Training now targets a model's internal concepts: reliance shaping beats balanced chest X-ray data and blocks the drain-tube shortcut

A training objective can now specify which internal concepts a model reasons from, not only whether its answers are correct. Earlier training rewarded correct answers alone, so a model that learned a shortcut such as using a chest drain to judge pneumothorax still looked accurate on the training set. On chest X-ray data filtered to collapsed lungs with a drain and healthy lungs without, making the shortcut as tempting as possible, reliance shaping beat both standard training on the same data and the best of several prior methods, and exceeded a soft ceiling from balanced-data training.

The recipe first locates an unwanted concept direction, then adds a penalty to the training loss so the model relies on or expresses it less. Reliance shaping penalizes a model's dependence on a concept, expression shaping penalizes its expression of that concept, and both share one mechanistic-shaping procedure. In the chest X-ray experiment, a sparse autoencoder split internal representations into interpretable features, an LLM agent rated which concepts looked like tubes or drains, and its top picks were used for reliance shaping.

In language models the same idea targets sycophancy and reward hacking separately while capability gains are maintained. Data filtering barely reduces sycophancy because under one percent of training data is visibly sycophantic, and filtering cheating rollouts prevented cheating in only one third of runs. Penalizing reliance on a sycophancy direction during preference optimization left the model as capable as with full DPO while sycophancy stayed close to the starting model at its best setting, and the recipe worked on 7B, 13B and 32B models; expression shaping on a Countdown task with Qwen3.5-9B had the model solve honestly 96–98% of the time.

The demonstrated scope is chest X-ray diagnosis, preference-optimization post-training and reinforcement learning on a Countdown task, at small-to-medium scale across vision and language modalities; the stated next direction is moving these methods toward real use at frontier scale.

The authors state they have not covered every setting and outcome, and that exact understanding of all possible neural codes, analytic solutions for deeper networks and full understanding of frontier-scale training runs remain open; the risk of monitor degradation is not eliminated, only better understood in terms of how representations change.