acceptodds
Under review as a conference paper at ICLR 2027

Mechanova: An Autonomous Scientist for Discovering Falsifiable Mechanisms of Immune-Cold-to-Hot Conversion in Tumors

Abstract

Autonomous scientific discovery requires agents to revise causal explanations in response to evidence and translate those revisions into new experiments. We introduce Mechanova, an autonomous scientist for discovering falsifiable mechanisms of immune-cold-to-hot conversion in tumors. Mechanova couples competing causal graphs to experimental evidence and counterevidence, selects discriminating experiments, revises mechanistic structure in response to prediction–observation discrepancies, and compiles surviving explanations into testable intervention programs. Its Mechanism Evidence Chain (MEC) grounds causal claims and licenses structural edits through context-matched experimental evidence. Biological and operational world models predict intervention effects and research-action outcomes, respectively. Provenance-aware memory and observation-conditioned hierarchical control sustain adaptive, long-horizon investigation, while offline policy learning improves execution across investigations. In a comparison on 125 paper-context tasks, the rule-assisted system achieves 60.8% mechanism identification and 52.0% pre-observation intervention agreement, exceeding the strongest completed comparator for each endpoint by 19.6 and 12.4 percentage points. A separate matched-prompt study on 88 source families isolates structural editing under identical downstream probability and action rules: enabling edits reduces native Brier by 25.3%, from 0.1119 to 0.0835 (paired difference −0.0284; 95% CI [−0.0439, −0.0146]). Graph-replacement controls further demonstrate that downstream outputs depend on the revised mechanism. Supporting studies raise evidence recall from 25.2% to 36.8% with zero measured unsupported assertions and useful execution completion from 50% to 100%. Mechanova establishes a mechanism-centered approach to autonomous scientific discovery in which evidence governs causal claims, structural revision improves prediction, and surviving explanations generate explicit experimental commitments.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
AI scientist, autonomous scientific discovery, agentic AI, mechanistic reasoning, causal mechanism revision, falsifiable hypotheses, cancer immunology, tumor microenvironment, immune-cold-to-hot conversion

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_spZYOpenReview ↗
Rating
1
Confidence
4(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work introduces Mechanova, an agent-based framework designed to study how immune-cold tumors can be converted into immune-hot tumors. It arranges evidence drawn from the literature into competing causal graphs, chooses informative experiments, updates causal relationships according to observed results, and produces intervention plans. The framework also incorporates statistical predictors for biological responses and research-action outcomes, along with evidence management and hierarchical agent execution

Critical strengths

The paper addresses an interesting problem of evidence-grounded scientific reasoning. The notable strength is the controlled evaluation of structural revision, where the authors use matched experimental settings and graph-substitution tests to assess whether revised mechanisms improve downstream predictions. The reported improvement in Brier score provides meaningful evidence for the value of structural editing. Also, the Mechanism Evidence Chain offers a systematic way to track experimental evidence and constrain causal claims, improving the interpretability of the proposed reasoning process.

Critical weaknesses

The paper's methodological positioning is unclear, and its evaluation does not convincingly support its central claims. Although Mechanova is presented as a unified autonomous scientific discovery framework, most components are evaluated independently on different tasks, without demonstrating their necessity or collective contribution to scientific discovery. The main benchmark shows only modest improvements over a simple title-only baseline (5.6% points), while the structural-revision experiments primarily demonstrate improved probabilistic prediction rather than causal correctness. In the archival intervention evaluation, Mechanova reports 10/10 successful intervention programs but 9 cases remain structurally unchanged, and success is assessed against curated historical outcomes through intervention-class matching. The results provide limited evidence that the proposed mechanism-revision approach actually contributes to improved intervention design.

Besides, the roles of the Biological and Operational World Models remain insufficiently justified. The BWM is not evaluated in the structural-revision pipeline, and the OWM is mainly evaluated on operational efficiency but not how it improves mechanism discovery. The supplementary Replogle experiments compare existing perturbation-prediction model configurations but have no clearly demonstrated connection to the central tumor immune-cold-to-hot task or the proposed causal reasoning framework. These experiments appear to be primarily comparisons of single-cell representations (HVG-based versus STATE embeddings), but not the method's contribution to scientific discovery.

Questions for authors

How are the Biological and Operational World Models integrated into the core mechanism-revision pipeline, and how their predictions influence causal revision or downstream intervention selection?

What is the intended contribution of the Replogle experiments?

How structural revision contributes to more accurate causal mechanisms or better intervention decisions?

Could the authors better explain the contribution of evidence-grounded reasoning to mechanism identification?

Reviewer_STRTOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper presents Mechanova, an LLM-based agent meant to act as an autonomous scientist for one question: why some tumors do not trigger an immune response, and how to change that. The agent represents competing explanations as causal graphs built from the literature, picks experiments to tell them apart, revises the graphs when results disagree with its predictions, and finally suggests a treatment combination. On author-built tasks derived from published studies, the paper reports gains over existing research agents, and further gains from allowing graph edits.

Critical strengths
  1. The paper addresses a problem: it aims to let the agent revise its causal explanations in response to evidence, rather than only rank fixed hypotheses. Stating predictions before results are seen is also a reasonable design choice.
  2. The paper tries to isolate the effect of graph editing with a controlled comparison. It also takes steps to avoid information leakage, such as hiding the target paper from retrieval.
Critical weaknesses
  1. Lack of end-to-end integration. The authors introduce MECHANOVA as a full autonomous scientist with a WHAT–HOW control protocol. However, from my understanding, there is no clear end-to-end integration of these two parts. I did not see how BWM and OWM are actually integrated into the scientific discovery cycle. It is really confusing how the overall pipeline works.
  2. Limited novelty in hypothesis generation. I really doubt the novelty of the generated hypotheses. According to the paper, the hypotheses are mainly generated by recombining existing causal relations extracted from the literature. If the hypothesis space is constrained by existing knowledge, how can the system generate novel hypotheses involving mechanisms or relations that are missing from existing studies?
  3. Highly constrained discriminating tests. The discriminating test component is also highly constrained. From my understanding, no new biological experiments are conducted in the evaluated pipeline. What the authors actually do is select from a predefined set of experiments and use historical observations to discriminate between hypotheses. This is quite different from open-ended scientific discovery. Also, the graph-to-readout rules are predefined, and their reliability is unclear. How are these rules validated, and how would the system handle inaccurate predictions?
  4. Difficulty in distinguishing different sources of errors. Another problem is that there are two possible sources of errors: (1) errors in the proposed causal mechanisms and (2) prediction errors from the graph-to-readout mapping in the discriminating test component. When the predicted outcome differs from the actual observation, how can the system distinguish the first type of error from the second? If the prediction model itself is inaccurate, the system may incorrectly revise an otherwise valid mechanism.
  5. Limitations of the experimental evaluation. The paper evaluates the proposed system entirely on author-constructed tasks, without testing on established benchmarks or sufficiently validating the constructed benchmarks. Moreover, the baseline comparison is not well controlled: external agents are not designed for this specific task, their outputs are truncated and processed through an additional LLM projector, and their compute budgets are not aligned. Finally, the main evaluation lacks comprehensive ablations, making it difficult to attribute the reported improvements to any proposed component.
  6. Poor presentation and organization. 1) The paper is difficult to understand without extensively consulting the appendix (e.g. task definitions, evaluation protocols, and key baselines). 2) Numerous newly introduced terms are insufficiently defined, and internal code identifiers appear directly in the manuscript, making parts of it read more like an engineering report than a research paper. 3) The experimental section compresses roughly a dozen separate studies into only a few pages, often presenting experimental counts before explaining the tasks themselves. 4) Figure 1 is overly dense and difficult to read, and the paper lacks a dedicated discussion of its limitations.
Questions for authors

see weakness

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

This paper proposes a new method of autonomous scientific discovery of immune-cold-to-hot conversion in tumours. Their approach constructs causal chains to ground the experiments proposed by the autonomous scientist. New experimental proposals are aware of historical edits. The authors report strong improvements over baselines.

Critical strengths
  • The topic is interesting and valuable.
  • The causal analysis perspective on the problem is interesting.
Critical weaknesses
  • The introduction is unclear to me. I found it difficult to follow, understand the main contributions of the paper, and understand the relation to prior work.
  • It is unclear to me how Mechanova expands on existing work. It seems that Mechanova is more focused on answering causal questions; however, previous methods mentioned in the Related Work section are also trying to answer causal questions. Hence, it is unclear how Mechanova is different.
  • The theoretical section is not well explained. For example, none of h, id_i, and v from equation 1 are defined. And the other variables are not well-explained. E.g., B_i is the "bottlenecks" of G. But exactly what does this mean? Mathematically, how are we representing a "bottleneck"? Similarly, Epsilon^{+,-,?} is defined as "the supporting, opposing, and unresolved evidence." But how are we defining "evidence?" Are these vectors? Booleans? Strings? It is not clear to me. In general, the whole section is underspecified and lacking clear definitions. These issues persist with equations 2, 3, and 4.
  • The experimental setup is not clear. The tasks are not described within the main paper, and the experimental procedure is not clear to me. Appendix F describes one data set of paper assessments, but I do not understand how these are used (it is not described within the paper). Line 323 says, "Mechanova achieves 60.8% mechanism and 52.0% intervention agreement." But intervention agreement is not yet defined. As the reader, I do not understand what the authors are measuring. The absence of an experimental setup description persists throughout Section 4. In the absence of a description of the experiments, I cannot assert their correctness, validity, or the strength of their claims, and I must assume a pessimistic view.
  • The introduction and title seem to promise an autonomous scientist; however, I see experimentation of a scientist performing experiments autonomously. I believe the experiments do not support the claims made by this paper.
Questions for authors
  • Where can I see a description of the experimental setup?
  • How were baselines chosen?
  • How does the theoretical formalism presented in this paper compare to that presented in previous papers?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.