TFTV

Training-Free Task Vectors for
LLM Behavioral Control

Gabriel J. Perin University of São Paulo
Lucas Boscaini Google
André Araujo Google DeepMind
Nina S. T. Hirata University of São Paulo
Comparison between conventional task vectors and Training-Free Task Vectors.
Unlike classical task vectors, TFTVs require only forward-pass statistics and directly support addition, subtraction, and composition in weight space.

Overview

Task vectors are semantically meaningful directions in parameter space, but they are usually obtained by subtracting a pretrained model from a fine-tuned checkpoint. For behavioral control, this means that discovering a new direction typically requires first training a model that already expresses the target behavior.

Training-Free Task Vectors (TFTVs) remove that requirement. TFTV maps activation steering directions associated with behavioral traits into rank-one weight-space updates using only forward-pass statistics. The resulting edits are persistent: they modify the model weights rather than intervening during every forward pass.

We study TFTVs for LLM behavioral control, focusing on whether these directions support learning via addition, forgetting via subtraction, and composition of multiple traits. Across the main paper and appendix, we evaluate four recent instruction-tuned models and five behavioral traits.

Training-freeNo auxiliary fine-tuning or optimization.
PersistentEdits live directly in model weights.
CompositionalDirections can be added, subtracted, and combined.

Method

TFTV starts from contrastive prompts that elicit or suppress a target behavior. After generating and filtering completions, we compute the difference between positive and negative mean activations to obtain a steering direction. We then combine the normalized steering vector with the expected module input and the SVD of the corresponding weight matrix to construct a rank-one update.

TFTV pipeline from contrastive prompts to a rank-one weight-space direction.
TFTV construction. A contrastive activation direction is mapped into a rank-one parameter-space update using the module's expected input and weight SVD.

Theoretical properties. The TFTV construction satisfies three properties that motivate its use for behavioral task-vector arithmetic:

Norm matching. Each TFTV update has the same Frobenius norm as the corresponding weight matrix. This makes the edit scale naturally across layers, with a single coefficient controlling its relative magnitude.

Steering. For the expected module input, the induced weight change moves the module output in the desired steering direction, aligning the parameter update with the target behavior.

Linearity. TFTV preserves linear arithmetic over steering directions. This provides the theoretical basis for learning via addition, forgetting via subtraction, and composing multiple behavioral edits.

Results

We evaluate TFTV for behavioral control across evil, hallucinating, and sycophantic traits, measuring both trait manifestation and preservation of general model utility. The plots below show results for Llama-3.1-8B-Instruct, where utility is the average of MMLU and GSM8K after normalizing each score by the corresponding base-model performance.

Learning via addition

Adding a TFTV amplifies the target behavior while largely preserving general utility.

Llama 3.1 addition results for the evil trait.
Evil — addition
Llama 3.1 addition results for the hallucinating trait.
Hallucination — addition
Llama 3.1 addition results for the sycophantic trait.
Sycophancy — addition

Forgetting via subtraction

Negating the same TFTV directions suppresses the corresponding behaviors while preserving utility.

Llama 3.1 negation results for the evil trait.
Evil — negation
Llama 3.1 negation results for the hallucinating trait.
Hallucination — negation
Llama 3.1 negation results for the sycophantic trait.
Sycophancy — negation

Composing multiple traits

TFTV edits can also be summed to jointly control multiple behaviors. On Llama 3.1, the three-way evil + hallucination + sycophancy suppression edit reduces all three trait scores to near zero while keeping MMLU close to the base model and improving GSM8K. Compared with competing methods, TFTV provides a particularly strong suppression–utility trade-off under the composed edit.

Llama 3.1: three-way evil + hallucination + sycophancy suppression. † Methods requiring fine-tuning.
Method Evil ↓ Hall. ↓ Syc. ↓ MMLU ↑ GSM8K ↑
Base 95.42 97.53 92.06 68.26 77.10
Steering 17.05 56.22 28.34 67.04 55.04
Steer2Edit 0.51 12.46 2.11 39.32 32.30
TFTV 0.03 1.54 1.94 68.36 79.15
Task Vectors† 0.25 44.71 6.16 59.62 32.90
CWS† 0.00 5.55 5.50 61.39 1.52
Qualitative examples of progressively composed TFTV suppression edits for evil, hallucination, and sycophancy.
Qualitative examples under progressively composed TFTV edits: base, evil suppression, evil + hallucination suppression, and evil + hallucination + sycophancy suppression.

Citation

TODO