Chen-Wei Kuo

Research profile

Research summary

Chen-Wei Kuo is an early-stage researcher at National Tsing Hua University (Taiwan), name-resolved from public OpenAlex bibliographic data. The visible record is compact and recent: three arXiv preprints, all posted in 2026, split evenly across three areas — reinforcement learning for battery energy storage dispatch, 4-bit weight-only quantization of large language models, and natural language processing for edge AI systems. The first layers action masking, projection, and advisory PPO to keep storage dispatch feasible under suppressed-spread market shifts; the second shows NF4 quantization preserves global BOS-sink attention-head ranks (ρs ≥ 0.980) while selected top-k head sets shift (0.619–0.793); the third presents ARID, a Jetson-deployed system extracting four structured fields from industrial maintenance work orders at 82.9% F1, 5.3s P50 latency, and 12.5W. Bibliometric indicators are correspondingly thin: h-index 0, i10-index 0, three works in the last three years, and only two distinct coauthors (Kuanlin Chen, Cheng-En Ou), with no paper flagged corresponding-author. The perfectly even field distribution (0.33 each) and mixed OpenAlex topic counts produce a focus score of 0, while reproducibility averages 55 (medium readiness on all three papers). The pattern reads as a technically versatile contributor to applied machine learning rather than a specialized line of work with measurable citation impact yet.

Research keywords

Highlights

• 4-bit quantization preserves global BOS-sink head ranks

  Across 3 small LLMs, NF4 weight-only quantization keeps global ranks of BOS-sink attention heads at ρs ≥ 0.980 while selected top-k head sets shift to 0.619–0.793 — a solid but deliberately narrow empirical study of attention-sink topology under post-training quantization.

• ARID: edge LLM extracts structured fields from maintenance work orders

  A deployable edge AI system distills a 4-bit LLM to pull a 4-field JSON from industrial maintenance work orders directly on Jetson hardware, reaching 82.9% F1 at 5.3s P50 latency under 12.5W — with candid scoping, though baselines are thin and no variance data is reported.

• Cost-aware RL dispatch for battery storage under market shifts

  Layers action masking, projection, and advisory PPO to keep battery energy storage dispatch feasible under suppressed-spread electricity market shifts. Evolve's review finds the approach honest but notes PPO loses to MPC, arbitrage profits are trivially small, and statistical support is thin.

Timeline

1. 2026First cataloged work appears on arXiv: cost-aware reinforcement learning with action masking and projection for battery energy storage dispatch under suppressed-spread market shifts.
2. 2026Publishes the BOS-sink topology study, showing global attention-head ranks survive 4-bit weight-only quantization while selected top-k head sets shift.
3. 2026Releases ARID, a deployable edge AI system for structured information extraction from industrial maintenance work orders, as last author.

Papers in the Evolve catalog

• 2026-09-20

  Cost-Aware Reinforcement Learning with Action Masking and Projection for Battery Energy Storage Dispatch under Suppressed-Spread Market Shifts

• 2026-09-20

  Global Ranks Survive, Selected Heads Shift: BOS-Sink Topology under 4-bit Weight-Only Quantization

• 2026-09-20

  ARID: A Deployable Edge AI System for Structured Information Extraction from Industrial Maintenance Work Orders

Frequent co-authors