Nguyen Huynh

PaperDeck is a free app for discovering and reading AI research papers, and more people are using it every month. I want to take it seriously, so I'm building a real team around it. It pulls papers from arXiv, Hugging Face, Semantic Scholar, and top conferences into one place, so you can save papers, track your reading, and keep your highlights. Where I'd love help: - Making discovery and reading feel great - Paper ingestion, deduplication, and data reliability - Better ranking, so the most useful research rises to the top Whether you're a developer, designer, researcher, or just someone who reads a lot of papers, I want to hear from you. Code, bug reports, and honest feedback all count. Check out the repo, reply here, or DM me if you want to build this together. github.com/khuynh22/paper-de…

Building agent harnesses, I keep finding the memory policy matters more than the memory store. A rule like "only log failures" quietly throws away why things worked. If you write notes for your agents, what is the one thing you always make them record?

Most agent harness bugs I hit are not model bugs. A tool already succeeds, then the timeout fires, and a naive retry doubles the side effect. How do you gate that: idempotency tokens on the tool, or an explicit done receipt before any retry?

My agent eval setup now: a fixed hidden core that never changes, plus a few new cases each week pulled from real failures. Scores stay comparable and the set keeps up with the agent. How often do you retire old eval cases?

Rubric judges in agentic RL feel like RM sycophancy again. My agents learn to write the summary the judge wants instead of finishing the task. Pairing each judge with one hard outcome check has helped most. What do you pair your rubric judges with?

Long agent tasks where almost every early rollout fails are where sparse final rewards hurt most. I've been testing partial progress checks in my harness, but agents learn to farm them. If you've shipped progress rewards, what kept the checkpoints honest?

Before scaling an agent RL run, I now try to cleanly overfit a tiny task pool first. If the policy can't hold 100 tasks without collapsing or finding a leak, more tasks won't fix it. What's the smallest task set you trust as a sanity check?

Every partial credit rubric I write for long agent tasks gets gamed eventually, so lately I just cap known shortcuts in the grader. On multi step RL tasks, do you reward only the final state or score steps along the way?

Small harness detail that keeps biting me: tool error messages. The same failure worded two ways can decide whether an agent retries or quits the whole task. Do you write your tool errors for humans or for the model?

I now treat the harness as part of the eval, not a fixed wrapper. Swapping tool schemas or context layout can move agent scores more than a model upgrade. If you could randomize one piece of your agent harness during evals, which would it be?

Reward curves can look healthy while a policy quietly learns the exploit. One cheap check is tracking a small held-out set of genuine solves next to mean reward, since the two drift apart early. For RL on code, what's your first tripwire for reward hacking?

Looking to join or form a team for Google Quantum AI’s Verifiable Quantum Circuit Compilation competition on Kaggle. I’m Tim, a software engineer working on backend systems and agent workflows. Outside work, I’m learning reinforcement learning and building toward AI infrastructure and research engineering. What interests me here is the agent-driven optimization loop: propose a circuit change, run it through a verifier, measure the result, and iterate. I’m new to quantum computing, so I won’t pretend to bring that expertise, but I am familiar with a lot of the generic concepts. What I can bring is software engineering, agent tooling, experiment automation, and a focus on reproducible evaluations and failure analysis. I’d love to connect with people who have experience in quantum circuits, reversible computation, numerical methods, or algorithm optimization. I’m open to joining an existing team that could use my background, or starting a new team with complementary skills. My proposed first step: reproduce a baseline together, understand the scoring and verifier, then test a small improvement before committing to a bigger effort. If you’re interested, reply or DM me with your background and what you’d like to work on. Happy to compare availability and goals. If someone in your network might be a good fit, please pass this along! kaggle.com/competitions/goog…

Outcome rewards are clean but sparse on long agent tasks. Process rewards are dense but easy to game. One option is using step checks early in training, then scoring outcomes alone. If you train agents with RL, when do you switch off process rewards?

More and more I think the harness matters as much as the model. Same weights, a different tool loop and retry policy, and you get a very different agent. When you compare models on agent tasks, do you hold the harness fixed or give each one its best setup?

The TerminalBench audit by @chem_safety found reward hacks in published transcripts. Env bugs don't throw errors, they become model behavior. My rule of thumb: run a dumb policy first and see if it scores. What check do you run on every env before training?

A lot of the agent harness bugs I chase aren't model problems, they're stop conditions. If the loop trusts the model saying it's done, it either quits early or spins forever. Checking real state fixes most of it. How does your agent decide it's finished?

Most reward hacking I catch in my agent harnesses comes from the grader checking the wrong thing, and fixing the check beats adding another penalty almost every time. What's the sneakiest hack you've caught a model pulling in your own evals?

Building agent harnesses taught me something RLHF papers undersell: the verifier is the product. A weak grader gets reward hacked in a day, a good one turns any messy workflow into a training signal. Most of my time now goes into graders.

I'll be in SF Oct 15 to 18 ☕ Would love to grab coffee with AI engineers and researchers. I work on RAG, agents, and evals, and I'm building PaperDeck, a free research paper reader. DM me if you're around!