acceptodds
Withdrawn from ICLR 2027

Meta-Delta: Evolving Knowledge through Global Review for LLM-Driven Optimization

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

LLM-driven evolutionary optimization accumulates experience about which changes improve a candidate solution. As search progresses, this experience must be interpreted in context: lessons may have limited scope, accumulated guidance may conflict, and repeated refinement may leave alternatives unexplored. We introduce **Meta-Delta**, a framework that jointly evolves candidate solutions and the knowledge guiding their generation. A **Delta Bank** stores structured **Delta Cards** with search lessons, applicability conditions, and supporting evidence. Routine local updates incorporate what each experiment teaches, while Global Review examines the Bank and broader search history to reconsider existing guidance and propose directions for subsequent testing. The updated Bank guides new experiments, whose outcomes further refine its knowledge. Empirically, Meta-Delta achieves the best or tied-best result on 10 of 11 tasks across mathematical, scientific, and systems optimization. On scientific law discovery, it achieves the highest mean numeric-fit score on SCILAWS-REAL, ties for the highest scientific-validity score at displayed precision, and achieves the highest mean structure score on SCILAWS-PARALLEL. Global Review ablations provide evidence that broader knowledge revision helps sustain progress, and case studies show how revised knowledge changes subsequent search and the structure of discovered scientific laws. These results support treating search knowledge as an evolving component of long-horizon optimization.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM-driven optimization, Evolutionary search, Knowledge evolution, Meta-agents

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_NjTuOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

Meta-Delta augments LLM-driven evolutionary optimization with a revisable knowledge bank. Structured Delta Cards record search lessons, applicability conditions, and supporting evidence. Routine updates incorporate individual experiments, while Global Review uses three complementary perspectives to reconsider accumulated guidance and propose subsequent directions. Evaluation spans 11 optimization tasks and two scientific-law discovery settings. The framework is well motivated and includes informative case studies. However, the current evidence leaves substantial uncertainty about computational fairness, robustness of the reported gains, and the specific contribution of knowledge revision.

Critical strengths
  • Careful evidence management. The distinction between supported and exploratory Cards, explicit evidence provenance, and separation of implementation failures from strategy failures are sensible design choices.
  • Broad empirical coverage. The evaluation spans mathematical, scientific, and systems optimization, with promising results on tasks such as Signal Processing and EPLB.
  • Informative qualitative analysis. The case studies trace Card revisions to subsequent plans and candidate changes, include an unsuccessful hypothesis, and acknowledge attribution limits.
  • Useful controlled evidence. The same-checkpoint continuation study provides additional support beyond selected trajectories, although its scope is limited.
Critical weaknesses
  1. Computational fairness and mechanism attribution remain unresolved. Methods are compared using equal outer-round budgets, while Meta-Delta can perform multiple candidate evaluations per round and additional Global Review calls. The released configuration explicitly treats outer iterations as the budget unit. Moreover, disabling Global Review simultaneously removes additional computation, broader historical context, exploratory suggestions, and joint Bank revision. This ablation therefore does not isolate the contribution of persistent knowledge revision.
  2. The strongest scientific-benchmark comparisons lack uncertainty estimates. In Table 1, the numeric-fit advantage over DeltaEvolve is 49.30 versus 49.12, and the structure-score advantage over the strongest external baselines is 48.02 versus 47.53. Relative to w/o GR, the structure-score improvement is only 0.21 percentage points, while scientific validity remains tied at 84.24. Without per-task differences and uncertainty estimates across seeds, the robustness of these margins is unclear. For the 11-task suite, best@3 reporting similarly leaves average performance and variability unresolved.
  3. The MMD results do not support the stated Global Review attribution. Section 5.2 counts earlier plateau attainment on both MMD tasks as a benefit of enabling Global Review. However, Full reaches both plateaus at round 1, whereas Algorithm 1 places Global Review after that round’s candidate generation and evaluation. Under the described procedure, these first-round outcomes cannot result from preceding Global Review. The timing or interpretation requires clarification.
  4. The incremental contribution needs a closer mechanism-level comparison. The paper appropriately discusses evolving experience libraries and applicability refinement, including ACE and AlphaOPT. A budget-matched generic global reflection or single-manager global update would help establish the added value of the proposed review structure. The present comparisons leave its contribution relative to simpler alternatives uncertain.
Questions for authors
  1. Under matched token or candidate-evaluation budgets, how does Meta-Delta compare with w/o GR and a simpler global-reflection control? Please report actual resource usage.
  2. How can Global Review explain the round-1 MMD results? Is there an initialization-stage review omitted from Algorithm 1?
  3. Can the authors clarify the actual stopping/trigger settings and final candidate-selection protocol, including how these correspond to Algorithm 1?
Reviewer_EGmEOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Meta-Delta, an LLM-driven evolutionary optimization framework that jointly evolves candidate solutions and the knowledge guiding their generation. It maintains structured Delta Cards to record search experience and employs Global Review to reconsider accumulated knowledge, resolve conflicting guidance, and explore new search directions. Experiments across mathematical, scientific, and systems optimization tasks show promising results compared with existing evolutionary optimization baselines.

Critical strengths
  1. The paper explicitly represents accumulated search knowledge through structured Delta Cards with applicability conditions and supporting evidence, enabling systematic knowledge maintenance during long-horizon optimization.
  2. The Global Review mechanism provides a principled workflow for reconsidering conflicting guidance and exploring alternative search directions when progress stagnates, complementing routine local knowledge updates.
Critical weaknesses
  1. The core components of Meta-Delta, including memory management, reflective reasoning, and multi-agent collaboration, have been explored in prior work. Their combination does not sufficiently establish a novel methodological contribution.
  2. Several closely related methods discussed in Section 2, including KSI, AlphaOPT, ACE, and PACEvolve, are missing from the comparisons, leaving the incremental contribution over existing knowledge-management approaches unclear.
  3. Global Review introduces additional LLM calls and token consumption, but experiments match only the number of search rounds. The reported gains may therefore stem from greater inference budgets rather than the proposed knowledge-revision mechanism.
Questions for authors
  1. How does Meta-Delta fundamentally differ from existing approaches to memory management, reflective reasoning, and multi-agent collaboration, beyond combining these components?
  2. Could the authors compare Meta-Delta with closely related methods such as KSI, AlphaOPT, ACE, and PACEvolve to better establish its contribution over existing knowledge-management approaches?
  3. Could the authors provide comparisons under matched LLM inference or token budgets to isolate the gains from Global Review itself, rather than from additional LLM calls?
Reviewer_VBoNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Meta-Delta, which improves LLM-driven iterative optimization by revising search experience. Delta Bank uses structured cards to store experience, applicable conditions, and experimental justification, updating the cards after each round of experiments. When the search stalls or knowledge conflicts occur, Global Review examines the knowledge base and history from three perspectives: exploration, refinement, and bottleneck diagnosis, then revises the cards and proposes directions for testing. Experiments cover 11 optimization tasks and two settings of SCILAWS. The method achieves noticeable improvements over external baselines in signal processing and EPLB, and a slight lead in aggregation metrics for scientific law discovery. The paper also analyzes the role of Global Review through disabling review, continued search from the same checkpoint, and case tracking.

Critical strengths
  1. Delta Cards clearly record applicable conditions, supporting experiments, and hypothesis status, facilitating the reuse and revision of search experience. Case studies also demonstrate how card revisions change subsequent candidate solutions.

  2. The continued experiment from the same checkpoint in Appendix A.5 supports the overall role of Global Review, with 5/5 runs escaping stall when enabled and 2/5 when disabled.

Critical weaknesses
  1. Compared to the long-term reflections of ReEvo and the knowledge updates of ACE/KSI, the independent value of global knowledge reassessment has not been fully validated. Existing ablations simultaneously disable global knowledge reassessment and exploratory card generation, failing to distinguish the contributions of each.

  2. The advantages under equivalent computational budgets have not been validated. Meta-Delta involves multi-role calls and in-round retries, but the paper does not report the actual token consumption, candidate evaluation counts, or total search time for each method.

Questions for authors
  1. How much benefit can global knowledge reassessment bring while retaining exploratory card generation and matching token budgets?

  2. Please report the actual LLM call count, token consumption, candidate evaluation count, and total search time for each method, and compare the solution quality under the same token budget.

  3. What additional benefits do the three specialized perspectives offer compared to a single ordinary global reflection? Please provide a comparison under the same token budget.

  4. In Table 1, both MMD tasks reached the plateau in round 1, while Algorithm 1 only performed a review after candidate evaluation. Please verify the first review round and explain why Section 5.2 attributes this result to Global Review.

  5. In SCILAWS-PARALLEL, enabling Global Review improved SS from 47.81 to 48.02. Can you provide the gain for each run to determine if this small improvement is stable?

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

Related papers

Discussion (0)

Sign in to comment.