Chih-Hsi Chen 陳致希

ML engineer & researcher — LLM agents, speech, and ML systems

📍 Taiwan → Pittsburgh, PA

👋 About

I'm an incoming M.S. student in Electrical and Computer Engineering at Carnegie Mellon University (starting January 2027). I received my B.S. in Computer Science and Information Engineering from National Cheng Kung University.

I build machine learning systems end to end — data, training, serving, and the evaluations that prove the result. Most recently I worked on long-term memory for LLM agents at GIGABYTE; before that I did speech research at Academia Sinica and built LLM tool-calling infrastructure at Phison. I currently collaborate remotely on LLM inference systems research with groups at UC Merced and UConn. I also contribute upstream to Apache Airflow and Kinto.

🔬 Research Interests

Long-term memory for LLM agents — knowledge graphs, agentic retrieval, and evidence curation
Efficient LLM systems — Mixture-of-Experts computation reuse, sparse-attention decoding on wafer-scale accelerators, and inference serving
Speech-language modeling — low-resource ASR, encoder fusion, and speech generation

📰 News

Jan 2027Starting my M.S. in ECE at Carnegie Mellon University.
Nov 2026Our DSA-NIFT system appeared in Poster Session I at ROCLING 2026 (NTHU). Poster
Sep 2026GRACE-Mem accepted at the EMNLP 2026 AKBC Workshop — selected as one of three Featured Papers and a Best Paper Award candidate.
Sep 2026Our first-author system paper was accepted at the ROCLING 2026 DSA-NIFT shared task (4th of 7 teams).
Jun 2026Joined GIGABYTE's LLM Algorithm Team as a Generative AI Engineering Intern.
Feb 2026Pull request Kinto #3636 merged, fixing a server crash on invalid input.
Jan 2026Graduated from NCKU with a B.S. in CSIE (GPA 4.15/4.30).
Nov 2025First-author Hakka ASR system paper published at ROCLING 2025.
Jul 2025Joined the Institute of Information Science, Academia Sinica as a Research Assistant.
Oct 2024Joined Phison Electronics' LLM Application Group as an ML Algorithm Engineer Intern.
20241st Place, Hsinchu × Meichu Hackathon (NXP Group).

📄 Publications

GRACE-Mem: Graph Retrieval with Agentic Curation of Evidence for Long-Term Conversational Memory

Heng Wang, Xuan Ni Chen, Pei-Yun Chang, Chih-Hsi Chen, Hsin-Yu Huang, Chienyi Lin, I-Ling Chung

EMNLP 2026 Workshop on Automated Knowledge Base Construction (AKBC) · Accepted · Featured Paper · Best Paper Award candidate

AgentMemoryKnowledge Graph

A long-term conversational memory system that combines graph retrieval with agent operations for evidence curation. With GPT-4o-mini it reaches 85.3% on LoCoMo and 73.2% on LongMemEval-S. I built the agent layer and graph ingestion and ran every experiment in the paper.

[Paper][Code]

NCKU_NLP at ROCLING-2026 Shared Task: Source-Aware Lexicon-Augmented Regression for Valence–Arousal Prediction

Chih-Hsi Chen, Hang Tan

ROCLING 2026 Shared Task (DSA-NIFT) · Accepted · 4th of 7 teams

NLPSentiment

Valence–arousal regression for Chinese new-immigrant reflection texts with no labeled in-domain data, using a MacBERT encoder, a 31-dimensional affective-lexicon feature vector, and source-aware arousal supervision. Test valence PCC 0.866. Finding: without target labels, how arousal is supervised matters more than adding lexical cues, and LLM-generated data buys correlation at the cost of calibration.

[Poster][Code & cue lists (Zenodo)]

The AS-SLAM System for Formosa Speech Recognition Challenge 2025

Chih-Hsi Chen, Pei-Jun Liao, Chia-Hua Wu, Pang-Cheng Wu, Hsin-Min Wang

ROCLING 2025 · Published

SpeechASRLow-resource

Low-resource ASR for the Dapu and Zhaoan Hakka dialects, fine-tuning Whisper-large-v3 with web-collected and TTS-generated speech. On our ablation set, Hanzi CER dropped from 20.10% to 2.54%.

[Paper]

🛠️ Selected Projects

Mixture-of-Experts expert-output cache in vLLM

Research collaboration · UC Merced

Studying whether cached expert outputs can replace expert recomputation in Qwen3-30B-A3B. I optimized the CPU cache path on an A100, cutting lookup from 10.07 to 4.23 µs per route and write from 22.06 to 8.98 µs per row, and replayed 21 eviction policies to find a per-key LRU cap that keeps lookup flat over 3,000 tasks.

Sparse-attention decoding on a wafer-scale accelerator

Research collaboration · UConn

Implemented query-aware page selection and sparse attention as Cerebras CSL kernels, driven by captured Qwen3 decode traces at 8K–64K contexts and checked bit-exactly against a host reference on the SDK simulator.

Model Context Protocol tool-calling platform

Phison Electronics

Nine MCP servers and two agent dispatchers for an internal pilot of about 200 R&D users. A dependency-ordered tool-routing token raised exact tool-sequence match from 64.0% to 88.3%.

Speculative decoding benchmark

GIGABYTE

Compared four speculative-decoding methods for a 27B Qwen model, raising decode throughput from 35.2 to 69.5 tokens/s (1.97×).

Hakka speech-generation service

Academia Sinica

A FastAPI + Redis + MinIO service shared by several partner labs, generating 100–500 speech clips per day.

Apache Airflow — Playwright UI end-to-end tests

Open source

Added Playwright end-to-end tests for the Airflow web UI.

[PR #59374]