Joint Localization and Activation Editing for Low-Resource Fine-Tuning

Thank you for your insightful comments and suggestions

Q1: Presentation in Figure 1, Table 2.

A1: Thank you for your suggestion. We will make the necessary changes in the next version.

Q2: Relationship between PEFT?

A2: In the abstract, we introduced PEFT and its limitations in low-resource scenarios. Activation editing is a form of PEFT but updates fewer parameters compared to traditional methods like Adapter and LoRA. Traditional methods are less suitable for low-resource scenarios, and our approach aims to tackle this issue. We will clarify this logic in the next version to avoid any confusion.

Q3: Provide a detailed comparative analysis between JOLA and LOFIT.

A3: Below, we provide a comparison focusing on methodological differences, intervention strategies, and empirical performance.

(1) Methodological differences: Unlike LoFiT’s rigid two-stage pipeline, JoLA unifies localization and intervention into a single end-to-end framework. This allows dynamic adaptation to task requirements, avoiding suboptimal head selection caused by decoupled optimization.

(2) Formula-level comparison

LoFIT: $z^{(l,i)}_t \leftarrow z^{(l,i)}_t + v^{(l,i)} \quad \text{(Additive bias)}$

Limitation: Static additive edits may insufficiently adapt complex tasks requiring both amplification and suppression of features.

JoLA:

$$z^{(l,i)}_t = (1 + g_m^{(l,i)} \cdot m^{(l,i)}) \odot z^{(l,i)}_t + g_a^{(l,i)} \cdot a^{(l,i)}$$

where, Scaling: $(1 + g_m^{(l,i)} \cdot m^{(l,i)})$, Bias: $g_a^{(l,i)} \cdot a^{(l,i)}$

Advantage: The hybrid operation enables fine-grained control over activations—scaling amplifies/suppresses existing features, while biases shift representations.

(3) Empirical Performance

• JoLA show robustness across 26 tasks under low-resource settings.
• JoLA is more parameter efficient than LoFIT, we will update the parameter counting as discussed in Review 2Bjt [Q2] in the new version of the paper.

Q4: Parameter issues

A4: As noted in the original papers for each baseline method, these methods are sensitive to hyperparameters, and we detail this in Appendix D.3. The following table shows the sensitivity of these methods, with the results highlighting the lack of robustness, which motivates our proposal for a more stable approach.

Different learning rate in RED

Q5: What happens on the physics task in Figure 2?

A5: Upon reviewing the generated text for the physics task in Figure 2, we found that in the low-resource setting, the generated outputs did not consistently follow the required format, which impacted the results. Specifically, the generated text lacked the necessary result options, leading to lower accuracy when evaluated using exact match criteria.

We appreciate your comments and the opportunity to clarify and improve our work.

Q1: Huge difference in hyperparameter selection among the baseline methods and JoLA. The reproducibility of our proposed JoLA

A1: As noted in the original papers for each baseline method, these methods are sensitive to hyperparameters, and we detail this in Appendix D.3. The following table shows the sensitivity of these methods, with the results highlighting the lack of robustness, which motivates our proposal for a more stable approach.

Different learning rate in RED

Q2: Providing a more detailed explanation of the statistical significance of the results.

A2: Thank you for your suggestion. We will incorporate the statistical significance; note though that the gap is large We agree that conducting a statistical significance analysis is important to strengthen the credibility of our results. The significance test shows a meaningful difference between our method and the baseline methods.

Q3: How does the JoLA method perform when given larger resources?

A3: Thank you for your suggestion. We conducted experiments on larger datasets, and as shown in the table, our method remains effective with 5,000 and 10,000 samples. However, using 20,000 and 100,000 samples, there is a slight gap between our method and LoRA. These differences are acceptable, as we update fewer parameters compared to LoRA.

Q4: Could the effectiveness of JoLA on the MLP layer be verified?

A4: We also applied the gating mechanism-based strategy to the MLP layer. As shown in the following table, our experimental results demonstrate that the JoLA method is also effective in the MLP layer. However, it empirically tends to be more effective when used in the attention layer.

Thank you for taking the time to review our paper and provide valuable insights.

Q1: Baseline setup seem either under-tuned or too weak to compare with?

A1: Thank you for your feedback on our baseline system setup.

(1) As noted in the original papers for each baseline method, these methods are sensitive to hyperparameters, and we detail this in Appendix D.3. The following table shows the sensitivity of these methods, with the results highlighting the lack of robustness, which motivates our proposal for a more stable approach.
Different learning rate in RED

(2) Baseline methods require specific fine-tuning for each task, and we have 26 tasks to evaluate in our method. Fine-tuning all tasks individually is impractical due to the large hyperparameter search space. Therefore, we performed hyperparameter selection using a grid search approach. For each task, we ran a grid search with five different hyperparameter configurations, which were chosen to explore a diverse range of parameter settings that could provide the best model performance. We performed this search over key hyperparameters (e.g., learning rate, selected head/layers/positions etc. highlighted in Appendix D.3), using a validation set to select the configuration that resulted in the best performance. The results presented here correspond to the best hyperparameter configuration selected for each task based on the validation set performance. The final model was evaluated with these hyperparameters, and we averaged the results across all tasks. As observed, this had little effect on their relative performance, with our method continuing to outperform the others.

Table: best hyperparameter configuration in LLaMA

(3) Given the small dataset (e.g., 200 samples in our setting), overfitting was a concern. To reduce overfitting's impact on the baseline, we used early stopping, which was not applied in the original implementation of the baseline systems. We also found that learning rate adjustment significantly affected the results. We evaluated four strategies, including linear schedule[1], Cyclic Learning Rate Schedule[2], Adaptive Heuristic Schedule[3] and Exponential Decay Schedule[4]. As shown in the following table, the exponential decay strategy proved most stable, so we used it for both the baseline and our method, as explained in Appendix D.1. The comparison of different learning rate decay strategies for JoLA and LoFIT is as follows. [1] Human-level control through deep reinforcement learning (Mnih et al., 2015) [2] Cyclical Learning Rates for Training Neural Networks (Smith et al., 2017) [3] A disciplined approach to neural network hyper-parameters (Smith et al., 2018) [4] An exponential learning rate sched- ule for deep learning (Li et al., 2019)

Different learning ratestrategies in JoLA

Q2: (same questions as Review QWX1 Review m3pE) parameter counting of the gating mechanism?

A2: We mentioned the parameter counting in Table 3 of Appendix A, using the same calculation method as ReFT and LoFIT, where trainable parameters are divided by the parameters of the base LLM. For example, we calculated the parameters for the SIQA task using the LLaMA-3-8B model. JoLA’s parameter count considered only the interventions. Interestingly, JoLA’s count matches LoFIT’s.

Q3: Some figures are hard to read, e.g., Figure 4 and Figure 8.

A3: Thank you for your feedback. Due to space constraints, we combined multiple images into one, which affected readability. In the next version, we will adjust the layout to improve clarity and readability.

Thank you for your feedback and suggestions.
Comment: The parameterized form of the hard-concrete distribution is not explicitly detailed. The trainable parameters of the gating mechanism are not clearly stated.
Response: The hard concrete distribution has two associated scalar parameters: a scale parameter and a temperature parameter. Following prior work on sparsification (e.g., Voita et al., 2019; Louizos et al., 2017), we train only the scale parameter and fix the temperature to 0.33. To clarify, these gates do not take any input – each gate is simply an instance of the hard concrete distribution with a single learnable parameter. We will clarify this in the new version of the paper.

• Q1: What exactly do the hidden states and bias term in Fig. 2 refer to?

• A1: Figure 2 illustrates the previously proposed forms of intervening in Transformer modules. The “Hidden states” approach follows REFT, applying interventions directly to the MLP hidden states. The “Bias” approach follows BitFit, modifying only the bias terms of attention, dropout, and layer norm activations. We will elaborate on these in Section 3.1 (currently lines 86–88) to clarify further.

• Q2: Could the reason for multiple components performing worse in Fig. 2 be overfitting?

• A2: We think it can be interpreted as a form of overfitting, though we do mitigate this using early stopping. More broadly, the results highlight that both deciding where we intervene (i.e., the choice and location of components) and having fewer interventions are important in our low-resource setting. As shown below, even when provided with more data (500 examples), the MLP+Attention intervention fails to match the performance of Attention-only using 200 examples across most datasets.

The same sample size as in Figure 2 (200)

• Q3: How many attention heads are typically completely shut off by the end of training?
• A3: To clarify, we are not shutting down any attention heads. Instead, we selectively choose which heads to apply interventions to. When both the offset gate $g_a$ and the multiplicative update gate $g_m$ are set to $0$, the head’s computation becomes identical to that of the original model. By the end of training, most gates are closed – for example, on OBQA, 86% of attention heads have $g_a = 0$, and 94% have $g_m = 0$ (See Figure 8).
• Q4: How does the gating mechanism's randomness manifest during inference?
• A4: During training, we model each gating variable (e.g., $g_a^{(l,i)}$ and $g_m^{(l,i)}$) as random, sampled from a hard-concrete distribution. However, during inference, we use their expected values, $\mathbb{E}[g_a^{(l,i)}]$ and $\mathbb{E}[g_m^{(l,i)}]$, instead of sampling from the distributions. This removes randomness, ensuring consistency and stability in inference. We briefly mention this in lines 165-166 and will clarify in the next version.