- Rating
- 2
- Confidence
- 2
- Informativeness
- 3(+1)
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The authors present DHAFT, a soft prompt tuning method for frozen LLMs. DHAFT uses linear CKA to compare layers based on 256 unlabeled training inputs. Then, agglomerative clustering is used to group the layers into G "stages". G is determined by silhouette score with bootstrap. At the beginning of each stage, a trainable soft prompt is inserted; at the end of each stage, the prompt is removed. This ensures that the sequence length remains constant at n+m.
The paper also introduces PSMI, an index derived from layer-wise gradient-times-attention saliency. PSMI quantifies both the amount of saliency mass that extends beyond the stage where a prompt is inserted and the variability of saliency within the stage.
The authors evaluate DHAFT on eight commonsense reasoning tasks using four different LLaMA-family models and five math-related tasks on Llama-2-7B, Llama-3-8B and Qwen3-8B. The gains on commonsense reasoning are minimal, the gains on math-related tasks are more substantial, primarily due to improvements on GSM8K.
Critical strengths
-
The proposed approach is relatively straightforward and computationally efficient. Only a small number of soft prompts are trained, and there is no additional per-sample inference cost other than the extra prompt tokens.
-
Using CKA rather than cosine distance to compare layers is logical, and the authors provide a clear explanation of the invariance property in Appendix A.5.
-
Removing the prompt at the end of each stage is a clean way to limit the sequence length. Algorithm 2 make the forward pass easy to understand.
-
The experiments involve multiple backbones and two task families, and seed variance is reported for the math results.
Critical weaknesses
- The saliency score is not specific to the prompt. Equation (1) sums the gradient-times-attention term over the entire attention matrix and then averages over all token positions. It does not restrict the calculation to attention involving prompt tokens. The method calculates flows between specific token groups. Therefore, I^l represents overall attention saliency at layer l, not "prompt influence". PSMI mainly reflects stage size. The histogram in Figure 7 indicates that most saliency values fall between 0.70 and 0.82, so after normalization, q is nearly uniform. For nearly-uniform q, the leakage term is approximately 1 - |G_i|/L, as stated by the authors in property (v). Therefore, the comparison of shallow/middle/deep in Figure 2(f) is primarily a comparison of how large the stage containing each insertion layer is.
- The relationship between PSMI and accuracy shown in Figure 2(g) is based on four points. These four configurations also vary in aspects other than PSMI. Therefore, there is insufficient evidence to conclude that higher PSMI corresponds to lower accuracy. Furthermore, PSMI is never reported for DHAFT, so the paper does not demonstrate that DHAFT actually reduces misalignment. As presented, the PSMI contribution is unrelated to the method.
- The math results appear excessively high, and the training data is not specified. Llama-2-7B at 66.1% on GSM8K is comparable to what MetaMath-7B reports with full fine-tuning on approximately 395K augmented examples.
- All baselines listed in Table 1 are taken from previous publications, whereas DHAFT is evaluated within the authors' own pipeline.
Questions for authors
- Figure 7 uses 100 of the 500 probing examples. Why?
- The seeds also affect the calibration sample. Did the partition or G vary across seeds?
- Typos: "Propogation" in Figure 1(c), and "Eq. equation 1" in several locations.
6 comments
Response to Reviewer 1BYm [2/4]
W1.2: Relationship Between PSMI and Reasoning Accuracy
The relationship between PSMI and accuracy shown in Figure 2(g) is based on four points. These four configurations also vary in aspects other than PSMI. Therefore, there is insufficient evidence to conclude that higher PSMI corresponds to lower accuracy. Furthermore, PSMI is never reported for DHAFT, so the paper does not demonstrate that DHAFT actually reduces misalignment. As presented, the PSMI contribution is unrelated to the method.
Response: We thank the reviewer for highlighting this important limitation. We agree that the original four-point PSMI–accuracy analysis does not isolate the effect of cross-stage prompt influence. We also recognize the need to directly evaluate PSMI for DHAFT.
Following your suggestion, we conduct a controlled attention intervention on the same Llama-3-8B Retain checkpoint, fixing model parameters, stage boundaries, prompt insertion locations, and evaluation settings while progressively restricting direct attention access to earlier prompts. We evaluate prompt-specific PSMI on 500 GSM8K probing examples and separately report independently trained DHAFT.
| Prompt Configuration | PSMI ↓ | GSM8K ↑ |
|---|---|---|
| Retain-All | 0.475 | 72.8 |
| Retain + Mask 25% | 0.442 | 73.0 |
| Retain + Mask 50% | 0.367 | 72.9 |
| Retain + Mask 75% | 0.255 | 73.2 |
| Retain + Mask All | 0.009 | 73.5 |
| DHAFT (Remove) | 0.011 | 74.0 |
As shown above, prompt-specific PSMI consistently decreases as direct cross-stage access is restricted, while GSM8K accuracy generally improves. Within the same Retain checkpoint, fully masking earlier prompts improves accuracy from 72.8% to 73.5% without retraining. Independently trained DHAFT achieves 74.0% accuracy with a substantially lower PSMI of 0.011, compared with 0.475 for Retain-All.
These results provide complementary evidence that restricting direct access to earlier prompts can improve reasoning performance. We accordingly interpret PSMI as a diagnostic measure rather than treating its correlation with accuracy as causal evidence.
The additional PSMI analysis in Section 4.3 (Table 7) is highlighted in red, with the corresponding discussion highlighted in blue. The revised PSMI formulation in Section 3.2 is highlighted in blue.
W1.3:
The math results appear excessively high, and the training data is not specified. Llama-2-7B at 66.1% on GSM8K is comparable to what MetaMath-7B reports with full fine-tuning on approximately 395K augmented examples.
Response: We appreciate the reviewer's concern regarding training data transparency and evaluation reproducibility. Following your suggestion, we clarify the training data, decoding settings, and answer extraction protocol underlying our mathematical reasoning results.
For GSM8K [1], DHAFT uses the official 7,473 training examples and 1,319 test examples, optimizing only stage-specific soft prompts while keeping the pretrained Llama-2-7B backbone frozen. The 256 calibration examples for CKA-based stage discovery are sampled exclusively from the training split.
Evaluation uses greedy decoding with a maximum of 512 new tokens. We extract the final numerical answer after removing numerical separators and trailing punctuation, counting invalid or unparseable outputs as incorrect. DHAFT is evaluated across three independent training seeds.
| Method | Params (%) | GSM8K ↑ |
|---|---|---|
| LoRA | 0.83 | 54.1 |
| DoRA | 0.84 | 55.4 |
| LoReFT | 0.03 | 58.9 |
| DHAFT (Ours) | 0.025 | 66.1 ± 0.28 |
DHAFT achieves 66.1 ± 0.28% GSM8K accuracy, exceeding the published LoReFT result by 7.2 percentage points with approximately 0.025% trainable parameters. The LoRA, DoRA, and LoReFT results are taken from their respective published evaluations.
For context, MetaMath-7B [2] reports 66.5% using approximately 395K augmented mathematical examples with full fine-tuning, whereas DHAFT uses only the original 7,473 GSM8K training examples for supervised adaptation with a frozen backbone. The training data, decoding settings, and answer extraction protocol in Sections 4.1–4.2 and Appendix A.6.2 are clarified in blue.
References
[1] Cobbe et al. Training Verifiers to Solve Math Word Problems. 2021.
[2] MetaMath: Bootstrap Your Own Mathematical Questions for Large Language Models. ICLR 2024.
Response to Reviewer 1BYm [3/4]
W1.4:
All baselines listed in Table 1 are taken from previous publications, whereas DHAFT is evaluated within the authors' own pipeline.
Response: We appreciate the reviewer's concern regarding the fairness of baseline comparisons. We agree that results obtained under different implementation and evaluation settings should be distinguished from strictly controlled comparisons.
For the original LLaMA-backbone results in Table 1, we retain the published baseline scores and clarify their provenance in Appendix A.6.1. DHAFT follows the established commonsense reasoning protocol of LLM-Adapters [1] and DoRA [2], using the combined training data of eight benchmarks and their corresponding task-specific test splits. We summarize the Llama-3-8B comparison below.
| Method | Params (%) | Avg. Acc. ↑ |
|---|---|---|
| LoRA | 0.70 | 80.8 |
| DoRA | 0.71 | 85.2 |
| LoReFT | 0.03 | 86.6 |
| HiRA (r=32) | 0.70 | 86.7 |
| DHAFT (Ours) | 0.021 | 86.9 |
DHAFT achieves 86.9% average accuracy with only 0.021% trainable parameters, exceeding the published LoReFT and HiRA (r=32) results by 0.3 and 0.2 percentage points, respectively. Given the differences in evaluation pipelines, we interpret these comparisons as evidence of parameter efficiency and competitiveness rather than attributing the small accuracy differences exclusively to DHAFT.
Additional Qwen3-8B Evaluation. To provide further empirical evidence beyond the previously reported LLaMA baselines, we extend our evaluation to eight commonsense reasoning benchmarks on Qwen3-8B, including the unadapted backbone and multiple adaptation baselines. We adopt the official Qwen3 chat template, Non-Thinking Mode, and candidate answer log-likelihood scoring, with baseline provenance explicitly distinguished in the revised manuscript. The results are summarized below.
| Method | Params (%) ↓ | BoolQ | PIQA | SIQA | HellaS. | WinoG. | ARC-e | ARC-c | OBQA | Avg. ↑ |
|---|---|---|---|---|---|---|---|---|---|---|
| Pretrained | — | 86.6 | 76.7 | 83.3 | 57.2 | 68.0 | 83.3 | 55.5 | 31.4 | 67.75 |
| Prefix-Tuning | 0.0216 | 80.6 | 84.5 | 80.0 | 90.3 | 79.5 | 89.6 | 75.9 | 75.6 | 82.00 |
| LoRA | 0.0234 | 85.1 | 87.8 | 82.4 | 93.4 | 84.5 | 92.0 | 79.6 | 81.8 | 85.83 |
| LoReFT | 0.0216 | 86.4 | 89.3 | 83.7 | 94.7 | 85.9 | 92.8 | 81.4 | 83.1 | 87.16 |
| HiRA (r=16) | 0.3588 | 86.3 | 89.1 | 83.7 | 94.8 | 85.8 | 92.9 | 81.8 | 83.5 | 87.24 |
| HiRA (r=32) | 0.7150 | 87.1 | 89.2 | 83.9 | 95.0 | 86.0 | 92.9 | 82.6 | 83.3 | 87.50 |
| DHAFT (Ours) | 0.0204 | 86.7 | 89.7 | 84.1 | 95.1 | 86.2 | 93.3 | 82.1 | 83.9 | 87.64 |
As shown above, DHAFT achieves 87.64% average accuracy with only 0.0204% trainable parameters, outperforming the unadapted backbone, LoRA, and LoReFT by 19.89, 1.81, and 0.48 percentage points, respectively. It also achieves a slightly higher average accuracy than HiRA (r=32), while using substantially fewer trainable parameters. These results provide additional evidence of DHAFT's competitive performance and parameter efficiency on a modern backbone.
The baseline provenance and evaluation protocol in Appendix A.6.1 are highlighted in blue. The Llama-3-8B comparison in Appendix A.6.2 (Table 9) and the additional Qwen3-8B results in Section 4.2 (Table 1) are highlighted in red, with evaluation configurations in Appendix A.7 (Table 13) clarified in blue.
References
[1] LLM-Adapters: An Adapter Family for Parameter-Efficient Fine-Tuning of Large Language Models. EMNLP 2023.
[2] DoRA: Weight-Decomposed Low-Rank Adaptation. ICML 2024.
Response to Reviewer 1BYm [1/4]
Dear Reviewer 1BYm,
We sincerely appreciate your insightful and constructive feedback, which has helped us refine the PSMI formulation, strengthen its empirical validation, and clarify the fairness and reproducibility of our evaluation. Below, we provide detailed point-by-point responses together with the corresponding revisions, with new experimental results highlighted in red and methodological clarifications highlighted in blue in the revised manuscript.
W1.1:
Response: We thank the reviewer for this important observation. We agree that the original saliency formulation includes non-prompt attention interactions, while stage size and injection depth may introduce structural bias into PSMI. Following your suggestion, we introduce two refinements to address these concerns.
(1) Prompt-Specific Saliency. We restrict gradient-weighted attention to task-token queries and the corresponding prompt-token keys, rather than aggregating over the entire attention matrix:
Iiℓ=H∣T∣∣Pi∣1∑h=1H∑t∈T∑p∈PiAtph,ℓ∂Atph,ℓ∂L.
Here, T and Pi denote the task-token and prompt-token positions, respectively. This formulation excludes task-to-task attention and explicitly measures the direct attention saliency associated with the target prompt.
(2) Stage-Size and Injection-Depth Correction. To account for structural bias in cross-stage leakage, we introduce a uniform-saliency reference over the layers accessible to each prompt under the unmasked Retain policy:
CSLobs(i)=∑ℓ∈RiIiℓ∑ℓ∈Ri∖GiIiℓ
CSLnull(i)=1−∣Ri∣∣Gi∩Ri∣
CSLcorrected(i)=CSLobs(i)−CSLnull(i)
Here, Gi denotes the assigned stage and Ri denotes the layers accessible to the target prompt under the unmasked Retain policy. The corrected metric is undefined when the total prompt-specific saliency is zero.
As shown above, the corrected leakage is zero across all three injection depths under uniform saliency. This confirms that the expected contribution of stage size and injection depth is removed by construction under the uniform reference, rather than being interpreted as prompt-specific misalignment. We therefore treat the PSMI–accuracy relationship as descriptive evidence and further examine direct cross-stage prompt access through the fixed-checkpoint intervention in W1.2.
The revised prompt-specific saliency formulation in Section 3.2 and the stage-size correction in Appendix A.4 (Table 8) are highlighted in blue and red, respectively, with the corresponding explanations highlighted in blue.