- Rating
- 3(+1)
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 3
Read review
Paper summary
This paper introduces ExLLM, an LLM-as-optimizer, evolutionary algorithm for large-scale discrete optimization: An LLM generates offspring per call when given one parent pair, and optionally the experience distilled from all evaluated candidates so far, with probability . At iteration , the next populations are selected from the complete candidate set based on a hybrid strategy: a half of candidates is decided based on the scalar fitness value, while the rest depends on the pareto optimality. Additionally, a feedback adapter normalizes objectives and formats constraints and expert hints for iteration.
ExLLM achieves strong results on the PMO benchmark (including single- and multi-objective optimization) when compared against strong baselines. It also generalizes to other domains.
Critical strengths
The empirical evaluation covers a broad set of baselines and tasks (both molecular design benchmark and problems from other domains), and ExLLM obtains strong results on most of the problems.
Critical weaknesses
- The diversity of the offspring generated in an LLM call is assumed, and is considered as a diversity-efficiency trade-off factor. However, from Figure 7, I see no clear pattern when varying : the ordering is not consistent across either backbones or tasks, and differences appear small. This might suggest that the diversity argument is weak, or different -s don’t really affect the results. Moreover, it might be more meaningful to report against the diversity metric rather than uniqueness, since the latter would still take non-duplicated but structurally similar molecules into account.
- It’s unclear how adapter works, and therefore how normalization is conducted. As a result, it’s also not clear to me how the history enters the experience: if the fitness score depends on the history, how stable or comparable the good / bad labels over iterations?
Questions for authors
- About the evolving experience: the evidence set is decided based on the fitness score only, while both fitness score and Pareto optimality are considered when selecting the next population. Is there a reason for using only the fitness score for the evidence set?
- Could you specify the normalization over objectives - is each objective normalized by a fixed, pre-defined range, or min/max computed from the context? If the latter, does the fitness score shift over time?
- Follow-up to 2: does the prompt for distilling experience receive numeric values, or only good/bad labels? If it’s the latter, the quantitive information would be discarded.
3 comments
Thank you for the clarifications. While the k-offspring ablation does not offer a consistent practical guideline for selecting k, I find the overall approach sound and the reported performance promising. Given the agreed revisions to the manuscript:
- clarifying how positive examples are selected for the experience update,
- clarifying how objectives are normalized across populations, and
- clarifying how the experience-generation prompt is constructed
I am happy to raise my score to weak accept.
Thank you so much for your positive assessment and constructive feedback. We greatly appreciate your recognition of the overall soundness and promising performance of ExLLM. We will incorporate the three agreed clarifications into the revised manuscript.
Response to Reviewer EBG4
We thank the reviewer for recognizing the breadth of the evaluation and the optimization results. We clarify the evidence for k-offspring generation, experience-set selection, and numerical feedback below.
Response:
The k-offspring ablation in Section 7.2.1 examines a trade-off rather than a monotonic improvement with k. Moving from k = 1 to k = 2 improves GPT-4o's Top-10 AUC from 0.644 ± 0.025 to 0.698 ± 0.038 on JNK3, and from 3.975 ± 0.088 to 4.015 ± 0.074 on the five-objective task. With Gemini-2.5-Flash, uniqueness also increases from 0.376 to 0.686 on JNK3 and from 0.300 to 0.390 on the five-objective task. Larger k gives less consistent gains and sometimes reduces performance.
Autoregressive generation allows later candidates to condition on earlier candidates within the same response, providing an opportunity to vary proposals. Although this does not guarantee structural diversity, our ablations support its benefits for exploration and final solution quality. Because generating multiple candidates also shares the input prompt across them, reducing repeated input tokens and LLM calls, which can lower API cost and running time.
As complementary evidence, with the same GPT-4o backbone, ExLLM without evolving experience achieves Top-10 fitness of 4.245 ± 0.121, compared with 4.076 ± 0.020 for MOLLEO on the five-objective task, and 18.165 versus 17.862 on PMO. The direct k ablation above isolates the effect of k; the MOLLEO comparison supports the effectiveness of the remaining k-offspring design.
Response:
Thank you for your question, we would like to clarifty that the implementation considers both fitness and Pareto optimality when selecting positive examples for the experience update. It samples 10 positive examples uniformly from either the fitness-based pool or the Pareto-based pool, choosing each pool with probability 0.5, as implemented in
algorithm/PromptTemplate.py. Both population selection and experience samples selection therefore use fitness and Pareto information. The current description of the positive examples as only “top-r by fitness” is incomplete, and we will correct it in the method section.Response:
Each objective uses a fixed, predefined normalization range or scale, following the task definition and known objective ranges. For molecular optimization, we follow MOLLEO's normalization. We do not compute min/max values from the current population or search history, so the same objective values always receive the same fitness throughout a run. The good/bad examples therefore remain comparable across generations.
Response:
The experience-generation prompt receives numerical values. Each selected candidate is included with its representation(such as SMILES in the molecular experiments), its raw value for each objective, and its aggregate fitness. When constraint feedback is available, the candidate's constraint values or diagnostic feedback are included as well. The experience update therefore has access to both the candidate and its quantitative evaluation.
The prompt then asks the LLM to summarize actionable patterns and pitfalls in a single evolving experience of approximately 200 words, integrating the previous experience and removing redundant or outdated content. Thus, numerical evidence is preserved in the experience-generation input, while the output deliberately compresses it into reusable guidance rather than repeating each candidate and its scores. We will clarify this distinction in the method section.