acceptodds
Withdrawn from ICLR 2027

Adaptive Evolutionary Pareto Search with Online Transferability Estimation for Resource-Aware LLM Adaptation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Adapting large language models to changing tasks and deployment requirements requires balancing accuracy, latency, and memory under limited evaluation resources. Although parameter-efficient fine-tuning (PEFT) reduces adaptation cost, selecting which layers to adapt together with rank and dropout creates a combinatorial multi-objective search problem. Under a fixed evaluation budget, reusing configurations from previous tasks can further waste evaluations when their usefulness on the new target is unknown. We present Adaptive Evolutionary Pareto Search with Online Transferability Estimation (AEPS-OT), a framework that allocates a fixed evaluation budget across multiple proposal mechanisms and historical configurations. A controller uses the current search state and objective weights to select among genetic algorithm (GA), differential evolution (DE), particle swarm optimization (PSO), and estimation of distribution algorithm (EDA) proposals, while a Pareto archive preserves search quality. To reuse experience from prior tasks, AEPS-OT first probes historical configurations on the target task and estimates transferability separately for accuracy, latency, and memory from rank agreement, stability, and performance gains measured on the target task. The resulting scores, combined using the current deployment preference, determine bounded source quotas, with all probes and transfers charged to the same 1,000-slot target budget. On Qwen2-0.5B and the Internet Movie Database (IMDb) dataset, five-seed independent audits give a mean hypervolume of 0.193424, higher than all fourteen external controls, including Bayesian optimization, fixed evolutionary operators, policy-learning baselines, classical multi-objective evolutionary algorithms and randomized search. In terms of mean audited hypervolume, AEPS-OT outperforms the base Adaptive Evolutionary Pareto Search (AEPS) by 4.4%. Search using target probes without subsequently adding historical configurations reaches 0.192410, indicating that evaluating historical configurations on the target task accounts for most of the observed transfer gain. These results support evaluating historical configurations on the current task before reusing them in resource-aware PEFT search under the evaluated single-model setting with one target task and four historical source tasks.

Primary area:
optimization
Keywords:
parameter-efficient fine-tuning, multi-objective optimization, adaptive evolutionary search, Pareto optimization, online transferability estimation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_GEeaOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how to select a fine-tuning configuration for large language models that balances accuracy, latency, and memory overhead, within a limited number of evaluation trials. AEPS-OT learns to select evolutionary operators and leverages configurations from past tasks to assist in the search process. Experimental results show that this method achieves the best average performance, although ablation studies indicate that most of the gains come from initializing the search with historical configurations.

Critical strengths

The experimental section in the paper provides a detailed validation of the results. Instead of directly using the best results obtained during the search process, the authors retrain the selected configurations and evaluate them on an independent test set, which makes the experimental results more credible.

Critical weaknesses
  1. The improvement brought by the new mechanism is still not clear. As shown in Table 2, using only historical configurations for initialization has already achieved most of the gains, and the additional benefits from subsequent transfer are relatively small. It's unclear whether it's necessary to introduce a scoring mechanism that weights by relevance, stability, and preference. Moreover, all experiments use the same number of transfer slots, which resembles simply selecting configurations under a fixed budget. This still doesn't demonstrate that the method can effectively determine how much transfer budget should be allocated. The authors need to conduct a fair comparison with simpler strategies such as randomly selecting historical sources or choosing configurations based solely on performance on the target task.

  2. There are many comparison methods, but the comparisons directly relevant to this paper remain insufficient. The experiments mainly compare general optimization algorithms, while PEFT configuration search methods such as AutoPEFT and BIPEFT mentioned in the paper were not included in the comparison. It's still difficult to assess the actual improvements of this work over existing approaches. The soundness of the resource-constrained experiments is also limited. All configurations checked in Appendix D meet the latency and memory requirements, and the final evaluation mainly focuses on accuracy, without demonstrating how trade-offs should be made under resource constraints. Simply adjusting the scale of hyper-volume calculation after the experiments end does not indicate that the method can adapt its search strategy according to deployment requirements.

Questions for authors
  1. Why are there no methods directly compared to AutoPEFT, BIPEFT, etc.? Could an experiment be added after reasonably considering the search space and training cost differences?

  2. If using the same initial configuration and transfer, does the complete scoring mechanism perform better than randomly selecting sources or selecting only based on performance on the target task? Are the improvements consistent across different random seeds?

  3. If delay or memory limits are tightened, making some high-accuracy configurations unavailable, does the proposed method still have advantages? Compared to search with fixed preferences, which different configurations would it select?

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

This paper proposes AEPS-OT framework for searching LoRA fine-tuning configurations under a limited evaluation budget. The framework combines an adaptive evolutionary search strategy with an online transferability estimation module. The adaptive search strategy selects suitable search operators and modes according to the current search state, while the transferability estimation module evaluates whether configurations from previous tasks are useful for a new target task. The proposed method aims to improve the Pareto trade-off among task accuracy, inference latency, and memory usage. Experiments are conducted on Qwen2-0.5B with IMDb as the target task, and the results show that AEPS-OT achieves the highest mean hypervolume among all the compared methods.

Critical strengths
  1. This paper addresses a practically important problem of resource-aware LoRA configuration search. It jointly optimizes task accuracy, inference latency, and memory usage.

  2. AEPS-OT combines adaptive evolutionary search with online transferability estimation. it provides a reasonable mechanism for selecting search strategies and reusing historical configurations.

  3. Ablation studies are conducted to verify the contributions of the main components.

Critical weaknesses
  1. This paper combines adaptive evolutionary search with online transferability estimation, but it is still unclear what fundamentally distinguishes this method from existing adaptive search and transfer-based optimization approaches.

  2. The PPO-based operator selection controller is only compared with uniform random selection. Since beating a random baseline alone cannot fully justify the additional complexity brought by reinforcement learning, the necessity and contribution of this component are not sufficiently supported.

  3. The paper reports negative source objective correlations and excludes some sources from transfer, but it lacks systematic experiments under strongly mismatched source tasks, so the robustness of the estimator to negative transfer is still unclear.

  4. The evaluation is only conducted on a single target task and a single model, which restricts the assessment of generalization across different tasks and model scales.

  5. AEPS-OT relies on pre-constructed source task memories, which is an upfront cost not required by non-transfer baselines. Since this cost is excluded from the budget-matched comparison, it is unclear whether the reported advantage still holds when the total computational budget is fully matched.

Questions for authors
  1. What are the fundamental differences between AEPS-OT and existing adaptive search methods and transfer-based optimization techniques?

  2. Can the authors compare the PPO-based controller with simpler adaptive selection strategies to better justify its additional complexity?

  3. How robust is the transferability estimator when the source and target tasks are substantially mismatched?

  4. Can the authors provide evidence that the method generalizes to a broader set of target tasks and model scales?

  5. Does AEPS-OT still show its reported advantage when the cost of source memory construction is included in a matched total computational budget?

Reviewer_4SjYOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary
  1. Framing the problem: Adapting a large language model with LoRA requires choosing which layers to adapt, plus the rank and dropout. Each candidate must be fine-tuned and measured for accuracy, latency, and memory, and the evaluation budget is limited. It is not known in advance whether configurations from earlier tasks will transfer to a new task. The authors formulate this as a budget-constrained multi-objective Pareto search problem.
  2. Proposing AEPS-OT: The method builds on Adaptive Evolutionary Pareto Search. A reinforcement learning controller chooses among four proposal mechanisms: genetic algorithm, differential evolution, particle swarm optimization, and estimation of distribution algorithm. The method adds online transferability estimation. A few historical configurations and random anchors are evaluated on the target task, and each source task is judged separately for accuracy, latency, and memory. Bounded slots for historical configurations are then allocated according to the current deployment preference. All probes and transfers count against the same evaluation budget.
  3. Experiments and conclusion: The experiments use Qwen2-0.5B with LoRA on IMDb sentiment classification, with BoolQ, MNLI, COPA, and RTE as source tasks, over five random seeds. AEPS-OT's mean audited hypervolume is 0.193424. This is higher than all 14 external control methods and about 4.4% above base AEPS. An ablation shows that most of the improvement comes from evaluating historical configurations on the target task. The authors conclude that such configurations should be evaluated on the current task before they are reused. This conclusion rests on a single-model, single-target-task setting.
Critical strengths

The paper does not optimize a single metric. It jointly optimizes validation accuracy, P95 latency at batch size one, and the inference memory working set. Latency and memory are both measured with a realistic deployment pipeline, which includes tokenization, host-to-device data transfer, and statistics from the CUDA memory allocator. The preference weights adjust automatically depending on whether the front meets its targets, so the search direction responds to specific deployment constraints. Under the constraints of latency no higher than 15 ms and memory no higher than 1600 MiB, AEPS-OT achieves the highest feasible accuracy among the four methods compared.

Critical weaknesses
  1. Lacks reliable Pareto front visualization. The figure uses only one seed. It shows only the search process, and it does not show the fronts of other methods. It also doesn't show the two-dimensional projections such as accuracy–latency or accuracy–memory.
  2. The generalization in the title does not match the scope of the experiments. The title says "LLM adaptation," but the experiments use only one 0.5B model and one binary sentiment classification task.
  3. The related work section mentions AutoPEFT and BIPEFT, yet neither is included as a baseline.
Questions for authors

See "Weaknesses" part

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

Related papers

Discussion (0)

Sign in to comment.