Skip to content
View ray24724919's full-sized avatar
2 followers · 5 following

Achievements

Achievement: Pull Shark

Achievements

Achievement: Pull Shark

Block or report ray24724919

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
ray24724919/README.md

Hi 👋, I'm Ray Chen（陳騰睿）

LLM alignment and inference-time control — and shipping the results as software people actually run.

  • 📄 Test-time alignment. Second author on Test-Time Alignment for Large Language Models via Textual Model Predictive Control (ICLR 2026), which treats generation as a predictive-planning problem: draft, score segment contributions with a reward model, re-plan. Preference alignment without touching the weights, aimed at long-form outputs where sequence-level rewards are too coarse.

  • 🔨 Now. Building agent tooling and AI infrastructure — local-first observability for multi-agent coding sessions, and the workflow systems around them.

  • 🎓 Background. M.S. Data Science & Engineering, National Yang Ming Chiao Tung University (2025) · B.S. Applied Mathematics, National Chengchi University (2023) · RL & Bandit Lab.

  • 🧭 Working on: Large Language Models · Test-Time Alignment · Reinforcement Learning · Edge AI Deployment

Selected work
pixel-agents-multi A local-first VS Code control room for Claude and Codex agents — pixel-office visualization, usage intelligence, timeline replay, and a handoff/executor queue. TypeScript
Plan2Align Research code for the ICLR 2026 test-time alignment work: algorithm variants, reward-model tooling, math-reasoning and translation experiments. Python
Baseline_of_Plan2Align Baseline implementations for the same work — ARGS, RAIN, and MetricX-24 evaluation. Python
CVAE_for_Human_Preference Conditional VAE experiments: contextual-bandit regret minimization, and BERT-embedding human-preference/reward prediction on WebGPT comparisons. Python
ViT_MIL Attention-based Multiple Instance Learning over ViT embeddings — bag-level classification from weak supervision only. Python
NCF_BERT4Rec News click prediction: NCF, NCF-Layer, and a Transformer that scores a candidate slate jointly rather than item by item. Python

Coursework
RL_from_Scratch Policy/value iteration, REINFORCE (vanilla/baseline/GAE), and DDPG. Python
Deep_RL_Labs TD learning (2048), DQN (Atari), PPO (Atari), and TD3 (CarRacing). Python
Convex_Optimization_Labs Convex analysis proofs, first-order methods, variance-reduced SGD/SVRG, and QP/conic programming with Gurobi. Python

Connect with me:

teng-ruei-chen 陳騰睿 ray24724919 ray24724919

Languages and Tools:

cplusplus pandas python pytorch scikit_learn

Pinned Loading

  1. Baseline_of_Plan2Align Baseline_of_Plan2Align Public

    Baseline implementations for the Plan2Align / TMPC work (ICLR 2026): ARGS, ARGS for translation, RAIN, and MetricX-24 evaluation.

    Python

  2. NCF_BERT4Rec NCF_BERT4Rec Public

    News click-through prediction over BERT article embeddings - NCF, NCF-Layer, and a Transformer (BERT4Rec) that scores the whole candidate slate jointly rather than item by item.

    Python

  3. pixel-agents-multi pixel-agents-multi Public

    A local-first VS Code control room for Claude and Codex agents - pixel-office visualization, usage intelligence, timeline replay, and a handoff/executor queue.

    TypeScript

  4. Plan2Align Plan2Align Public

    Research code for "Test-Time Alignment for LLMs via Textual Model Predictive Control" (ICLR 2026). Generation as predictive planning: draft, score segment contributions, re-plan.

    Python

  5. ViT_MIL ViT_MIL Public

    Attention-based Multiple Instance Learning with ViT Embeddings for bag classification

    Python