acceptodds
Under review as a conference paper at ICLR 2027

TraceSeek: Scale-Aware Evidence Seeking for Progression Learning from Single-Cell Population Snapshots

Abstract

Disease processes are often observed through ordered population snapshots, where multiple progression explanations can remain compatible with the same data. We introduce TraceSeek, a framework for progression learning that, under a fixed Scientific Query Program, selects which atlas view to inspect and at which inferential unit a claim may be supported. TraceSeek couples a progression-claim layer with a typed analytical topology, a query-conditioned budgeted controller, and surrogate Marginal Utility Distillation (MUD), enabling evidence refinement, comparison, aggregation back to valid units, and abstention when evidence remains insufficient. In controlled evaluations across eleven disease atlases with a shared typed tool interface and a deterministic query compiler, TraceSeek improves scale-selection diagnostics, reduces invalid-unit and false-support proxy rates, and achieves higher mean family-specific Task Score than ReAct and other tool-using agents. On a 54-episode multi-atlas progression-chain suite, TraceSeek improves Task Score from 0.606 to 0.840 and path completeness from 0.500 to 0.917 relative to ReAct. Ablations show that topology constraints, active scale selection, and MUD contribute materially to these gains. An Alzheimer's disease case study further shows stable microglial early-response localization under leave-one-donor-out analysis and stage-associated DAM structure.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
single-cell genomics, disease progression, evidence seeking, multi-scale analysis, scientific agents

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.25 over 4 reviews, above 2% of the others.
Reviewer_bLMYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TraceSeek, a framework for disease progression inference from ordered single-cell population snapshots. It formulates the problem as budgeted evidence seeking over a typed analytical topology and uses a scientific query program to constrain the query process. The method includes a progression-claim layer that maintains support, conflict, exclude, and abstain decisions; a typed analytical topology that uses action masks to rule out invalid inferential unit transitions and avoid pseudoreplication; a budgeted query-conditioned controller that actively selects views, inferential units, operators, and fidelity; and marginal utility distillation that uses cached counterfactuals to train the controller offline. Validation is conducted on multiple disease single-cell atlases and task families, comparing against baselines such as ReAct, and reporting improvements in task score, path completeness, and inferential unit accuracy. Ablations, transfer evaluation, and an Alzheimer's disease case study further demonstrate the potential of the method.

Critical strengths
  1. Inferential unit validity and scale selection are enforced by the environment, not by prompts. The paper uses a typed analytical topology and action masks to rule out illegal inferential unit transitions before decision making, preventing cell-level observations from being treated as donor-level independent evidence, and explicitly records support, conflict, exclude, and abstain. Pseudoreplication and invalid aggregation are core statistical failure modes in single-cell disease progression analysis. Even if the absolute scores need independent replication, this formulation already provides a reusable audit framework for inference.

  2. Scale and unit selection are learned under budget constraints, with ablations supporting each component. The active scale controller makes sequential decisions over views, inferential units, operators, and fidelity, and is trained offline with marginal utility distillation. Ablations show that removing topology constraints, active scale selection, or MUD all cause clear drops. This turns evidence path selection from a heuristic into a learnable policy and explains why the method may outperform general tool-using baselines such as ReAct.

Critical weaknesses
  1. Limited independent validation of scientific correctness beyond path-based scoring. Several Task Score components explicitly reward predefined evidence paths, and the MUD objective includes handcrafted family, unit, and teacher-action bonuses. This raises the concern that reported gains partly reflect alignment with the benchmark's structural priors rather than improved scientific claim correctness. In addition, the progression-chain evaluation contains many episodes derived from only a few unique queries, and paired uncertainty at the independent-query level is not reported, which reduces confidence in the paper's central reliability claim.

  2. Limited evidence for the distinct algorithmic contribution of MUD. MUD is closely related to cost-aware sequential decision-making and contextual bandit approaches, but the paper does not clearly establish its distinction from these paradigms or compare against a matched cost-aware learning baseline. The typed topology, hard action masks, and fixed query program already encode strong scientific constraints. Although ablations support individual components, it remains unclear how much additional benefit MUD provides over established decision-making methods under the same interface and training data. This also limits the evidence for end-to-end agent applicability.

Questions for authors
  1. Task Score partly rewards predefined evidence paths, and MUD uses handcrafted utility bonuses. Can you provide an evaluation of scientific claim correctness independent of path scoring, or show whether false support and inappropriate abstention improve after controlling for path completeness?

  2. The progression-chain evaluation contains many episodes derived from only a few unique queries. Can you report paired performance differences and confidence intervals at the level of independent queries, rather than repeated evaluations of the same query?

  3. The claim layer and abstention are central components, but the main atlas evaluation does not clearly isolate their contributions. Can you report coverage-risk, false-support, and abstention rates, together with an ablation removing the claim layer under the same evaluation setup?

  4. What distinguishes MUD from cost-aware contextual bandits, budgeted sequential decision-making, or active hypothesis testing? Can you compare it with a cost-sensitive learned policy trained on the same cached observations and action space to better isolate its algorithmic contribution?

Reviewer_7DtPOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Focusing on progression learning from signle-cell population snapshots, this paper proposes a framework, TraceSeek,which can selects which atlas view to inspect and at which inferential unit a claim may be supported, under a fixed Scientific Query Program. The TraceSeek consists of a progression-claim layer , a query-conditioned budgeted controller, and surrogate Marginal Utility Distillation. The paper conducts the experiments across eleven disease atlases, showing that TraceSeek outperforms ReAct and other tool-using agents.

Critical strengths
  1. The framework proposed in this paper, TraceSeek, is quite meaningful, aiming to solve a critical problem in progression learning from single-cell population snapshots.
Critical weaknesses

1.The motivation of the paper lacks a clear hierarchical structure (writing issues) The problem, "Multiple compatible explanations", is raised in the abstract. The problems,"Fixed coarse summaries" and "Naive fine-grained tests" are both raised in the first paragraph of Introduction. This two problems are mentioned again with "Exhaustive exploration" to introduce Figure 1,where they are summarized as "Limits of fixed analysis",which I think does not clearly describe all three failure modes,particularly exhaustive exploration.

2.The Explanation of "Progress ambiguity" is not enough In Figure1(a), there are three population snapshots(Early, Intermediate, Late) connected by two arraws. So why three ordered snapshots can lead to Multiple compatible explanations? Since this is the main problem mentioned in the first sentence of the abstract, it would be better to provide a clear and accessible explanation.

3.The construction of the typed analytical topology is unclear. How to get the E\mathcal{E} in H\mathcal{H} from a scientific dataset? I didn't find clear explanation in 3.2 Problem Formulation and Appendix A: Datasets and preprocessing.

4.Symbol duplication(Writing Issue) The symbol denoting biological scale is the same as the symbol denoting trajectory, both are τ\tau.

5.It is unclear how the SQP induces a typed mask over candidate actions. The paper only introduces an example in Section 3.3 ,but does not provide a detailed explanation of the rule used to obtain Atq\mathcal{A}_t^q

6.Lack of explanation on how surrogate utility target used for the main-table controller is computed. Where is the definition of ΔR\Delta R? Where are the calculation rule of the three bonus?

7.One-Step Utility.I consider that the main task of the problem is a long-term one. However, the surrogate utility target seems to be one-step utility.

8.The experimental setup is insufficiently clear. In the ablation experiments, "w/o Active Scale Selection","w/o Topology Constraints","w/o Inference Constraints",and the "w/o Aggregate-Back",but there is no detailed explanation of what is actually being ablated. These terms are mentioned only once or twice in the Experiment section, without any explanation.

Questions for authors
  1. How to get the E\mathcal{E} in H\mathcal{H} from a scientific dataset?
  2. What do "w/o Active Scale Selection","w/o Topology Constraints","w/o Inference Constraints",and the "w/o Aggregate-Back" mean?A detailed explanation of how they are implemented in the experiments is expected.
  3. How does the surrogate utlity target calculate,expecially the Δ(R)\Delta(R) and three bonus?And what is the reason for designing it this way?
Reviewer_PXtXOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper addresses the following problem: Given only snapshot data from different patients at different stages of a disease, how can a system automatically identify relevant evidence and assess the extent to which it supports a hypothesis about disease progression?

The proposed method provides an end-to-end workflow that starts with a question, introduces a new loss function to estimate the value of candidate actions and guide action selection, and ultimately produces a conclusion.

Strengths

  1. The paper presents a complete end-to-end system that covers data processing, information retrieval, and conclusion generation.

Weaknesses

  1. The benefit of using a graph is unclear. The graph component appears to serve primarily as a data preprocessing step, and the paper does not describe the processing procedure in sufficient detail to support reproducibility or reuse.
  2. The action formulation essentially maps an input to a decision about which data to retrieve. However, the paper does not adequately address complex scenarios. Given the relatively shallow network architecture, stronger justification and empirical evidence are needed to establish its effectiveness in such settings.
  3. The contribution to action selection appears limited to a function for evaluating candidate actions, without further methodological innovation. The novelty and practical benefit of this function are not sufficiently established.
  4. The current experiments do not isolate or demonstrate the specific benefits of the proposed system orchestration.
Critical strengths

The paper presents a complete end-to-end system that covers data processing, information retrieval, and conclusion generation.

Critical weaknesses
  1. The benefit of using a graph is unclear. The graph component appears to serve primarily as a data preprocessing step, and the paper does not describe the processing procedure in sufficient detail to support reproducibility or reuse.
  2. The action formulation essentially maps an input to a decision about which data to retrieve. However, the paper does not adequately address complex scenarios. Given the relatively shallow network architecture, stronger justification and empirical evidence are needed to establish its effectiveness in such settings.
  3. The contribution to action selection appears limited to a function for evaluating candidate actions, without further methodological innovation. The novelty and practical benefit of this function are not sufficiently established.
  4. The current experiments do not isolate or demonstrate the specific benefits of the proposed system orchestration.
Questions for authors
  1. What is the most complex input scenario considered by the paper’s action selection mechanism?
  2. The paper does not describe the overall algorithm in sufficient detail to make its specific design clear.
Reviewer_wkSrOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

It was very difficult to get a concrete idea of what this paper actually did, from both the abstract and introduction.

Critical strengths

It is very hard to see the strengths.

Critical weaknesses

The writing is too unclear to understand what is going on.

Questions for authors

This paper seems to introduce a system with a particular capability. Could the authors explain in one paragraph what the target audience is, who would benefit from the tool that is created, and how this would improve their workflow?

It is possible that I have missed the value of this paper. A clear and concise answer to this question would help me a lot.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.