Heads up! Large Language Models Can Perform Tasks Without Your Instruction via Selective Attention Head Masking

Many thanks for your appreciation and valuable remarks on our work. We would like to answer and clarify your concerns as below.

W1. Theoretical analysis

Since our conclusions are primarily derived from experiments and observations rather than strict theoretical analysis, we sincerely apologize for the shortcomings in this regard. Nevertheless, some studies [1][2] have provided insights that support the validity of our approach, in which we regard attention heads as computational nodes in the information flow and analyze their functionality.

[1] Yao, Y., Zhang, N., Xi, Z., Wang, M., Xu, Z., Deng, S., & Chen, H. (2024). Knowledge circuits in pretrained transformers. arXiv preprint arXiv:2405.17969.
[2] Hao, Y., Dong, L., Wei, F., & Xu, K. (2021, May). Self-attention attribution: Interpreting information interactions inside transformer. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 35, No. 14, pp. 12963-12971).

W2. Larger scale experiments

In Sec. 3.4, we stated that "Larger model shows better head functionality," primarily based on the criterion of whether behavior switching can be achieved. This is because, in models smaller than 3B, masking out certain attention heads does not reliably trigger behavior switching. To further support this claim, we conducted additional experiments on Qwen2.5-32B-Instruct and found that it also exhibits behavior switching in translation tasks (see Q1).

Q1. Large model & MoE model

We are very sorry, but we currently do not have the resources to conduct experiments on a model as large as 72B. However, we have performed training on a 32B model (Qwen2.5-32B-Instruct). Due to time constraints, we only trained head masks for en to other six languages translation task. Generally, we found that applying masks can still improve model performance on the 32B model.

• PPL

• ROUGE-L

For MoE models, we selected Phi-3.5-MoE-instruct (16×3.8B, activating 6.6B) for head mask training on the en-zh translation task. We observed that while head masking in the MoE model could trigger behavior switch, it did not lead to performance improvements. This might be because MoE model parameters are more inclined toward memory. We propose another two possible explanations: (1) The primary functionality of MoE might reside in the FFN, with the router playing a role similar to attention head masking for function selection. (2) The task-level mask and the token-level router may not work well together. Improving the effectiveness of head masks in MoE models is an important direction for our future research.

• PPL

• ROUGE-L

Q2. Stimulate reasoning ability

After the release of DeepSeek-R1, we attempted to use R1 to generate high-quality reasoning chains for simple tasks and then trained head masks on these chains in an effort to activate reasoning capabilities in other instruct models. Unfortunately, our results were not as promising as we had hoped. Whether we applied the mask directly or after explicitly prompting the model to "think step by step," the model was still unable to achieve high-quality reasoning solely by selecting specific attention heads. (It may still require parameter modification to stimulate its potential reasoning capabilities, such as LIMO[3])

For the masks we obtained, they tended to discard only a small number of attention heads (less than 10% in a 7B model), and for complex tasks like mathematics, almost all attention heads were retained. This leads us to believe that reasoning is a complex process involving the coordination of multiple functions at different stages of generation. As a result, it may not be feasible to activate reasoning abilities solely through head masking at this level of "task granularity".

[3] Ye, Y., Huang, Z., Xiao, Y., Chern, E., Xia, S., & Liu, P. (2025). LIMO: Less is More for Reasoning. arXiv preprint arXiv:2502.03387.

Thank you for your thoughtful comments and suggestions on our work. We would like to answer and clarify your concerns as below. As the text length is constricted, we provide the results in the anonymous github (rebuttal.pdf).

First, we would like to clarify that the primary focus of this paper is on the interpretability and functionality of LLMs - specifically, the phenomenon that "tasks can be directly executed by selecting attention heads". We find this phenomenon very interesting, and most of the paper is devoted to validating its generality and attempting to explain its underlying mechanisms. (We have even observed the same phenomenon even in multimodal speech models! See our response to b411) Although using masks can indeed enhance model performance in some cases, we sincerely hope that you will pay more attention to our contribution regarding the interpretability of model functionality.

W1. Training Overhead

It is indeed unavoidable to train a mask for each task on every model to identify its functional heads. However, because our training only involves trainable parameters equal to the number of attention heads, the process is both fast and requires little memory (only 44 minutes and 16.5GB of VRAM on llama3.1-8b). We have also compared this approach with full fine-tuning and LoRA fine-tuning in terms of resource consumption. (Note that since fine-tuning modifies the model parameters, these two methods would undoubtedly achieve better performance. Nevertheless, we are not proposing a PEFT method, and we have neither modified nor added to the model’s original parameters.)

W2. Lack of Experiments

Initially, we chose to test on 100 samples because we wanted to use more training samples to locate the attention heads required for the translation task more accurately, and the evaluation set was just used to verify whether the model can perform translation. But, indeed, it's our bad to use such a small number of samples for the quantitative evaluation. As it is on the XNLI dataset, we nonetheless evaluated on the full IWSLT2017 evaluation set in the paper (see Table 7 and App. A) without incorporating any additional training data.

We would also like to argue that with a smaller train-test ratio, we can still achieve results similar to those reported in the paper. Please refer to the experimental results obtained after re-dividing the dataset with an 8:2 ratio.

For other benchmarks: although most benchmark tasks include explicit "instructions" such as options, we have still managed to conduct evaluations on some datasets. We achieved better performance on winogrande and hellaswag.

W3. Efficiency Analysis

Taking the llama3.1-8b model and the en-zh translation task as an example, we masked out 192 (18.75%) attention heads, and pruned the corresponding Q-head and $W_O$ matrices. Since the attention heads account for only 16.7% of the model’s total parameters, and the pruning only affects Q and O, we actually only saved 2.5% of the parameters. As some studies focusing on pruning can remove 20-30% of the parameters with only minimal performance loss, we would like to clarify again that our work is not aimed at achieving inference efficiency, but rather at locating and interpreting the model’s modularity.

Q1. Head Pattern

In App. C.2 and Figure 7 we have some discussions on the head patterns, and you may check it out. We also have some other findings: 1. Similar tasks tend to share more attention heads， for example, 42 translation tasks share 177 common attention heads, but 35 simple tasks share only 5 heads. 2. Due to sparsity, as the model size increases (from Qwen2.5-1.5B to 14B), each attention head tends to be responsible for fewer tasks on average.

Q2. Performance on few-shot learning

We conducted some supplementary experiments on simple tasks; please take a look at the results. Summary:

1. When using the mask obtained from 0-shot training for 5-shot inference, the performance ranking is: instructed 5-shot > mask 0-shot > mask 5-shot > 5-shot.
2. For the tasks in Table 2, using a mask trained with 5-shot and then performing mask 5-shot inference yields better performance than instructed 5-shot.
3. We combined multiple simple task datasets into a hybrid dataset, where each sample may come from different tasks but includes 5-shot examples from its corresponding task. After training the mask on this hybrid dataset, the model was able to enhance its 5-shot context learning performance (including tasks that were not seen during training).

For our more response please refer to the anobymous github. Sorry for the inconvience!

Thank your for your revisions and comments. We would provide our clarifications and responses below.

First and foremost, we would like to clarify certain statements in your Summary, Methods and Experimental Designs sections. Our experiments are not limited to language translation tasks; they also include 35 simple tasks (Lines 166-169, left column; Table 2, 4, 6, 9, etc.), which account for nearly half of all our experiments. However, this aspect was not mentioned in the review. Additionally, as a paper focused on the functionality and interpretability of LLMs, we have dedicated substantial portions of the text to exploring and analyzing the mechanism that how model "switches behavior through selective attention heads". While it is true that using head masking can improve performance in certain scenarios, we would greatly appreciate it if you could focus more on the paper’s contributions to model functionality and interpretability.

If there are any aspects of the paper that remain unclear, we are always here to provide further clarification and assistance.

W1. Benefits

As stated above, our paper primarily contributes to the interpretability of models. We find the phenomenon of "enabling the model to perform tasks without instructions by selecting attention heads" to be very intriguing. We propose an efficient method, applicable to a wide range of tasks, for identifying which attention heads are crucial for the task. This phenomenon is not only observable across various architectures and sizes of LLMs (Section 3), but also applies to multimodal models, including those for speech (see below). These attention heads form interdependent "functional pathways" responsible for executing corresponding functions, thus helping us better understand how model modularity manifests in attention heads (Section 4). Additionally, by selecting these attention heads, the model often achieves a certain level of performance improvement (Section 5).

• Speech model: test on SALMONN

W2/Q1. Degrades

After using the mask, the model undergoes a behavioral change and can no longer perform the original task as intended. As a result, the general performance will inevitably decrease. However, since the mask does not modify or add/remove any of the model’s parameters, we can simply stop using the mask to restore the model's original instruction capabilities. This process is straightforward and flexible.

We also conducted some mask experiments on general benchmarks. Please see our response to uwMV.

Q2. activation experiments

In the paper, we conducted an experiment (Line 370, right column) to investigate the relationship between attention head activation (focus on special tokens) and the mask. Our findings suggest that the correlation between them is not strong. We discuss this phenomenon further in Appendix C.3.2, where we suggest that the "attentive function" of the attention heads does not play a major role in their selection.