acceptodds
Withdrawn from ICLR 2027

MEMIG: Self-evolving Memory Construction for Antibody Optimization

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Memory has become central to self-evolving LLM agents, which store and retrieve experience from past rollouts to guide future decisions. Existing systems represent this experience as free-form natural language and maintain it by rewriting, which fits tasks whose experience is narratable. However, in scientific discovery and optimization, experience is often a quantity measured by an oracle under a specific context. Rewriting it as text leaves no index for where a record should be, replaces a measurement with a description, and cannot express how interventions interact. To address this, we argue that agent memory for biomolecule optimization should instead be indexable, quantitative, and structured, and we introduce MEMIG. This memory construction framework realizes these three properties. In detail, MEMIG addresses each insight with a domain-grounded identifier and accumulates the quantitative effect measured by a property oracle, so that updates remain local and retrieval becomes an exact lookup. Next, MEMIG connects insights whose interventions are applied together and measures their epistasis by decomposing each evaluated set. Our results show that MEMIG improves performance in most settings across in silico antibody optimization tasks.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
antibody optimization, agent, memory, self-improvement, drug discovery

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

this paper introduce MEMIG as a memory system which addresses some shorcomings of existing works and achive good numerical results.

Critical strengths

The studied topic is important.

numerical results are good.

Critical weaknesses
  1. line 54-73, this figure is a bit liked generated by codex 6. (no evidence, just by experience), the same as figure 3.

  2. line 226-239, it is unusual to put results direcly into the main method part.

  3. The method is not very theoretical sound.

  4. the main method is pretty descripitive, not very sound. no formulations. The proposed memory system is not clearly described.

  5. do not use color in ur submission, not very formal.

  6. Using "four steps" in your main method is somehow vague. consider at least use psedo codes.

  7. Only GLM results. How about on opus 5.5?

  8. the position of related work is strange somehow, and some important works are missing.

Questions for authors
  1. what is the main method?

  2. what about the results on more strong base models such as codex 6.1

  3. what is the std of your experiments.

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

MEMIG is a memory framework for LLM-driven antibody optimization. It stores oracle-measured mutation effects under biological identifiers, connects mutations through measured pairwise epistasis, and retrieves relevant records to generate natural-language guidance for subsequent optimization runs. Its main contribution is integrating domain-specific indexing, quantitative evidence accumulation, and mutation interaction information into reusable agent memory without updating model weights. The authors evaluate MEMIG on multi-property antibody optimization and computational antibody–antigen binding optimization using two LLM backbones. Comparisons with search and agent-memory baselines, component ablations, and additional controls providing baselines with the same decomposition measurements support its empirical utility. Validation is entirely computational; the experiments do not fully isolate the contribution of graph organization from the information it stores.

Critical strengths
  1. Well-motivated domain-specific memory design. MEMIG integrates biological identifiers, quantitative mutation effects, and pairwise epistasis into reusable memory. This provides a practical approach to exploiting historical measurements and mutation interactions in antibody optimization.

  2. Empirical support for the overall framework. Evaluation across two LLM backbones and two optimization tasks, together with component ablations and controls supplying baselines with the same decomposition measurements, supports MEMIG’s practical utility in the tested computational settings.

Critical weaknesses
  1. The methodological novelty beyond domain-specific memory design remains unclear. Existing systems such as A-MEM, Zep, and HAGE already explore graph-based memory organization and retrieval. While MEMIG introduces biologically grounded identifiers and oracle-measured epistasis relations, the paper does not clearly establish what additional mechanisms it contributes to memory updating, retrieval, or experience reuse. This limits the evidence for a broader methodological advance beyond its antibody optimization application.

  2. The experiments do not isolate the contribution of the graph mechanism from the information it stores. Removing epistasis edges also removes pairwise mutation measurements, so the resulting performance drop does not independently establish the benefit of the proposed graph organization and retrieval mechanisms. Although Table A4 supplies baselines with the same decomposition measurements, it does not establish whether a simpler structured-memory approach could achieve comparable results. This leaves the source of the reported gains insufficiently resolved.

Questions for authors
  1. Beyond biological identifiers and epistasis-based relations, what specific mechanisms distinguish MEMIG from existing graph-memory methods in memory updating, retrieval, or experience reuse?

  2. Can a simple structured-memory baseline retaining identical measurements, using conventional field indexing and relational queries, achieve comparable performance under matched information access and computational budgets? What operational differences would distinguish this baseline from MEMIG?

  3. What benefits does LLM-generated memory interpretation provide over directly supplying structured records? Have the authors compared direct JSON/tabular input, deterministic template rendering, and LLM-generated interpretation using identical retrieved records, in terms of optimization performance, additional costs, and fidelity to the underlying measurements?

Reviewer_raM3OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper argues that self-evolving LLM agents for biomolecular optimization need structured, quantitative memory rather than free-text narratives. It proposes MEMIG, which indexes mutational effects by biological context (property/germline/region/AHo-mutation) and tracks pairwise epistasis in a co-occurrence graph, retrieving top-voted quantitative records and rendering them into prompt text via an LLM curator. Experiments on OAS multi-property antibody optimization and SAbDab CDR-H3 binding show gains over stateless baselines and prior narrative-memory methods.

Critical strengths
  1. Well-motivated identifier design. I agree with the argument that narrative memory replaces measurements with text descriptions and lacks addresses, which forces full rewrites. The fix here is clean: because the identifier is both the addressing key and the aggregation key, local updates and exact retrieval come from the same mechanism rather than two separate modules bolted together. In contrast, the IDs in ACE and DC-Cu really are just opaque tags. This mapping from motivation to design is the most convincing part of the paper.

  2. The step-wise curves and budget controls are clean. Running concurrent rollouts at each memory step alongside a synchronous vanilla LLM control, while also showing unsmoothed curves, genuinely separates memory accumulation from base-model priors.

Critical weaknesses
  1. The LLM curation layer systematically flips optimization directions, undermining the core thesis. Although the paper argues numeric records beat LLM narratives, the agent only sees curator-rendered text. In the authors' own showcase examples, both curator and agent misinterpret sign directions: Appendix A.4 labels beneficial liability reductions (-2.85 for to K, -3.31 for to T) as an "avoid-mutation" and a "penalty," and in A.1 the agent treats a favorable -2.66 immunogenicity reduction as "disfavored" and acts on it. The agent has no fallback to inspect raw numbers. Table 2's faithfulness metrics (0.931 adoption, 0.909 helpfulness) miss this entirely because retrieval already pre-filters candidates by wild-type residue and net votes.

  2. The epistasis graph lacks both design consistency and empirical support. The first is context-free edges. The paper argues throughout that mutation effects depend on biological context and need context-based addressing, yet graph edges are keyed only by mutation pairs, and epistasis is averaged across different starting sequences and backgrounds. Epistasis is background-dependent by definition, so averaging it into a single scalar strips away context in the exact way the paper criticizes narrative memory for. Nodes are indexed by germline and region, while edges are not, and this asymmetry is never discussed. Second is that current ablation does not support the edges. It is unclear whether removing epistasis also drops the pairwise decomposition oracle calls; if it does, the drop in Table 1 mixes up "fewer edges" with "fewer evaluations" and cannot be attributed to memory structure. As a structural contribution, it is also missing the simplest baseline: a flat lookup table storing mean effects keyed only by property and mutation, with no hierarchical IDs, no edges, and no LLM curation. This baseline is cheap to run and necessary to tell whether the proposed structure is actually needed.

Questions for authors
  1. Why do the AHo region annotations appear systematically shifted from standard boundaries? In the main text and appendix examples, several mutations at positions that belong to FR2 or FR3 under standard Honegger & Plückthun numbering are labeled as CDR1 or CDR2. Since region is both an aggregation key in the identifier and explicitly passed to the curator as the authoritative region for CDR-vs-framework reasoning, could you clarify your boundary definitions or confirm if this is a bug? If it is a bug, both the aggregation buckets and the biophysical explanations in the curated text would be affected.

  2. Why does the identifier ablation cause a large drop on clusters where removing germline and region should be a no-op? The ablation removes the germline and region fields. On two of the tested clusters, the paper notes that germline is constant across starting complexes, and the binding task only mutates CDR-H3 so the region is fixed too. If each cluster has its own independent memory, dropping two constant fields changes neither aggregation nor retrieval. Is memory actually shared across clusters, or does the remaining cluster account for the entire drop?

  3. Have you tested whether a frozen memory generalizes to unseen sequences? Right now memory is accumulated online on the same sequences being evaluated, which is a transductive setup. Since the paper claims that AHo indexing and germline grounding allow insights to transfer across antibodies of different lengths, a natural check is to build memory on one split, freeze it, and evaluate on held-out sequences or across clusters. This takes no extra oracle budget and would be a much stronger test of whether the memory captures biological regularities rather than instance-specific statistics.

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

Related papers

Discussion (0)

Sign in to comment.