base_model:
- nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16
base_model_relation: finetune
library_name: transformers
pipeline_tag: text-generation
tags:
- nemotron
- agentic-search
- grpo
- web-search
- reinforcement-learning
- post-training
model-index:
- name: nemotron35-rl-step32-20260925
results:
- task:
type: text-generation
name: Agentic web search
dataset:
name: BrowseComp
type: browsecomp
metrics:
- type: accuracy
value: 45.45
name: Accuracy
youdotcom/nemotron35-rl-step32-20260925
A post-trained version of NVIDIA Nemotron 3.5 Lightning (30B-A3B MoE) built for agentic web search. Starting from the base model, it was trained with SFT and then GRPO, using the You.com Search and Contents APIs as tools, to answer hard, multi-hop questions that take many rounds of searching and reading.
On BrowseComp it raises accuracy from 36.97% to 45.45% (+8.5 points, +23% relative) and cuts average tool calls per task from ~168 to 115 (−32%). It got more efficient even though nothing in training rewarded efficiency. The post-trained model gets stuck in fewer search loops, makes better use of each search and page fetch, and is less likely to abandon a correct answer it has already found.
The whole run used fewer than 5K SFT traces and about 1.5K RL tasks, and it took three weeks. It was a joint project of You.com, NVIDIA and CoreWeave.
Training Procedure
Base model and tools: NVIDIA Nemotron 3.5 Lightning 30B-A3B. During data generation, SFT, RL and evaluation, the agent ran in a NeMo Gym harness with two tools: You.com Search and You.com Contents (page fetch).
Training data: Questions came from OpenSeeker, hard subsets of ASearcher, and You.com web-graph QA. The web-graph QA questions are multi-turn and grounded in objective facts from mid-popularity pages in the You.com web graph and the pages they link to. All data was decontaminated against BrowseComp.
Difficulty filtering: The base model attempted each question 8 times with You.com search, and an LLM judge scored every attempt. That came to about 26K agent runs over 3,298 questions. Questions were routed by pass rate:
- Correct traces went to SFT.
- Questions solved 1 to 7 times out of 8 went to RL.
- Questions solved 0 of 8 times were held out.
SFT: 4,932 traces the model generated itself, in its own harness format, trained with the NVIDIA NeMo Automodel recipe. Took about 1 hour on B200s.
RL: GRPO on 1,480 tasks the model sometimes solved. The reward was binary: an LLM judge hosted on CoreWeave Dedicated Inference marked each final answer right or wrong. There was no efficiency reward. vLLM served policy rollouts, and the run took about 8 hours on CoreWeave.
Evaluation
| Model | BrowseComp accuracy | Avg. tool calls / task |
|---|---|---|
| Nemotron 3.5 Lightning (base, default search tool) | 36.97% | ~168 |
| This model (SFT + RL, You.com search) | 45.45% (574/1,263) | 115 |
Evaluation ran on NVIDIA NeMo Gym, the same agent environment used in training. We forked NeMo Gym's BrowseComp harness and made improvements to the agent framework, including <specific changes, e.g. You.com Search and Contents tool integration, …>. Answers were scored by an LLM judge ().
The baseline is NVIDIA's published configuration, so the gain reflects the switch to You.com search, the agent framework changes, the new training data and post-training together, not SFT + RL alone. SFT alone added about 3 points.
On the 684 tasks that appear in both the base and post-trained runs, 104 went from wrong to right. On those 104 tasks, the mean number of tool calls fell from about 252 to about 50.
Acknowledgements
Built by You.com with NVIDIA (model, NeMo Gym, SFT) and CoreWeave (RL training, inference).