Column: Insights from the world’s premier AI conference

NeurIPS 2025 may be the first instance of the conference that will be remembered both for the papers that were accepted and for the ones that were rejected. For the first time in its history, “resource constraints” forced the rejection of 400 papers previously marked for acceptance by reviewers. 

Traditionally, the stereotype around NeurIPS (at least for me) was built around two ideas: the conference as a talent marketplace  swarming with industry recruiters, and as a gatekeeper to impactful, math-heavy theoretical AI research. 

It is apparent that the identity of the conference has evolved significantly over time, now including a special focus on applications and interdisciplinary AI (natural sciences, healthcare, society, econometrics) and domain-specific areas. 

Unsurprisingly, efficiency was a common theme in the main track, spanning topics from 3D reconstruction in robotics, to “LLMOps” (Large Language Model Operations, e.g. training, inference, reasoning, etc.) and beyond. The conference thus continues to commensurably reflect the reality that the AI frontier is just as much about engineering and value creation at the use-case level as it is about fundamental research.

AI Sweden’s own contributions at NeurIPS reflect this shift toward practical research. During the 2025 conference we presented papers on how to train models on financial transaction and network data without compromising privacy, alongside work on profiling sensitive data leakage in SOTA model inference attacks.

This graph provides an amazing high-level overview of the main focus areas of NeurIPS2025. There are key clusters emerging in topics like LLM (Large Language Model) evaluation, benchmarking and general capability developments (reasoning, RAG, coding, agentic), multimodal models (specifically Multimodal LLMs and how to benchmark them), diffusion models, reinforcement learning and causal inference, learning theory, graph neural networks, and efficiency-focused “LLMOps”, all in the context of large-scale learning systems.

These clusters clearly represent the very core of what the research community is currently occupied with, and this is most clearly reflected in the conference’s Best Paper selections. 

Here are my personal observations:

• Reasoning capabilities are at the center of the discussion. 2025 undoubtedly became a year of transition from train-time scaling to test-time scaling. While some research showed the limitations of RLVR (Reinforcement Learning from Verifiable Rewards), the narrative was countered by practical breakthroughs such as GPT-5.2 surpassing the 50 percent threshold on the ARG-AGI-2 benchmark, a landmark achievement for the industry. Likewise, another Best Paper runner-up showed that model depth is a key enabler for leveraging reinforcement learning (RL) to attain new capabilities. In summary, the reasoning topic (and especially, reasoning efficiently) is currently probably the number one question in the research community, and will in all likelihood continue to be a key focus in the foreseeable future.
   
• Algorithmic mode collapse and the societal implications of uniformity. The implications of untethered scaling and alignment methods on society are now tangible. One awarded paper found that language models have a tendency to produce homogeneous outputs both individually, but also collectively, in a way that humans do not. The risk is clear: if we continue aligning LLMs as we are doing now, we risk ending up with mode-collapsed tools that will end up “homogenizing” human thought as well. These thoughts resonate well with the continued criticism in the community that our internet-scale training data is already exhausted, and that differences between models are attributable mostly to data filtering during pre-training and alignment methods in post-training. To me it is noteworthy that this is not a deeply technical paper, but rather a principled study that contributes key data and makes key observations from a fairly high point of view.
   
• AI research is maturing. Robustness and operational efficiency are keeping model architecture research alive, and proving that approaches work at scale is now commonplace. For example, the “Gated Attention” paper managed to considerably increase training robustness and efficiency after processing over 3.5 trillion tokens to address a key mechanistic failure in the standard Transformer architecture with a fix reminiscent of older Recurrent Neural Network (RNN) models. It is telling that the Best Paper award in this case went to a “boring” engineering fix. The Alibaba Qwen team didn't invent a new paradigm, instead they “just” fixed a leaky valve in the attention mechanism. It’s a signal of maturation to rigorous industrialization, showing that 1) practical considerations are important, and 2) they can often still be addressed by relatively simple means. The “burden of empirical proof” is still high, at times perhaps even prohibitively so for academia.

My takeaway from the invited speaker talks is that they represent a welcomed shift from the focus on scaling, serving rather as both an introspective reflection on value, and whether the “AI ship” is being steered in the right direction in the first place. 

My personal highlights: Richard Sutton (recipient of the 2024 Turing Award) pushes continuous learning as an enabler for the next frontier. Both Yejin Choi and Melani Mitchell focus on core cognition topics and highlight thoughts very reminiscent of Moravec’s paradox for AI a.k.a the “Jagged Frontier” of AI, a notion that is becoming painfully relevant as model capabilities increase.

Personally, I believe that putting these points in the context of the key focus areas of the conference also tells us what a conference like NeurIPS envisions the future of AI to be: a clear bet on “on-the-fly” adaptability, with weights encoding not only knowledge, but capabilities for attaining that knowledge in the first place. Continuous learning, still an underrepresented subset of research, is almost certainly a key piece of the puzzle here, and gives new context to questions surrounding robustness and learning dynamics - it is also my pick for the “next frontier to scale.” Self-supervised RL in this context represents another piece of the future puzzle. In my opinion, still somewhat missing from the wide gamut of topics in this year’s conference are methods related to (for example hierarchical) memory mechanisms beyond in-context learning and RAGs (Retrieval-Augmented Generation). 

I believe we will certainly continue to struggle to develop and trust the right benchmarks and, at a core level, the basic methods to measure the cognitive performance of models. As capabilities continue to increase, the risk also increases of misjudging the “jagged frontier” of these capabilities for a smooth one, suggesting that AI safety is becoming as much about the underlying technical research as it is about its adoption. This is a point I have personally emphasized in my own work for the better part of three years now, and I am quite satisfied to see it represented in these discussions at NeurIPS this year.

NeurIPS 2026 will undoubtedly be larger, and the growing pains, both logistical and technical, will persist. But as the field pivots from “training bigger” to “thinking longer”, the definition of progress changes with it. In particular, I look forward to seeing where we all land on the (as of yet) unresolved question of reasoning, and how much of a role this will actually play in the continued performance scaling of models. Until then, it's back to watching the leaderboards.

See a short summary of the NeurIPS 2025 talks here.