Urgent.News

What's breaking now, across hundreds of outlets.

AI

Nvidia used 20,000 training examples to fix its AI’s mistakes. Its coding score jumped.

Nvidia used Factory’s Droid coding-agent harness to generate roughly 20,000 rows of post-training data for Nemotron 3 Ultra, the 550-billion-parameter The post Nvidia used 20,000 training examples to fix its AI’s mistakes. Its coding score jumped. appeared first on The New Stack .

Nvidia used 20,000 training examples to fix its AI’s mistakes. Its coding score jumped.

Nvidia utilized a network of 20,000 training instances to rectify errors in its AI following post-training. This enhancement resulted in a significant boost to the coding accuracy of Nemotron 3 Ultra, an expansive 550-billion-parameter open model unveiled in June. In a detailed case study, Nvidia's Factory revealed how they deployed its Droid agent within an isolated network and monitored the harness to capture and analyze recurring errors in task execution, tool invocation, and error rectification.

Droid leveraged Nemotron to maintain a consistent task plan throughout its operations. The examination uncovered issues with tool calls and argument structuring, as well as situations where the model persisted in functioning despite encountering errors. Factory also assessed Nemotron using bespoke datasets, such as Legacy Bench, to pinpoint specific choices contributing to unsuccessful task completion.

Jian Zhang, Nvidia's director of applied research for Nemotron post-training, confirmed that Factory worked alongside them to refine the model's strategic planning, tool utilization, and error recovery mechanisms. In parallel, Nvidia adopted PivotRL, a reinforcement learning technique introduced in March, which trains models on pivotal choices within existing supervised fine-tuning trajectories, eliminating the need for extensive multi-turn rollouts.

Additionally, PivotRL demonstrated performance on par with end-to-end reinforcement learning while requiring merely one-quarter of the rollout turns on coding tasks. Compared to standard supervised fine-tuning on comparable datasets, PivotRL achieved a 4.17% boost in accuracy across four agentic domains and a 10.04% improvement on out-of-domain non-agentic tasks.

Nvidia had previously integrated PivotRL into production-scale post-training for Nemotron 3 Super. For Ultra, the company provided Factory with revised checkpoints for additional evaluation. To mitigate harness-specific overfitting, Nvidia varied Droid's configuration and integrated data from diverse agent environments. Nvidia's NeMo RL documentation outlines a multi-stage pipeline combining reinforcement learning with verifiable rewards, human feedback, and multi-teacher on-policy distillation.

The training configurations encompass OpenCode, Codex, and an environment labeled as agent006, alongside Droid-specific PivotRL data. The distillation process involves transferring capabilities from teacher models trained across various technical domains into the student model. Nvidia has released its reinforcement learning blends along with guidelines for replicating the pipeline on GB200 NVL72 hardware.

The Terminal-Bench 2.0 score of Nemotron 3 Ultra rose from 38.5% to 49%. Preliminary checkpoints of Nemotron 3.5 Super VL improved from 52.11% to 60.4% on Terminal-Bench 2.1. The extent of Factory's contribution to these benchmark improvements remains undisclosed, and no ablation study has been conducted to isolate its impact on either benchmark gain or individual tool-use error rates.

Hugging Face's Harness Hub team adopted an alternative approach by employing reinforcement learning to train distinct LoRA adapters for Pi, OpenCode, and Hermes without altering the harnesses or base model weights. A version of Qwen 3.8 27B, optimized for Pi, experienced a Terminal-Bench 2.1 score increase from 39.5% to 55.8% and a 55% reduction in tool calls per task.

The team also devised safeagents, a rule-based proxy that scrutinizes tool calls to thwart reward hacking during training. Neither Nvidia nor Hugging Face has published comparable assessments to quantify the performance and computational trade-offs associated with harness-specific adaptation versus multi-harness post-training.

Rewritten by urgent.news; may contain errors — check the original.

Read the original at thenewstack.io →

Report a problem with this story

More in AI

Spectral AI: 3 Key Takeaways for Long-Term Investors

Recurring software leases, not device sales, are what could turn Spectral AI's research base into lasting margins, and the article leaves that transition unproven.

  • Spectral AI manufactures first 30 DeepView AI units for burn center deployments
  • Applied for New Technology Add-on Payment to increase Medicare reimbursement
  • Commercial agreements include minimum 3-year terms with recurring software subscriptions

Fact-check AI energy claims before they steer your architecture

You will hear AI called an environmental disaster. Many of the numbers behind that claim do not survive a calculator, a units conversion, or a question about what you are actually comparing.

  • Andy Masley found a figure in "Empire of AI" off by a factor of approximately 4,500.
  • Developers should focus on infrastructure constraints, not individual AI prompt impacts.

More from Saturday 21 November →