acceptodds
Withdrawn from ICLR 2027

Pilot: Transferring LoRA Adapters with One-Time Meta-Training

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

LoRA adapters trained on one base model may not retain their task performance when transferred to an updated version. Transferring them requires recovering their task behavior while preserving the target model's instruction-following ability. We propose , a transfer operator trained once for a source and target model pair and reused across adapters. first constructs a closed-form approximation of an adapter's effect on the target from unlabeled activations, then applies learned low-rank corrections. Matching intermediate module outputs alone does not ensure correct or complete answers, so the corrections receive feedback through the frozen target model. During supervised meta-training, these corrections are optimized for accurate, complete task responses, with a KL penalty that limits changes to the target model's behavior on general instructions. Once trained, the operator transfers held-out adapters from the same task family between models with matching architectures without further labels or gradient computation. For different architectures, -X constructs a transfer basis from source-generated answers using a target backward pass, avoiding iterative fine-tuning for each adapter. Experiments within and across model families show that transferred adapters achieve task accuracy competitive with supervised target adaptation while largely retaining the target model's instruction-following ability.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
LoRA adapter transfer, Meta-learning, Parameter-efficient fine-tuning

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_HUP6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PILOT addresses LoRA transfer from a source base model to a target model while retaining the target model’s instruction-following behavior. For matching architectures, it fits an adapter-specific, preservation-aware low-rank basis from unlabeled task and general-instruction activations, then applies shared low-rank input/output operators and direction-wise gains. These corrections are meta-trained once per fixed model pair and task family using answer-level supervision through a frozen target and a KL retention term; a new source adapter then needs no task labels or target-model gradients at transfer time. PILOT-X replaces module alignment with source-generated answers and a per-adapter target backward sweep for heterogeneous architectures.

Critical strengths

1.A coherent and practically relevant transfer design. The closed-form basis has a clear rank-constrained objective that explicitly penalizes interference on general-instruction activations. For the fixed model pair and task family, it addresses a real setting in which new adapters have no available ground-truth labels.

2.Thorough experimental controls and candid self-scrutiny. The authors proactively conduct several experiments that could weaken their own conclusions and report the results honestly. In Table 1, with 4,096 meta-training examples, PILOT achieves mean seven-task accuracies of 79.02 / 88.49 / 77.95 across the three same-architecture pairs, with IFEval scores of 65.13 / 80.96 / 68.15, close to the unmodified targets. The paper also reports three-seed variability, held-out adapter tests, ablations, and a shared-correction control that sometimes outperforms PILOT in task accuracy.

Critical weaknesses

The motivation is somewhat overstated. The introduction motivates transferring large collections of adapters when their original training data are unavailable. However, PILOT’s one-time stage still requires 4,096 labeled examples from the same task family (equal to the training-set size of the LoRA reference) and must be repeated for each model pair and task family. If labels from that task family are available, directly training a target-side LoRA takes only 5–8 minutes according to the appendix; PILOT does not save GPU time at the evaluated scales.

In Appendix F, Table 12, the shared target-only LoRA achieves 80.98 versus PILOT’s 79.80 on Qwen3.5-2B and 80.67 versus 78.84 on Llama. PILOT’s remaining advantage in this comparison is mainly lower gen-KL; its IFEval is higher on Qwen3.5-2B, but not on Llama. Its more distinctive use case is therefore when the available labels do not cover the incoming adapter’s task: in the right block of Table 12, the closed-form transfer achieves 79.21, versus 76.72 for the target-only LoRA. Yet this advantage largely comes from the closed-form initialization itself; the learned operators add only 0.50 points on the group-B tasks.

Questions for authors

The abstract and introduction motivate transfer to an "updated version" of a base model, but all same-architecture pairs are Base→Instruct models sharing one pretraining run, and all other pairs cross model families. Could you evaluate at least one genuine same-family version upgrade (e.g., an older to a newer release within Qwen or Llama)?

Reviewer_vsW8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

PILOT transfers LoRA adapters between base models while trying to preserve the target's instruction following. It combines a closed-form initialization from unlabeled activations with low-rank corrections meta-trained for a model pair and task family. The neat thing is that subsequent same-architecture transfers require no ground-truth task labels or gradients. PILOT-X extends the approach across architectures using source-generated answers and a backward sweep-oass. Experiments cover 7 commonsense benchmarks, a few model pairs, and adapters with different ranks and training configurations.

Critical strengths
  1. The paper addresses a useful problem: reusing task adapters after a base-model change while retaining the target's instruction-following capabilities. I find learning a correction once and reusing it across adapters is an interesting approach, and the combo with the closed-form initialization makes sense and shows coherence
  2. Experiments include multiple model pairs, 3 seeds, and tests of different adapter ranks and training configurations; that diversity is quite decent. The different interventions and correction controls help explain what the learned correction actually contributes. I also appreciate that the authors report results that limit their claims, including the competitive but simpler controls and the small gains on unseen task groups
Critical weaknesses

Task coverage is way too narrow to establish the broad usefulness and corresponding claims of the proposed method convincingly: The 7 task-accuracy benchmarks are all commonsense QA tasks, and the main held-out adapters differ from the meta-training adapters only by seed! Sure, additional rank and training-configuration experiments broaden the evaluation, but they don't improve or address task diversity. The disjoint-task-group experiment is relevant, yet the learned correction adds only 0.3-0.6 percentage points over the closed-form initialization on the new group's tasks. Task content and answer format also change together in that split. For a reusable transfer method, I would need to see stronger evidence on a substantially different task family before recommending acceptance... evaluating math or code as retention metrics does not show that adapters for those tasks can be transferred successfully

The results leave the practical advantage of the full method over simpler alternatives very unclear. When meta-training labels cover the adapter tasks, the shared target-only LoRA in Table 12 achieves higher task accuracy than PILOT. It has lower generation KL on both evaluated models and better IFEval on Qwen, but on Llama the simpler control achieves both higher accuracy and similar IFEval. Where does the discrepapncy come from; any idea? Also, when the task groups differ, transferring the source adapter helps, but most of the benefit is already present in the closed-form initialization. These controls leave limited evidence that the learned correction provides a substantial advantage in the setting where task-specific information from the incoming adapter is needed. The reported per-adapter basis construction also takes more GPU time than target LoRA training, so the practical benefit rests on avoiding labels and iterative optimization at transfer time - a real benefit, but results are not convincing and broad enough imo.

Questions for authors
  1. Have you evaluated task performance after transferring math, code, or open-ended generation adapters, beyond the capability-retention tests? Such results more informative about the method's usefulness
  2. Given Table 12 and the task-group experiment, when does the learned correction provide a substantial benefit over both the closed-form initialization and the shared target-only LoRA?
Reviewer_tPBEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces PILOT, a method for transferring LoRA adapters between models. The method consists of two steps. The first step uses a closed-form regression to fit the adapter's effect on a target model based on unlabeled activations with a penalty that keeps the update’s effect on general instructions small. The second step uses a model, trained once per pair of source-target models and family of tasks with two source adapters and using labeled data, to correct the basis from the first step. The paper also introduces PILOT-X to handle different architectures. It replaces the regression with gradients from the target model computed with the answers from the source model with the adapter. Experiments on seven commonsense tasks with Qwen3.5, Llama-3.2 and Gemma-4 show accuracy comparable to a LoRA trained on the target, while keeping IFEval close to the original model.

Critical strengths

The closed-form basis is simple and effective on its own, recovering most of the accuracy of the adapter without labels and with little IFEval degradation The measurement of retention next to the accuracy is a good way to evaluate the problem and shows that the target LoRA loses about half of the target’s IFEval score Broad set of baselines

Critical weaknesses

The paper’s controls suggest that the meta-model learns the task and not the transfer. From Table 12 we can see that a target LoRA trained with the same KL objective is more accurate than PILOT when the meta-training labels cover the adapter’s tasks. Table 10 shows that the meta-modelimproves the meta-training tasks more than the adapter’s own tasks. The authors themselves note that when the meta-training labels cover the adapter’s tasks, a target LoRA with the same objective is simpler and more accurate. Also, Table 1 compares PILOT against LoRA without KL which favors PILOT, since PILOT is trained with the KL penalty and the baseline is not. This matters because the main contribution from the paper is a transfer method that is trained once and reused. If the correction mostly encodes task knowledge, this claim would not hold.

Questions for authors

When should PILOT be used instead of training a target LoRA with the same KL penalty? The meta-training needs as many task labels as training a LoRA and takes more time to build the basis than it takes to train the LoRA. (This could change my assessment.) How does PILOT’s accuracy change with the number of adapters used for meta-training? (Clarification.) Could you repeat the Table 10 experiment with more seeds and with adapters trained on a single task? (Clarification.) Could you report the KL-regularized target LoRA of Table 12, for all pairs, in the main tables? (This could change my assessment.) Could you report significance tests for PILOT vs. LoRA? Some of the gaps are small and might not be significant. (Clarification.) The evaluation covers only multiple-choice tasks. How would the objective be adapted to other types of tasks such as open-ended generation? (Future work.)

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

Related papers

Discussion (0)

Sign in to comment.