acceptodds
Under review as a conference paper at ICLR 2027

Fulcrum: Predicting Layer Potential Before RL-Training

Abstract

Reinforcement learning (RL) is a central paradigm for post-training large language models, yet it typically updates all parameters at substantial compute and memory cost. Prior research has shown that layers differ in their ability to reshape the final representation. Concentrating updates on high-potential layers may therefore achieve performance comparable to full-parameter training at lower cost with limited training resources. The challenge is to identify these layers before RL begins. To address this challenge, we propose Fulcrum, a pre-RL layer assessment and selection method. It measures a layer’s potential as the maximum gain of a change in its representation as it propagates through the downstream network. This gain is characterized by the spectral norm of the downstream Jacobian and efficiently estimated using power iteration based on vector–Jacobian products (VJPs). Fulcrum further converts the resulting discrete layer potential ranking into a contiguous high-potential interval for training. Identifying this interval takes approximately 10.9 seconds for an 8B model. In Group Relative Policy Optimization (GRPO) experiments with Qwen3-1.7B, 4B, and 8B, updating only 27.8%–35.7% of all layers matches full-parameter training across ten benchmarks. On Qwen3-8B, peak GPU memory use falls by 56.2%. Experiments with supervised fine-tuning (SFT) and LoRA fine-tuning of the SDXL U-Net further demonstrate Fulcrum’s applicability across training objectives, architectures, and update strategies.

Primary area:
reinforcement learning
Keywords:
Jacobian Analysis, Layer Assessment, Training-Free Profiling, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_oixgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies how to select layers for updating before reinforcement learning begins. Fulcrum measures the amplification of representation changes using the spectral norm of each layer’s downstream Jacobian. It then removes a linear depth trend and selects a contiguous interval of layers to train. Experiments on three Qwen3 model sizes show that updating ten layers approaches the average performance of full-parameter training while reducing memory use. The paper also includes single-layer training experiments, an ablation that freezes the selected interval, and experiments on SFT and SDXL.

Critical strengths
  1. Selecting trainable layers before RL is a useful problem. A reliable method could reduce the cost of both layer-by-layer trial runs and subsequent training.
  2. The reported performance and memory results are encouraging. Across three model sizes, training only ten layers retains most of the average improvement from full-parameter training. On Qwen3-8B, peak memory falls from 69.96 GB to 30.62 GB. The performance drop when the selected interval is frozen also provides some evidence that these layers matter.
Critical weaknesses
  1. The paper lacks direct controls showing that Fulcrum is better than simple selection rules. There is no comparison with fixed middle intervals or random contiguous intervals under the same ten-layer budget and update scheme. This comparison is particularly important here: the selected interval [13,22] for Qwen3-4B is exactly the middle ten layers, and [12,21] for Qwen3-8B is also close to the center. The current results therefore do not establish whether the proposed score is useful or whether simply training middle layers would give similar results. The ablation in Appendix A.3 freezes the selected interval and trains all remaining parameters. This shows that excluding these layers can hurt performance, but it does not answer the question above. The paper also does not isolate the effects of the raw spectral norm and the detrending step. In addition, the five repeated runs are evaluations of the same checkpoint; training uses only one seed. It is therefore unclear whether the differences between methods are stable across training runs.

  2. The metric’s definition and implementation are not clearly reconciled, and its connection to training gains remains insufficiently tested. Under Eq. (5), for a sequence of length (T), the final-layer Jacobian is the identity only for matching token positions. Averaging over all causally reachable position pairs therefore gives (2I/(T+1)), rather than (I). This conflicts with the statement in Section 4.2 and Appendix A.2 that the final layer’s spectral norm is always 1. There may be an additional normalization that is not described, but the authors need to clarify what was actually computed. Section 3.4 also omits the complete power-iteration procedure, iteration count, and probe-vector settings, making the estimator difficult to check. More broadly, a large possible change in hidden representations does not establish that parameter updates can realize that direction or that doing so improves reward. The authors acknowledge that this connection has not been tested. The current single-layer experiments cover only ten selected layers in one model and provide limited evidence for the score’s predictive value.

Questions for authors
  1. Clarification: Do the experiments compute the aggregated Jacobian exactly as defined in Eq. (5)? If so, how is this consistent with a final-layer spectral norm of 1? Please provide the formula actually used and pseudocode for the estimator, and clarify whether the exact and VJP-based methods compute the same quantity.
  2. Additional evidence needed: Under the same ten-layer budget and training setup, does Fulcrum outperform a fixed middle interval, random contiguous intervals, and selection using the spectral norm without detrending? If these results are available, please also report variation across independent training seeds. These comparisons would directly affect my assessment.
3 comments
AuthorsOpenReview ↗

Reply to Reviewer oixg

We thank the reviewer for the detailed review of our paper.


W1. Fulcrum differs from simple middle selection.

Fulcrum is not a "middle-layers" rule. Although the selected intervals happen to lie at or near the center on Qwen3-4B and Qwen3-8B. It is an outcome of the potential profile, not an assumption of the method. In contrast, on the other two models in our experiments, Qwen3-1.7B and the SDXL U-Net, the selected layers are not in the middle.

Table R1. Fulcrum-selected intervals vs. middle intervals.

ModelLayersFulcrumMiddleSame?
Qwen3-1.7B28[6,15][9,18]✗
Qwen3-4B36[13,22][13,22]✓
Qwen3-8B36[12,21][13,22]✗
SDXL U-Net282 ResNet blocks (first up block)Middle block✗

Table R2. Fulcrum vs. the middle interval performance on Qwen3-1.7B-Base. Results are averaged over five evaluation seeds.

LayersMathCodeReasoningLanguageAvg.
Middle [9,18]49.79 ± 0.6839.31 ± 1.6430.91 ± 1.0342.14 ± 1.2942.38 ± 0.36
Fulcrum [6,15]49.92 ± 0.3040.33 ± 1.4532.99 ± 1.1641.72 ± 0.9042.97 ± 0.40

W2. Writing errors in the paper.

We thank the reviewer for the careful derivation.

(1) Incorrect formula. Eq. (5) contains a typo in the normalization coefficient: it should be TnT_n but not Tn(Tn+1)/2T_n(T_n+1)/2. Fulcrum sums the cotangent at all target positions, and then averages over source positions. For the final layer, the exact Jacobian reduces to the identity, yielding an exact spectral norm of 1

(2) Incorrect wording. Section 3.4 incorrectly describes the estimator as "power iteration". The implementation performs no iteration: it estimates σl\sigma_l with a Gaussian sketch, in which PP fixed-seed standard Gaussian probes are placed as cotangents on the final representation,a single round of VJPs yields Sₗ = UᵀĴₗ for all layers at once, and σl\sigma_l is estimated as σₘₐₓ(Sₗ)/√P. There is therefore no iteration count or convergence criterion. The Gaussian sketch provides a computationally efficient proxy for the spectral norm used in layer ranking, rather than an exact evaluation of the full Jacobian's spectral norm.

We note that this estimator is a standard randomized sketching technique from numerical linear algebra [1], used solely to accelerate the computation of σl\sigma_l. It is not a methodological contribution of this paper. Our contribution lies in defining layer potential through the downstream Jacobian and in turning it into a training-free layer-selection rule.

We sincerely apologize, thank the reviewer for pointing them out, and we will correct them in the paper. We will also add the pseudocode to the appendix.

Q1.

Please see our response to W2.

Q2.

To further examine the effectiveness of the selected interval, we provide an additional interval ablation:

Table R3. More interval training on Qwen3-1.7B-Base. Results are reported across five seeds.

LayersMathCodeReasoningLanguageAvg.
[0,9] (Raw spectral norm)48.31 ± 0.9139.53 ± 3.0431.95 ± 2.2742.12 ± 1.1442.04 ± 0.61
[4,13]49.11 ± 0.3838.88 ± 1.8731.90 ± 1.2341.48 ± 0.8142.09 ± 0.55
[6,15] (Fulcrum)49.92 ± 0.3040.33 ± 1.4532.99 ± 1.1641.72 ± 0.9042.97 ± 0.40
[8,17]49.82 ± 0.8140.62 ± 1.2132.05 ± 1.4940.77 ± 0.7142.61 ± 0.52
[9,18] (Middle)49.79 ± 0.6839.31 ± 1.6430.91 ± 1.0342.14 ± 1.2942.38 ± 0.36
[18,27]45.86 ± 0.4937.31 ± 1.2830.91 ± 1.4838.31 ± 1.3039.65 ± 0.59
Full49.57 ± 0.5940.62 ± 1.9132.17 ± 0.5443.50 ± 0.9343.09 ± 0.33
Base37.30 ± 0.7533.22 ± 1.6419.88 ± 0.8628.60 ± 1.3431.26 ± 0.35

The Fulcrum-selected interval achieves the best average performance among all tested partial-training intervals, approaching full fine-tuning.

Fulcrum offers a practical contribution: a training-free criterion that identifies a small set of layers within seconds, enabling RL post-training to approach full-parameter performance with substantially lower memory requirements. We welcome further discussion and would be grateful if the reviewer could reconsider their evaluation in light of these corrections and additional results.


Reference

[1] Halko, N., Martinsson, P.-G., & Tropp, J. A. (2011). Finding structure with randomness: Probabilistic algorithms for constructing approximate matrix decompositions. SIAM Review, 53(2), 217–288.

As the rebuttal deadline approaches, we would appreciate hearing your thoughts on our responses. We have clarified your issues concerning the equations and wording, results are consistent across seeds, and presented evidence that our method differs from simply mid-selecting.

In light of these clarifications and results, we respectfully believe that the weak-reject assessment merits reconsideration. We hope the author can engage in the discussion and let us know whether any concerns remain. We would be happy to provide further clarification.

If you remain inclined toward rejection, we would appreciate specific, actionable feedback for the valuable weaknesses thus we could address them to improve our work.

Reviewer_oixgOpenReview ↗

I thank the authors for the corrections and additional experiments. However, they do not resolve my central concerns.

I thank the authors for the corrections and additional experiments. However, they do not resolve my central concerns.

W1: The new comparison is only on Qwen3-1.7B, and it does not provide the direct controls I requested under the same ten-layer budget and training setup: fixed middle intervals, random contiguous intervals, and raw-spectral-norm selection without detrending across all models. Showing that Fulcrum differs from the middle on Qwen3-1.7B does not address the fact that on Qwen3-4B the selected interval is exactly the middle ten layers and on Qwen3-8B it is very close to the middle. The reported average gap on Qwen3-1.7B is small (42.97 vs. 42.38), and [8,17] is close (42.61); with only evaluation-seed variation and a single training seed, this is not sufficient to establish that Fulcrum is better than simple selection rules. Table R3 also does not include random contiguous intervals or independent training runs.

W2: The response concedes that Eq. (5) has a typo and that Section 3.4 incorrectly describes the estimator as power iteration; the actual implementation is a Gaussian-sketch proxy, not an exact spectral norm. This does not reconcile the metric’s definition with the reported final-layer spectral norm of 1, nor does it establish that the ranking proxy and the defined quantity are the same object. Correcting the formula and adding pseudocode is necessary, but it does not resolve the soundness concern about what was actually computed and whether this score predicts training gains.

Therefore, these responses do not change my assessment.

Reviewer_6X1uOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Fulcrum, a method for selecting a subset of LLM layers for RL training without first running RL. Fulcrum estimates each layer’s potential by measuring the maximum amplification of a perturbation to its representation as it propagates through the downstream architecture. Experiments on Qwen3-1.7B, 4B, and 8B show that updating only 27.8%–35.7% of layers matches the performance of full-parameter RL training while reducing peak memory usage.

Critical strengths
  1. The proposed method is quite intuitive and reasonable.

  2. Selecting a subset of layers for training is an interesting research question, although its practical utility appears limited.

Critical weaknesses
  1. The efficiency evaluation focuses solely on peak memory usage during RL training. However, training only a subset of layers does not reduce rollout costs, which can dominate total runtime. The memory savings may therefore yield limited wall-clock improvements. The authors should report end-to-end training time to clarify the practical benefits.

  2. The experiments are limited to dense Qwen models with at most 8B parameters. Evaluation on larger models, mixture-of-experts (MoE) architectures, and other model families would help establish the method’s scalability and generality.

Questions for authors
  1. How does the computational cost of identifying candidate layers scale with model size—is it approximately linear or superlinear? For trillion-parameter models, this overhead could become substantial and outweigh the subsequent training savings, particularly if the resulting model only matches baseline performance. The authors should quantify the selection overhead and clarify whether the method provides a net computational benefit.

  2. What motivates the choice of N=100N=100 input sequences for layer selection? How sensitive are the selected layers and downstream performance to this choice, for example, when N=200N=200 or 300300? Please also clarify how the input sequences are selected and whether the results are robust to different samples.

5 comments
AuthorsOpenReview ↗

Reply to Reviewer 6X1u

We thank the reviewer for finding Fulcrum intuitive and reasonable, and for the constructive questions on its practical benefits.


W1. Practical Benefit

We agree that reducing the number of trainable layers does not directly reduce rollout costs in GRPO, and Fulcrum does not establish an end-to-end speedup in this setting. The practical benefit of Fulcrum lies in two aspects.

1. Lowering the hardware barrier for RL. We would like to clarify the motivation behind Fulcrum. Our primary goal is not to reduce training time, but to lower the hardware barrier for RL research, as the GPU memory limitation is often determines whether researchers can conduct RL training at all. For example, full-parameter GRPO on Qwen3-8B requires either a data-center GPU with more than 96 GB (e.g., H200 NVL, 33.7k USD) or two 96 GB GPUs (~32k USD), whereas Fulcrum with simple optimizations fits on a single 80 GB GPU (A100 80GB, ~7–15k USD), reducing hardware cost by ~3×. Similarly, Qwen3-4B with Fulcrum fits on a single consumer RTX 5090 (32 GB, 1,999 USD), instead of the 16,000 USD RTX PRO 6000 required for standard full-parameter training, an ~8× reduction.

Fulcrum requires only about 10 seconds for layer selection before training, without modifying the model architecture or introducing additional modules. We have packaged Fulcrum as a tool in the supplementary material, welcome to try it.

2. Speeding up other post-training. Beyond GRPO, Fulcrum also benefits post-training settings where computation is dominated by model updates rather than rollout generation. On SDXL, it reduces training time by 34.0% and FLOPs by 39.5% (Table 5).

W2. Scalability and Generality

Fulcrum is broadly applicable by design. It only requires the mapping from each layer’s representation to the final representation of a base model without introducing component specific to any architecture or training loss. Our experiment on SFT/SDXL U-Net/Lora confirm its Generality in Section 4.5. Specifically,

(1) Other model families. Fulcrum applies to all dense architectures. Moving to a different model family does not significantly change the internal structure of a dense model. Accordingly, other families such as Gemma show the same trends, supporting its generality across model families.

(2) MoE architectures. MoE models show the same trend as dense models, but the advantage is not significant. The Jacobian is a first-order approximation, and it may not capture the specific contribution of each layer well in this family of architectures.

Full RL experiments on models substantially larger than 8B remain beyond our current computational resources, which we will state as a limitation.

Q1. Computational Cost (second-level)

Scaling. For a fixed profiling configuration, Fulcrum's selection cost grows linearly with the number of parameters PP, as each profiling prompt requires only one forward pass and several backward passes.

Measured runtime. In practice, the measured selection time is on the order of seconds, which is negligible.

Table R1. Runtime (s) of Fulcrum (from Table 3, Section 4.4).

Model1.7B4B8B
Standard (N=100)5.69.010.9
Fastest1.00.60.5

“Standard” uses the default setting of 100 profiling prompts; “Fastest” uses the fewest prompts and probes that still yield the same selected interval.

Fulcrum therefore provides a clear net benefit: substantial memory savings at essentially no selection cost.

Q2. Why N=100

The profiling sequences can be randomly sampled from any datasets. Our experiments suggest that the selected intervals are relatively stable across different profiling samples.

We chose N = 100 based on experimental testing: we had started from 1,000 sequences, following the setting of J-space[1] and progressively reduced N. N = 100 is a conservative setting that still yields stable intervals.As shown in Table R1, far fewer sequences already identify the same intervals (Fulcrum (Fastest), 0.5–1.0 s), so N = 100 leaves a large safety margin. Users who prefer more sequences can increase N at little extra cost.

We hope these responses clarify the practical benefits of Fulcrum and address the reviewer's concerns. We warmly welcome further discussion, which will help us further improve this work. We sincerely believe this work can benefit the development of the RL community. We would be grateful if the reviewer would consider re-evaluating our work.


Reference [1] Gurnee et al. (2026), Verbalizable Representations Form a Global Workspace in Language Models, Transformer Circuits Thread. https://transformer-circuits.pub/2026/workspace/

Thank you for the detailed responses.

Regarding the claim that your method lowers the hardware barrier for RL, I remain concerned that the current experiments do not sufficiently support such a broad conclusion. Hardware requirements are closely tied to the underlying RL training infrastructure. To substantiate this claim, I would expect evaluations across multiple training frameworks (e.g., VeRL, AReaL) and different hardware configurations. Alternatively, the claim could be qualified to reflect the specific settings evaluated.

Could you also clarify which GPU models were used in your 4B experiments?

We thank the reviewer for the quick respond and this constructive and reasonable suggestion.

Our current experiments use VeRL as the RL training framework (Section 4.1). We will investigate Fulcrum under other frameworks, such as AReaL, and report basic results as soon as they become available. Based on the coming results, we will consider qualifying our claims, as your suggestion.

Regarding hardware, we had tested Qwen3-4B across four GPU configurations: RTX 5090, A100, RTX PRO 6000, and RTX 4090 (48 GB) and didn't find any notable differences in experimental results. Specifically, the Qwen3-4B layer-selection experiments reported in the main paper were conducted on a single RTX 5090, while the Qwen3-8B experiments used an RTX PRO 6000. The parts of rebuttal experiments with different random seeds were conducted on B200 GPUs.

Let's keep in touch! :-)

Reviewer_3z3HOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper targets the problem that RL training of large language models is very memory-intensive.

The authors propose a method named Fulcrum, where only a few layers are selected for RL training. To this end, the memory cost of RL training can be largely reduced. They also demonstrate experimentally that performance does not degenerate much compared to full-parameter training.

In the proposed method, the selected layers are estimated using the spectral norm to assess the influence that changing a given layer will have on the final result.

Critical strengths

This paper tackles an important problem: LLM training is very memory-costly, and many researchers or organizations cannot afford such training costs. So the research question is valuable.

The analysis in this paper makes sense. It includes both theoretical analysis and experimental results; the claims in this paper are supported by both kinds of results.

Critical weaknesses

The first weakness is that I think there is a gap between the analysis performance and the experimental performance. The theoretical analysis may only indicate the influence of the change to certain levels on the changes of the final layer. What the relationship is between the changes to the final layer and the final performance in experiments still remains unclear.

The second thing is that I may need some experimental results to demonstrate that it's necessary to select continuous layers rather than scattered layers in the model, or maybe just select layers with top potential values.

Questions for authors

Does this proposed method also save other computation costs besides the memory cost, such as training time or training FLOPs?

The spectral norm only picks the largest singular value. How about using the Frobenius norm, or the sum of the top singular values?

Have you done comprehensive parameter tuning for all the methods, like learning rate tuning for all the methods, including the baseline?

Why should the motivation of this paper focus on reinforcement learning rather than general model training? I know the SFT experiment is also included here, but I think the analysis might not limit the methods to reinforcement learning. This seems like a more general method.

3 comments
AuthorsOpenReview ↗

Reply to Reviewer 3z3H (part1)

We sincerely thank the reviewer for the thoughtful comments and for recognizing the importance of the problem. Below, we address each point and welcome further discussion.


W1. Relationship between changes in the final representation and model performance.

We agree with the reviewer that the ability to change the final representation cannot be directly equated with performance gains. This is precisely why we term this quantity ''layer potential'' rather than ''layer effectiveness/contribution''. It quantifies how sensitive the final representation is to parameter changes in a given layer during fine-tunin, without presuppose any causal relationship between potential and performance.

The relationship between potential and fine-tuning performance is validated by experiments, mostly through these two:

(1) Single-layer training (Section 4.2). Compare with the single-layer training performance, we find that detrended potential match with each layer’s contribution to the full-parameter gain.

(2) Complementary interval ablation (Appendix A.3). Training Fulcrum's high-potential interval (1/3 parameters), outperforms training the remaining 2/3 low-potential parameters.

Since the submission, we have further extended more experiments data:

Table R1. More single-layer Comparison on Qwen3-8B-Base.

GroupMetric0234579101213141618192228
Fulcrumrlr_l−1.60−7.97−13.56−18.62−18.66−12.78−7.45−10.203.159.8822.8150.3739.1919.66−2.24−10.97
Rank111632343530152310951271225
Single-layerAvg. Perf.51.9051.4352.0452.5152.5252.9052.7953.3754.3354.4755.1855.0454.5654.1252.9751.24
C(l)C(l)0.310.270.330.370.370.410.400.460.550.570.640.620.580.530.420.25

It proves that the detrended potential predicted before training correlates strongly with measured performance (Spearman ρ=0.75,p<0.001ρ = 0.75, p < 0.001).

Table R2. More interval training on Qwen3-1.7B-Base. Results are reported across five seeds.

LayersMathCodeReasoningLanguageAvg.
[0,9]48.31 ± 0.9139.53 ± 3.0431.95 ± 2.2742.12 ± 1.1442.04 ± 0.61
[4,13]49.11 ± 0.3838.88 ± 1.8731.90 ± 1.2341.48 ± 0.8142.09 ± 0.55
[6,15] (Fulcrum)49.92 ± 0.3040.33 ± 1.4532.99 ± 1.1641.72 ± 0.9042.97 ± 0.40
[8,17]49.82 ± 0.8140.62 ± 1.2132.05 ± 1.4940.77 ± 0.7142.61 ± 0.52
[9,18] (Middle)49.79 ± 0.6839.31 ± 1.6430.91 ± 1.0342.14 ± 1.2942.38 ± 0.36
[18,27]45.86 ± 0.4937.31 ± 1.2830.91 ± 1.4838.31 ± 1.3039.65 ± 0.59
Full49.57 ± 0.5940.62 ± 1.9132.17 ± 0.5443.50 ± 0.9343.09 ± 0.33
Base37.30 ± 0.7533.22 ± 1.6419.88 ± 0.8628.60 ± 1.3431.26 ± 0.35

These results consistently show that layers with higher potential lead to better fine-tuning performance.

Understanding how changes in the final representation translate into task performance is beyond the scope of this work and remains an interesting direction for future interpretability research. We will clarify this in the revision.

W2. Contiguous interval vs. scattered top-potential layers.

On Qwen3-4B and Qwen3-8B, the ten layers with the highest detrended potential are the selected contiguous intervals ([13, 22] and [12, 21]; see Tables A2–A3), but differ on Qwen3-1.7B, where we additionally trained the scattered top-10 layers:

Table R3. Contiguous interval vs. scattered layers.

LayersCriterionAvg.
[6–9,12–15,25,26]Scattered top-1039.92 ± 0.33
[6,15]Contiguous (Fulcrum)42.97 ± 0.40

Beyond accuracy, a contiguous interval is also more practical in the real training.

AuthorsOpenReview ↗

Reply to Reviewer 3z3H (part2)

Q1. Benefit beyond memory.

Beyond memory savings, Fulcrum also reduces FLOPs and training time, yielding 35% fewer update FLOPs on Qwen3-8B and reduces training time and FLOPs on SDXL by 34.0% and 39.5%, respectively. We agree that the end-to-end speedup is not substantial in GRPO because rollout generation dominates the total runtime. We expect larger gains in RL settings where policy updates account for a greater share of computation, potentially including PPO. We are actively testing PPO.

We would also like to emphasize that our primary goal is not to push benchmark numbers or to improve efficiency, but to make RL post-training accessible to a broader community. In practice, GPU memory capacity is often the primary barrier, determining whether RL training is feasible at all. Fulcrum can lower the hardware tier without changing the model structure. For example, full-parameter GRPO on Qwen3-8B requires either a data-center GPU with more than 96 GB (e.g., H200 NVL, 33.7k USD) or two 96 GB GPUs (~32k USD), whereas Fulcrum with simple optimizations fits on a single 80 GB GPU (A100 80GB, ~7–15k USD), reducing hardware cost by ~3×. Similarly, Qwen3-4B with Fulcrum fits on a single consumer RTX 5090 (32 GB, 1,999 USD), instead of the 16,000 USD RTX PRO 6000 required for standard full-parameter training, an ~8× reduction.

Fulcrum takes only seconds before training to provide this reference. The tool is included in the supplementary material, and we warmly welcome the reviewers to try it.

Q2. More matrix/spectral measures.

We did analyze and test several metrics after modeling the Jacobian, including the Frobenius norm and the sum of the top-kk singular values, based on their empirical correlation with fine-tuning performance.

Table R4. Comparison of Jacobian metrics.

MetricLayer ρ (8B)Interval ρ (1.7B)1.7B4B8B
Spectral norm0.75*0.75*[6,15][13,22][12,21]
Frobenius0.190.53[6,15][15,24][14,23]
Nuclear−0.58*0.60[12,21][22,31][20,29]
Spectral radius−0.300.62[7,16][17,26][17,26]
Top-10−0.300.62[7,16][17,26][17,26]

The spectral norm exhibits the strongest correlation with fine-tuning performance and consistently identifies the best-performing intervals.

Q3. Optimization of learning rate.

We tuned the learning rate before the main experiments. Representative results are shown below. Both Fulcrum and full-parameter training performs best with a learning rate of 1 × 10⁻⁶. We therefore use it for all methods in the main experiments (Appendix A.1). All baselines were trained with reasonable hyperparameter settings, without signs of training collapse or under-training.

Table R5. Learning-rate sweep on Qwen3-1.7B-Base (greedy decoding, average accuracy).

LRFulcrumFull FT
3 × 10⁻⁷42.4942.51
5 × 10⁻⁷42.26–
1 × 10⁻⁶45.2244.86
2 × 10⁻⁶43.52–
3 × 10⁻⁶43.08–

We are currently running additional learning-rate experiments, including for the baselines, and will update this thread with the results once they are available.

Q4: Why RL rather than general training?

This work originates from our own needs: make RL post-training cheaper. When surveying existing layer-selection strategies, we found that they do not transfer well to RL. Its optimization signal is defined jointly by policy-generated trajectories, rewards, and advantage estimates, and it drifts as the policy updates. As a result, the loss-based criteria used in supervised settings (gradients, Fisher information, influence functions) cannot be defined in a stable way before training begins.

Fulcrum avoids this by measuring a structural property of the base model that does not depend on any objective. This is also why, as discussed in our response to W1, Fulcrum quantifies each layer’s influence only up to the final representation: beyond that point, the reward and loss in RL are complex and hard to quantify before training.

We agree with the reviewer that it is a general method. In the revision, we will explain Fulcrum as a general layer-selection method for post-training, with RL as its primary motivating use case.

We hope these clarifications and additional results address the reviewer’s concerns, and we would be grateful if the reviewer would consider re-evaluating our work. We are happy to provide any further details.


Reference

*[1] Zijian Zhang, Rizhen Hu, Athanasios Glentis, Dawei Li, Chung-Yiu Yau, Hongzhou Lin, and Mingyi Hong. Is one layer enough? Training A single transformer layer can match full-parameter RL training. arXiv preprint arXiv:2607.01232, 2026. *

As the rebuttal deadline is coming, we would appreciate hearing your thoughts on our responses. We believe our work is really valuable for the community and hope that further discussion can help resolve your concerns and give you reason to reconsider rating. If any questions remain, we would be happy to clarify and discuss them further. The feedback and discussion would help us improve the quality of our work. We look forward to hearing from you.

Comments on the paper

1 comment

Dear Reviewers,

The authors have posted responses. As the discussion period is ending soon, could you please take a look and reply shortly, or at least acknowledge that you have read it?

Thank you!

AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.