acceptodds
Under review as a conference paper at ICLR 2027

Zero: Programming Long-Horizon Mathematical Research Across Agents and Tools

Abstract

We introduce Zero, a new framework for programming long-horizon mathematical research. Zero represents research procedures as executable workflows. A meta-agent constructs and revises these workflows, while a runtime executes them across agents and mathematical tools. Researchers can start from a problem, provide procedural knowledge, or augment an existing research system. Zero lifts an existing system's workflow into its executable representation, where the different stages can be modified, reassigned to different resources, or extended with new capabilities. Across four studies, Zero (i) proves optimality of a previously open discrete-packing instance and finds novel constructions for more than 15 related variants; (ii) improves the state of the art for the five-color almost-coloring problem of the plane, lowering the removed fraction from the previous state-of-the-art value of 3.74% to below 3.5%, with the construction validated by exact-arithmetic geometric checks; (iii) resolves two open questions in Lagrangian relaxation; and (iv) extends existing research harnesses with formal verification and open-solver backends. Together, these results suggest a shift from programming individual mathematical agents to programming the mathematical research process itself.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
AI for mathematics, agentic workflows, mathematical research automation, meta-agents, formal verification

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_PBrqOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ZERO represents mathematical research procedures as executable workflows of agents, shell commands, and researcher interactions. A meta-agent constructs or modifies these workflows using the task description, researcher input, and retrieved examples. The paper studies packing, coloring, Lagrangian relaxation, and extensions of existing research systems. It reports new constructions and counterexamples, supported by exact checks and, for some results, Lean formalization. The main contribution is a framework for organizing and reusing research procedures across tools and agents.

Critical strengths
  1. The mathematical results give concrete evidence that the system can support useful research. These include the optimal packing for the previously unresolved n = 15 instance, improved coloring constructions, and counterexamples answering two questions on Lagrangian relaxation (Sections 3.1–3.3). The explicit certificates and reported Lean checks strengthen the evidence for these results. These outcomes are a substantial part of my positive assessment.

  2. The framework supports several stages of research within one executable procedure, including search, proof development, deterministic checks, and human decisions. Section 3.4 gives concrete examples of adapting existing systems to add formal verification and replace commercial solvers. This demonstrates practical value for researchers who need to combine tools with different interfaces and validation requirements.

Critical weaknesses
  1. The methodological contribution is not sufficiently clear. Sections 2.1–2.3 explain the workflow primitives, skills, and retrieval component, but give limited detail on how the meta-agent chooses and revises a research procedure. For example, Section 2.3 describes the case corpus and BM25 retrieval without showing how retrieved cases affect important workflow decisions. The paper also gives little analysis of which design choices help across the studies. This makes it difficult to identify the new, reusable knowledge contributed by the framework beyond the implementation and individual results.

  2. The evaluation does not clearly isolate the benefit of ZERO’s workflow construction. Table 1 provides a useful comparison on one packing instance, but does not report repeated-run success rates or separate the effects of orchestration and task-specific search improvements. In Section 3.2, ZERO receives GPT Pro guidance, while it is unclear whether the comparison methods have equivalent resources. The remaining studies mainly demonstrate capabilities and extensions. This limits the evidence that the reported gains will carry over to new tasks under comparable conditions.

Questions for authors
  1. Could you trace one reported case from the initial workflow to the successful result, identifying the decisions made by the meta-agent, changes prompted by execution feedback or retrieved cases, and decisions supplied by the researcher? Which workflow change was important for making progress? This would clarify the contribution of automated workflow design.

  2. For Table 1 and the planar coloring comparison, please clarify which models, tools, starting artifacts, researcher input, and budgets were shared across methods. Is there evidence from the existing runs or a focused comparison that separates the benefit of workflow construction from the underlying coding agents and task-specific implementation changes? This would affect my assessment of how broadly the reported benefits apply.

2 comments

We thank the reviewer for their positive assessment and constructive questions.

Weakness 1: Methodological contribution

Our contribution is making the research procedure itself programmable, i.e., choosing and combining models, tools, verification methods, and reusable workflows, rather than fixing a particular agent harness or strategy.

We deliberately use simple BM25 retrieval with the source-observatory skill to identify relevant methods. ZERO can incorporate existing systems or modify their internal stages, representing procedures from evolutionary search to prover–verifier loops and FirstProof multi-agent systems (Appendix F).

Workflows can also be reused: in Section 3.1.2, ZERO incorporated its successful n=15n=15 packing workflow into a higher-level research procedure, discovering constructions for rectangles, cuboids, triangles, and tetrahedra without prescribing the instances.

Weakness 2: Evaluation and ablations

We conducted additional ablations keeping the underlying coding agent, runtime, tools, and task inputs fixed:

  • Without /design-workflow: The agent failed to construct an executable packing workflow or improve the published planar coloring result.
  • Without /augment-system: The agent only partially reproduced AutoOPT, failing to adapt SCIP's branch-and-bound procedure to achieve the required numerical accuracy.

We believe these ablations provide more direct evidence for the contribution of ZERO's workflow mechanisms compared to the end-to-end comparisons from before.

Question 1: Packing workflow trace

For n=15n=15, the researcher provided only the objective and existing Java search code, without prescribing a procedure.

Retrieval suggested SCIP-based geometric optimization [1] and verification-and-refinement [2]. The SCIP case study was particularly useful in suggesting that an initial local search could identify promising candidates before finer optimization. ZERO also considered AlphaEvolve-style approaches but deemed them less suitable for the discontinuous objective.

The meta-agent constructed a multi-round workflow with agents exploring pairings, optimizing solver formulations, improving efficiency, and synthesizing findings. A key choice was pursuing independent searches in separate contexts (two hours per agent using /goal) before synthesizing results.

During execution, one agent discovered promising rectangle pairings, which another passed as fixed inputs to CP-SAT. This yielded the optimal side length 50, illustrating the value of separating exploratory search from finer optimization.

The workflow itself remained unchanged during the run, the important methodological choice was its initial design, while adaptation occurred within the agents' search strategies.

Subsequently, the researcher specified a higher-level procedure, select problems, prove bounds, and search for constructions, without specifying the instances. Using /implement-workflow, ZERO implemented this procedure, reusing and adapting the original packing workflow to new geometries (Section 3.1.2).

Question 2: Comparison conditions

The comparisons shared task information and starting artifacts, while allowing different methodological choices.

Packing: All methods had 24-hour wall-clock budgets. ZERO used GPT-5.5 and achieved side length 50, versus 51 for the agent-based baselines. Token usage differed as reported in Table 1.

Planar coloring: ZERO and both Codex baselines used GPT-6 Astra as their underlying coding model, with the same task information and coding tools. ZERO achieved 3.49% removed area versus 3.57% for both baselines.

[1] Berthold et al., Global Optimization for Combinatorial Geometry Problems Revisited in the Era of LLMs. https://arxiv.org/abs/2601.05943

[2] Huang and Yang, Winning Gold at IMO 2025 with a Model-Agnostic Verification-and-Refinement Pipeline. https://arxiv.org/abs/2507.15855

Thank you for the helpful response. The packing example clarified how retrieved cases informed the initial workflow and which decisions came from the meta-agent versus the researcher. I also appreciate the additional ablations examining the role of the workflow skills.

I maintain my rating of 3. With more concrete implementation details, I believe this work could serve as a useful tech report and contribute valuable practical knowledge to the community.

Reviewer_cemdOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2(+1)
Read review
Paper summary

The authors propose ZERO, an agentic framework to solve open math problems. A meta-agent designs a harness and tackles open problems, where the input of the meta-agent may vary between modes. There are different modes in the ZERO framework, which is problem-led (where the description of the problem is given), research-directed (where the research direction is given), and system augmentation (where the original system is given and ZERO is asked to improve it). They successfully solve multiple different problems, including the packing problem that has been open for more than two decades.

Critical strengths

First the paper is easy to follow with relevant diagrams and arguments in the right place. Also there are real solutions to open problems that show that these harnesses work, without doubt.

Critical weaknesses

Two weaknesses:

  1. There is no comparison with existing harness engineering methods, so it is very hard to understand that this framework is needed, if this is better than others. Also there is no ablation - whether certain parts of the harness is needed, whether one mode is better than other mode or not etc. These points, I believe can be tackled given time, are making me doubt that "I understand this harness works but is this what we would prefer to do compared to others"? Addressing this concern and justifying the design of the framework will make the paper into a clear accept, but currently due to the concern I propose weak reject.

  2. The results that open math problems are solved would have been impressive six months ago. Now it is happening too much, and it is happening everywhere, just by plugging in a problem to GPT-6 Astra. So it is unclear to me whether these problems are solved due to a clever harness design or just the capability of these models (which goes back to point 1 about not enough ablation).

Questions for authors
  1. What was the challenge while doing this project, and what were the efforts to overcome these? Are these techniques , if exists, applicable to different areas in ML or agentic discovery or harness engineering?
2 comments

We thank the reviewer for their positive assessment and constructive questions. We agree that the contribution of ZERO and its distinction from existing harness-engineering approaches should be clearer.

Relation to other harness-engineering methods. Our goal is not to introduce another fixed research harness for math, but to provide mathematicians with a framework for constructing and adapting such harnesses. While many harness-engineering approaches rely on iterative optimization through scored evaluations (e.g. Meta-Harness [1]), such feedback loops are often prohibitively expensive for frontier mathematical research (i.e. a single task evaluation can cost hundreds of dollars). ZERO instead allows a mathematician, together with a meta-agent to construct a task-specific research workflow directly from the objective and available procedural knowledge, without requiring repeated evaluations of complete research runs. Moreover, ZERO can represent existing research systems as editable workflows: Appendix F demonstrates this with three FirstProof multi-agent systems, while Section 3.4 shows how existing systems can be modified and extended. This allows mathematicians to incorporate new methods while retaining control over their research process.

Baselines and further ablations. We additionally evaluated two recent frameworks, JIT-Agent [2] and MechMath Agent Team [3]. We evaluated both frameworks on the packing and almost-coloring problems. For a controlled comparison, we evaluated all methods using the same underlying model, task specifications, available tools, and computational and time budget

MethodPacking ↓Almost coloring ↓
ZERO503.49 %
JIT-Agent [2]n.f.3.73 %
MechMath Agent Team [3]513.60 %

We also compare ZERO with existing approaches in Table 1, including ShinkaEvolve and the Agentic Researcher. Additionally, in Section 3.2.1 we compare ZERO with a generic Codex agent and a Codex setup with a domain expert, both using GPT-6 Astra. Both achieve 3.57% removed area, compared with 3.49% for ZERO. Since the three modes address different research settings, we focus our ablations on their underlying mechanisms rather than comparing the modes directly.

Furthermore, we assessed the contribution of ZERO's workflow-construction mechanisms by ablating the corresponding skills while keeping the underlying coding agent, runtime, available tools, and task inputs unchanged.

  • In packing, the meta agent failed to construct an executable workflow without /design-workflow.
  • In planar coloring, it failed to improve upon the published result without /design-workflow.
  • For AutoOPT, removing /augment-system led to a partial reproduction, but the agent failed to adapt SCIP to the required branch-and-bound procedure and could not reproduce the reported numerical results to the desired accuracy.

Frontier model capabilities and ZERO's contribution. We agree with the reviewer that frontier models have become substantially more capable at mathematical research. We also tested GPT-6 Pro directly on packing, almost-coloring, and Lagrangian relaxation. GPT-6 Pro neither solved packing nor improved the published almost-coloring result. GPT 6 Pro did, however, produce a correct counterexample for Lagrangian relaxation, although its proof contained minor gaps and it did not provide independent numerical or Lean certificates. ZERO thus complements frontier-model reasoning by aligning the computational research process with mathematicians' intentions, including how solutions are searched for, verified, and formalized.

Question 1. A particular challenge was preserving researcher-given procedures and validation requirements while coordinating heterogeneous tools and existing systems. As demonstrated by the AutoOPT case study, adapting a system can require more than substituting individual components/solvers: reproducing the global-search stage required implementing the two-pass BnB-PEP procedure and a specialized SCIP constraint handler. A similar approach, explicitly representing research stages, incorporating deterministic validation, and modifying individual stages in response to failures, could also transfer to other areas of ML and agentic scientific discovery.

[1] Meta-Harness: End-to-End Optimization of Model Harnesses, https://arxiv.org/abs/2603.28052

[2] JIT-Agent: Scaling Harness Intelligence via Just-in-Time Harness Evolution, https://arxiv.org/abs/2608.25593

[3] MechMath Agent Team: LLM Driven Agents for Mathematical Research, https://arxiv.org/abs/2607.04394

It is interesting that Zero can outperform existing frameworks and GPT6 pro. These experiments were good enough to change my score, because they show that using the framework has benefits compared to other methods, and it can actually lead to mathematical discovery that frontier models cannot.

Reviewer_9fbyOpenReview ↗
Rating
1
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents ZERO, a framework that represents long-horizon mathematical research procedures as executable workflows. A meta-agent constructs or edits workflows, while a stateful runtime executes agent, shell, and human-pause nodes organized into sequences, branches, and bounded loops. ZERO supports three entry modes: designing a procedure from a problem, implementing a researcher-specified procedure, and augmenting an existing research system. It also retrieves relevant case studies and repositories and can expose components of external systems as editable workflow stages.

The paper validates ZERO through four heterogeneous research programs. It finds an optimal 50-by-50 packing for the unresolved n=15 discrete rectangle-packing instance and reuses the procedure on related shapes and dimensions. It improves published almost-coloring constructions and expands the certified continuum of off-diagonal six-colorings. From a researcher-specified search/prove/formalize loop, it constructs counterexamples to two questions about extensions of Geoffrion's theorem and produces Lean developments without admissions. Finally, it augments VALG with formal verification and replaces commercial solver dependencies in AutoOPT with an open stack. Validation includes exact packing and geometry checks, independent numerical checks, Lean builds and axiom inspection, and human review.

Critical strengths
  1. The demonstrated outcomes are unusually substantive. Settling a long-open packing instance, improving multiple geometric constructions with exact checks, answering open optimization questions with formally verified counterexamples, and repairing/extending existing research systems together provide compelling evidence that explicit workflows can coordinate agents and tools over long horizons. The paper generally distinguishes computational evidence from proof and uses validation appropriate to each domain.

  2. ZERO's representation is simple but useful: agent, deterministic shell, and researcher-pause nodes make control flow, validation gates, and provenance explicit while allowing complete external systems to be wrapped or decomposed. Reuse is demonstrated rather than merely asserted, especially when the packing workflow becomes a component of a broader research program and when existing harness stages are locally replaced.

Critical weaknesses
  1. The case studies do not isolate the effect of ZERO from the capabilities of the underlying frontier models, tools, researchers, and task-specific engineering. Only the n=15 packing study has a direct comparison, and even there the baselines differ in workflow, model/resource allocation, and synthesis. The other studies lack repeated runs, ablations of the meta-agent and runtime, or matched baselines implementing the same researcher procedure without ZERO. Consequently, the impressive discoveries validate the complete sociotechnical process but do not establish that ZERO itself is the causal source of improved research performance.

  2. Reproducibility and evaluation remain difficult to assess. The most consequential runs are expensive and apparently single-shot; researcher intervention and approvals vary across studies; different frontier models are used; and the full source is promised only upon acceptance. The paper reports outcome certificates well, but provides limited aggregate information about failed workflow revisions, total human time, model/tool costs, sensitivity to prompts, and how often the meta-agent constructs an effective workflow. This makes reliability and transfer to new researchers or domains uncertain.

Questions for authors
  1. For each case study, can you provide a complete accounting of workflow revisions, failed branches/runs, wall-clock time, tokens/API cost, CPU/GPU/solver resources, and researcher interventions? This would materially affect the assessment of autonomy, efficiency, and reproducibility.

  2. Can you run a matched ablation in which the same models, tools, budget, and researcher inputs are used without ZERO's explicit workflow representation/runtime (for example, a single persistent coding agent or a manually scripted pipeline)? This is needed to isolate the framework's contribution from the surrounding resources.

  3. Will the artifact release include the exact workflows, meta-agent prompts/skills, pinned dependencies, raw execution traces, unsuccessful attempts, and all compact verifiers needed to reproduce every headline result? If any component cannot be released, which claims can an independent reviewer reproduce now?

1 comment

We thank the reviewer for their thoughtful review and positive assessment of the mathematical results and demonstrated reuse of ZERO. We address the main concerns below.

Researcher involvement. We agree that researcher involvement varies across our studies. This variation is deliberate, as ZERO is designed not only to construct research procedures from mathematical objectives, but also to implement researcher-specified procedures and extend existing research systems. Appendix B (Table 2) distinguishes researcher inputs, agent contributions, and validation procedures for each study. We do not interpret these studies as measuring a uniform degree of autonomy.

Baselines and comparisons. Beyond the packing case study for n=15n=15, we also compare ZERO in the almost-coloring problem (Section 3.2.1) against generic Codex with GPT-6 Astra and Codex with GPT-6 Astra assisted by a domain expert. Both baselines underperform ZERO in the /design-workflow setting.

We additionally evaluated two recent frameworks, JIT-Agent [1] and MechMath Agent Team [2], on both packing and almost coloring. For a controlled comparison, all methods used the same underlying model, task specifications, available tools, and computational and time budgets.

MethodPacking ↓Almost coloring ↓
ZERO503.49%
JIT-Agent [2]n.f.3.73%
MechMath Agent Team [3]513.60%

These comparisons provide further evidence of ZERO's effectiveness across different mathematical research tasks under matched resources.

Additional ablations. We also conducted resource-matched ablations of ZERO's meta-agent skills while leaving the runtime intact. Removing /design-workflow caused workflow construction to fail altogether in the packing study. In almost coloring, the ablated system performed worse than generic Codex with GPT-6 Astra and failed to improve upon the published baseline.

We additionally ablated /augment-system in the AutoOPT case study. While the system correctly used all three open-source solvers, it failed to adapt SCIP to the required branch-and-bound setting and was therefore unable to reproduce the numerical results of the AutoOPT paper to the desired accuracy.

Together, these comparisons and ablations provide additional evidence for the contribution of ZERO's meta-agent skills to the observed research outcomes.

Cost and repeated runs. We agree that repeated, resource-matched evaluations would be valuable for estimating success probabilities and further isolating the contributions of models, workflows, and researcher interventions. Due to the substantial cost of frontier mathematical research runs, our evaluation instead focuses on demonstrating research capabilities through independently verifiable mathematical results, complemented by the controlled comparisons and ablations above. We therefore do not claim that the case studies establish average success rates or isolate the causal contribution of every individual component.

Reproducibility and artifact availability. We will release the complete ZERO implementation and experimental setup upon acceptance, including workflow definitions, meta-agent prompts and skills, dependencies, and verification code. Currently, we cannot redistribute the full research corpus because the underlying papers and repositories are subject to different licenses. Importantly, this restriction does not prevent independent verification of our mathematical results: the constructions, proofs, certificates, and validation procedures provided in the paper and supplementary material allow the mathematical claims to be checked independently.

[1] JIT-Agent: Scaling Harness Intelligence via Just-in-Time Harness Evolution, https://arxiv.org/abs/2608.25593

[2] MechMath Agent Team: LLM Driven Agents for Mathematical Research, https://arxiv.org/abs/2607.04394

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.