Teaching NVIDIA's Nemotron 3.5 Super-VL ML Engineering

Abstract

Autonomous Machine learning research remains hard, even for frontier models. World class MLEs know how to explore promising approaches, design useful experiments, interpret their results, and revise their plans when the experimental evidence changes. These decisions unfold over hours of work, with each experiment affecting what the agent should do next.

With early access to NVIDIA Nemotron-3.5-Super-VL open model, we investigated how training on long horizon research based MLE tasks could improve this capability. Using a large dataset of long-horizon MLE tasks, we explored two sources of supervision: selected decisions from teacher trajectories and Nemotron’s own successful guided attempts, an approach we call self-context distillation (SCD).

Both approaches improved the reported AutoResearchExam score. Extending SCD with retrospective lessons and replayed action examples produced our strongest model: 0.358 versus 0.286 for Base, a 24.8% relative improvement, using the same Terminus-2 harness. Alongside these results, we examined how to preserve the evidence behind a decision, remove dependencies on privileged guidance, and combine lessons from experience with concrete examples of research actions.

In the corresponding comparison using the same tasks, harness, and time horizon, Muse Spark 1.3 scores 0.258 and Qwen 3.8 Max scores 0.353. Our final model therefore reaches a score comparable to Qwen 3.8 Max on this evaluation.

Methodology and Experimental Setup

Evaluation Dataset

AutoResearchExam is developed for evaluating how models perform over extremely long-horizons in a variety of machine-learning tasks[8]. Out of the original 29 total tasks, for this experiment we focus on 22 CPU tasks, as the other 7 need dedicated H100 GPUs, along with 14-18 hour sandboxes.

Each task has its own default time limit: 4 hours on 19 tasks, 5 hours on 1, and 6 hours on 2. We leave the limits unchanged, and we raise the turn limit to 1,000 so that time, not turns, is the constraint. The agent submits once and gets no feedback from the grader. A task's score is the grader's reward between 0 and 1, and we report the mean over the 22 tasks.

Our focus is whether post-training improves Nemotron’s performance within the existing Terminus-2 workflow, using the prescribed task limits and a single final graded submission. The public ARE leaderboard evaluates a different setting, with a 24-hour horizon and repeated submissions with feedback. Our results therefore measure endpoint solution quality under this specific protocol and are not directly comparable to the public leaderboard.

Data Curation Methodologies

We built our training datasets from collections of long-horizon machine learning engineering tasks spanning optimization, training, alignment, data curation, and related research work. We adopt two methodologies, learning from a stronger teacher model and learning from self context. Both methods use data from our pool of tasks.

The teacher data came from an existing rollout archive covering 465 tasks. For Nemotron’s own attempts, the collection program drew on a broader pool of approximately 850 internally developed tasks. These figures describe the source collections; the training examples were selected downstream through task exclusions, quality review, and sampling.

Across both routes, the central process was the same: collect research experience, identify useful learning signals, reconstruct the context available at each decision, and train on selected model responses. The routes differed in how those demonstrations were generated and prepared.

We screened training sources for overlap with ARE using task identifiers, instruction-text comparisons, and review of suspected variants. Checks of the retained SCD and reflection datasets, including replay examples, confirmed that they didn't contain any eval task contamination.

Learning from a Stronger Teacher

The first route extracted useful decisions from existing Qwen 3.8 Max teacher trajectories. We examined the preceding observations, the teacher’s reasoning, and the action it took next, with the aim of retaining examples that connected evidence to a meaningful research decision. Each training example preserved the context needed to interpret the selected response. When the full history exceeded the training window, we combined relevant earlier evidence with recent observations. Only the selected model response contributed to the training loss; the preceding history supplied context.

We call the resulting approach Curated Decision SFT. It explored how existing research trajectories could provide reusable supervision for future tasks. The final dataset contained 400 training examples with approximately 2.19 million supervised tokens.

Learning from Nemotron’s own Attempts

The second route generated demonstrations using Nemotron itself, with additional guidance during collection. This approach builds on context distillation: generating useful behavior with teachers who have additional context, then training the model to reproduce it without that context. Work such as Memento No More applies a related idea to internalizing hints for agent tasks. Because Nemotron both generated the demonstrations and learned from them, we call this technique self-context distillation, or SCD.

Our recipe had four steps:

Attempt the task with help. Add terminal-use guidance and, for selected previously failed tasks, short hints aimed at the observed failure.

Select useful attempts. Apply outcome, completion, grading, and trajectory-length filters, alongside comparisons with earlier attempts on the same task.

Remove the help. Reconstruct the training context without the guidance or hints. Exclude steps whose reasoning depends on information supplied only by a removed hint.

Supervise selected actions. Extract model responses from the retained trajectories and train on those responses, with the preceding observations supplying context.

Removing guidance required more care than deleting a block of text. A correct action could still be explained using a fact that the student would no longer see. We therefore checked whether the retained decision remained supported by its visible context. The final SCD dataset contained 3,365 training examples with approximately 7.44 million supervised tokens.

Both selected teacher decisions and Nemotron’s own guided attempts provided useful supervision, reaching reported scores of 0.31 and 0.31 respectively. Because the datasets and training configurations differed, these results do not establish which approach is more effective under a matched budget.

We concentrated subsequent experiments on SCD to pursue one research direction in greater depth, building on the collection and training workflow already in place. This provided a natural setting for investigating how Nemotron could learn from its own attempts through both action demonstrations and retrospective lessons.

Extending SCD with Reflection

We extended the SCD model with lessons it generated from its own research attempts. Inspired by Retrospection-Only Fine-Tuning (ROFT), we investigated whether these retrospective lessons could complement supervision from successful actions. We extended SCD because it supplied both Nemotron’s own transcripts and the successful action examples needed for this experiment.

Reviewing past attempts

We compared 64 pairs of attempts at the same training task, each with a better and a worse result. An LLM judge found a weaker approach in 77% of pairs and insufficient testing in 61%. It also identified problems with terminal use, runtime, and submission requirements.

We used these findings to ask the SCD model to review its own runs and explain what it would do differently next time. We checked that each lesson followed the required format and matched what happened in the run. An LLM judge rated the lessons; the SCD model wrote them.

For example, one planner failed all 16 local tests, each with a four-minute limit. The model switched to a simpler method and fixed its implementation. One later simulation finished in about a minute. We kept the lesson it wrote afterwards: fix the runtime problem before making further model changes, then check that the solution still finishes on time when several tests run together.

Training on Reflections

Our final recipe combined 972 reflection examples with 500 action examples replayed from the first-stage SCD dataset. Starting from the SCD checkpoint, we trained for one epoch at a learning rate of 3e-5, over 184 steps.

For reflection examples, the transcript supplied context while supervision targeted the model’s retrospective response. For replay examples, supervision targeted the original reasoning and next action. This kept concrete examples of agent behavior present alongside the new retrospective lessons.

Reflection-only variants regressed relative to their SCD parent in our diagnostic evaluation. Combining reflections with action replay performed better than those variants and produced the highest score in the reported 22-task comparison. The diagnostic evaluation used a 19-task subset of ARE rather than 22-task comparison.

Learning rate and training mixture changed together, and the comparison did not include a matched replay-only control. The result therefore supports the combined recipe without isolating reflection’s or replay’s individual contribution.

We call this model Base + SCD + Reflection. Reflection was incorporated during training; evaluation used the same Terminus-2 harness without an additional reflection loop.

Results

We evaluated all four models on the 22 CPU tasks from AutoResearchExam (ARE), Bespoke’s benchmark of extended machine learning engineering tasks. Scores measure final submission quality under task-specific limits of four to six hours.

* Each task scores between 0 and 1.

** Relative improvements are calculated from unrounded means. Standard errors describe variation across tasks, not uncertainty across repeated evaluations.

Base + SCD + Reflection scored higher than Base on 16 tasks, lower on four, and equally on two.

Qwen 3.8 Max has 20× Nemotron’s total parameter count: 2.4 trillion versus 120 billion. This comparison concerns total parameters, rather than active parameters or inference cost.

The Model and Agent

Our base model is NVIDIA's Nemotron-3.5-Super-VL 120B-A12B: a hybrid Mamba-attention mixture-of-experts model with 120B total and 12B active parameters. Every post-trained model in this post is a LoRA adapter (rank 32, alpha 64) on the attention and Mamba projection layers: 224 adapter tensors, 205 MB, which is 0.08% of the 249 GB of base weights.

We use the Terminus-2 harness, a terminal agent that returns JSON with its analysis, its plan, and the keystrokes to send. We use sampling temperature 1.0, top-p 0.95, and give the model up to 32,768 output tokens per turn. This is the standard harness also used by AutoResearchExam.

Future work

A central next question is how to select and identify the research experiences that a model can learn from, using its own rollouts and possible guidance from a teacher. We used failure analysis to guide hints, continuation rollouts, and retrospective supervision. A natural extension is to make collection adaptive: identify recurring weaknesses in the current model and gather demonstrations that directly address them.

This could include alternative approaches from a shared starting point, recovery from failed experiments, and decisions about verification, resource allocation, and completion. Where environment state can be restored, short continuations from consequential decision points could provide targeted supervision without repeating an entire research attempt.

Data curation would then extend beyond selecting successful trajectories. Useful decisions can occur within unsuccessful attempts, while successful runs can contain redundant or unsupported actions. Future datasets could retain verified sequences that connect an observed problem to a diagnosis, an executable response, and a measured outcome, while preserving the evidence available to the model at each step.

As the model improves, the collection strategy could evolve with it. Comparing these adaptive curricula under matched budgets would help establish which kinds of experience transfer to new research problems, and whether those gains support sustained progress over longer research horizons.

Teacher-generated supervision remains an important direction. More capable teachers, teachers with complementary strengths, or better selection of their decisions could produce stronger demonstrations. Future experiments could compare these sources under matched training budgets and investigate whether teacher demonstrations and self-generated recovery examples provide complementary benefits.

Conclusion

For open-ended tasks without golden trajectories, a useful training signal lies beyond the final solution. Rollouts record which approaches were tried, what the tests revealed, and how the agent responded when something failed. Using Bespoke’s research environments and trajectory data, we found useful supervision in both teacher demonstrations and Nemotron’s own guided attempts.

Turning those attempts into training examples required careful choices: which actions to teach, what context to preserve, and how to combine retrospective lessons with examples of successful work. We tested those choices through the resulting agent’s performance. Our strongest model, Base + SCD + Reflection, reached 0.358 compared with 0.286 for Base - a 24.8% relative improvement across the 22 CPU tasks in our ARE evaluation, using the same Terminus-2 harness.

These results show a practical path for specializing NVIDIA Nemotron through supervised post-training. Bespoke’s research tasks supplied the experience; trajectory analysis and data preparation made that experience useful for training. Together, they helped an open-weight model produce stronger solutions to demanding machine learning engineering problems.

Acknowledgements

We thank NVIDIA for providing early access to Nemotron-3.5-Super-VL, which made these experiments possible.

If you use this work, please cite it as:

@misc{bespoke_teaching_nvidias_nemotron_ml_engineering},
    author = {Patel, Romir* and Kumar, Mrinal* and Ramesh, Anirudha and Parulekar, Advait and Devic, Siddartha and Pimpalgaonkar, Shreyas and Dimakis, Alex and Sathiamoorthy, Maheswaran},
    title = {Teaching NVIDIA's Nemotron 3.5 Super-VL ML Engineering},
    howpublished = {https://www.bespokelabs.ai/blog/teaching-nvidias-nemotron-ml-engineering},
    note = {Published: 2026-10-15},
    year = {2026}
}

@misc{bespoke_teaching_nvidias_nemotron_ml_engineering},
    author = {Patel, Romir* and Kumar, Mrinal* and Ramesh, Anirudha and Parulekar, Advait and Devic, Siddartha and Pimpalgaonkar, Shreyas and Dimakis, Alex and Sathiamoorthy, Maheswaran},
    title = {Teaching NVIDIA's Nemotron 3.5 Super-VL ML Engineering},
    howpublished = {https://www.bespokelabs.ai/blog/teaching-nvidias-nemotron-ml-engineering},
    note = {Published: 2026-10-15},
    year = {2026}
}

@misc{bespoke_teaching_nvidias_nemotron_ml_engineering},
    author = {Patel, Romir* and Kumar, Mrinal* and Ramesh, Anirudha and Parulekar, Advait and Devic, Siddartha and Pimpalgaonkar, Shreyas and Dimakis, Alex and Sathiamoorthy, Maheswaran},
    title = {Teaching NVIDIA's Nemotron 3.5 Super-VL ML Engineering},
    howpublished = {https://www.bespokelabs.ai/blog/teaching-nvidias-nemotron-ml-engineering},
    note = {Published: 2026-10-15},
    year = {2026}
}

References

References

[1] Askell, A. et al. A General Language Assistant as a Laboratory for Alignment. arXiv:2112.00861, 2021.

[2] Snell, C. et al. Learning by Distilling Context. arXiv:2209.15189, 2022.

[3] Light, J. et al. Shockingly Simple Self-retrospection Improves Agentic Models Without RL. arXiv:2609.35741, 2026.

[4] Alakuijala, M. et al. Memento No More: Coaching AI Agents to Master Multiple Tasks via Hints Internalization. arXiv:2502.01562, 2025.

[5] Bhargava, A. et al. Prompt Baking. arXiv:2409.13697, 2024.

[6] Ye, T. et al. On-Policy Context Distillation for Language Models. arXiv:2602.12275, 2026.

[7] Shenfeld, I. et al. Self-Distillation Enables Continual Learning. arXiv:2601.19897, 2026.

[8] Bespoke Labs. Introducing AutoResearchExam. September 10, 2026.

[9] Zelikman, E. et al. STaR: Bootstrapping Reasoning With Reasoning. arXiv:2203.14465, 2022.

[10] Wang, S. et al. Hindsight Hint Distillation: Scaffolded Reasoning for SWE Agents from CoT-free Answers. arXiv:2605.11556, 2026.

[11] Penaloza, E. et al. Privileged Information Distillation for Language Models. arXiv:2602.04942, 2026.