Twitxy Saved

Daniel Domínguez

Verified account This account is verified. Learn more

@danidominguez

AI pedagogical alignment + digitally mediated learning. Professor @uned. Visiting Scholar @AIResponsibly @nyuniversity. Fellow @EDEN_DLE. Member @CoLabUNED.
Madrid daniel-dominguez.com Joined July 2008
510 Following 982 Followers 4.1K Posts

Latest posts 5

Pinned post

Daniel Domínguez Verified account @danidominguez
Slides from today's presentation at @AIResponsibly on analyzing patterns of pedagogical drift in LLMs. We used a mixed approach in a two-step process, combining a behavioral audit via CoT first to inform mechanistic diagnostics and causal interventions for mitigating misalignment. Paper and associated datasets in progress 🚀
drive.google.com/file/d/16lkjUI…
drive.google.com LLMs_Pedagogical_Alignment_NYU_March 26_2026.pdf
Daniel Domínguez Verified account @danidominguez
Glad to see our work on pedagogical alignment of LLMs highlighted in the latest @AIResponsibly newsletter. Without explicit auditing of how models teach, we risk deploying systems that are accurate but pedagogically rigid. Newsletter, mailchi.mp/nyu/rai-sep202…
Daniel Domínguez Verified account @danidominguez Dec 15, 2025
Excited to share our new paper on the pedagogical alignment of LLMs with @jlfvindel & Laura M. Guerrero, just published in Education and Information Technologies!

As LLMs become more integrated into learning environments, we have become adept at using benchmarks to evaluate their safety and general knowledge capabilities. Yet, we largely ignore the pedagogical logic driving these systems. While we understand the models' outputs, we don't fully grasp how they teach. We view this evolution in the field as a necessary shift, changing the question from "Is the model correct?" to "Is the model teaching effectively?"

In this paper, we argue that standard benchmarking is no longer sufficient and that there is a need to make pedagogical alignment explainable. And we argue that a model is not "aligned" unless it can adapt its instructional strategy to the learning needs of each user and the circumstances of each learning context.

To address this issue, we conducted two studies analyzing GPT-4’s performance as a tutor (via Khanmigo) and an instructional designer. We examined whether an LLM’s behavior aligns with established learning theories, such as active learning, behaviorism, and social learning. The results reveal a persistent "behaviorist bias" across both contexts.

In Study 1, while students frequently prompted the AI for active learning support, seeking critical thinking and inquiry-based scaffolding, the model consistently defaulted to behaviorism. Nearly 72% of the model's outputs aligned with behaviorist principles (structured reinforcement), compared to ~24% for active learning and ~10% for social learning.

In Study 2, we evaluated the model as an instructional designer using our Pedagogical Alignment Distance (PAD) metric. We found that even when we enriched the prompts with specific "skills" data, the model only improved its alignment with behaviorism. Scores for active and social learning remained negative, showing that simply adding more educational data does not automatically improve pedagogical flexibility.

Together, these findings suggest that without explicit pedagogical auditing, we risk deploying systems that are "safe" but pedagogically rigid—favoring rote instruction over the discovery and collaboration users actually seek.

📄 Full paper: Explainable pedagogical alignment: auditing LLMs with learning theories, doi.org/10.1007/s10639…

P.S. Thanks @OpenAI for the support through the OpenAI Researcher Access Program (ID: 0000002609).
View quoted post on X
Daniel Domínguez Verified account @danidominguez
Séb Krier Verified account @sebkrier May 12
If anyone builds it, everyone thrives. Over the past decade, a lot of important work on AI alignment has focused on avoiding harm. But freedom from harm isn't the same as freedom to flourish.

In this paper, we introduce 'Positive Alignment'. A positively aligned agent is one that helps us navigate our own value trade-offs, builds our resilience, and acts as a scaffold for human flourishing. Doing this without slipping into top-down, technocratic paternalism is the great design challenge of our time.

We think a lot more research is now needed to explore this frontier: how do we align models that actively help us thrive?

Amazing work by @RubenLaukkonen, @drmichaellevin, @weballergy, @verena_rieser, @AdamCElwood, @996roma, @FranklinMatija, @shamilch, @_fernando_rosas, @scychan_brains, @matybohacek, @sudoraohacker, and others.

arxiv.org/abs/2605.10310
View quoted post on X