NeurIPS has long been the stage where foundational questions are challenged, new empirical frontiers are revealed and where the next era of AI and machine learning begins.
This year, Solvd joins the global AI research community with work that pushes forward three essential areas of modern machine learning: scalable reinforcement learning, explainable generative modeling and architectures that adapt flexibly to real-world constraints.
Across two research papers and participation in the Generative AI in Finance Workshop, Solvd’s work highlights a central message: AI systems must be interpretable, robust and capable of reasoning across long horizons—and scaling is a key enabler to that progress.
DiCoFlex: Model-Agnostic Diverse Counterfactuals with Flexible Control
Counterfactual explanations play a central role in explainable AI, helping users understand how a model’s prediction might change under different conditions. But most existing techniques come with familiar limitations: they require continuous access to the underlying model, involve slow optimization for each example and cannot easily adapt to new constraints without retraining.
DiCoFlex introduces a new approach: a model-agnostic, conditional generative framework that produces multiple diverse counterfactuals in a single forward pass. Built using conditional normalizing flows trained only on labeled data, DiCoFlex enables:
- Fast, single-step counterfactual generation
- Real-time, user-controlled constraints (such as sparsity or actionability)
- Model-agnostic deployment with no access to internal weights
- High diversity and validity, outperforming existing baselines
This work demonstrates a practical, scalable direction for counterfactual explanations—one that adapts to real-world demands and supports more transparent, interactive AI systems. This research was delivered by Oleksii Furman, Ulvi Movsum-zada, Patryk Marszałek, Maciej Zieba and Marek Śmieja representing Wrocław University of Science and Technology, Jagiellonian University, Tooploox and Solvd.
1000 Layer Networks for Self-Supervised RL: Scaling Depth Can Enable New Goal-Reaching Capabilities
Reinforcement learning has traditionally relied on shallow architectures, constrained by the belief that RL signals are too sparse to guide large models effectively. Solvd’s second NeurIPS contribution challenges that assumption directly.
In 1000 Layer Networks for Self-Supervised RL, our research team demonstrates that very deep neural networks—up to 1,024 layers—can be trained stably and effectively in a self-supervised, goal-conditioned RL setting. Using contrastive objectives and GPU-accelerated simulation, the work reveals that:
- Scaling depth leads to dramatic performance improvements, from 2X to over 50X in complex locomotion, navigation and manipulation tasks.
- Emergent behaviors appear at critical depth thresholds, including upright walking, maze-aware navigation and even acrobatic strategies impossible for shallow models.
- Deeper networks learn richer representations, capturing environmental structure more accurately and enabling better long-horizon reasoning.
- Depth outperforms width, achieving better results with far fewer parameters and unlocking the benefits of larger batch sizes.
This research provides empirical evidence that reinforcement learning can scale — and that scaling unlocks qualitatively new capabilities.
This research was delivered by Kevin Wang, Ishaan Javali, Michał Bortkiewicz, Tomasz Trzciński and Benjamin Eysenbach, representing Warsaw University of Technology, Princeton University, Tooploox and Solvd. We are also honored to be recognized for this research with the “Best Paper Award” at NeurIPS 2025.
Outcome
Both papers were presented at NeurIPS 2025 in San Diego, California, alongside Solvd’s participation in the Generative AI in Finance Workshop, which explored the practical implications of explainable and scalable AI systems in financial applications.
Together, these efforts build on the conversations Solvd advanced earlier at Money20/20 around responsible AI adoption, long-horizon reasoning, and translating research progress into real-world enterprise impact.
The papers can be found on arXiv and in the official NeurIPS proceedings, continuing Solvd’s mission to drive AI research that is rigorous, transparent and deeply connected to real-world needs.
