CAFT (Concept Ablation Fine-Tuning)
Persona Vectors: Monitoring and Controlling Character Traits in Language Models — inherited
A prior training-time intervention (Casademunt et al., 2025) that zero-ablates a concept's activation-space projection during finetuning to prevent the model from learning unwanted correlations along that direction. Compared against preventative steering using the paper's own persona directions, CAFT is effective for evil and sycophancy (whose base-model projections are already negative, so zero-ablation acts like positive preventative steering) but ineffective for hallucination, and requires no tunable coefficient, unlike preventative steering.
Preventing a model from learning an unwanted correlation during training doesn't have to mean actively pushing its activations somewhere — an earlier method took the opposite approach, simply erasing a chosen direction from the model's activations at every step of training instead. That erasure-based technique is borrowed here and re-tested as a direct comparison point for the paper's own approach.
The erasure mechanism, and why it needs no tunable strength setting, comes first. What follows is a result that's trait-dependent rather than uniform: the erasure-based method matches the paper's own steering approach on some traits but falls short on others, for reasons traceable to where a model's untouched activations already sit before any training begins.
What CAFT is
Concept Ablation Fine-Tuning (CAFT) is a prior training-time intervention, introduced by Casademunt et al. (2025), for preventing a model from learning an unwanted correlation along some direction during finetuning. Rather than steering activations in a chosen direction, CAFT zero-ablates a concept's projection at every forward pass throughout training: it "uses sparse autoencoder latents and PCA directions to zero-ablate specific concepts during finetuning, preventing models from learning unwanted correlations, and thereby controlling generalization behavior" (persona-vectors, §"7 Related work", p. 12). Its own paper demonstrates the technique on gender bias, where the goal is indifference to an axis of variation regardless of which pole a given example falls on, not movement away from one specific pole — and it needs no tunable coefficient, since ablation to zero has only one setting.
Re-implemented as a baseline against preventative steering
The persona-vectors paper directly re-implements CAFT to compare it against its own Preventative steering method, applying both ablation (CAFT) and directional steering across all layers during finetuning, using the paper's own extracted persona directions for Evil (persona trait), Sycophancy, and Hallucination (closed-domain fabrication) (persona-vectors, §"J.4 Comparison with CAFT", p. 47). The result is trait-dependent: "when applying CAFT using our extracted persona directions, we observe that it works well for evil and sycophancy, where the base model exhibits strong negative projection values. However, for hallucination, where the base model's projection is already near zero, we find CAFT to be far less effective" (persona-vectors, §"J.4 Comparison with CAFT", p. 47). Preventative steering, by contrast, is reported to be effective across all three traits.
Where each tool is the right one
The paper's own read on the comparison is that CAFT's apparent success on evil and sycophancy is closer to an artifact of where the base model already sits than to a genuine advantage: "CAFT's effectiveness in the evil and sycophancy cases stems from the fact that forcing activations to have zero projection effectively acts as positive preventative steering" given those directions' negative base-model projections (persona-vectors, §"J.4 Comparison with CAFT", p. 47). CAFT's real advantage lies elsewhere — cases like gender bias, where the intervention should be indifferent to an axis rather than push away from one pole of it, a use case where directional preventative steering does not apply. The tradeoff runs the other way on tuning: "CAFT operates without any tunable parameters," while preventative steering "requires selecting an appropriate steering coefficient" (persona-vectors, §"J.4 Comparison with CAFT", p. 47).