--- title: How to Run an Autoresearch Workflow with RL Agent Skills... description: A walkthrough of running an autonomous ML research workflow using AI coding agents (Codex with GPT 5.5), NVIDIA NeMo RL, and NeMo Gym on a Brev GPU instance.... canonical: https://daily.dev/posts/how-to-run-an-autoresearch-workflow-with-rl-agent-skills-and-nvidia-nemo-wki2g3ref twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: How to Run an Autoresearch Workflow with RL Agent Skills and NVIDIA NeMo | daily.dev og:description: A walkthrough of running an autonomous ML research workflow using AI coding agents (Codex with GPT 5.5), NVIDIA NeMo RL, and NeMo Gym on a Brev GPU instance.... og:url: https://daily.dev/posts/how-to-run-an-autoresearch-workflow-with-rl-agent-skills-and-nvidia-nemo-wki2g3ref og:image: https://api.daily.dev/og/posts/WKI2g3ReF.png og:image:alt: How to Run an Autoresearch Workflow with RL Agent Skills and NVIDIA NeMo og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # How to Run an Autoresearch Workflow with RL Agent Skills and NVIDIA NeMo **[NVIDIA Developer](https://daily.dev/sources/nvidiadev)** · 14 min read · 1 upvotes · 0 comments ## Summary A walkthrough of running an autonomous ML research workflow using AI coding agents (Codex with GPT 5.5), NVIDIA NeMo RL, and NeMo Gym on a Brev GPU instance. The workflow demonstrates three agent capabilities: full-stack autonomy (environment setup, dependency resolution, GPU management), goal-driven autoresearch (baseline profiling, hypothesis generation, experiment iteration), and paper-to-code (reading a research paper and implementing the OAPL off-policy RL algorithm). In a concrete example, the agent trained a Qwen3-VL-2B-Instruct model on a custom visual counting task, improving accuracy from 25% to 96.9%. Three reusable agent skills (brev-etiquette, session-memory, autoresearch) encode operating context and institutional knowledge to make the research loop reproducible. Best practices include setting explicit time/compute budgets, staged human review, and treating the agent as a capable junior colleague rather than a fully autonomous researcher. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How much did accuracy improve when training Qwen3-VL-2B-Instruct on a visual counting task with NeMo RL? Accuracy improved from a 25.0% out-of-the-box baseline to 96.875% after an SFT campaign using LoRA on a custom NeMo Gym star-counting environment. Mean reward rose to 0.992188 at the best checkpoint, from an initial 0.705208. Remaining errors were near-miss off-by-one counting mistakes rather than formatting failures, suggesting hard-case data augmentation as a next step. _daily.dev surfaces practical writeups like this for engineers benchmarking agent-driven RL fine-tuning results._ ### Why would GRPO be replaced with SFT for training a NeMo Gym RL environment? A prior autoresearch session found that GRPO did not improve accuracy on a star-counting visual task, while a switch to SFT (via LoRA) proved effective, eventually reaching 96.875% exact accuracy. This choice was made by an autonomous coding agent after profiling results rather than assumed upfront, illustrating how algorithm selection can be data-driven mid-campaign. _Compare training algorithm trade-offs for your own RL tasks by following coverage like this on daily.dev._ ### What is OAPL and how does it compare to DAPO for off-policy reinforcement learning? OAPL (Optimal Advantage-based Policy Optimization with Lagged Inference policy) is a long-lag off-policy RL method that treats the training-inference policy mismatch as part of the training objective rather than eliminating it. On a Qwen3-1.7B model with the DAPO-math-17K dataset, OAPL reached higher accuracy with fewer training steps than the GRPO-variant DAPO, though it used noticeably more tokens per response. _daily.dev helps researchers track emerging off-policy RL techniques like OAPL as they compare training approaches._ ## Similar posts on daily.dev - [How to Train Scientific Agents with Reinforcement Learning](https://daily.dev/posts/how-to-train-scientific-agents-with-reinforcement-learning-uwezkfknl) · NVIDIA Developer · 1 upvotes · 0 comments - [How to Build an AI Agent That Runs its Own LLM Experiments with autoresearch](https://daily.dev/posts/how-to-build-an-ai-agent-that-runs-its-own-llm-experiments-with-autoresearch-brly8l88g) · freeCodeCamp · 3 upvotes · 1 comments - [Build an AI Scientist for Life Science Discovery with NVIDIA BioNeMo Agent Toolkit](https://daily.dev/posts/build-an-ai-scientist-for-life-science-discovery-with-nvidia-bionemo-agent-toolkit-plwbrgvwb) · NVIDIA Developer · 1 upvotes · 0 comments - [Conversational LLM Evaluations in Minutes with NVIDIA NeMo Evaluator Agent Skills](https://daily.dev/posts/conversational-llm-evaluations-in-minutes-with-nvidia-nemo-evaluator-agent-skills-ewmza9ujj) · Hugging Face · 1 upvotes · 0 comments - [Andrej Karpathy’s 630-line Python script ran 50 experiments overnight without any human input](https://daily.dev/posts/andrej-karpathy-s-630-line-python-script-ran-50-experiments-overnight-without-any-human-input-pjqwmdwbv) · The New Stack · 45 upvotes · 0 comments --- Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#reinforcement-learning](https://daily.dev/tags/reinforcement-learning), [#vlm](https://daily.dev/tags/vlm), [#nvidia-nemo](https://daily.dev/tags/nvidia-nemo) [View this post on daily.dev](https://daily.dev/posts/how-to-run-an-autoresearch-workflow-with-rl-agent-skills-and-nvidia-nemo-wki2g3ref) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"How to Run an Autoresearch Workflow with RL Agent Skills and NVIDIA NeMo","url":"https://daily.dev/posts/how-to-run-an-autoresearch-workflow-with-rl-agent-skills-and-nvidia-nemo-wki2g3ref","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/how-to-run-an-autoresearch-workflow-with-rl-agent-skills-and-nvidia-nemo-wki2g3ref"},"datePublished":"2026-07-14T16:01:27.789Z","dateModified":"2026-09-14T06:50:34.793Z","description":"A walkthrough of running an autonomous ML research workflow using AI coding agents (Codex with GPT 5.5), NVIDIA NeMo RL, and NeMo Gym on a Brev GPU instance....","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/75353bfdaeacaf189a9a7f6febcb5fd8?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/75353bfdaeacaf189a9a7f6febcb5fd8?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"NVIDIA Developer","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"NVIDIA Developer","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/86e45aab42ba48ce83103d01b1119910","url":"https://daily.dev/sources/nvidiadev"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/how-to-run-an-autoresearch-workflow-with-rl-agent-skills-and-nvidia-nemo-wki2g3ref","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai-agents,reinforcement-learning,vlm,nvidia-nemo","timeRequired":"PT14M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"NVIDIA Developer","item":"https://daily.dev/sources/nvidiadev"},{"@type":"ListItem","position":3,"name":"How to Run an Autoresearch Workflow with RL Agent Skills and NVIDIA NeMo"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/how-to-run-an-autoresearch-workflow-with-rl-agent-skills-and-nvidia-nemo-wki2g3ref#faq","mainEntity":[{"@type":"Question","name":"How much did accuracy improve when training Qwen3-VL-2B-Instruct on a visual counting task with NeMo RL?","acceptedAnswer":{"@type":"Answer","text":"Accuracy improved from a 25.0% out-of-the-box baseline to 96.875% after an SFT campaign using LoRA on a custom NeMo Gym star-counting environment. Mean reward rose to 0.992188 at the best checkpoint, from an initial 0.705208. Remaining errors were near-miss off-by-one counting mistakes rather than formatting failures, suggesting hard-case data augmentation as a next step. daily.dev surfaces practical writeups like this for engineers benchmarking agent-driven RL fine-tuning results."}},{"@type":"Question","name":"Why would GRPO be replaced with SFT for training a NeMo Gym RL environment?","acceptedAnswer":{"@type":"Answer","text":"A prior autoresearch session found that GRPO did not improve accuracy on a star-counting visual task, while a switch to SFT (via LoRA) proved effective, eventually reaching 96.875% exact accuracy. This choice was made by an autonomous coding agent after profiling results rather than assumed upfront, illustrating how algorithm selection can be data-driven mid-campaign. Compare training algorithm trade-offs for your own RL tasks by following coverage like this on daily.dev."}},{"@type":"Question","name":"What is OAPL and how does it compare to DAPO for off-policy reinforcement learning?","acceptedAnswer":{"@type":"Answer","text":"OAPL (Optimal Advantage-based Policy Optimization with Lagged Inference policy) is a long-lag off-policy RL method that treats the training-inference policy mismatch as part of the training objective rather than eliminating it. On a Qwen3-1.7B model with the DAPO-math-17K dataset, OAPL reached higher accuracy with fewer training steps than the GRPO-variant DAPO, though it used noticeably more tokens per response. daily.dev helps researchers track emerging off-policy RL techniques like OAPL as they compare training approaches."}}]} ```