acceptodds
Withdrawn from ICLR 2027

Expectation Reasoning for Exploring Functional Forms of Dynamics

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Network reconstruction becomes particularly challenging when both the interaction structure and the dynamics governing the observations are unknown. We investigate whether an LLM-driven scientific process can explore effective dynamical response laws that improve reconstruction under a fixed graph-reconstruction backend. A closer examination of the research process reveals two limitations of proxy-driven optimization. First, improving the fitting proxy does not necessarily improve downstream network recovery. Second, the proxy-optimized realization provides insufficient evidence to determine whether a proposed dynamical mechanism is actually realized, whether it has a meaningful effect, or whether it remains worth investigating. Thus, successful proxy optimization can support numerical realization without providing an adequate basis for evaluating the scientific status of the proposed mechanism. We introduce Expectation Reasoning (ER), which separates numerical realization, evidence formation, epistemic revision, and solution commitment within the scientific-search process. Across two truth-blind CausalDynamics comparison cohorts totaling 20 cases, the recorded ER trajectories preserve contradictory, partial, and unresolved evidence across research rounds and exhibit a clearer separation between epistemic revision and candidate commitment. Downstream effects remain criterion-dependent: ER reduces fixed-threshold structural error most clearly in the primary cohort, while AUROC and AUPRC do not improve and the resulting reconstructions become substantially sparser. Taken together, our findings support separating scientific evidence and revision from proxy-driven numerical realization, rather than treating optimization outcomes as sufficient scientific decisions.

Primary area:
learning on time series and dynamical systems
Keywords:
network reconstruction, dynamical systems, interaction law discovery, structure learning, scientific discovery, large language models

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_rJyEOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper aims to solve a classic problem inferring the mechanics and topology from the dynamics observation. This is a very challenging problem. This paper proposed to introduce LLMs to leverage their ability in symbolic regression, and then fit the functions of network dynamics and recover the network topology. Based on this, this paper has two negative findings and try to introduce a pipeline empowered by the idea of auto scientific discovery, to adjust the functions by the loop. The paper reported the results from both positive and negative perspectives.

Critical strengths
  1. The paper explores a potentially valuable direction by investigating whether LLMs’ capabilities in symbolic regression can be transferred to the problem under study.
  2. The authors are transparent about the negative results, explicitly acknowledging the two limitations discussed in the paper.
Critical weaknesses
  1. The paper does not meaningfully advance research on the problem it addresses. Although reporting negative results is valuable, the findings are largely unsurprising: decoupling network dynamics from network topology restricts the search space, making the approach more locally optimal solutions.
  2. The proposed solution to the identified limitations also appears to have limited effectiveness. Moreover, its description is vague and remains largely at the conceptual level, making it difficult for readers to understand how the method is implemented in practice.
  3. The paper does not clearly describe the experimental implementation, the precise problem formulation, or the meaning of the evaluation metrics. These omissions make it difficult to interpret the experimental results in depth.
Questions for authors

Please refer to the weakness.

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

This paper studies proxy-driven discovery of causal systems via LLMs whereby the LLM proposes a candidate function whose parameters are fitted using observable data. In particular, the authors highlight two issues with this approach:

  1. Minimizing MSE over the observable data does not guarantee reconstruction of the true causal graph (in fact the correlation is often negative)
  2. Minimization frequently suppresses non-linear components of the proposed model (e.g. |f'(z)| is constant) To address this, authors propose a method forcing LLM's to pre-register hypotheses accompanying their proposed function classes which can be tested in addition to minimizing parameters via MSE loss. Multilple rounds of testing are done at which point the model commits to a single function class
Critical strengths
  • The authors convincingly show numerical fitting frequently minimizes MSE by finding flat, inactive parameter regimes (suppressing the proposed nonlinear mechanism in ~30% of cases) or by trading off functional gain.
Critical weaknesses
  • Although the proposed method improves SHD, it gets both worse precision and recall. Much of the gain from SHD is potentially due to fewer edges being proposed in highly sparse graphs
  • A simple baseline regularizing the MSE loss to prompote more sparsity could also be considered
Questions for authors
  • can the authors include the specific prompts used in the appendix + an example of the agent workflow?
  • can you try a baseline with the standard MSE loss + a l1 penalty on G_off?
Reviewer_AYYQOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the inconsistency between proxy-based fitting objectives and actual scientific goals in LLM-guided response-law search, proposing the Expectation Reasoning (ER) framework. This method separates numerical realization, evidence formation, epistemic revision, and solution commitment. It continuously refines the Research State by specifying pre-outcome Expectations and incorporating experimental results. The paper validates the method on two truth-blind CausalDynamics cohorts totaling 20 cases.

Critical strengths

S1. The paper distinguishes between two types of mismatches: proxy–task mismatch and proxy–mechanism mismatch. It proposes introducing pre-outcome Expectations and targeted experiments in addition to numerical fitting, providing a more granular hypothesis verification mechanism for LLM-driven scientific exploration.

S2. ER separates evidence accumulation, epistemic revision, and final candidate selection, allowing the retention of contradictory, partially supporting, and unresolved evidence without forcing immediate replacement of candidate solutions. This design enhances the traceability of automated research processes and provides a new approach to uncertainty management in LLM scientific reasoning.

Critical weaknesses

W1. The core methodology has limited innovation and its distinction from existing automated scientific hypothesis testing methods is not clearly defined. Existing methods like POPPER generate testable expectations and execute sequential falsification experiments using an LLM agent, and HEP also proposes an auditable research process based on hypothesis, evidence, and belief updates. In contrast, ER primarily reorganizes numerical fitting, evidence recording, and final decision-making without proposing a clear new inference algorithm or reliability guarantee. While the coupling of agent optimization and graph re-estimation has certain unique characteristics, the paper has not fully demonstrated that this process separation brings capabilities that existing methods cannot achieve.

[1] Huang et al., Automated Hypothesis Validation with Agentic Sequential Falsifications, ICML 2025.

[2] Takahara and Mizoguchi, Toward Auditable AI Scientists: A Hypothesis Evolution Protocol for LLM Agents, 2026.

W2. The experimental design cannot adequately demonstrate the independent contributions of each module in ER. ER and the Inherited process use different numerical optimization budgets, allocating a 176-trial fixed numerical base plus additional targeted trials and 256 trials, respectively, for each Function Family. Furthermore, the candidate pools generated by the two methods are not identical. The paper also lacks independent ablation analysis of pre-outcome Expectations, evidence revision, and final decision separation. Therefore, the existing results cannot determine whether the performance changes stem from the new inference mechanism, the different candidate generation process, or differences in experimental resource allocation.

W3. Improvements in structural error are mainly accompanied by excessive sparsity, and the actual network recovery capability lacks persuasiveness. In CD011–CD020, the number of off-diagonal edges predicted by ER decreased from 193 to 13, and the number of true positives decreased from 57 to 3, with an off-diagonal recall rate of only 0.33%. Its off-diagonal SHD was 90.7, even slightly worse than the empty off-diagonal SHD reference of 90.0. Meanwhile, the AUROC and AUPRC of both experimental groups were lower than the original method, and the improvement in SHD in the second group was not significant. This indicates that the reduction in structural error mainly reflects a reduction in false positives, and cannot prove that ER more accurately recovers the true network relationships.

Questions for authors

Q1. Compared with existing scientific agent frameworks such as POPPER and HEP, what new reasoning mechanism does ER propose? Besides separating the recording of hypotheses, evidence, and final decisions, can the authors prove that this design brings a substantial improvement in scientific reasoning ability?

Q2. Why was a control group not conducted with the same set of candidate functions and computational budget?

Q3. Can the authors remove the pre-outcome Expectations, Expectation–Reality comparison, and Research State update separately to verify the independent contribution of each module to the final result?

Q4. When the off-diagonal recall of ER is only 0.33% and the SHD is close to the empty off-diagonal SHD reference, how can it be proven that it truly improves network reconstruction? Can the authors compare it with the empty off-diagonal SHD reference and the original method after threshold calibration, and report the Precision–Recall tradeoff?

Q5. The paper reports 100% same-question correspondence between Expectations and Reality records, but does this only indicate that the system correctly executed the pre-defined process, rather than that the scientific judgment is more accurate? Especially in CD029, ER retained several unresolved issues but still selected a candidate with an SHD of 273, while the best value in the same candidate pool was only 130. How can the authors ensure that the evidence of uncertainty truly constrains the final decision?

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

Related papers

Discussion (0)

Sign in to comment.