acceptodds
Under review as a conference paper at ICLR 2027

CAST-LoRA: Cross-Attentive Soft-Transition MoE-LoRA for Multi-Task Adaptation

Abstract

Parameter-efficient multi-task fine-tuning is important for deploying large language models (LLMs) across diverse tasks under limited training, storage, and serving budgets. However, existing Low-Rank Adaptation (LoRA) methods for multi-task adaptation still have the following limitations that require further improvement in performance: (1) training a single adapter across all tasks induces severe task interference in multi-task scenarios. (2) routing over a flat expert pool neglects dedicated shared capacity, failing to capture common knowledge across tasks. To address these problems, this paper proposes CAST-LoRA, a novel multi-task Mixture-of-Experts Low-Rank Adaptation (MoE-LoRA) framework. By clustering training data to guide early routing and then annealing cluster-guided assignments into input-dependent soft routing, CAST-LoRA stabilizes task-to-expert specialization and mitigates task interference. CAST-LoRA further introduces role-separated task-specific, entropy-aligned shared, and floor-preserved shared experts, providing dedicated capacity for routing uncertainty and reusable cross-task knowledge. Additionally, a cross-attentive task encoder extracts compact task representations with rather than attention interactions, where is the source-sequence length and is the hidden dimension. Extensive experiments with 10 baselines on 7 benchmarks demonstrate that CAST-LoRA achieves the optimal average accuracy, reduces gradient conflict to an optimal level of 1.09%, cuts task-encoder latency by 86.9%, cuts task-encoder peak memory by 23.8%, lowers expert-parameter redundancy by 26.8%, and increases effective expert directions by 26.8%.

Primary area:
foundation or frontier models, including LLMs
Keywords:
low-rank adaptation (LoRA), mixture-of-experts (MoE), expert routing, multi-task 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 2 reviews, above 27% of the others.
Reviewer_54L6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CAST‑LoRA, a multi‑task MoE‑LoRA framework for parameter‑efficient multi‑task adaptation. The core designs include annealed cluster‑guided soft routing, role‑separated expert bank, and a single‑query cross‑attentive task encoder. The authors conduct extensive experiments across seven benchmarks against ten baselines.

Critical strengths
  1. Annealed cluster‑to‑soft routing is a practically valuable design. By only leveraging semantic clusters as temporary training guidance and gradually transferring control toward input‑dependent soft routing, the method avoids requiring explicit task labels during inference. This circumvents a common practical limitation of many existing task‑conditioned MoE‑LoRA approaches.
  2. Overall empirical performance is competitive. On seven multi‑task benchmarks, CAST‑LoRA achieves the best average accuracy and substantially lowers gradient conflict compared with strong MoE‑LoRA baselines, demonstrating its effectiveness for mitigating multi‑task interference. Comprehensive ablation and causal intervention studies are provided to validate different components.
  3. The cross‑attentive task encoder brings favorable complexity reduction. Replacing full self‑attention based task encoder with a single‑query cross‑attention design reduces attention complexity from O(L2d)O(L^2d) to O(Ld)O(Ld), yielding lower latency and peak memory for the encoder module, especially for long input sequences.
Critical weaknesses
  1. Concerns about parameter‑efficiency and unfair comparison against PEFT baselines. The paper positions CAST‑LoRA as a parameter‑efficient fine‑tuning method. The total trainable parameters reach 60.34M (3.76% of the Qwen2‑1.5B backbone). Compared with vanilla LoRA, LoRA+, MixLoRA and LoRAMoE, which only maintain LoRA‑expert parameters without such a heavy standalone encoder, the parameter budget is significantly higher.
  2. Training‑side time efficiency is insufficiently reported. The paper thoroughly reports inference latency of the isolated task encoder, but largely omits training overhead for the full model. CAST‑LoRA introduces multiple computation overheads: annealing interpolation between cluster assignments and learned routing, auxiliary regularization losses for shared experts, backward pass of the heavy cross‑attentive encoder, and offline prompt clustering preprocessing.
  3. Unexplained mismatch: seven evaluation benchmarks while selecting 6 task‑specific experts. Empirical ablation shows that setting Kt=6K_t=6 obtains peak average performance, while increasing to 7 task‑specific experts degrades performance. Given there are seven downstream evaluation benchmarks, it is natural to wonder why adding one extra task‑specific expert does not bring benefits. One plausible hypothesis is that some benchmarks contain highly semantically similar tasks, leading to redundant expert capacity and higher risk of expert under‑specialization. The paper does not provide further analysis on this phenomenon. I encourage the authors to compute pairwise task / benchmark semantic similarities, analyze the behavior when Kt=7K_t=7, and explain the performance drop.
Questions for authors

See Weakness

3 comments

Part Ⅰ

Dear Reviewer 54L6,

Thank you for recognizing the practical value of annealed routing and for raising important questions about parameter budget, end-to-end cost, and expert granularity. CAST-LoRA targets two specific limitations of multi-task adaptation: conflicting task updates and insufficiently differentiated shared versus task-specific knowledge. Its contribution is supported by 1.09% gradient conflict versus 5.07% for MixLoRA, 26.8% lower expert-update redundancy at matched capacity, and the highest seven-benchmark mean accuracy (64.95%) (Tables 1–2, Fig. 2).

  • W1. “Parameter efficiency and unfair comparison with PEFT baselines.”

Against the same-budget HMoRA-LW, CAST-LoRA is clearly better. Against lighter MixLoRA, the larger budget buys a large reduction in interference and redundancy. CAST-LoRA and HMoRA-LW [1] both use 60.34M trainable parameters, but CAST-LoRA improves Avg from 62.56 to 64.95 (+2.39). Thus, the gain is not merely from a larger budget.

MethodTrainable paramsMean Accuracy
HMoRA-LW [1]60.34M62.56
CAST-LoRA60.34M64.95

Compared with MixLoRA [2], CAST-LoRA uses a larger budget, but it reduces gradient conflict from 5.07% to 1.09% and expert redundancy by 26.8%. More importantly, Flat8-TOP4 holds expert count, trainable budget, and four active branches fixed: CAST-LoRA reduces pairwise update cosine from 0.414 to 0.303 (−26.8%), and expert-output cosine from 0.467 to 0.343 (−26.5%) (Fig. 2), while improving the mean accuracy from 64.33 to 64.59. This is the intended trade-off: more capacity is used to separate task-specific and shared directions, not to simply increase the adapter size.

  • W2. “Training-side efficiency is insufficiently reported.”

The encoder is substantially cheaper, and the complete model remains deployable without task labels; the end-to-end training-time trade-off is real and is now quantified. On a single RTX PRO 6000 Blackwell Max-Q (BF16), with identical profiling workloads:

Whole-model metricLoRAMoE[4]HotMoE[3]CAST-LoRA
Training time (s/optimizer update)1.6932.8803.400
Peak allocated training memory (GiB)15.0825.588.79
Multiple-choice latency (ms/batch, B=4)70.21117.0583.95
Generation latency (s/request, 1,024 input + 32 output tokens)1.8252.8152.143

CAST-LoRA is not the fastest end-to-end, but it has lower peak training memory and lower inference latency than HoTMoE while achieving better accuracy and much lower conflict. The isolated encoder latency drops by 86.9%. End-to-end, compared with HotMoE [3], CAST-LoRA uses 8.79 vs. 25.58 GiB peak training memory, 83.95 vs. 117.05 ms MC latency, and 2.143 vs. 2.815 s generation latency. Training latency is 3.400 vs. 2.880 s/update. The encoder’s benefit is largest for long prompts. The end-to-end training cost is the price of broader adaptation and routing. We will add these measurements.

MiniLM embeddings and spherical k-means over 1,820 training-task centroids are a one-time offline preprocessing stage; their complete wall-clock cost was not recorded and is not included above. No cluster labels or MiniLM inference are required at serving time. Thus, the intended trade-off is improved multi-task separation and shared-knowledge utilization at the cost of a manageable increase in training latency.

Part Ⅱ

  • W3. “Six task experts for seven benchmarks; why does a seventh expert not help?”

The six experts partition 1,820 training tasks, not seven evaluation datasets; adding a seventh expert produces less balanced specialization without an accuracy gain. In a matched 10,000-update comparison:

Diagnostic6 task experts7 task experts
Seven-benchmark accuracy64.9564.18
Task-expert gate-mass CV (lower is more balanced)0.06150.2112
Mean shared routing mass29.732%29.729%

Six experts are selected on validation data and are better than seven; the seventh expert increases routing imbalance without improving accuracy. At 10,000 updates, six experts give 64.95 vs. 64.18 for seven experts. The load CV of task-expert routing increases from 0.0615 to 0.2112 (3.44×), while shared weight remains almost unchanged (29.732% vs. 29.729%). Several benchmarks are semantically close: ARC-E/ARC-C centroid similarity is 0.985, and MMLU/MMLU-Pro is 0.883. Expert count should match training-task structure, not the number of evaluation benchmark names. The seventh expert adds 7.6% parameters but lowers accuracy by 0.77.

In summary, CAST-LoRA’s principal benefit is measurable mitigation of task interference and more effective allocation of specialized/shared capacity, with competitive accuracy at a clearly reported parameter/latency cost.

References

[1] HMoRA: Making LLMs More Effective with Hierarchical Mixture of LoRA Experts. ICLR, 2025.

[2] MixLoRA: Enhancing Large Language Models Fine-Tuning with LoRA-based Mixture of Experts. arXiv, 2024.

[3] Hybrid Routing for a Mixture of LoRA Experts. AAAI, 2026.

[4] LoRAMoE: Alleviating World Knowledge Forgetting in Large Language Models via MoE-Style Plugin. ACL, 2024.

Dear Reviewer, thank you for your time and valuable feedback. We are writing to gently follow up on our submitted rebuttal. We wonder if our response has addressed your concerns and clarified the issues raised.

If you have any remaining questions, please feel free to let us know—we would be more than happy to provide further clarification. If our response has resolved your concerns, we would be very grateful if you could kindly consider updating your score.

Thank you again for your constructive review.

Reviewer_U32HOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CAST-LoRA is a multi-task MoE-LoRA method for a single 1.5B backbone. A learned query cross-attends once to the task description, giving O(Ld) instead of O(L²d). Six task-specific experts and two shared experts are combined, with 1,820 FLAN tasks clustered offline into six groups whose labels form a routing scaffold annealed away during training. Entropy-alignment and floor-preservation losses act on the shared-expert gates.

Critical strengths
  1. The paper is well-written. The methods are well-presented with words easy to follow and dedicated figures. The paper is also well-organized and the introduction part gives a good start.
  2. The evaluation is very compressive. This paper includes five seeds, ten baselines including the MoE-LoRA methods, the low-rank reparameterisations and an FFT reference, plus a Flat8-TOP4 control that holds expert count, budget and active branches fixed.
  3. The task encoder is a measured design. Replacing per-token cross-attention with a single learned query cuts the encoder from 38.28 to 5.01 ms at 18.89M of 60.34M trainable parameters. The saving is measured on the isolated module rather than end-to-end, which is a fair criticism, but the claim of O(Ld) instead of O(L²d) is at least backed by a wall-clock number.
Critical weaknesses
  1. The main-table margin is small. CAST-LoRA achieves 64.95 in overall performance against MixLoRA's 64.36 and is only 0.59 points improvement on a seven-benchmark mean. Meanwhile, the matched Flat8-TOP4 version attributes only 0.26 to expert organisation. The performance enhancement is considered quite margin.
  2. The shared-expert ablations are non-monotoney. Table 5 suggests that full CAST-LoRA achieves 70.96. However, the w/o entropy alignment version achieves 70.06, while the w/o floor preservation achieves 70.48 and with both removed ony achieves 70.55. It’s quite strange that adding a single type of expert will hurt the model, since the optimization directions are the same for the two experts.
  3. CAST-LoRA is similar to existing method. The cluster scaffold uses MiniLM-L6-v2 task centroids and spherical k-means from an external encoder, so CAST-LoRA sees a task-similarity signal that MixLoRA and the others never get. However, such methods have been explored by many existing works like MoANT (CoLM2026).
Questions for authors
  1. Table 5's “w/o both” row (70.55) is better than either single removal (70.06, 70.48). Can you explain why it happens?
  2. On which split was Appendix C's configuration selected (six experts, 1,000 annealing steps, one encoder layer)? If it is the same seven benchmarks, the configuration is tuned on the evaluation set. Relatedly, does the scaffold still help if the MiniLM cluster labels are replaced by task indices or a random partition into six groups?
  3. The 86.9% latency reduction is measured on the isolated task encoder. What is the end-to-end difference in training step time and in serving latency?
  4. What makes your method different from existing work?
5 comments

Part Ⅰ

Dear Reviewer U32H,

Thank you for your careful reading and constructive comments. We first clarify the main claim of CAST-LoRA: it targets two key problems in multi-task MoE-LoRA—task interference and shared-knowledge extraction—and improves both while achieving the best average accuracy. In particular, CAST-LoRA reduces gradient conflict to 1.09% vs. 5.07% for MixLoRA (78.5% lower) and reduces expert-parameter redundancy by 26.8% relative to a matched flat expert pool. With the same 60.34M trainable budget as HMoRA-LW [1], CAST-LoRA improves Avg from 62.56 to 64.95 (+2.39). We will revise the unclear descriptions accordingly.

  • W1(a). “The improvement over MixLoRA is small.”

The gain is competitive and consistent for this line of work, and its significance is clearer when combined with interference and redundancy metrics. CAST-LoRA achieves 64.95 vs. 64.36 over MixLoRA (+0.59) and ranks first on five of seven benchmarks. More importantly, it reduces gradient conflict from 5.07% to 1.09% and expert redundancy by 26.8% at matched expert organization. The gain is also not merely from a larger budget: against HMoRA-LW [1], which has the same 60.34M trainable parameters, CAST-LoRA improves Avg by +2.39 (64.95 vs. 62.56).

Reported gains in recent multi-task LoRA work are of a similar scale:

StudyBaseline AccuracyMethod AccuracyGain
CAST-LoRA(Ours), Qwen2-1.5B64.3664.95+0.59
MoANT [2], LLaMA-3 8B83.8084.01+0.21
MoANT [2], LLaMA-2 13B83.8184.37+0.56
SAMoRA [3], Llama-3.1 8B86.7787.64+0.87
SAMoRA [3], Qwen3 8B90.9891.71+0.73
SMoRA [4], Llama-2 13B60.461.1+0.70

Thus, +0.59 is within the normal range, while the main advance is the joint improvement in accuracy, task interference, and expert redundancy.

  • W1(b). “Flat8-TOP4 is only 0.26 points behind.”

Flat8-TOP4 reaches comparable accuracy but learns much more redundant experts. CAST-LoRA achieves comparable accuracy with clearer specialization. On the comparison suite, CAST-LoRA is 64.59 vs. 64.33 for Flat8-TOP4. Under matched experts, trainable parameters, and active branches, CAST-LoRA lowers mean pairwise expert-update cosine from 0.414 to 0.303 and expert-output cosine from 0.467 to 0.343 (Figure 2). The accuracy is comparable, but the expert bank is substantially less redundant, which is exactly the design goal.

Matched metricFlat8-TOP4CAST-LoRA(Ours)
Mean expert-update cosine ↓0.4140.303 (−26.8%)
Mean expert-output cosine ↓0.4670.343 (−26.5%)
Mean Accuracy64.3364.59 (+0.26)
  • W2/Q1. “The shared-expert ablation is not monotonic.”

The full design is best; the non-monotonic Table 5 is caused by an imprecise label, not by a flawed mechanism. All four runs retain both shared experts. The “w/o floor preservation” run removed both the hard floor in Eq. (5) and the floor penalty in Eq. (6), while retaining entropy alignment. The “w/o entropy alignment” run set the floor to zero and disabled the entropy term. The full model uses the hard floor plus both regularizers and is best on all four benchmarks (70.96). We will correct the labels and Eq. (6) description.

MechanismsHard floorEntropy alignmentFloor penaltyMean accuracy
Full CAST-LoRA0.02OnOn70.96
Hard floor only0.02OffOff70.06
Entropy alignment only0OnOff70.48
Neither mechanism0OffOff70.55

These interventions test functional contribution and non-interchangeability, which is the motivation for role separation. The slightly higher unrestricted score than either single-mechanism setting (70.55 versus 70.06/70.48) indicates that either constraint alone can misallocate capacity. The joint design scores highest (70.96).

Part Ⅱ

  • W2/Q1.(Continued) “The shared-expert ablation is not monotonic.”

The floor penalty is sparse but the trained branch is causally important. During 100 post-annealing updates, the floor-preserved gate fell below 0.02 at 0.322% of locations, with RMS shortfall 1.51e-5. Yet, with total shared mass preserved, transferring the floor-preserved branch’s mass to the entropy-aligned branch lowers the five-benchmark mean from 71.83 to 68.66 (−3.17). This shows that the two shared branches are functionally distinct, even when the hinge penalty is rarely active.

  • W3/Q4. “CAST-LoRA is similar to existing methods such as MoANT.”

CAST-LoRA differs in how semantic structure is used and in the explicit separation of shared capacity. MoANT adds centroid-similarity scores to routing logits. CAST-LoRA uses cluster one-hot assignment only as an early training scaffold, then anneals it to input-dependent soft routing and removes cluster labels at inference. It further separates entropy-aligned and floor-preserved shared experts and uses a single-query task encoder.

AspectMoANTCAST-LoRA
Semantic guidanceCentroid similarity added to logitsCluster one-hot interpolated early, then annealed to learned routing
Shared capacityNo separate shared roles specifiedEntropy-aligned + floor-preserved shared experts
Task representationEmbedding-to-centroid similaritySingle-query cross-attentive summary + hybrid routing

Table 3 directly tests the transition: full model 63.23/79.2 (Avg/consistency), w/o annealing 62.58/68.8, w/o cluster scaffold 62.48/64.6.

  • Q2a. “Which split determined the configuration?”

The configuration was selected on validation data before test evaluation, not on the seven test benchmarks. Six task experts, 1,000 annealing steps, and one encoder layer each gave the highest validation mean (61.51%). We will clarify Appendix C.

ExpertsVal. AvgEncoder layersVal. AvgAnneal stepsVal. Avg
261.12161.51100061.51
360.43261.12200060.30
458.28460.43400059.75
558.15858.67500058.52
661.511658.74——
759.35————
  • Q2b. “Task-index or random grouping instead of MiniLM clusters.”

Semantic clusters are better than random grouping on all seven benchmarks. CAST-LoRA uses group labels to guide routing among six task experts early in training. FLAN contains 1,820 distinct task IDs, and these IDs alone do not specify which tasks should share an expert. Using them for routing supervision therefore requires a mapping from 1,820 tasks to six groups. So we compared two such mappings under the same routing procedure: random assignment and clustering based on MiniLM embeddings. Random grouping gives 61.96 Avg vs. 63.12 for semantic clusters, with semantic clusters better on ARC-E, ARC-C, OBQA, MMLU, MMLU-Pro, SWAG, and CSQA.

Part Ⅲ

  • Q3. “End-to-end training and server latency.”

The encoder is substantially cheaper, and the complete model remains deployable without task labels; the end-to-end training-time trade-off is real and is now quantified. On a single RTX PRO 6000 Blackwell Max-Q (BF16), with identical profiling workloads:

Whole-model metricLoRAMoE[6]HotMoE[5]CAST-LoRA
Training time (s/optimizer update)1.6932.8803.400
Peak allocated training memory (GiB)15.0825.588.79
Multiple-choice latency (ms/batch, B=4)70.21117.0583.95
Generation latency (s/request, 1,024 input + 32 output tokens)1.8252.8152.143

CAST-LoRA is not the fastest end-to-end, but it has lower peak training memory and lower inference latency than HoTMoE[5] while achieving better accuracy and much lower conflict. The isolated encoder latency drops by 86.9%. End-to-end, compared with HotMoE [5], CAST-LoRA uses 8.79 vs. 25.58 GiB peak training memory, 83.95 vs. 117.05 ms MC latency, and 2.143 vs. 2.815 s generation latency. Training latency is 3.400 vs. 2.880 s/update. The encoder’s benefit is largest for long prompts. The end-to-end training cost is the price of broader adaptation and routing. We will add these measurements.

MiniLM embeddings and spherical k-means over 1,820 training-task centroids are a one-time offline preprocessing stage; their complete wall-clock cost was not recorded and is not included above. No cluster labels or MiniLM inference are required at serving time. Thus, the intended trade-off is improved multi-task separation and shared-knowledge utilization at the cost of a manageable increase in training latency.

Thank you again for the constructive questions.

References

[1] HMoRA: Making LLMs More Effective with Hierarchical Mixture of LoRA Experts. ICLR, 2025.

[2] MoANT: Mixture-of-Rank-One-Experts with Semantic-aware Intuition for Multi-task Large Language Model Finetuning. CoLM, 2026.

[3] SAMoRA: Semantic-Aware Mixture of LoRA Experts for Task-Adaptive Learning. Findings of ACL, 2026.

[4] Each Rank Could be an Expert: Single-Ranked Mixture of Experts LoRA for Multi-task Learning. KDD, 2026.

[5] Hybrid Routing for a Mixture of LoRA Experts. AAAI, 2026.

[6] LoRAMoE: Alleviating World Knowledge Forgetting in Large Language Models via MoE-Style Plugin. ACL, 2024.

Comments on the paper

3 comments

Dear Reviewer PkMd,

Thank you for your careful reading of the shared-gate equations and for highlighting the need to distinguish configuration selection from final evaluation. We clarify the floor ablation and configuration selection below. However, the core contribution of our method is role-separated multi-task adaptation: task experts reduce interference, while two non-interchangeable shared paths preserve useful cross-task transformations. The evidence includes 1.09% versus 5.07% gradient conflict (CAST-LoRA versus MixLoRA), 26.8% less expert-update redundancy than a matched flat pool, and 64.95% mean accuracy across seven benchmarks (Tables 1–2, Fig. 2).

  • W1/Q1. “Floor regularization ablation is hard to reconcile with Eq. (5).”

Eq. (5)’s hard floor and Eq. (6)’s penalty are distinct mechanisms. The Table 5 “w/o floor preservation” run removed both, which explains the 0.48-point drop. With the hard floor active, the hinge penalty is usually zero, as you correctly note. The missing detail is that the 70.48 run removed the hard floor from the forward pass and the penalty from the objective, while retaining entropy alignment. This changes computation even when the penalty is zero, because tokens with very low shared weight can give the branch almost no influence.

MechanismsHard floorEntropy alignmentFloor penaltyMean accuracy
Full CAST-LoRA0.02OnOn70.96
Hard floor only0.02OffOff70.06
Entropy alignment only0OnOff70.48
Neither mechanism0OffOff70.55

We measured the full model over 14,285,824 valid routing locations. The floor-preserved gate fell below 0.02 at 0.322% of locations, with RMS shortfall 1.51e-5. The penalty is sparse. Its value is shown by causal interventions: preserving total shared mass but transferring the floor-preserved branch’s mass to the entropy-aligned branch lowers the mean from 71.83 to 68.66 (−3.17). Thus, the trained shared branches are not interchangeable. We will relabel Table 5 and correct Eq. (6).

The substantive result is that the learned shared paths have distinct, useful functions, rather than that the tiny hinge alone yields a 0.48-point gain. With total shared gate mass preserved, transferring the floor-preserved expert’s mass to the entropy-aligned expert reduces five-benchmark accuracy 71.83→68.66 (−3.17 points). Swapping gates or expert B matrices causes −3.04 / −3.10 points (Table 6). Removing both shared experts in a separate training run costs 0.63 points. These interventions test functional contribution and non-interchangeability, which is the motivation for role separation. The slightly higher unrestricted score than either single-mechanism setting (70.55 versus 70.06/70.48) indicates that either constraint alone can misallocate capacity. The joint design scores highest (70.96).

  • W2/Q2. “Which data selected expert count, annealing duration, and encoder count?”

These choices were selected on validation data before test evaluation, not on the seven reported test sets. The validation mean was computed over validation splits of MMLU, MMLU-Pro, ARC-C, ARC-E, OpenBookQA, CommonsenseQA, and SWAG. Six experts, 1,000 annealing steps, and one encoder layer each achieved the highest validation mean (61.51%). We will revise Appendices C and F to make this protocol explicit.

ExpertsVal. AvgEncoder layersVal. AvgAnneal stepsVal. Avg
261.12161.51100061.51
360.43261.12200060.30
458.28460.43400059.75
558.15858.67500058.52
661.511658.74——
759.35————

In summary, the corrected analysis strengthens the central conclusion: the method’s value comes from differentiating task-specific and reusable shared updates, alongside reduced measured task interference, not from overstating the effect of a near-zero hinge term. Thank you for identifying the necessary reporting corrections.

Dear Reviewer, thank you for your time and valuable feedback. We are writing to gently follow up on our submitted rebuttal. We wonder if our response has addressed your concerns and clarified the issues raised.

If you have any remaining questions, please feel free to let us know—we would be more than happy to provide further clarification. If our response has resolved your concerns, we would be very grateful if you could kindly consider updating your score.

Thank you again for your constructive review.

Thank you for the clarifications, which improve the reporting, however, the corrected floor ablation shows non-monotonic differences is small, so the evidence for the mechanism remains weak, overall I decide to maintain my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.