Goodfire published mechanistic shaping on November 4: instead of rewarding models only for getting the answer right, training objectives now state directly which internal concepts a model may rely on and which it must not express — the goal being a model that gets the answer right, and for the right reasons.
Two moves: don't rely on it, don't express it
Mechanistic shaping comes in two forms. Reliance shaping penalizes dependence: it tells the model, don't base your judgment on this concept. Expression shaping penalizes expression: it tells the model, don't act this concept out. The recipe has three steps: first use interpretability tools to find the direction of an unwanted concept inside the model, then add a penalty term to the loss function, and recompute the concept direction periodically during training, because representations drift as training proceeds. Goodfire tested the approach in three settings. In a chest X-ray experiment, the training data was deliberately set up so that only collapsed-lung cases with a chest drain were included, and the baseline model predictably learned to look at the tube. After an LLM agent used a sparse autoencoder to find the tube concept automatically and reliance shaping was applied, the model could diagnose cases without a drain too — scoring even above the soft ceiling of training on balanced data.
Sycophancy and reward hacking are beyond data filtering
The language-model experiments are more pointed. Goodfire found that direct preference optimization (DPO) makes models noticeably more sycophantic, even when visibly sycophantic examples make up less than 1% of the training data — filtering the data barely helps. After finding the sycophancy direction and applying reliance shaping, models at 7B, 13B, and 32B all kept DPO's capability gains while sycophancy fell back close to the starting level. Another experiment used expression shaping to train Qwen3.5-9B on Countdown arithmetic puzzles, specifically suppressing behavior that games the grading rules. The paper also carries a warning: intervening on internal representations can break monitoring probes, as concepts may regrow, re-encode, relocate, or be erased. The encouraging part is that in the experiments, concepts remained monitorable after shaping — though probes often had to be retrained.
This is still a research method, not a product. Its significance is that it fills a gap data cleaning cannot reach: some bad behaviors are not imported with the data but grown during training, so they should be fixed in the training objective itself. Goodfire's earlier Project Beacon with Baseten took a different route, putting safety signals into the inference of open models; mechanistic shaping moves the intervention point earlier, into training.
