Chaoyue He
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Chaoyue He has a total of 16 co-authors. Here are the top co-authors by number of shared publications:
- Chunyan Miao
School of Computer Science and Engineering, Nanyang Technological University, Singapore, Singapore
19 shared publications - Xin Zhou
Nanyang Technological University
19 shared publications - Di Wang
Nanyang technological university, Singapore, Singapore
18 shared publications - Hong Xu
Nanyang technological university, Singapore, Singapore
18 shared publications - Wei Liu
Alibaba Group, Hangzhou, China, China
16 shared publications Chunyan Miao
2 shared publications
Latest Publications
- Open Access
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 13178-13185
PCA-OS: A Planetary Climate Adaptation Operating System
Predictive climate machine learning is increasingly good at forecasting hazards, but hazard maps alone do not decide what to do, where, when, for whom, and under which futures. We argue that climate ML remains insufficient for adaptation unless interventions become first-class, versioned, and auditable objects. Many climate digital twins still prioritize state estimation and simulation, whereas adaptation requires intervention observability, counterfactual effect estimation, and constrained portfolio choice. We propose PCA-OS (Planetary Climate Adaptation Operating System), a decision-support operating abstraction built on an intervention-aware global causal knowledge graph. PCA-OS standardizes schemas, versioned updates, query primitives, and audit interfaces across three core objects: (1) an Adaptation Intervention Ledger recording measurable interventions with provenance and uncertainty; (2) a Causal Effect Atlas storing scenario-indexed, spillover-aware estimands, identification assumptions, diagnostics, and sensitivity bounds; and (3) a Robust Portfolio Decision Layer optimizing intervention portfolios under budget, equity, and no-harm constraints. Foundation models and intervention-aware world models should support, not replace, identification-aware causal analysis by surfacing candidate confounders, mechanisms, and spillover pathways for human review. We also outline AdaptBench, an evaluation suite where systems can fail for inequitable or maladaptive recommendations despite high predictive accuracy. The result is a field-level provocation: move climate ML from read-only hazard intelligence to auditable decision support for adaptation.
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 13178-13185
Open AccessPCA-OS: A Planetary Climate Adaptation Operating System
Predictive climate machine learning is increasingly good at forecasting hazards, but hazard maps alone do not decide what to do, where, when, for whom, and under which futures. We argue that climate ML remains insufficient for adaptation unless interventions become first-class, versioned, and auditable objects. Many climate digital twins still prioritize state estimation and simulation, whereas adaptation requires intervention observability, counterfactual effect estimation, and constrained portfolio choice. We propose PCA-OS (Planetary Climate Adaptation Operating System), a decision-support operating abstraction built on an intervention-aware global causal knowledge graph. PCA-OS standardizes schemas, versioned updates, query primitives, and audit interfaces across three core objects: (1) an Adaptation Intervention Ledger recording measurable interventions with provenance and uncertainty; (2) a Causal Effect Atlas storing scenario-indexed, spillover-aware estimands, identification assumptions, diagnostics, and sensitivity bounds; and (3) a Robust Portfolio Decision Layer optimizing intervention portfolios under budget, equity, and no-harm constraints. Foundation models and intervention-aware world models should support, not replace, identification-aware causal analysis by surfacing candidate confounders, mechanisms, and spillover pathways for human review. We also outline AdaptBench, an evaluation suite where systems can fail for inequitable or maladaptive recommendations despite high predictive accuracy. The result is a field-level provocation: move climate ML from read-only hazard intelligence to auditable decision support for adaptation.
- 9 July 2026
- Published by Springer Nature
- p. 527-541
An Expert-In-The-Loop Design Utilising Knowledge Graphs to Prompt LLMs in Professional Writing
- 9 July 2026
- Published by Springer Nature
- p. 527-541
An Expert-In-The-Loop Design Utilising Knowledge Graphs to Prompt LLMs in Professional Writing
- Open Access
- 29 April 2026
- Published by MDPI AG in Preprints.org
Robots That Last: Long-Horizon Robotics Should Optimize Persistent Autonomy
This position paper argues that long-horizon robotics should optimize persistent autonomy, not only longer reset-based episodes. Real deployments require robots that remain safely useful over days to months while accumulating memory, adapting to evolving human preferences, recovering from inevitable failures, and managing constrained physical and computational resources. Many embodied AI evaluations still inherit the logic of episodic reinforcement learning---where environments are frequently reset and hidden human labor is often unreported---but continuous operation exposes vulnerabilities in state continuity, resource coupling, recovery, and maintenance. Although long-term autonomy is not conceptually new, recent progress in generalist robot policies, open robot datasets, and language-conditioned control makes persistence a primary machine-learning evaluation target rather than a deferred downstream systems-engineering concern. As base policies grow more competent, the practical bottlenecks of autonomy concentrate in memory staleness, hidden intervention burden, recovery loops, and maintenance debt. To align evaluation with these realities, we propose a persistent-autonomy scorecard and a layered benchmark blueprint centered on long-run service utility, intervention burden, recovery quality, proactive usefulness, memory hygiene, uptime, and wear-adjusted throughput. By treating persistence as the fundamental scientific object, modern robot learning can focus on systems that turn calendar time into compounding competence rather than relying on isolated task success.
- 29 April 2026
- Published by MDPI AG in Preprints.org
Open AccessRobots That Last: Long-Horizon Robotics Should Optimize Persistent Autonomy
This position paper argues that long-horizon robotics should optimize persistent autonomy, not only longer reset-based episodes. Real deployments require robots that remain safely useful over days to months while accumulating memory, adapting to evolving human preferences, recovering from inevitable failures, and managing constrained physical and computational resources. Many embodied AI evaluations still inherit the logic of episodic reinforcement learning---where environments are frequently reset and hidden human labor is often unreported---but continuous operation exposes vulnerabilities in state continuity, resource coupling, recovery, and maintenance. Although long-term autonomy is not conceptually new, recent progress in generalist robot policies, open robot datasets, and language-conditioned control makes persistence a primary machine-learning evaluation target rather than a deferred downstream systems-engineering concern. As base policies grow more competent, the practical bottlenecks of autonomy concentrate in memory staleness, hidden intervention burden, recovery loops, and maintenance debt. To align evaluation with these realities, we propose a persistent-autonomy scorecard and a layered benchmark blueprint centered on long-run service utility, intervention burden, recovery quality, proactive usefulness, memory hygiene, uptime, and wear-adjusted throughput. By treating persistence as the fundamental scientific object, modern robot learning can focus on systems that turn calendar time into compounding competence rather than relying on isolated task success.
- Open Access
- 27 April 2026
- Published by MDPI AG in Preprints.org
From Super-Apps to Agent Economies: Delegated AI Requires Transaction Closure
This position paper argues that delegated AI must be evaluated and governed by transaction closure, rather than mere task completion. Agent economies emerge when AI systems execute externally binding commitments—such as purchases, bookings, or procurement orders—on behalf of authorized principals. Safe delegation at this boundary requires environments that support bounded mandates, reversible lifecycle states, verifiable receipts, and contestable failures. While AI-native platforms rapidly advance action closure, mature super-app ecosystems and enterprise suites currently offer the clearest testbeds for the payment, refund, and liability semantics required for true transaction closure. This work contributes a formal definition of transaction closure, minimal schema objects for transaction-ready delegation, a comparative ecosystem analysis, a principal-facing mandate-card model, and ClosureBench—a benchmark design for evaluating authorization, split states, recovery, and portability. The long-run goal is contestable transaction closure: portable infrastructure where delegated commitments can be inspected, challenged, and repaired beyond any single host.
- 27 April 2026
- Published by MDPI AG in Preprints.org
Open AccessFrom Super-Apps to Agent Economies: Delegated AI Requires Transaction Closure
This position paper argues that delegated AI must be evaluated and governed by transaction closure, rather than mere task completion. Agent economies emerge when AI systems execute externally binding commitments—such as purchases, bookings, or procurement orders—on behalf of authorized principals. Safe delegation at this boundary requires environments that support bounded mandates, reversible lifecycle states, verifiable receipts, and contestable failures. While AI-native platforms rapidly advance action closure, mature super-app ecosystems and enterprise suites currently offer the clearest testbeds for the payment, refund, and liability semantics required for true transaction closure. This work contributes a formal definition of transaction closure, minimal schema objects for transaction-ready delegation, a comparative ecosystem analysis, a principal-facing mandate-card model, and ClosureBench—a benchmark design for evaluating authorization, split states, recovery, and portability. The long-run goal is contestable transaction closure: portable infrastructure where delegated commitments can be inspected, challenged, and repaired beyond any single host.
- Open Access
- 24 April 2026
- Published by MDPI AG in Preprints.org
Beyond Human Measure: ASI Should Be Guided by Open-World Alignment
ASI should be guided by open-world alignment. While human labels, preferences, and benchmarks remain indispensable, they are too narrow to serve as master objectives for frontier systems. The road to ASI is inherently a frontier supervision problem. As humans become structurally weak supervisors, the technical challenge shifts from imitating judgments to building evaluation loops answerable to reality, explicit constraints, and post-deployment evidence. We critique human-signal monism— the assumption that human-facing signals sufficiently proxy overall system quality—and diagnose its failure modes, including evaluator weakness and proxy over-optimization. As a constructive alternative, open-world alignment centers human intent while coupling it to verifier ecologies: layered task verifiers, hard constraints, uncertainty gates, and monitoring. This framework contributes an operational ASI definition, an actionable formalization, a coding agent case study, an analysis of open-world alignment’s own failure modes, and the Open-World Evaluation Card reporting artifact. Ultimately, ASI must be guided by objectives that remain answerable to the world.
- 24 April 2026
- Published by MDPI AG in Preprints.org
Open AccessBeyond Human Measure: ASI Should Be Guided by Open-World Alignment
ASI should be guided by open-world alignment. While human labels, preferences, and benchmarks remain indispensable, they are too narrow to serve as master objectives for frontier systems. The road to ASI is inherently a frontier supervision problem. As humans become structurally weak supervisors, the technical challenge shifts from imitating judgments to building evaluation loops answerable to reality, explicit constraints, and post-deployment evidence. We critique human-signal monism— the assumption that human-facing signals sufficiently proxy overall system quality—and diagnose its failure modes, including evaluator weakness and proxy over-optimization. As a constructive alternative, open-world alignment centers human intent while coupling it to verifier ecologies: layered task verifiers, hard constraints, uncertainty gates, and monitoring. This framework contributes an operational ASI definition, an actionable formalization, a coding agent case study, an analysis of open-world alignment’s own failure modes, and the Open-World Evaluation Card reporting artifact. Ultimately, ASI must be guided by objectives that remain answerable to the world.
- Open Access
- 23 April 2026
- Published by MDPI AG in Preprints.org
Harness Engineering for Language Agents: The Harness Layer as Control, Agency, and Runtime
Language agents that act through tools, files, browsers, APIs, and persistent sessions are shaped by more than the base model or a single prompt. Their reliability depends on a harness layer that determines which instructions remain authoritative, what actions are available, how state is carried forward, and how failures are handled over time. This paper argues that this layer warrants explicit treatment in NLP. We propose and operationalize a working decomposition of the harness layer as control, agency, and runtime (CAR); situate harness engineering in the arc from software engineering through prompt and context engineering; and audit 63 harness-relevant works, suggesting a meaningful visibility gap between academic papers and public engineering notes. We further argue that many reported agent gains may be partly harness-sensitive rather than purely model-driven, and propose HarnessCard as a lightweight reporting artifact. Grounded in papers, benchmarks, protocols, and engineering notes collected through April 21, 2026, we argue that progress in language agents should report not only the model, but also the harness layer that turns capability into governed action.
- 23 April 2026
- Published by MDPI AG in Preprints.org
Open AccessHarness Engineering for Language Agents: The Harness Layer as Control, Agency, and Runtime
Language agents that act through tools, files, browsers, APIs, and persistent sessions are shaped by more than the base model or a single prompt. Their reliability depends on a harness layer that determines which instructions remain authoritative, what actions are available, how state is carried forward, and how failures are handled over time. This paper argues that this layer warrants explicit treatment in NLP. We propose and operationalize a working decomposition of the harness layer as control, agency, and runtime (CAR); situate harness engineering in the arc from software engineering through prompt and context engineering; and audit 63 harness-relevant works, suggesting a meaningful visibility gap between academic papers and public engineering notes. We further argue that many reported agent gains may be partly harness-sensitive rather than purely model-driven, and propose HarnessCard as a lightweight reporting artifact. Grounded in papers, benchmarks, protocols, and engineering notes collected through April 21, 2026, we argue that progress in language agents should report not only the model, but also the harness layer that turns capability into governed action.
- Open Access
- 22 April 2026
- Published by MDPI AG in Preprints.org
Elevate Before You Eliminate: Firms Should Redesign High-Risk Roles Before Any AI-Attributed Layoffs
This position paper argues for a rebuttable presumption against AI-attributed layoffs, challenging the hardening managerial default that equates model-addressable task exposure with inevitable worker redundancy. We demonstrate that while generative AI compresses routine workflow substrate, it simultaneously expands a role's elevation space---the critical human layer of judgment, exception handling, orchestration, and institutional accountability. We formalize this dynamic to show that apparent automation ceilings are frequently local artifacts of frozen job designs rather than immutable technological frontiers. Driven by an emerging strategic pattern where firms explicitly use labor cuts to self-fund AI investments, we advance an elevate-first rule: organizations must systematically attempt workflow redesign, paid upskilling, internal mobility, and apprenticeship preservation before declaring headcount redundant. To operationalize this standard within the AI research and deployment community, we propose the Workplace AI Transition Card for transparent transition reporting and introduce an Elevation Impact Factor (EIF). Ultimately, we argue that beneficial AI must be evaluated not merely by throughput or substitution rates, but by its capacity to move the human value frontier outward and sustain long-term epistemic resilience.
- 22 April 2026
- Published by MDPI AG in Preprints.org
Open AccessElevate Before You Eliminate: Firms Should Redesign High-Risk Roles Before Any AI-Attributed Layoffs
This position paper argues for a rebuttable presumption against AI-attributed layoffs, challenging the hardening managerial default that equates model-addressable task exposure with inevitable worker redundancy. We demonstrate that while generative AI compresses routine workflow substrate, it simultaneously expands a role's elevation space---the critical human layer of judgment, exception handling, orchestration, and institutional accountability. We formalize this dynamic to show that apparent automation ceilings are frequently local artifacts of frozen job designs rather than immutable technological frontiers. Driven by an emerging strategic pattern where firms explicitly use labor cuts to self-fund AI investments, we advance an elevate-first rule: organizations must systematically attempt workflow redesign, paid upskilling, internal mobility, and apprenticeship preservation before declaring headcount redundant. To operationalize this standard within the AI research and deployment community, we propose the Workplace AI Transition Card for transparent transition reporting and introduce an Elevation Impact Factor (EIF). Ultimately, we argue that beneficial AI must be evaluated not merely by throughput or substitution rates, but by its capacity to move the human value frontier outward and sustain long-term epistemic resilience.
- Open Access
- 20 April 2026
- Published by MDPI AG in Preprints.org
The One-Person Laboratory Should Be a First-Class Unit of Evaluation in Dry-Lab AI Research
This position paper argues that in software-defined dry-lab AI research, the one-person laboratory (OPL) is the relevant minimum accountable unit under compressed coordination and should be treated as a first-class unit of evaluation wherever bounded verification and public contestability hold. We develop three propositions. P1 (descriptive): public research-agent systems and laboratory-shaped benchmarks suggest that the minimum efficient research unit is moving downward in parts of AI research. P2 (causal, conditional): the relevant gains are narrower than common ``AI scientist'' rhetoric implies---not general scientific superiority, but lower iteration latency per admitted claim, stronger provenance, higher replayability, clearer responsibility, and better retention of negative branches when abstention and disclosure are enforced. P3 (normative, conditional): the community should therefore evaluate and support OPLs as claim-producing laboratories rather than only models or PDFs, while simultaneously building public execution interfaces, trace-linked claim standards, benchmark sandboxes, and access institutions. Our empirical anchor is a purposive structured interpretive read of representative public systems and benchmarks; it is not a leaderboard and does not estimate prevalence, causal impact, or superiority. We do not claim that \opl{}s replace strong teams, justify broad scientific claims from a single run, or cleanly extend to wet-lab, clinical, or human-subject domains. The paper's contribution is a falsifiable governance position: if laboratory-shaped systems fail to cohere, if OPL-style runs do not improve admitted-claim speed or auditable process quality, or if access remains closed, the thesis should be weakened or reversed.
- 20 April 2026
- Published by MDPI AG in Preprints.org
Open AccessThe One-Person Laboratory Should Be a First-Class Unit of Evaluation in Dry-Lab AI Research
This position paper argues that in software-defined dry-lab AI research, the one-person laboratory (OPL) is the relevant minimum accountable unit under compressed coordination and should be treated as a first-class unit of evaluation wherever bounded verification and public contestability hold. We develop three propositions. P1 (descriptive): public research-agent systems and laboratory-shaped benchmarks suggest that the minimum efficient research unit is moving downward in parts of AI research. P2 (causal, conditional): the relevant gains are narrower than common ``AI scientist'' rhetoric implies---not general scientific superiority, but lower iteration latency per admitted claim, stronger provenance, higher replayability, clearer responsibility, and better retention of negative branches when abstention and disclosure are enforced. P3 (normative, conditional): the community should therefore evaluate and support OPLs as claim-producing laboratories rather than only models or PDFs, while simultaneously building public execution interfaces, trace-linked claim standards, benchmark sandboxes, and access institutions. Our empirical anchor is a purposive structured interpretive read of representative public systems and benchmarks; it is not a leaderboard and does not estimate prevalence, causal impact, or superiority. We do not claim that \opl{}s replace strong teams, justify broad scientific claims from a single run, or cleanly extend to wet-lab, clinical, or human-subject domains. The paper's contribution is a falsifiable governance position: if laboratory-shaped systems fail to cohere, if OPL-style runs do not improve admitted-claim speed or auditable process quality, or if access remains closed, the thesis should be weakened or reversed.
- Open Access
- 13 April 2026
- Published by MDPI AG in Preprints.org
Let Papers Flow: AI Conferences Should Embrace Submission Explosion via Autonomous Review Pipelines
As AI tools accelerate literature search, experimentation, and drafting, AI conferences face a growing, structural submission explosion. This position paper argues that to survive this era of accelerated research, venues must formally adopt autonomous review pipelines as their core control plane. By deploying machine-first, human-governed systems to run the scalable first pass of review—handling routine evidence construction, integrity checks, and routing—conferences can stop treating abundance as a pathology. Making this automated control plane the spine of the review process preserves scarce human attention for high-stakes escalation, calibration, and edge cases. Furthermore, we show that establishing this infrastructure carries a profound downstream implication: it enables conferences to decouple baseline technical certification from seasonal curation, ultimately absorbing more submissions, shortening time-to-feedback, and tightening quality standards without breaking the human review commons or requiring perpetual committee inflation.
- 13 April 2026
- Published by MDPI AG in Preprints.org
Open AccessLet Papers Flow: AI Conferences Should Embrace Submission Explosion via Autonomous Review Pipelines
As AI tools accelerate literature search, experimentation, and drafting, AI conferences face a growing, structural submission explosion. This position paper argues that to survive this era of accelerated research, venues must formally adopt autonomous review pipelines as their core control plane. By deploying machine-first, human-governed systems to run the scalable first pass of review—handling routine evidence construction, integrity checks, and routing—conferences can stop treating abundance as a pathology. Making this automated control plane the spine of the review process preserves scarce human attention for high-stakes escalation, calibration, and edge cases. Furthermore, we show that establishing this infrastructure carries a profound downstream implication: it enables conferences to decouple baseline technical certification from seasonal curation, ultimately absorbing more submissions, shortening time-to-feedback, and tightening quality standards without breaking the human review commons or requiring perpetual committee inflation.
- Open Access
- 8 April 2026
- Published by MDPI AG in Preprints.org
Remote-Capable Knowledge Work Should Default to AI-Enabled Flexibility
This position paper argues that remote-capable knowledge work should default to AI-enabled flexibility because the workflow-integrated foundation-model stack changes the coordination economics that once favored daily co-presence. By foundation-model stack, we mean systems that combine natural-language interaction, multimodal capture, long context, retrieval, transcription, translation, and increasingly bounded tool use inside everyday workflows. Their organizational significance is not generic automation but the accumulation of artifact capital: durable, queryable, reusable traces such as transcripts, summaries, decisions, tickets, code comments, and retrieval layers. The argument rests primarily on capabilities that are already widely deployed---transcription, summarization, retrieval, translation, drafting, and code assistance---with bounded agents treated as an amplifying but not necessary extension. Rather than eliminating the office, this shift supports selective co-presence, reserving in-person time for tasks with high tacitness, high coupling, or high relational stakes, including apprenticeship, conflict repair, trust formation, and early-stage synthesis. Because the same systems can also intensify surveillance, skill atrophy, and compute-related emissions, we outline a machine-learning research agenda centered on team-level evaluation, privacy-preserving memory layers, scaffolded AI for learning, carbon-aware routing, and pro-agency workflow design.
- 8 April 2026
- Published by MDPI AG in Preprints.org
Open AccessRemote-Capable Knowledge Work Should Default to AI-Enabled Flexibility
This position paper argues that remote-capable knowledge work should default to AI-enabled flexibility because the workflow-integrated foundation-model stack changes the coordination economics that once favored daily co-presence. By foundation-model stack, we mean systems that combine natural-language interaction, multimodal capture, long context, retrieval, transcription, translation, and increasingly bounded tool use inside everyday workflows. Their organizational significance is not generic automation but the accumulation of artifact capital: durable, queryable, reusable traces such as transcripts, summaries, decisions, tickets, code comments, and retrieval layers. The argument rests primarily on capabilities that are already widely deployed---transcription, summarization, retrieval, translation, drafting, and code assistance---with bounded agents treated as an amplifying but not necessary extension. Rather than eliminating the office, this shift supports selective co-presence, reserving in-person time for tasks with high tacitness, high coupling, or high relational stakes, including apprenticeship, conflict repair, trust formation, and early-stage synthesis. Because the same systems can also intensify surveillance, skill atrophy, and compute-related emissions, we outline a machine-learning research agenda centered on team-level evaluation, privacy-preserving memory layers, scaffolded AI for learning, carbon-aware routing, and pro-agency workflow design.