November 2026 Summaries

AI21’s Forward-Deployed Researchers and Deel’s Emet AI lab post-trained NVIDIA Nemotron 3.5 Super VL for global payroll support, a domain requiring accurate, complete, locally compliant answers across changing laws, contracts, and company policies. Using Deel’s EmBench Payroll benchmark, which evaluates factual, coverage, and stylistic criteria, the team found that the base model often possessed relevant knowledge but produced incomplete or overly cautious answers. They first applied near-policy supervised fine-tuning (SFT) to 412 high-quality, hint-guided outputs generated by the model itself, then used self-distillation policy optimization (SDPO) with privileged guidance to reinforce desired answer characteristics through fresh on-policy rollouts. The combined approach increased the model’s criteria-met score from 51.3% to 72.0% and its fully correct-answer rate from 10.0% to 33.9%, surpassing Claude Opus 5.5 by roughly eight points on the reported quality measure while costing about 50 times less per million tokens.

AI21 Labs and Alice post-trained NVIDIA’s Nemotron 3.5 Super VL over three weeks to improve enterprise-agent resistance to indirect prompt injection and other malicious requests while preserving task completion. Using Alice’s simulated Slack and GitHub reinforcement-learning environments, the team first applied supervised fine-tuning with successful trajectories rewritten in Nemotron’s own style to reduce distribution drift, repetitive reasoning, and overfitting from a relatively small dataset. It then used a modified GRPO online reinforcement-learning approach that removed standard-deviation reward normalization and added length and repetition penalties to better handle dense rewards and control excessively long outputs. Validation scores rose from 35.19 for the base model to 76.2 after training, while out-of-distribution performance on Alice’s CHASM security benchmark increased from 7.54 to 33.96. The authors report a 41-point resilience improvement over the base model, a 10-point advantage over GPT-6.1 Sol, and estimated inference costs 25 times lower, presenting the work as evidence that specialized open models can be trained to combine security-focused behavior with effective agent performance.