00 for toxicity are the weighted disclosed-minus-withheld differences.The standardized differences weight novice, intermediate, and expert contrasts by combined focal counts of 7, 7, and 9, respectively, over 23 participants.
- A Aggregate Results: The weighted differences are calculated as ∆std = (7∆y,N + 7∆y,I + 9∆y,E)/23 for outcome y.The displayed cell means produce the reported values subject to rounding of retained means.
- A.1 Condition Mapping: The focal presentations varied capability-claim disclosure rather than runtime interference.Both used the same documented rate-limited opponent and RAI configuration.
- A.1 Condition Mapping: The study did not retain numeric target or realized APM, RAI event logs, launch configuration, checkpoint and replay identifiers, or the exact disclosure script.Participant-level assignments and the baseline implementation were also not retained.
- A.1 Condition Mapping: The optional renderer queue reset was treated as distinct from RAI.This distinction appears in the condition-mapping documentation.
B Personas … E StarCraft II Persona Card Transcriptions
The paper treats synthetic StarCraft II persona outputs as bounded design aids rather than participant evidence, while separating runtime control, capability disclosure, and ethical documentation. It transcribes six persona cards but emphasizes unresolved provenance, contestable assumptions, and the need for factorial follow-up studies.
- B Personas: The synthetic comparison labels and grouping are ambiguous because run-level mappings were not retained and novice and intermediate cards were collapsed.The reported “No Disclosure” label coexists with separate constrained-baseline and high-capability nondisclosure configurations.
- B Personas: Synthetic StarCraft II outputs are reported as model-generated summaries, not participant observations or evidence about player populations.The authors did not retain raw generations, run identifiers, group sizes, complete prompts, seeds, sampling settings, or generation logs, preventing generation-level reproduction or validation.
- B.1 Implications for Use: Synthetic outputs cannot expand the human evidence base, but can identify competing predictions, test formats, and expose missing measures before recruitment.Their means and rationales depend on predefined expertise categories, named heuristics, prompts, and model settings, with unresolved contradictions in labels, grouping, and result direction.
- C.1 Separating Control and Framing Effects: A factorial follow-up should independently vary RAI and capability disclosure using exact scripts, launch manifests, replays, event logs, and participant-level outcomes.This design would estimate effects of runtime interference, framing, their interaction, and different cooldown and content-detection settings.
- D Ethics and Societal Impact Statement: Ethics approval and participant debriefing were completed, but the approval identifier and underlying review record were not retained.Excerpts were deidentified and screened to reduce re-identification risk, while synthetic outputs remained separate from participant evidence.
- E StarCraft II Persona Card Transcriptions: The appendix transcribes six detailed StarCraft II Persona Cards but excludes chat-assistant cards and does not provide a complete Persona Cards dataset.Row-level synthetic interactions, complete prompts, and persona backstories were not retained.
- E StarCraft II Persona Card Transcriptions: A capability claim coincided with outcome-specific evaluations despite a shared documented technical configuration, without identifying a causal mechanism.The finding motivates separate specification of policy capability, release control, and participant-facing information.
- E StarCraft II Persona Card Transcriptions: The six persona cards embed contestable researcher judgments about APM bands, cognitive heuristics, and strategy, without demographic fields or a claim of demographic neutrality.These fields are treated as hypotheses rather than measured player attributes, and incomplete prompts and backstories limit conclusions about the broader generation process.