acceptodds
Under review as a conference paper at ICLR 2027

What Should We Work on Next? Node-Level Compute Allocation Across Open Scientific Questions

Abstract

Rapid advances in AI capabilities have enabled AI systems to contribute to research on open scientific questions. There are many open scientific questions that AI could help investigate, but the computational cost of frontier models is high. These costs motivate careful allocation of compute across open scientific questions. Two key considerations for allocation are the scientific importance of each question and its difficulty relative to the AI system’s capabilities. However, research-level open questions often require prolonged investigation, leaving limited outcome-based feedback for the difficulty estimates needed to guide compute allocation. This is particularly problematic for question-level allocation, as substantial compute may be invested before sufficient feedback is available to reassess the investment. Therefore, in this paper, we propose node-level compute allocation over evolving AND/OR directed acyclic graphs (DAGs), which record intermediate research progress and dependencies among research tasks. Finer-grained feedback from these tasks guides allocation across nodes from all questions, even before the underlying questions are resolved. We illustrate this allocation process through a case study involving five research-level open questions in quantum error correction (QEC). In controlled comparisons on four mathematical problems, the node-level scheduler ranked first in two of the three value settings and second in the remaining one, giving the best overall performance among the four conditions.

Primary area:
foundation or frontier models, including LLMs
Keywords:
test-time compute, scientific reasoning, metareasoning, resource allocation, AI for science, open problems

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_EU9hOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies how to allocate a shared compute budget across multiple open scientific questions. Building on an evolving AND/OR research DAG, it schedules concrete proof and exploration tasks using model-estimated utility gains, forecast output-token costs, and a stagnation penalty. A five-question quantum error correction case study illustrates the scheduling process. In controlled experiments on four mathematical problems, node-level scheduling ranks first in two of three value settings with DeepSeek-V4-Pro, whereas the question-level variant ranks first in all three settings with Claude Sonnet 5. Both algorithmic schedulers reduce the paper's measured lock-in waste relative to corresponding model self-scheduling baselines.

Critical strengths

1. The paper addresses an important and well-motivated research problem.

Allocating limited compute across open scientific questions requires deciding when to continue an investigation and when to redirect effort, despite sparse and delayed outcome feedback. This is a meaningful challenge that is not adequately captured by allocation across short, readily verifiable problems. The proposed framework makes intermediate research progress available to a global scheduler, and the QEC case study provides a concrete illustration of this setting. The problem formulation has value independently of whether the particular allocation score is ultimately optimal.

Critical weaknesses

1. The distinctive probability-guided allocation mechanism is not sufficiently justified or isolated experimentally. The scheduler relies on Research Agent estimates of current and post-task local probabilities, but the paper does not evaluate whether these estimates provide reliable task rankings. More specifically, Section 3.1 defines the post-task estimate conditional on obtaining the task's stated result, while Equation (3) uses the corresponding hypothetical probability change as a predicted utility gain without specifying the likelihood of that result. The expected-utility interpretation therefore needs clarification; this is not, by itself, proof that the score is mathematically incorrect. More importantly, the experiments do not compare the full method with a matched cost- and stagnation-aware scheduler that omits probability propagation. As a result, they do not establish that the distinctive probability-based component explains the observed improvements.

2. The evaluation does not yet support strong conclusions about robust reductions in unproductive research compute. The controlled study uses four mathematical problems and reports no run-to-run uncertainty for its milestone-score AUC, although the timing of a successful intermediate result can materially affect this metric. The QEC experiment illustrates a scheduling trajectory but does not provide a controlled comparison. In addition, the reported lock-in waste measure marks tokens spent on the third or later consecutive task on the same question without a new rubric milestone as waste. The cutoff is not justified or sensitivity-tested, and the measure depends on task segmentation and milestone timing; necessary, extended proof attempts can therefore be classified as waste. This does not invalidate the reported improvements in the defined metric, but it limits their interpretation as reductions in genuinely unproductive research computation.

Questions for authors
  1. Probability semantics. How are the current and post-task local probabilities elicited? Is the post-task estimate conditional on the stated result or an expectation over all task outcomes, and is there evidence that the estimates yield useful relative task rankings?

  2. Component attribution. Under matched DAGs, candidate tasks, and budgets, how does the full method compare with a value-/cost-aware scheduler retaining stagnation penalties but omitting probability propagation? What is the effect of removing the stagnation penalty separately?

  3. Lock-in and rubric validity. Why is the third consecutive milestone-free task the cutoff for lock-in waste? Do the conclusions persist with other cutoffs or alternative task segmentations? How were the reference-solution milestone weights selected?

3 comments
AuthorsOpenReview ↗

Clarification of probability semantics and task allocation

Consider a simple research tree with root AA and two child nodes, BB and CC. Here, AA is an AND node: this proof route requires both BB and CC, and the implication B∧C⇒AB\land C\Rightarrow A has already been established. Human experts assign a scientific value VAV_A to the root. Based on the current evidence, the local probabilities of BB and CC are 0.30.3 and 0.40.4, respectively. Under the independence approximation, the current root probability is therefore P(A)=0.3×0.4=0.12P(A)=0.3\times0.4=0.12, giving a current allocation utility of 0.12VA0.12V_A. For an OR node, under the same independence approximation, the probability that at least one alternative proof route holds is P(A)=1−(1−qB)(1−qC)P(A)=1-(1-q_B)(1-q_C), which equals 0.580.58 for the same current probabilities.

Continuing with the AND example, each child has a concrete next research task specified by a prompt. Conditional on obtaining the respective stated task results, we predict that the local probability of BB would become 0.50.5, while that of CC would become 0.70.7. We also predict the token costs of these tasks as T^B\widehat{T}_B and T^C\widehat{T}_C, respectively. To decide which task should receive compute next, we evaluate the two hypothetical updates separately. If the task for BB produces its stated result, we replace BB's probability with 0.50.5 while keeping CC's current probability at 0.40.4. The root probability becomes 0.5×0.4=0.200.5\times0.4=0.20, yielding a utility gain of (0.20−0.12)VA=0.08VA(0.20-0.12)V_A=0.08V_A. If the task for CC produces its stated result, we keep BB's current probability at 0.30.3 and replace CC's probability with 0.70.7. The root probability becomes 0.3×0.7=0.210.3\times0.7=0.21, yielding a utility gain of (0.21−0.12)VA=0.09VA(0.21-0.12)V_A=0.09V_A.

Dividing each gain by the corresponding predicted token cost gives the base allocation scores 0.08VA/T^B0.08V_A/\widehat{T}_B and 0.09VA/T^C0.09V_A/\widehat{T}_C. Thus, although scientific values are assigned to root questions, individual tasks can be scored according to their predicted effects on root probabilities and utility, allowing the scheduler to determine where to allocate compute next.

This scoring rule also captures an intuitive consideration in allocating research effort. At an AND node, all prerequisites along a proof route must hold jointly. If one necessary prerequisite is difficult to establish and is currently assigned a low probability of being established, then even if another prerequisite is easy to prove, the value of advancing it remains constrained by the difficult prerequisite: completing the easy part still leaves the overall route dependent on the unresolved requirement. In the example above, if advancing BB increases its local probability by ΔqB\Delta q_B, the resulting gain in root utility is VAqCΔqBV_A q_C\Delta q_B. The lower the current probability of CC, the smaller the root-level benefit from the same amount of progress on BB. This allows the scheduler to account for the contribution of local progress within the complete proof route.

At an OR node, the branches represent alternative proof routes, each of which can independently support the target claim; establishing any one route is sufficient. Difficulty in advancing one route therefore does not create the same prerequisite bottleneck as in an AND structure, nor does it suppress the value of another viable route in the same way. Under the independence approximation, the gain in root utility from advancing BB is VA(1−qC)ΔqBV_A(1-q_C)\Delta q_B. When the current probability of the alternative route CC is low, advancing BB can still provide a substantial benefit. When CC is already highly promising, however, further progress on BB provides a smaller additional benefit because the target already has a promising alternative route. Probability propagation thus incorporates both dependencies among jointly required premises and the availability of alternative routes into task scores, which are then combined with predicted token costs to guide the next allocation decision.

We do not introduce a separately estimated probability of task completion. Our modeling rationale is that the likelihood of obtaining a task's stated result is related to its difficulty, and predicted token cost already serves as a proxy for task difficulty in the allocation score. Accordingly, in this heuristic design, we compare candidate tasks using their predicted utility gains conditional on obtaining their stated results, divided by their predicted token costs, without multiplying by an additional, separately estimated probability of task success.

AuthorsOpenReview ↗

Stagnation penalty S: design rationale and replay analysis

We thank the reviewer for raising this concern about the stagnation penalty S. This comment prompted us to replay and examine the AI proof attempts and task scheduling process in our existing experiments. Through this replay analysis, we found that removing S would leave the currently reported performance curves unchanged.

During the initial exploratory phase of this project, we used a relatively strict accumulation rule: whenever a task failed to produce the result specified by its prompt, we accumulated S. However, in these early attempts, we observed that this strict mechanism could cause the agent to abandon potentially valuable research branches prematurely, as these branches often require sustained exploration and the accumulation of intermediate results before the original objective can be achieved.

We view this as a tradeoff between breadth and depth in research exploration. A stricter penalty encourages the agent to move to other branches sooner, while a more permissive mechanism allows it to investigate the current branch more deeply. Accordingly, the current implementation adopts a more permissive mechanism: in addition to completing the original objective, meaningful intermediate progress can also justify resetting S, provided that the relevant task settlement requirements are met. This choice aims to reduce premature abandonment of promising research branches.

The reviewer’s suggestion has helped us assess the role of this mechanism in our current experiments more accurately. Based on the replay results above, we do not attribute the gains reflected in the reported performance curves to S. In future versions of the paper, we will reduce our emphasis on S and its contribution, and identify better designs for stagnation detection, penalties, and reset rules, together with a systematic evaluation of the mechanism’s independent contribution, as directions for future work.

AuthorsOpenReview ↗

Lock-in and rubric validity

We thank the reviewer for asking about how lock-in waste is measured. The main purpose of this metric is to compare the shares of research output tokens devoted to consecutive milestone-free tasks across scheduling strategies under the same measurement rules. Using a finer or coarser segmentation of the research records, as well as counting from the second, third, or fourth consecutive milestone-free task onward, affects the absolute values of this metric; our focus is on the relative comparison between strategies evaluated under the same rules. To examine this, we recomputed the metric using cutoffs of two and four consecutive milestone-free tasks while keeping the original experimental scope, budgets, and task segmentation unchanged. As shown below, when averaged over the three value settings, algorithmic scheduling has a lower share than the corresponding model self-scheduling baseline for both models, with and without a DAG. This is consistent with the direction of the comparison at the original cutoff of three. Furthermore, we find that under model self-scheduling, the models tend to remain focused on the same question more persistently with a DAG than without one, even when multiple consecutive tasks yield no new rubric milestones. This pattern appears for both models at all three examined cutoffs: the average share of sustained investment without new milestones is consistently higher with a DAG than without one. This suggests that, in our experimental implementation, dependency graphs may reinforce continued investment in the current question, while explicit scheduling helps reduce such sustained investment without new milestones. Regarding record granularity, dividing the same research records into finer units, while keeping the total token counts and milestone assessments unchanged, makes it easier for each method to reach the same consecutive milestone-free cutoff, thereby increasing the fraction of tokens counted. We expect that, when a consistent refinement rule is applied across methods, this primarily changes the absolute percentages, while the relative ordering among methods remains broadly similar. Our focus is precisely on this relative comparison under a common measurement convention.

Average lock-in waste counting from the second consecutive task:

DeepSeek-V4-ProClaude Sonnet 5
ModelAlgorithmModelAlgorithm
With DAG45.2%9.8%66.7%16.0%
Without DAG23.4%9.4%52.5%12.5%

Average lock-in waste counting from the fourth consecutive task

DeepSeek-V4-ProClaude Sonnet 5
ModelAlgorithmModelAlgorithm
With DAG27.8%0.0%39.0%0.0%

Our scoring rubric was manually developed by the authors based on the key proof steps in the published reference solutions. We selected key intermediate conclusions and final results as milestones and assigned scores to them, with the total score for each problem normalized to 1. These manually assigned weights serve as a scoring convention for recording which key proof components the model has completed. | Without DAG | 5.0% | 0.0% | 28.9% | 0.0% |

Reviewer_ozAMOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies how to allocate AI research compute across multiple open scientific questions under a limited budget. The authors record intermediate propositions, proof paths, and research evidence in an evolving DAG, use concrete next-step research tasks as the units of allocation, and dynamically schedule them by combining expert-assigned scientific value, predicted progress propagated through the research graph, token cost estimates, and stagnation penalties. The paper illustrates scheduling trajectories using five questions and compares four allocation methods on four mathematical problems with verifiable reference solutions. The experiments show that node-level scheduling generally performs better with DeepSeek-V4-Pro, whereas the question-level variant performs better with Claude Sonnet 5.

Critical strengths
  1. I strongly agree with the premise of this paper: as AI begins to participate in long-term scientific research, how to invest limited resources in research directions worth pursuing is itself an important question. Open questions differ in value, their difficulty relative to the model is unknown, and terminal feedback is very sparse; continually adjusting resource allocation using evidence from intermediate research tasks directly addresses these challenges. This problem formulation and its implications for AI research workflows are my primary reasons for supporting acceptance.
  2. The research dependency graph connects local results to the ultimate questions, while information about value, cost, and stagnation jointly guides the selection of the next task. Scoring based on verifiable milestones and reporting the budget-normalized area under the curve make it possible to measure when valuable progress occurs. Lock-in waste decreases substantially for both research models, supporting the practical value of the explicit scheduling mechanism as a whole.
Critical weaknesses
  1. The existing evidence primarily validates the scheduling scheme as a whole; the additional benefits of the node-level design and the conditions under which it is applicable have not yet been sufficiently clarified.
Questions for authors
  1. Does the output token budget in §4.2.1 include the Forecast Agent's predictions and scheduling calls such as those for generating candidate tasks and updating qq and qafterq_{\mathrm{after}}? If so, approximately what proportion do these overheads account for? If they are counted separately, could you clarify the accounting conventions used for each condition? This would help assess the gains of the full method in comparisons under equal budgets.
3 comments
AuthorsOpenReview ↗

Token Budget Accounting and Forecasting Overhead

Yes. The output token budget in §4.2.1 includes all output tokens used for research, the Forecast Agent’s predictions, and scheduling operations, including generating candidate tasks and updating qq and qafterq_{\mathrm{after}}. In our implementation, a task’s token cost is predicted through a separate model call, so the output tokens consumed by these Forecast calls can be measured separately. In contrast, qq, qafterq_{\mathrm{after}}, and the prompt for the next task are mostly returned together with the research results when the model completes its current research task. It is therefore difficult to accurately separate the token overhead of generating this scheduling information from the research task’s own output and report a reliable overhead share for these components.

The separately measurable output token shares of Forecast calls are shown below:

RunOutput token for Forecast calls
DeepSeek / Equal13.68%
DeepSeek / Ascending6.26%
DeepSeek / Descending11.27%
Claude / Equal7.12%
Claude / Ascending5.46%
Claude / Descending4.19%
Reviewer_ozAMOpenReview ↗

Thanks for the authors’ response.

I am quite favorable toward this work mainly because the authors conduct a meaningful study grounded in real-world scenarios, which distinguishes it from existing benchmark-chasing papers. Nevertheless, as other reviewers have also pointed out, this work lacks sufficient and rigorous experiments.

Furthermore, the authors have not fully addressed my question: the additional benefits of the node-level design and the conditions under which it is applicable have not yet been sufficiently clarified.

For this reason, I will keep my score unchanged.

I have gone through your work again from start to finish.

I still appreciate that your research does not follow the crowd merely to chase benchmark rankings, but is driven by practical real-world needs. This is highly commendable. After careful consideration, I have raised my score to 4. That said, I hope you can refine the manuscript organization and add supplementary experiments in the revised version.

All the best.

Reviewer_D94SOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies how to split a shared compute budget across multiple scientific research questions when final outcomes provide sparse feedback. It uses node-level allocation where each question is kept as an evolving DAG of claims and proof routes. Tasks are ranked using expert-assigned question values, model-estimated probability, token costs, and accumulated expenditure without progress. For evaluation the work comprises a trajectory over five quantum error correction questions and performs controlled comparisons on four mathematical problems with published solutions. The study compares the proposed node-level scheduling approach with question-level and model scheduling both with and without DAGs where it is found node-level scheduling achieves a higher budget-normalized score with DeepSeek-V4-Pro while question-level scheduling achieves higher with Claude Sonnet 5.

Critical strengths
  1. The allocation problem is well motivated and the corresponding formula is clear and intuitive. The proposed set of claim dependencies, next-step tasks and cross-question scheduling makes the idea of research progress informing decisions before a question is resolved as actionable and intuitive. Components like stagnation decay and mandatory OR branch for exploration address noticeable failure models.

  2. The research contains revealing analysis around failed approaches and unresolved dependencies. The descending-value setting shows how self-scheduling models may over-invest in a hard question with significant upside. For example DeepSeek without a DAG spends 88% of its research tokens on Q1, while both Claude self-scheduling conditions spend their entire budgets there without earning credit. The authors report settings where the proposed method loses and acknowledge sensitivity to the granularity of the DAGs.

Critical weaknesses
  1. The main claim is centred around the benefit of node-level allocation but the evidence towards this consists of only four mathematical problems and contains no reported uncertainty across repeated runs under 3 value settings. The 5 QEC questions have no baseline so they provide a qualitative case study rather than comparative evaluation. The question-level scheduling outperforms node-level scheduling with Claude Sonnet 5 while DeepSeek-V4-Pro wins in node-level settings making it difficult to establish a consistent advantage for the proposed granularity especially when considering the limited number of questions and lack of reported repeated-run results. The paper’s explanation that stronger research models favor node-level allocation remains difficult to justify given the limited results supporting this claim.

  2. The framing of the work is broader than its evaluation. The paper motivates the problem of deciding which open-ended scientific questions to pursue but the method applies to a fixed set of expert-selected questions with preassigned values focused on proof-oriented tasks. This is a reasonable way to keep the papers scope realistic but the paper should make this explicit during its motivation. Otherwise this transition makes it unclear how much of the original research-prioritization problem is addressed. The controlled experiments further depart from this motivation as they are applied to four known answer mathematical problems with ~1M output token budgets and evaluation milestones from published solutions.

Questions for authors
  1. How many independent runs were completed for each experiment condition and what was the variance across the runs?

  2. Is there any additional evidence that supports the claim that stronger research models benefit more from node-level allocation?

  3. How were the QEC questions verified as “closed”?

3 comments
AuthorsOpenReview ↗

Does stronger research models benefit more from node-level allocation?

We thank the reviewer for raising this question. Our intended application is a setting in which current AI models face open scientific questions at the research level and cannot necessarily solve all of them. Under a limited compute budget, the scheduler must identify questions that have high scientific value and offer a reasonable prospect of progress within the model’s capabilities. For example, in our QEC case study, the model continued to explore the fifth question but ultimately did not resolve it. We recognized that this question was particularly challenging for the model only after the experiment, when we reviewed the research results and scheduling process. The replay shows that, despite relatively active exploration of this question early on, the scheduler subsequently prioritized other questions and concentrated its resources on the fifth question after the other four had been resolved. This trajectory illustrates how the scheduling mechanism adjusts resource allocation in response to research feedback, providing a qualitative example of its role in research-level investigations.

However, even a single research iteration on these original scientific questions incurs substantial computational cost. Using them directly for controlled experiments involving more conditions and further exploration would require considerable time and funding. We therefore selected four mathematical problems that were more manageable in scope and had reference solutions, aiming to reduce experimental costs while retaining the key challenges of allocating research resources.

During model selection, we also conducted experiments with GPT-5.6sol and found that it could solve all four problems very quickly. For this model, the task set was too easy to fully examine the role of assessing task difficulty, comparing the value of further investment, and reallocating compute across questions. We therefore selected models that were less capable on this task set and still faced substantial difficulties, seeking to approximate a setting in which a model can make progress on some questions while other questions or research routes may exceed its current capabilities. In this setting, identifying valuable questions on which progress is feasible, and deciding when to continue an investigation or move to another question, become consequential scheduling decisions.

Our central consideration is therefore the relationship between model capability and task difficulty. By adjusting both the experimental tasks and the models used, we aim to study this resource allocation problem at a manageable cost. These observations from model selection primarily explain the rationale for our experimental design; they do not establish the general claim that stronger research models benefit more from node-level allocation. Determining how model capability affects the benefits of node-level allocation requires more systematic comparisons using tasks that remain appropriately challenging for the models being evaluated.

AuthorsOpenReview ↗

Verification of the Closed Results

We thank the reviewer for asking how the QEC results were verified. The five QEC questions were proposed by a coauthor who specializes in quantum error correction. After the AI completed its work, the results were reviewed by this coauthor. Only after this human review did we consider the first four questions to be “closed.”

For the fifth question, the model continued its investigation but did not produce a complete solution, so we retained its “open” status. Here, “open” means that the question remained unresolved within the research process and with the model and resources used in this study.

AuthorsOpenReview ↗

Repeated Runs on DeepSeek-V4-Pro

We thank the reviewer for asking about run-to-run variability. We supplemented the DeepSeek-V4-Pro experiments in following table with two additional runs per condition, giving three runs per condition in total. The updated table reports the mean and sample standard deviation, with the sample variances also provided below. The node-level scheduler has the highest mean score under the equal and descending value settings. Under ascending values, its mean score is very close to that of model self-scheduling without a DAG (1.710 versus 1.717). The run-to-run differences reflect the stochastic nature of the model’s research process. For example, different runs can generate different research routes in the first step, leading to different subsequent research trajectories. Even when attempting to prove the same subproblem, some runs produce a valid proof, whereas others contain errors and therefore receive no credit for the corresponding rubric milestone. These differences in route selection and proof success contribute to the variation in scores across runs.

Budget-normalized score R̄ (Eq. (5)). Each entry reports the mean ± sample standard deviation over three runs.

ModelValuesNode-level schedulerQuestion-level schedulerModel self-scheduling (DAG)Model self-scheduling (no DAG)
DeepSeek-V4-ProEqual (1, 1, 1, 1)0.99 ± 0.220.80 ± 0.160.62 ± 0.130.54 ± 0.09
DeepSeek-V4-ProAscending (1, 2, 3, 4)1.71 ± 0.150.51 ± 0.160.50 ± 0.101.72 ± 0.09
DeepSeek-V4-ProDescending (4, 3, 2, 1)0.63 ± 0.090.50 ± 0.110.28 ± 0.110.12 ± 0.02

Due to the limited budget, the multi-run evaluation currently covers DeepSeek-V4-Pro only. We will also try to complete the additional runs for Claude Sonnet 5 before the rebuttal period ends.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.