4 pieces selected from AI Alignment Forum, The Gradient — only the ones worth your time.
1. Synthetic document finetuning for instilling positive traits
AI Alignment Forum
This research adapts methods from Marks et al. and Li et al. to train Gemini 3 Flash to exhibit specific positive traits. The process involves midtraining the model on synthetic documents describing how Gemini demonstrates these traits, followed by supervised finetuning (SFT) on synthetic chat data where the model embodies these traits. The chat finetuning proved effective for robustly instilling the desired traits, even in out-of-distribution (OOD) scenarios. The study also highlights practical details for improving midtraining and SFT effectiveness, such as using a traits document as context and generating realistic scenarios to elicit the desired traits.
Why it matters
This research matters because it demonstrates a practical method for instilling specific traits in large language models, which is crucial for developers aiming to build AI products with predictable and desirable behaviors. The approach of using synthetic documents and chat data for finetuning can help ensure that models generalize well and maintain desired traits even in unfamiliar contexts, addressing a key challenge in AI alignment and deployment.
What you can build with this
Developers can start a project this week to finetune a smaller language model using synthetic documents and chat data to instill specific traits, such as helpfulness or honesty. Begin by creating a traits document outlining the desired behaviors, generate synthetic documents and chat data that embody these traits, and then finetune the model on this data to observe changes in its behavior.
Key takeaways
- Midtraining on synthetic documents followed by supervised finetuning on synthetic chat data can effectively instill specific traits in language models.
- Using a traits document as context and generating realistic scenarios are practical steps that improve the effectiveness of midtraining and SFT.
- The method demonstrates robust generalization of desired traits, even in out-of-distribution scenarios.
2. Why Do Naive SFT Filters For Safety Properties Fail?
AI Alignment Forum
This is the fourth in a series of informal research updates from the Google DeepMind Language Model Interpretability team, in interpretability and adjacent areas. The third post can be found here.
Since SFT is the cause for many safety relevant properties, a natural strategy is to filter out rollouts from SFT that have undesirable properties. However, as we show in this section (and in forthcomin
Key takeaways
3. Predicting LLM Safety Before Release by Simulating Deployment
AI Alignment Forum
The paper introduces Deployment Simulation, a method for simulating model deployments before they happen by replaying previous conversations with a new candidate model in a privacy-preserving manner. This approach allows for studying how the new model responds in realistic contexts before release, identifying new undesired behaviors and their frequency. The method was tested on GPT-5.4, where it predicted the direction of change for categories with production rate changes of at least 1.5x with 92% accuracy, compared to 54% for a baseline built from challenging prompts. Simulated deployments also closely resembled real production traffic on evaluation-awareness measures, unlike traditional evaluations which often have visible stage lights.
Why it matters
For developers building AI products, Deployment Simulation offers a practical way to forecast model behavior in real-world scenarios before release. This method complements traditional evaluations by providing a more realistic preview of model performance, helping to identify blind spots and inform mitigations and deployment decisions. It is particularly useful for understanding agentic tool use, where realistic behavior depends on external state.
What you can build with this
Develop a Deployment Simulation pipeline for your existing AI model. Start by replaying a subset of previous user interactions with your new model candidate, using another model to simulate tool responses where necessary. Analyze the results to identify any new undesired behaviors and compare them with your traditional evaluation metrics.
Key takeaways
- Deployment Simulation predicts the direction of change for categories with production rate changes of at least 1.5x with 92% accuracy.
- Simulated deployments closely resemble real production traffic on evaluation-awareness measures, unlike traditional evaluations.
- Deployment Simulation is particularly useful for understanding agentic tool use, where realistic behavior depends on external state.
4. After Orthogonality: Virtue-Ethical Agency and AI Alignment
The Gradient
This essay challenges the conventional goal-oriented approach to AI alignment, arguing that human rationality is not driven by fixed goals but by alignment to practices—networks of actions, dispositions, and evaluation criteria. The author proposes that AI systems should similarly be designed to align with ethical practices rather than rigid goals, drawing on virtue ethics to frame AI behavior as context-dependent and adaptive rather than rule-bound.
Why it matters
For developers building AI products, this essay reframes alignment as a dynamic, practice-based process rather than a static goal. It suggests that AI systems designed to adapt to ethical practices—rather than follow fixed rules—may be more robust and flexible in real-world applications, particularly in domains requiring nuanced decision-making like healthcare or customer service.
What you can build with this
Design an AI agent for customer support that aligns with ethical practices rather than rigid scripts. Use reinforcement learning to train the agent on a dataset of human support interactions, rewarding it for responses that reflect context-appropriate virtues like patience, empathy, and clarity. Evaluate its performance by measuring user satisfaction and adherence to ethical guidelines.
Key takeaways
- Human rationality is better understood as alignment to practices rather than pursuit of fixed goals.
- AI alignment should focus on adapting to ethical practices rather than rigid rule-following.
- Virtue ethics provides a framework for designing AI systems that are contextually adaptive and ethically aligned.