AI research agents explored training methods on safety-related issues of the 10 type and improved model performance on target benchmarks.
Find out the training methods for each problem.
Anthropic instructed AI research agent Claude to deal with all types of alignment failures.Claude searched for literature, proposed techniques and data for training, trained models and repeated the process to test the results.
The improvement scale was measured by using the benchmarks of 3 to 5 for each problem, and by how much the difference to the theoretical fulfillment was reduced.
Check the restrictions before using the method.
Conditions were set for the exploration.Anthropic excluded methods that undermined the general capabilities of the target model and methods that destroyed Claude's own alignment directly to the target model.
In response to follow-up responses, Claude used activation steering, which linked the proposed technique to the generation of training data by adjusting the internal expression of the model during production and creating non-follow-up responses as training examples.
10 type assessment confirmed improvement
In Anthropic measurements, Claude found a method to improve the benchmark performance for all types of 10 alignment failures without reducing the overall capacity of the measured subject.
In an experiment focused on fraudulent behavior, Claude tested more than one 150 technique, which reduced the 82 percentage of the difference to the full point.
Apply to unrepresented assessments and large models.
The methods selected from the target benchmark were also effective in the alignment assessment that Claude did not see during the investigation. The evaluation tool Petri, which simulates hostile multi-turn conversations, also proved to be effective.
The method was also effective in models with the size of 4.7, which was twice as large as the size of the objects under investigation.
Explore early large-scale models in training methods.
Anthropic further improved Claude Sonnet 5’s initial checkpoints for the Claude Opus 4.8, which has not yet received most of the alignment training for the main series.
In this experiment.60by time.50More than one solution was tested and initial checkpoints reached alignment values closer to the original model. The training examples used for the chosen technique were about templates or public datasets.2000It was a matter.