Two years into ELOQUENCE, the project’s story is increasingly shaped by moments where ideas turn into shared tools, and collaboration turns into real impact. In this interview, the IDIAP team looks back on the advances that mattered most, the lessons learned from working across disciplines and cultures, and the questions that now define the future of responsible conversational AI.
Q: Which result, insight, or technological advancement do you feel has had the strongest impact on the project so far?
This year we saw SDialog evolve from early releases to a stable, documented, feature-rich open-source framework – with repeated version bumps, expanding capabilities, and publication backing. SDialog represents a concrete contribution of Idiap to open-source tools for conversational AI. SDialog is an MIT-licensed Python toolkit for “synthetic dialog generation and analysis with large language models (LLMs).” It provides a full end-to-end workflow: from building conversational agents (LLM-based), to simulating users, generating multi-agent dialogues, and evaluating them.
Its core aims are to:
- Standardize dialog schema (import/export in JSON) to make dialog data formats interoperable.
- Allow persona-driven multi-agent simulation (agents with contexts, “thoughts”, tools, etc.), orchestration (controlling flow/behavior), and flexible scenario configuration.
- Provide built-in evaluation tools (metrics, LLM-as-judge), and enable mechanistic interpretability (e.g. inspecting and steering per-token activations).
- Be backend-agnostic: i.e. interoperate with many LLM backends (OpenAI, HuggingFace, Ollama, AWS Bedrock, Google GenAI, etc.).
SDialog lowers the barrier for researchers and engineers to build, test, and benchmark conversational systems – especially when data is scarce or when one wants reproducible synthetic data at scale. As such, it can be an important building block for research in dialog systems, user simulation, evaluation, and interpretability.
The ELOQUENCE vision emphasises community collaboration: converting existing dialog datasets to SDialog format; contributing new components (personas, orchestrators, metrics, backend adapters); building benchmarks; improving interpretability; and sharing documentation/tutorials.
Q: What have you learned through ELOQUENCE that will be valuable for your future work or research?
Through ELOQUENCE, we have learned the value of working within a genuinely multidisciplinary, multicultural, and multilingual consortium, where scientific and industry partners bring complementary expertise and perspectives. The project has shown how essential continuous communication is – from regular technical meetings to informal exchanges – in creating an environment where collaboration becomes natural and feedback flows early and often. We have also experienced how a consortium can operate as a true support network: whenever one partner encounters challenges in advancing their work package, others step in with technical advice, resources, or alternative approaches. This spirit of collective problem solving, combined with exposure to diverse methodological backgrounds and real-world industrial constraints, has given us a deeper appreciation for how robust, responsible AI solutions emerge in practice. These lessons will be invaluable for our future work, shaping how we approach collaboration, multilingual AI development, and cross-domain research moving forward.
Q: From your perspective, what do you see as the most important opportunities and responsible AI challenges emerging right now? How do you think ELOQUENCE can help address them?
From our perspective, the most important opportunities right now lie in harnessing the growing power and sophistication of large language models (LLMs) to build multilingual, cross-cultural, and context-aware conversational AI that truly serves diverse populations, safety-critical domains, and under-resourced languages. At the same time, the rapid rollout of LLMs raises serious responsible-AI challenges: hallucinations, bias, lack of transparency, cultural or linguistic unfairness, lack of robustness in complex or critical contexts, and general “trust paradox” – their outputs may sound fluent and plausible but not necessarily accurate or aligned with real-world facts or diverse values.
ELOQUENCE represents a concrete and timely opportunity to address these challenges responsibly. By combining multilingual speech & language understanding technologies, hybrid knowledge-grounded models, and rigorous evaluation frameworks, ELOQUENCE aims to produce conversational agents that are explainable, trustworthy, bias-controlled, and aligned with inclusive European values, even when trained on limited data and operating across many languages. Its deliverables all contribute to building a foundation for socially responsible, robust conversational AI.
In short: the opportunity is to move beyond “toy/chatbot” AI and leverage LLMs for real-world, high-stakes, cross-cultural applications. The challenge is ensuring those systems remain safe, fair, transparent, and aligned. ELOQUENCE helps bridge that gap – not by ignoring the power of LLMs, but by adding layers of evaluation, grounding, multilingual fairness, and ethical oversight that are essential for responsible AI deployment today.
What emerges from this journey is a clear narrative: meaningful progress in conversational AI is built together. Open tools like SDialog, continuous dialogue across partners, and a strong commitment to multilingual fairness and transparency are helping ELOQUENCE move beyond experimental systems toward AI that can be trusted in real-world settings. ELOQUENCE shows that responsibility is not a constraint on innovation, but the path that makes it last.
