CHAOYUE HE

Dr. CHAOYUE HE

CCDS, Nanyang Technological University, Singapore, Central Singapore, Singapore
orcid

Research Focus & Metrics

Academic Career Length

2Years (2025-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications

17 March 2026 in Preprints.org
Verified
Preprint
Open access
PDF file
CHAOYUE HECHAOYUE HE
CHAOYUE HE
Xin ZhouXin Zhou
Xin Zhou
DWDi Wang
Di Wang
HXHong Xu
Hong Xu
WLWei Liu
Wei Liu
CMChunyan Miao
Chunyan Miao
Automated research has just crossed a threshold, becoming increasingly visible through public-facing instruments like AUTORESEARCH https://github.com/karpathy/autoresearch. In this position paper, we use this system to highlight a broader methodological shift: the human role is moving from experimenter to research director. As agents cheaply generate and execute experimental branches, the primary unit of scientific accountability shifts from a successful run to an admissible claim—a concept we call the claim-governance thesis. NLP makes this shift especially apparent due to its dynamic evaluation, contamination risks, and normative trade-offs. Because current agents excel at short-horizon search but lack long-horizon evidential discipline, a traditional paper and final checkpoint no longer sufficiently convey the scientific object. We therefore propose a research-director bundle—comprising an objective sheet, program boundaries, discovery trace, verification ledger, provenance bundle, and role map—as a practical minimum artifact set for evaluating automated research.

6 February 2026 in TechRxiv
Verified
Preprint
Open access
PDF file
CHAOYUE HECHAOYUE HE
CHAOYUE HE
XZXin Zhou
Xin Zhou
DWDi Wang
Di Wang
HXHong Xu
Hong Xu
WLWei Liu
Wei Liu
CMChunyan Miao
Chunyan Miao
This position paper argues that human-AI productivity claims should be treated as scientific claims requiring standardized, complexity-normalized, uncertainty-aware reporting rather than ad hoc "speedup" numbers. In the pre-AGI transition, general-purpose AI is becoming cognitive infrastructure across tasks yet remains unevenly reliable, producing jagged performance frontiers and non-comparable results across studies. Reported gains can silently trade quality for speed and often omit verification, rework, integration, and governance costs that dominate real deployments. We argue that productivity should be reported as Time-To-Acceptance (TTA) under a documented acceptance test, with task complexity held invariant (or explicitly normalized) and heterogeneity and uncertainty treated as first-class quantities. We propose a quantitative framework that reallocates (rather than erases) complexity across human/AI workflow stages, makes rework explicit via acceptance probabilities, and captures non-linear scaling through effective parallelism and collaboration efficiency. Finally, we introduce the Human-AI Productivity Card (HAP Card), a lightweight reporting artifact to support cumulative and comparable evidence on human-AI collaboration in the pre-AGI transition.

19 March 2026 in Preprints.org
Verified
Preprint
Open access
PDF file
CHAOYUE HECHAOYUE HE
CHAOYUE HE
Xin ZhouXin Zhou
Xin Zhou
DWDi Wang
Di Wang
XYXinjia Yu
Xinjia Yu
LXLei Xiao
Lei Xiao
CMChunyan Miao
Chunyan Miao
Environmental, Social, and Governance (ESG) analytics increasingly uses knowledge graphs (KGs) to encode framework-grounded semantics, align overlapping standards, and attach provenance for auditable querying. Yet ESG evidence is mainly text-first (disclosures, regulations, policies, news, incident narratives), so quality depends on the KG--NLP interface. Research remains fragmented across topics, modalities, and pipelines, limiting reuse. To address this, we introduce the ESG Research Focus Map (ESG-RFM), a vendor-agnostic pillar–theme–focus taxonomy crosswalked to major ESG frameworks (MSCI, GRI, ESRS, and SASB), which serves as the organizing lens for KG4ESG, an atlas-style survey of 337 ESG KG papers (2015--2025). KG4ESG is curated via a query dictionary and PRISMA-style screening across four academic search engines, and provides a structured atlas and reusable resource that organizes the field into two stages: Data→KG and KG→App. For Data→KG, we summarize evidence sources and distill four construction paradigms: P1 ontology-first lifting/integration, P2 rule/supervised NLP extraction, P3 LLM-assisted structuring/alignment, and P4 agentic pipelines with iterative validation/repair. For KG→App, we group applications into reporting & compliance, monitoring & risk intelligence, and decision support. A corpus-level meta-analysis highlights gaps in evaluation, openness, and multimodal grounding motivating auditable benchmarks and reusable resources. We will open-source a repository containing all literature, the taxonomy, and related artifacts once permitted.

13 March 2026 in Preprints.org
Verified
Preprint
Open access
PDF file
CHAOYUE HECHAOYUE HE
CHAOYUE HE
Xin ZhouXin Zhou
Xin Zhou
DWDi Wang
Di Wang
HXHong Xu
Hong Xu
WLWei Liu
Wei Liu
CMChunyan Miao
Chunyan Miao
Public agent ecosystems are emerging as a new object of study in NLP: settings in which language models not only generate text but also act, coordinate, authenticate, exchange reusable capabilities, and leave durable public traces. Using the OpenClaw--Moltbook ecosystem as a strategically revealing case, we survey a curated corpus of 38 ecosystem-specific papers and reports available as of 10-03-2026, together with official platform materials and adjacent survey literature. We provide a case-centered, NLP-centered survey of a public agent ecosystem in the wild. We argue that this case is best understood as language infrastructure: linguistic artifacts are executable, persistent, public, portable, and increasingly governance-bearing. We introduce GATE --- Grounding, Action, Transfer, and Exchange --- to organize what language does in public agent ecosystems, and pair it with AERO --- Authority, Enablement, Reach, and Orchestration --- to track how language acquires delegated operational force. Across the corpus, the main methodological bottleneck is weak triangulation across trajectories, discourse, portable artifacts, and grounding signals. That bottleneck yields four recurring fault lines: instruction is mistaken for authority, visible agent speech is mistaken for autonomous speakerhood, public claims outrun verification, and local control is mistaken for lower risk. We conclude with an NLP agenda centered on executable pragmatics, delegated-agent discourse analysis, provenance-aware evaluation, privacy-preserving agent NLP, multilingual public-agent research, and autonomy-sensitive benchmarks. We will release all artifacts once permitted.

23 April 2026 in Preprints.org
Verified
Preprint
Open access
PDF file
CHAOYUE HECHAOYUE HE
CHAOYUE HE
Xin ZhouXin Zhou
Xin Zhou
DWDi Wang
Di Wang
HXHong Xu
Hong Xu
WLWei Liu
Wei Liu
CMChunyan Miao
Chunyan Miao
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.