- Rating
- 3
- Confidence
- 2
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
The paper recognizes a scale-representation issue in using weight gradients to make decisions about which parts of the model are important and which ones are not. The idea is that if a product is preserved by scaling (), then the gradients change without the input-output function changing. If we just measure the gradients at then we lose reproducibility. We should seek a scoring system that is invariant to such scaling transformations. The paper recommends some alternatives and performs experiments to verify their fidelity.
Critical strengths
The paper recognizes an interesting perspective on how many interventions are made to machine learning models. Gradient-based decision making is common, yet is performed under a moving ground, which threatens its reproducibility. If the measurement is not invariant to transformations that preserve the function, how can we trust it? The paper does a good job building the argument and is quite convincing as to its validity.
A variety of balancing methods are proposed. From simple rescaling to a two-sided projection mechanism. These are interesting techniques that solidify the coordinate system in which gradients are measured and ensure reproducible decision-making.
Experiments are done in a sound and coherent way.
Critical weaknesses
- The writing is a bit dense. Most of the messages are coming across fine, but I would reduce the verbosity in certain parts so that the arguments are simplified.
- I am a bit uncertain about the correct way to think about this problem. I agree with the premise: a balanced model is the correct starting point if we want gradient-based decision making to be reproducible. But one thing that concerns me is whether balancing changes the gradients in ways that can negatively affect our decisions. The model was not trained with dynamics that these new gradients outline, so should we be making decisions based on them?
- It is not clear to me which scoring method to choose from the experiments. I think having an evaluation metric of how good a scoring method is would be nice. That metric now is just success / failure in a task. But that metric changes for each model / task, so it is hard to know when to try what.
- I would include some of the figures from the appendix to the main section. They helped me understand and digest the results.
Questions for authors
N/A
4 comments
| Score | Minimum scalar-gauge mask Jaccard | Held-out KL fidelity ↑ | Accuracy drop from dense (pp) ↓ |
|---|---|---|---|
| Raw, checkpoint coordinate | 0.189 | 0.623 ± 0.028 | 5.87 ± 2.39 |
| Raw, rescaled coordinate | 0.189 | 0.504 ± 0.028 | 0.45 ± 0.01 |
| CB | 1.000 | 0.620 ± 0.028 | 5.73 ± 2.09 |
| PN | 1.000 | 0.634 ± 0.028 | 3.35 ± 1.30 |
| QS | 1.000 | 0.583 ± 0.029 | 6.97 ± 1.31 |
| MG | 1.000 | 0.712 ± 0.030 | 0.27 ± 0.08 |
| FS | 1.000 | 0.724 ± 0.013 | 0.20 ± 0.12 |
MG and FS perform strongly on all three criteria: identical selections across the tested scalar coordinates, the highest mean held-out fidelity, and the smallest mean accuracy drops. FS retains higher mean accuracy than raw scoring in either tested coordinate, in all seven checkpoint means.
W3(b). Controlled ablations that isolate the scoring rule.
We added controlled experiments to separate the effects of the scoring rule, layer scale, and selection policy. To isolate the scoring rule, every method uses the same checkpoint, calibration examples, deletion budget, and per-layer L2 normalization, following Michel et al. (2019; https://arxiv.org/abs/1905.10650). We divide each layer's head scores by their L2 norm before global ranking. This separate study uses all eight encoder/vision checkpoints, with 128 GLUE training examples or 256 CIFAR-10 training images per draw (seeds 7, 11, 13), full validation/test evaluation, and no recovery.
Each entry averages Q–K/V–O within a draw, then reports mean ± sample standard deviation over the three draws. Dense accuracy is fixed. Random uses ten uniform masks and reports their mean ± sample standard deviation; the same masks are shared across calibration draws.
Held-out accuracy (%) after 15% head removal.
| Model | Dense | Raw | CB | PN | QS | MG | FS | Random |
|---|---|---|---|---|---|---|---|---|
| DistilBERT SST-2 | 91.06 | 90.60 ± 0.17 | 90.62 ± 0.14 | 90.52 ± 0.18 | 90.58 ± 0.18 | 90.56 ± 0.53 | 90.98 ± 0.29 | 90.45 ± 0.26 |
| BERT SST-2 | 92.43 | 92.13 ± 0.09 | 92.14 ± 0.06 | 92.03 ± 0.26 | 92.18 ± 0.18 | 92.32 ± 0.11 | 92.24 ± 0.07 | 91.63 ± 0.36 |
| RoBERTa SST-2 | 94.04 | 94.06 ± 0.12 | 93.98 ± 0.00 | 93.86 ± 0.32 | 93.98 ± 0.20 | 93.96 ± 0.18 | 94.15 ± 0.20 | 93.50 ± 0.56 |
| ELECTRA SST-2 | 95.18 | 94.61 ± 0.21 | 94.71 ± 0.12 | 94.76 ± 0.24 | 94.97 ± 0.17 | 94.69 ± 0.13 | 95.26 ± 0.18 | 94.62 ± 0.43 |
| RoBERTa MNLI | 88.10 | 87.33 ± 0.19 | 87.30 ± 0.11 | 87.07 ± 0.21 | 87.13 ± 0.38 | 87.63 ± 0.06 | 87.48 ± 0.09 | 85.46 ± 1.76 |
| RoBERTa QNLI | 92.68 | 92.15 ± 0.18 | 92.08 ± 0.30 | 91.95 ± 0.18 | 91.95 ± 0.41 | 92.51 ± 0.02 | 92.48 ± 0.08 | 89.52 ± 1.85 |
| RoBERTa MRPC | 91.18 | 90.69 ± 0.44 | 90.48 ± 0.43 | 89.34 ± 0.32 | 89.87 ± 0.19 | 90.69 ± 0.25 | 89.30 ± 0.51 | 84.90 ± 5.94 |
| ViT CIFAR-10 | 98.52 | 79.17 ± 0.70 | 79.17 ± 0.70 | 87.31 ± 5.14 | 79.00 ± 0.59 | 97.81 ± 0.10 | 97.56 ± 0.11 | 94.05 ± 7.25 |
The benefit of measuring head suppression persists with per-layer score normalization held fixed. On ViT, accuracy is 97.81 ± 0.10% for MG and 97.56 ± 0.11% for FS, versus 79.17 ± 0.70% for raw scoring. The paired MG-minus-raw gain is 18.64 ± 0.74 pp across the same three draws. MG and FS also exceed the random-control mean on every checkpoint.
The budget comparison uses the same normalization and averages the seven GLUE checkpoints equally within each draw. The final two columns compute method-minus-raw differences within the same draw, using unrounded accuracies, before calculating their mean and standard deviation.
GLUE mean accuracy (%), with paired gains in percentage points.
| Heads removed | Raw | CB | MG | FS | MG − Raw (pp) | FS − Raw (pp) |
|---|---|---|---|---|---|---|
| 10% | 91.74 ± 0.14 | 91.78 ± 0.17 | 91.83 ± 0.15 | 91.82 ± 0.02 | +0.10 ± 0.13 | +0.08 ± 0.12 |
| 15% | 91.65 ± 0.07 | 91.62 ± 0.06 | 91.76 ± 0.09 | 91.70 ± 0.11 | +0.11 ± 0.05 | +0.05 ± 0.08 |
| 30% | 90.48 ± 0.41 | 90.54 ± 0.37 | 91.02 ± 0.11 | 91.28 ± 0.27 | +0.54 ± 0.32 | +0.80 ± 0.64 |
MG and FS achieve higher mean GLUE accuracy than raw scoring at all three budgets. The gains increase at 30% removal, where FS improves all seven checkpoint means, with positive macro gains in every calibration draw (0.30, 0.59, and 1.52 pp).
A further controlled experiment varies the selection policy while keeping the underlying scores fixed. With per-layer score normalization fixed, changing coordinates still changes raw masks in 42/42 comparisons. A policy that retains at least one head per layer also leaves all 42/42 raw masks changed across coordinates and allows an 8.58 pp accuracy gap. MG and FS preserve every mask under both policies, removing the coordinate dependence that remains after these safeguards.
W3(c). Which score to choose, and its cost.
For head pruning, start with MG. Use FS when you want to measure complete deletion directly. Use QS when your question concerns loss sensitivity to small changes in the effective matrix or adapter.
| Your question | Recommended score | Measurement and cost |
|---|---|---|
| How sensitive is the loss to slightly weakening a head? | MG | Mean absolute per-example derivative with respect to the head-output multiplier at 1. One forward/backward pass per calibration minibatch scores all heads. |
| What changes if I remove a whole head? | FS | Output-logit change after deletion. One dense forward evaluation plus one per head. |
| How sensitive is the loss to a small change in the effective matrix or adapter? | QS | Largest local loss response to a unit product change. Factor gradients and two small Gram-system solves. |
| How can I compare parameter gradients in consistent scalar units? | CB or PN | CB equalizes paired weight norms; PN cancels reciprocal scales in the score. No per-head backward passes. |
The guarantees match these uses: MG/FS are invariant under transformations preserving the same head intervention; QS covers admissible invertible factor-basis changes under Theorem 2; CB/PN cover reciprocal scalar changes. QS extends the guide to general factor-basis changes: in the six-adapter study reported in our response to jjmy, Q3, it preserves the selected modules and post-pruning accuracy across all 204 equivalent factorizations.
On one A100 40GB, scoring 128 examples takes 0.11–0.39 seconds for MG and 2.73–19.39 seconds for FS across the seven GLUE checkpoints (median of three repeats after warm-up; loading/tokenization excluded). Batched MG preserves per-example absolute values and reproduces serial-gradient masks. These costs make the choice concrete: MG scores all heads at low cost; FS directly measures complete head removal through additional forward evaluations.
W1 / W4. Simpler writing and figures in the main text.
We have shortened Secs. 2–3, introduced the framework with rewrite → verify outputs → compare decisions, and explained each remedy's purpose before its formula. Submitted Figure 3 (summary diagnostics) and Figure 6 (ELECTRA mask substitutions) are now Figures 2 and 3 in the main text. The absolute-performance and score-selection tables appear beside the claims they support, with full results in Appendix D.18–D.21.
Thank you for your responses! The rebuttal clears up my concerns on the scoring method and I think including such a discussion in the main paper would help a lot. I do appreciate the discussion on the effect of balancing on the potential disconnect from what the model trained on. I am not fully convinced that balancing is a completely neural operation, but the initial experiments do seem favorable. I will maintain my slightly positive score due to the above, as I do further find the writing and presentation as requiring substantial simplification and improvement towards publication.
Thank you for the questions about balancing and score selection. For head pruning, we recommend mask gradients (MG) for fast scoring and function-space ablation (FS) for directly measuring complete head removal. The added controlled experiments test the effect of balancing, compare the scores under matched conditions, and support a practical guide to score selection and cost.
W2. Does balancing disconnect the measurement from the model that was trained?
Balancing preserves the trained function; pruning asks how that function responds when a head is weakened or removed. Canonical balancing (CB) puts paired weights on a common scale before measuring their gradients. Changing units changes the gradient entries, while expressing the same perturbation in those units preserves its first-order loss effect.
Multiply head h's output by mh: 1 keeps the head, and 0 removes it. MG measures how sensitive the loss is to slightly reducing mh from 1. FS sets mh to 0 and measures how much the output logits change. Our transformations preserve the function at every value of mh, so these scores measure the same operation on the trained checkpoint before and after balancing.
We verified this in 1,400 paired loss/derivative comparisons across seven checkpoints, five suppression levels, and scalar, balanced, orthogonal, and general-basis coordinates. A separate 128-example training subset evaluates the curves. Predictions agree throughout; maximum mean-loss discrepancy is 5.55×10−7 nats/example. The agreement holds from keeping the head at full strength to removing it completely.
We also tested the effect on actual decisions. CB preserves the raw checkpoint mask in 15/16 original encoder/ViT settings; the remaining case swaps one head (0.34 pp accuracy change). In the three-draw, layer-normalized study below, CB-minus-raw accuracy is −0.03 ± 0.06 pp, averaged over eight checkpoints. Balancing therefore closely tracks the checkpoint criterion while making its scalar units reproducible.
ELECTRA shows how the audit leads to an effective correction. Raw checkpoint gradients and CB choose the same mask, which removes all final-layer heads and yields 50.92%. With the checkpoint, calibration data, and budget fixed, changing only the per-layer score normalization raises CB to 95.18%. MG and FS reach 94.72% and 94.95% under the original global rule. This controlled ablation isolates the contribution of layer scale; W3(b) then gives every method the same normalization to isolate the scoring rule.
W3(a). A common metric for evaluating a score.
We evaluate a score with three questions:
The fidelity measure has the same interpretation across tasks: rank heads by their scores, delete each head on held-out inputs, and compare the score ranking with the ranking of the resulting KL changes:
q_h &= \mathbb{E}_{x\in D_{\mathrm{heldout}}} D_{\mathrm{KL}}\\!\left(p_{\theta}(\cdot\mid x)\\,\Vert\\,p_{\theta\setminus h}(\cdot\mid x)\right), \\\\ \rho_{\mathrm{KL}} &= \operatorname{Spearman}_{h}(s_h,q_h). \end{aligned}$$ Here $s_h$ is the head score, $q_h$ is the measured change in output distribution, and higher $\rho_{\mathrm{KL}}$ means better agreement between their rankings. The reference uses separate inputs and KL divergence; FS uses calibration inputs and logit changes. Fidelity tests the score's predictions about individual heads; accuracy tests the complete pruning decision. The global 15% study uses seven GLUE checkpoints and 128 training examples per calibration draw (seeds 7, 11, 13), with no recovery. A separate 128-example reference set measures fidelity and never selects masks. We first average Q–K/V–O within checkpoint, then weight checkpoints equally. ± denotes sample standard deviation across the three calibration-draw averages, for fixed checkpoints and evaluation splits; the same convention applies to the macro table below. PN denotes path normalization.