acceptodds
Withdrawn from ICLR 2027

Continual Graph Memory for Mathematical Research Agents

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Using frontier agent harnesses to tackle mathematical research problems has emerged as an effective means of advancing mathematics. However, solving frontier problems in mathematics may require a massive number of agents working in parallel for extended periods to construct proofs, thereby generating an enormous volume of intermediate proof results. Organizing these intermediate results throughout a long-horizon proof-search process and reusing knowledge gained from prior explorations remain major challenges. We present Ansatz, a mathematical research agent built around CONTINUAL GRAPH MEMORY, a graph-based, evolvable, cross-problem mathematical research memory system that explicitly organizes the entire proof search process and reuses information from exploration trajectories of previous problems. Specifically, we develop a unified graph memory that represents all intermediate exploration results, including facts, plans, and counterexamples, together with edges that explicitly represent the relationships among them; dependency-aware retrieval supplies precisely targeted local context; an evidence-sensitive curator updates the research frontier and distills lessons from prior attempts; and scoped recall surfaces earlier statements and negative findings for local re-proving rather than uncritical reuse. Experiments cover runs across all ten First Proof Second Batch problems, together with four component studies. Ansatz reports closure on all ten research tasks, demonstrating its ability to sustain and resume long-horizon mathematical search. Beyond these problems, Ansatz also produces solutions to the Jamison caterpillar conjecture and Erd˝os Problems 289, 348, and 488 without human intervention, and makes partial progress on several open problems, illustrating its strong ability to solve open mathematical research problems.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Graph Memory, Mathematical Research Agents

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 2 reviews, above 9% of the others.
Reviewer_3qdMOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The reviewer uses "human" to denote my hand-written opinions in the review and the reviewer uses "Opus" to denote Opus 5.5. The reviewer uses the incognito mode so no data will be shared to Claude.

Human:

This paper proposes a harness that incorporates continual graph memory, which involves multiple workers and verifiers. The authors propose curators to organize the graph, revisits stale annotations and record structured lessons. The scoped recall retrieves fact statements, and other findings with provance. The proposed system reports a 10/10 on FirstProof B2.

Critical strengths

Human:

  1. This paper shows good self-review audit.

  2. The proposed system demonstrates good performance in open problems.

Critical weaknesses
  1. Human: the provided link cannot be accessed.

  2. Human: Experiments have major flaws.

    • The grading schemes on First Proof B2 are not clear. It's not clear whether the authors invited expert mathematicians, and it's not clear how many of them are hired to grade each problem. Also, it's not clear whether the invited graders have the same standard as the official evaluation.

    • The proposed results on First Proof B2 (10/10) are too good to be true, especially with the weak model GPT 5.6.

    • The contribution states "An auditable empirical study of persistent mathematical research." But no proofs on First Proof B2 are submitted anywhere. The supplementary is empty. The reviewer is farmilar with all proofs in First Proof B2, but he cannot audit the proofs because they are not provided.

    • Minor revisions are not shown in the Table 1.

    • The compute may not be matched so it's an unfair comparison. The proposed method is a multi-worker system, so it might use significantly more tokens than baseline.

    • No enough ablations to isolate each part of the proposed system.

    • Danus cannot reach 7/10 as reported. The reviewer tried Danus but it could be much worse than 7/10 in First Proof B2. Detailed settings must be provided to support this results.

Opus also raised similar concerns which are verified by the reviewer:

The grading protocol for the headline 10/10 is not described, and may not match the baselines'. The official team results in Table 1 come from the benchmark's expert review. The Ansatz, Danus, Codex, and plain GPT-5.6 columns are labeled "Our reported results," but the paper never says who graded them. It doesn't say whether expert mathematicians reviewed them or under what rubric.The wording elsewhere suggests these may be self-assessed. The abstract and introduction say Ansatz "reports closure" or "local closure," and §3.2 defines "accepted" as accepted by the system's own LLM verifier. If the ✓ marks for Ansatz reflect internal acceptance rather than independent expert grading on the benchmark's scale, the main comparison is not valid. The authors should state exactly who graded each of the 40 new submissions, whether graders were blinded to the system, and release the final proofs.

Baseline comparisons are confounded by model and compute. All of the authors' runs use GPT-5.6 Sol, while the official teams used GPT-5.5 Pro or Gemini 3.1 Pro. So "substantially outperforms the official participating teams" largely reflects a newer backbone. Among the same-backbone systems, no compute is reported: tokens, cost, wall-clock time, number of worker sessions. Ansatz runs five worker lanes, a fresh verifier per submission, a curator, and strategy consultation. A gap from 4/10 (plain) to 10/10 may simply reflect orders of magnitude more inference. The introduction promises "matched resource and memory diagnostics from the earlier configuration," but these do not appear anywhere in the paper. Each system is also run once per task, with no indication of variance.

  1. Human: novelties might not be significant. The proposed system contains graph-based verifications, which has already been included in Danus, it's not sure what is the new contribution.

Opus also has concerns about the paper's novelty and ablation study:

The ablations do not support the memory contribution, which is the paper's title and stated novelty. Several problems compound here: Ablation A (flat notebook instead of the fact graph) shows no difference, 7/7 vs. 7/7. The authors acknowledge that graph structure of facts gives no measurable benefit. Ablation B (no exploration graph or curator) differs by a single task in a single run. Ablations C and D test the independent verifier and strategy consultation, which are not memory components. The independent verifier is inherited from Danus. The A/B task subset is exactly the set of tasks Danus already solves: {01, 02, 05, 06, 07, 08, 09}. The three tasks where Ansatz goes beyond Danus (03 Bernoulli, 04 Dilation, 10 von Neumann) are excluded from A and B. Task 04, the hardest and the one only Ansatz solves, is excluded from every ablation. So the ablations say nothing about which component produces the gains the paper highlights. The "continual" and "cross-problem" mechanisms are never evaluated. These are scoped recall and lessons. No experiment compares recall off vs. on, measures transfer across a sequence of projects, or shows that lessons change curator behavior. The scope setting used in the First Proof runs is not even reported.

Human: due to severe concerns, the reviewer decides to give a one to this paper.

Questions for authors

Human:

  1. Who graded those submissions in Table 1, and using what protocol? What's the expertise and focus domains of each grader?

  2. Since First Proof B2 solutions are already released, how can the authors prevent data contamination?

  3. What is the key novelty against Danus and why is it useful?

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

This paper presents Ansatz, a mathematical research agent that uses persistent graph memory to organize proof search and retrieve experience. The memory separates verifier-accepted facts from exploratory material, while a curator maintains the exploration state and records reusable lessons. The authors report closure on all 10 First Proof Second Batch tasks and solutions to four open mathematical problems.

Critical strengths
  1. Well-motivated problem formulation. The paper treats mathematical research memory as more than a record of past interactions. It distinguishes three types of information: verifier-accepted mathematical facts, exploratory materials, and experience that can be reused across problems. This formulation captures the different evidential status and purposes of information accumulated during mathematical research.
  2. Detailed system specification. The roles of workers, verifier, curator, and main agent are clearly separated. Appendix A also provides useful implementation details. The explicit acknowledgment that LLM acceptance is not formal verification is also valuable.
Critical weaknesses
  1. Evidence for the benefits of the proposed memory extensions is limited

As their important baseline, Danus already includes fact-graph memory and shared exploratory memory. Ansatz’s main within-project extensions are the typed exploration graph and the curator that maintains the exploration state. While these additions are well motivated, the experiments provide limited evidence of their benefits over simpler ways of organizing and maintaining the same information.

For example, replacing the fact graph with a flat fact notebook preserves the 7/7 success rate on the evaluated subset (Table 2), which seems to diminish the contribution and necessity of explicit dependency edges. Removing the exploration graph and curator jointly reduces performance from 7/7 to 6/7, yet it is about a single task difference on one single run. It also leaves unclear whether the benefit comes from the graph representation, active curation, or their interaction. Both studies also exclude Tasks 03, 04, and 10, which are particularly discriminative in Table 1.

Targeted comparisons that keep the same exploratory information while varying its organization and curation, under comparable budgets, would make the evidence more convincing.

  1. The experimental setup and evidence for carrying experience across problems are unclear

Cross-project experience transfer is a central claim of the paper, but the experimental description does not establish what information was available across problems or whether that information improved subsequent performance.

Section 3.5 defines off, batch, and all, with off as the default. However, the FPSB setup does not specify which mode was used, the order in which tasks were run, or which other projects were visible to each task. Appendix A.5 further notes that batch membership does not enforce chronological visibility. It is therefore unclear what information from other projects was accessible during a given run.

Moreover, Figure 3 shows how often the memory tools were called. However, it does not show whether the retrieved information was actually used, where it entered the reasoning process, or how it generally helped solve the problems. The counts themselves therefore provide limited evidence for the practical benefit of memory reuse.

Also, the paper lacks a clear matched comparison with and without access to historical memory, which would help establish a clear conclusion of whether cross-project memory improves performance.

Questions for authors
  1. Does the stated GPT-5.6 Sol backbone also apply to the verifier? What reasoning settings were used?

  2. Given that LLM verifiers can miss substantive proof errors[1], how confident are the authors that the higher reported pass rates reflect better solutions rather than more permissive grading from LLM? Were the final FPSB solutions graded under the original benchmark’s independent expert protocol?

  3. Was cross-project recall enabled in the FPSB experiments? It would be helpful to specify the memory mode, task order, and source projects available to each run.

  4. Is there a matched comparison with and without access to historical memory? How does this access affect performance or resource usage?

  5. Since each recall call can return multiple records, how often were those records actually used? How much were they contributing to the solving process?

  6. Is there further evidence separating the benefits of the exploration graph from those of the curator (beyond their joint ablation)?

  7. Section 3.5 specifies BM25, while Figure 3 describes memory_recall as embedding-based. Which implementation was used, and if embeddings were used, how were they combined with BM25?

  8. What per-task computational budgets were assigned to each agent system? How were these budgets enforced, and how were they allocated across different tools and stages?

[1] Barkallah, Slim, et al. "Pseudo-Formalization for Automatic Proof Verification." arXiv preprint arXiv:2605.20531 (2026).

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

Related papers

Discussion (0)

Sign in to comment.