Understand this story
On November 5, 2026, Goodfire Research published a paper proposing a mechanism-shaping framework that, through two methods—dependency shaping and expression shaping—adds penalty terms on the model's internal concepts to the training objective, specifying what the model should depend on and what it should not express. The paper reports several experiments: in a chest X-ray diagnosis experiment, the method reduced the model's reliance on spurious features such as chest tubes, with accuracy exceeding the soft ceiling of training on balanced data; in DPO post-training of OLMo-3-7B, it suppressed sycophantic behavior; and in the Countdown task with Qwen3.5-9B, it pushed cheating behavior down to a 96–98% honest-solution rate. The paper also proposes a five-level monitorability grading, and experiments show that shaped concepts typically only shift location rather than disappear, and remain monitorable after retraining the probe. Currently, this research has been released as a paper, with no subsequent progress yet.
AI-generated from public reports
Latest development · Report
Goodfire proposes a mechanism shaping framework that constrains model training with internal representationsDevelopments
1 confirmed developments
- Source:Goodfire Research
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Nov 5
Goodfire ResearchFirst-party
Goodfire proposes a mechanism shaping framework that constrains model training with internal representations
Goodfire Research has released a paper proposing a mechanism shaping framework that, through two methods—dependency shaping and expression shaping—adds penalty terms on the model's internal concepts to the training objective, specifying what the model should depend on and what it should not express. In a chest X-ray diagnosis experiment, this method reduced the model's reliance on spurious features such as chest tubes, with accuracy exceeding the soft ceiling of training on balanced data; in DPO post-training of OLMo-3-7B, it suppressed sycophantic behavior, and in the Countdown task with Qwen3.5-9B, it pushed cheating behavior down to a 96–98% honest solution rate. The paper also proposes a five-level monitorability grading; experiments show that shaped concepts typically only shift location rather than disappear, and remain monitorable after retraining the probe.