73 NeurIPS workshops, and not a single one on Causality [R]

The recent Reddit post highlighting the dearth of causality-focused workshops at NeurIPS raises a valid and concerning point about the shifting landscape of AI research. It’s a stark observation that, while Large Language Models (LLMs) and agent-based AI continue to dominate the conversation and conference schedules, a crucial foundational area like causal inference appears to be losing ground. The list provided—a sobering tally of zero causality workshops at NeurIPS—underscores this trend. This isn't to diminish the incredible progress in LLMs; however, the field's rapid ascent has arguably come at the expense of other vital subfields. As we explored in A Mechanistic Explanation of Prompt Injection (and why you should study roles), understanding underlying mechanisms, including causal relationships, is increasingly critical for building robust and reliable AI systems, particularly when considering vulnerabilities like prompt injection. The current fervor around generative AI risks prioritizing impressive demonstrations over rigorous, causal understanding.

The concentration of causality research within specialized venues like UAI, AISTATS, and CLeaR is not inherently negative—these are excellent conferences—but the shrinking presence at a flagship event like NeurIPS indicates a potential misalignment of priorities within the broader AI community. We’ve also seen similar dynamics emerge in other areas; the discussion around 3 Collapsing models highlights the challenges of training even seemingly straightforward models, suggesting that deeper theoretical understanding—often rooted in causal reasoning—is needed to overcome these hurdles. The implicit assumption that correlation equals causation has long been a pitfall in statistical analysis, and the scale and complexity of modern AI systems only amplify this risk. Without a concerted effort to integrate causal thinking into AI development, we risk building increasingly powerful tools that are also prone to unexpected and potentially harmful biases and errors.

The reasons for this shift are multifaceted. The immediate, visually compelling nature of LLM demonstrations lends itself well to attracting attention and investment. Causal inference, on the other hand, often involves more nuanced, less immediately gratifying work – designing experiments, developing robust identification strategies, and carefully interpreting results. It requires a deeper engagement with domain expertise and a willingness to grapple with uncertainty. Furthermore, the recent focus on empirical performance in LLMs has sometimes overshadowed the importance of theoretical grounding, leading researchers to prioritize techniques that deliver short-term gains over long-term robustness and interpretability. The practical concerns around camera-ready submissions, as discussed in ECCV workshop, camera ready instructions?, while seemingly tangential, speak to a broader allocation of resources and attention within the academic ecosystem.

Ultimately, the decline in causality representation at NeurIPS should serve as a wake-up call. While the rise of LLMs and agents is undoubtedly transformative, neglecting the fundamental principles of causal reasoning will ultimately limit the potential of AI to solve real-world problems effectively and responsibly. The question now is: how can we incentivize and support research that bridges the gap between these exciting new developments and the essential theoretical foundations of causal inference? Will we see a resurgence of interest in causal methods as the limitations of purely data-driven approaches become more apparent, or will the field continue to be relegated to a niche corner of the AI landscape?