OpenNovelty.org

DIVERSE: Disagreement-Inducing Vector Evolution for Rashomon Set Exploration

ICLR 2026 Conference SubmissionAnonymous Authors
OpenReview Score: 6.0Download Report PDF
Rashomon SetRashomon EffectFeature-wise Linear Modulation (FiLM)CMA-ESModel MultiplicityPredictive MultiplicityNeural Network Diversity
Abstract:

We propose DIVERSE, a framework for systematically exploring the Rashomon set of deep neural networks, the collection of models that match a reference model’s accuracy while differing in their predictive behavior. DIVERSE augments a pretrained model with Feature-wise Linear Modulation (FiLM) layers and uses Covariance Matrix Adaptation Evolution Strategy (CMA-ES) to search a latent modulation space, generating diverse model variants without retraining or gradient access. Across MNIST, PneumoniaMNIST, and CIFAR-10, DIVERSE uncovers multiple high-performing yet functionally distinct models. Our experiments show that DIVERSE offers a competitive and efficient exploration of the Rashomon set, making it feasible to construct diverse sets that maintain robustness and performance while supporting well-balanced model multiplicity. While retraining remains the baseline to generate Rashomon sets, DIVERSE achieves comparable diversity at reduced computational cost.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper introduces DIVERSE, a framework for exploring the Rashomon set of deep neural networks by augmenting pretrained models with FiLM layers and using CMA-ES for gradient-free search. It resides in the 'Evolutionary and Modulation-Based Exploration' leaf, which currently contains only this paper as its sole member. This places the work in a notably sparse research direction within the broader 'Rashomon Set Exploration and Discovery Methods' branch, suggesting it addresses a methodological gap in how researchers systematically generate diverse yet accurate model variants without retraining.

The taxonomy reveals neighboring exploration strategies in sibling leaves: 'Ensemble and Dropout-Based Rashomon Sampling' contains four papers leveraging ensemble methods or dropout mechanisms, while 'Visualization and Interactive Navigation Tools' focuses on visual analytics for model comparison. DIVERSE diverges from these by emphasizing evolutionary optimization in a learned modulation space rather than ensemble aggregation or interactive user guidance. The broader parent branch excludes theoretical characterization (handled in 'Rashomon Set Characterization and Theory') and downstream applications (in 'Rashomon-Aware Model Selection'), positioning DIVERSE squarely as a discovery method rather than a theoretical or application-driven contribution.

Among twenty-four candidates examined through limited semantic search, none clearly refute the three identified contributions. The DIVERSE framework itself was assessed against ten candidates with zero refutable overlaps, while the gradient-free FiLM-CMA-ES combination examined nine candidates with similar results. The modulation space concept reviewed five candidates, again finding no clear prior work. These statistics suggest that within the examined scope, the specific combination of FiLM-based parameter modulation with evolutionary search for Rashomon set exploration appears novel, though the limited search scale (twenty-four papers, not hundreds) means undiscovered prior work remains possible.

The analysis indicates promising novelty given the sparse taxonomy position and absence of refuting work among examined candidates. However, the limited search scope and the paper's isolation as the sole member of its leaf warrant caution: broader literature beyond top-K semantic matches may reveal related modulation or evolutionary approaches. The contribution appears most distinctive in its specific technical combination rather than in addressing an entirely unexplored problem space, as the broader Rashomon exploration field contains multiple active research directions.

Taxonomy

Core-task Taxonomy Papers
50
3
Claimed Contributions
24
Contribution Candidate Papers Compared
0
Refutable Paper

Research Landscape Overview

Core task: Exploring the Rashomon set of deep neural networks. The field examines the phenomenon that many distinct models can achieve nearly identical predictive performance, forming what is known as the Rashomon set. The taxonomy reveals several complementary perspectives: theoretical branches characterize the structure and size of these sets (e.g., Infinite Rashomon Ratio[15], Rashomon Capacity[49]), while methodological branches develop techniques to discover and navigate them, ranging from evolutionary approaches (Multimodal Evolutionary[11]) to user-guided exploration (User Guided Rashomon[6], Personalized Rashomon[4]). Parallel branches address practical concerns such as model selection under multiplicity (Rashomon Balancing Methods[9], Rashomon Informed Decisions[12]), empirical studies across domains (Medical Rashomon[19], MXenes Rashomon[47]), and the implications for interpretability (Feature Interaction Cloud[23], Consensus Feature Attribution[45]) and fairness (Fairwashing Detection[32], Robust Recourse[44]). Additional branches capture related phenomena in inverse problems (PINN Multiple Solutions[7], Inverse Photonic Design[8]) and model equivalence through transformations (DNN Transformations[10], DNN Functional Equivalence[40]). Recent work highlights tensions between discovering diverse high-performing models and ensuring their practical utility. Some studies focus on sampling or enumerating the Rashomon set to reveal hidden trade-offs (Many Good Models[5], Simpler Models[1]), while others leverage this multiplicity for robustness or personalization (Robust Active Learning[17], Dropout Rashomon[25]). DIVERSE[0] sits within the evolutionary and modulation-based exploration branch, emphasizing algorithmic strategies to efficiently traverse the solution landscape—contrasting with more interactive approaches like User Guided Rashomon[6] or theoretical characterizations such as Unique Rashomon Sets[14]. This positioning reflects a growing recognition that understanding model multiplicity requires not only acknowledging its existence (Breiman Rashomon Occam[3]) but also developing scalable methods to explore and exploit it, bridging the gap between theoretical insights and engineering practice (ML Engineering Paradox[28]).

Claimed Contributions

DIVERSE framework for Rashomon set exploration

The authors introduce DIVERSE, a method that augments pretrained models with FiLM layers and uses CMA-ES to search a latent modulation space, generating diverse model variants without retraining or gradient access. This framework enables systematic exploration of functionally distinct models that maintain similar accuracy.

10 retrieved papers
Gradient-free exploration method combining FiLM and CMA-ES

The authors develop a novel approach that combines Feature-wise Linear Modulation layers with Covariance Matrix Adaptation Evolution Strategy to explore the Rashomon set without requiring model retraining or gradient computations, offering a computationally efficient alternative to existing methods.

9 retrieved papers
Modulation space for controlled hypothesis space exploration

The authors define a continuous modulation space by freezing network weights and FiLM projection layers while searching over latent vectors z. This design enables fine-grained adjustments to internal representations and defines a low-dimensional region of the hypothesis space where functionally distinct variants can be realized.

5 retrieved papers

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Within the taxonomy built over the current TopK core-task papers, the original paper is assigned to a leaf with no direct siblings and no cousin branches under the same grandparent topic. In this retrieved landscape, it appears structurally isolated, which is one partial signal of novelty, but still constrained by search coverage and taxonomy granularity.

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

DIVERSE framework for Rashomon set exploration

The authors introduce DIVERSE, a method that augments pretrained models with FiLM layers and uses CMA-ES to search a latent modulation space, generating diverse model variants without retraining or gradient access. This framework enables systematic exploration of functionally distinct models that maintain similar accuracy.

Contribution

Gradient-free exploration method combining FiLM and CMA-ES

The authors develop a novel approach that combines Feature-wise Linear Modulation layers with Covariance Matrix Adaptation Evolution Strategy to explore the Rashomon set without requiring model retraining or gradient computations, offering a computationally efficient alternative to existing methods.

Contribution

Modulation space for controlled hypothesis space exploration

The authors define a continuous modulation space by freezing network weights and FiLM projection layers while searching over latent vectors z. This design enables fine-grained adjustments to internal representations and defines a low-dimensional region of the hypothesis space where functionally distinct variants can be realized.

Table of Contents