Hi NVIDIA NIM team,
I’m a researcher working on an open-source LLM-guided fuzzing framework
and would like to ask whether my account’s rate limit could be raised
above the default 40 RPM — any headroom you can grant would be helpful,
and the higher the better.
About the project
==================
FuzzPilot --- Off-hot-path LLM control for AFL++ via recipe-guided
mutation and micro-campaign validation.
GitHub: https://github.com/qiaozhiyi/fuzz_agent
License: see LICENSE in the repo (open source)
Paper: arXiv preprint in preparation; targeting a software-engineering
venue (ISSTA / FSE / TSE class)
The core idea is that the LLM never appears on the per-mutation hot
path of AFL++. Instead, a controller invokes the model only at
coverage-plateau boundaries to propose a structured "recipe" (op
weights + dictionary tokens + focus ranges). Each recipe is validated
empirically in a short isolated micro-campaign before it is promoted
to the main fuzz loop, and a native AFL++ custom mutator dispatches
on the promoted recipe at full native throughput.
The repo is fully reproducible: target binaries, AFL++ commit pin,
Ghidra extractor, manifest, run scripts, and decision logs are all
under version control.
Why NVIDIA NIM
===============
After evaluating several OpenAI-compatible providers, NIM is the
cleanest fit for an academic reproducibility story:
- One API surface across many open-weight model families
(Llama 4, Nemotron, GLM, DeepSeek, Qwen3, GPT-OSS).
- No per-token billing --- researchers replicating my paper can use
their own NIM account without surprise costs.
- No credit-card friction for academic users.
- OpenAI-compatible, so FuzzPilot's existing \`model_api\` schema
works out of the box.
I'm standardizing the paper's reference configuration on
\`meta/llama-4-maverick-17b-128e-instruct\`, with
\`openai/gpt-oss-120b\` as a documented fallback. Both endpoints
resolve through integrate.api.nvidia.com/v1/chat/completions.
When the paper drops, NVIDIA NIM will be cited as the reference
model backend in the Evaluation and Reproducibility sections, and
the repo's README will include a "Powered by NVIDIA NIM" pointer
to build.nvidia.com plus the Llama 4 Maverick model card.
Why I’d benefit from more RPM
==============================
The paper's evaluation matrix is 11 runs × 4 h wall-clock, organized
as 4 parallel AFL++ campaigns per experiment. Steady-state load is
well under 40 RPM, but plateau bursts (multiple campaigns hitting a
plateau in the same window, each spawning a small set of specialized
agents --- coordinator, plateau-diagnosis, scheduler, dictionary,
format, mutator, corpus, crash-triage) regularly push above it.
Any increase would help. As a concrete data point, \~100 RPM lets the
4-parallel matrix run as designed without serializing campaigns and
doubling the wall-clock; \~200 RPM gives comfortable headroom for
future ablation runs and reproductions by other researchers.
Account info
Single user, non-commercial, results released alongside the paper.
I’m happy to share manifests, sample agent_decisions.jsonl traces, or
anything else useful. Thanks for considering the request.
best,
qiaozhiyi