New AI Safety Training Boosts Model Alignment Results
Collaborations with Alice improve model resilience and reduce unnecessary refusals for key open-source architectures.
Updated on Oct. 15, 2026 in Artificial Intelligence
Live Poll
Do you believe artificial intelligence can be made safe while remaining powerful and useful?
Alice has released new safety results from collaborations with Inflection AI and AI21 aimed at improving model alignment and resilience. The effort involved training NVIDIA Nemotron models with adversarial datasets to mitigate risks while preserving core performance.
Why it matters
As enterprises look to deploy increasingly powerful open models, they must move beyond surface-level safety filters to bake behavioral policies directly into the underlying architecture. This approach aims to reduce unnecessary model refusals while hardening systems against complex adversarial threats.
Inflection AI achieved a 30% increase in Pi model alignment scores while holding unnecessary refusal rates at 1.6%. Simultaneously, AI21 saw its NVIDIA Nemotron 3.5 Super VL model gain 41 points in resilience against indirect prompt injection.
The players
Alice
A safety-focused firm that provides adversarial datasets and testing services to 8 of the 10 leading foundation model labs.
Inflection AI
A developer of AI models and the Pi conversational assistant that focuses on large-scale model fine-tuning.
AI21
A research-led AI company that has secured $336 million in funding to develop and post-train large-scale open models.
The details
Alice uses more than 20 realistic personas for red-team exercises to simulate diverse adversarial scenarios. AI21 leveraged reinforcement-learning environments—a technique where models learn through trial and error to maximize rewards—to harden the NVIDIA Nemotron 3.5 Super VL model against indirect prompt injection. These adversarial datasets are used during the post-training phase to refine how the model processes user requests while maintaining its foundational performance.
Timeline
October 15, 2026: Alice announced the results of its collaborations with Inflection AI and AI21.
2017: AI21 began training language models.
2021: AI21 expanded its focus on language model development.
The Tech Race
These results follow a trend of companies shifting from generic safety layers to deep architectural integration within the NVIDIA Nemotron model family. This strategy seeks to close the gap between frontier-grade performance and the safety requirements of enterprise-ready AI.
Developers and companies currently building on the NVIDIA Nemotron architecture can expect these safety refinements to reach production environments soon. These improvements will first benefit users of enterprise-grade AI assistants that require high reliability and resistance to prompt manipulation.
The takeaway
The move toward embedding adversarial defenses directly into model training suggests a maturing approach to AI safety. Developers should track the forthcoming performance benchmarks for the NVIDIA Nemotron 3.5 Super VL to see if these safety gains persist at scale.
Further reading
For broader coverage of recent progress in model safety, visit Artificial Intelligence.
Live Poll
Do you believe artificial intelligence can be made safe while remaining powerful and useful?
