Base Labs @baselabs

Read the full standard update here: aiuc-1.com/research/2026-…

We've collaborated with @aiunderwriting on integrating new controls for open-weight models into the standard for agent security, safety and reliability. Our focus at Base Labs is to build a frontier safety stack for open-source models. To deploy these models with confidence, organizations need safety controls at three layers: the organizational, application and model. The new AIUC-1 model-layer controls cover the safe use and deployment of open-weight models, including: - Pre-deployment testing of model changes - Verification of model source, integrity and license - Cyber evaluations on modified open-weight models We believe openness is an advantage for safety: open weights can easily be tested, monitored and verified. Standards are how that becomes the norm, and they're another piece of the stack we're excited to work toward at Base Labs. Thank you to @RuneKvist , @RajivDattani and the AIUC-1 team – we look forward to expanding our technical collaboration.

A core principle of our research agenda is following projects through to completion. Here we extend our original study of iterative learning, and draw insights about factual competition during such learning and general model performance. See the full study (and code and checkpoints) here: labs.baseten.co/articles/why-d…

Then, the ultimate question – does this iterative learning with more ordinary facts prevent model 'frying'? In a deceiving way, no. On standard benchmarks (that can be used with talkie) such as ARC-Easy, model performance is largely similar across experiments. What does change is the ability of the model to recall pre-1931 facts. This links to more recent work by Tan and colleagues that suggests that models fine-tuned to study a behavior often lose capabilities that standard evaluations miss.

New weekly Rollout from the Base Labs team. We ask: why do models forget facts that they iteratively learn? This is an extension of our NeurIPS workshop submission and uses talkie, a 13B model trained only on pre-1931 text (Levine et al., 2026).

Building the safety stack for open-source models is a priority for us. We’re partnering with @GoodfireAI to bring activation-based probes to inference, as an extra measure to prevent unwanted behavior.

Baseten @baseten

a week ago

Open models can come from anywhere. Securing them shouldn’t depend on where you get them. Today we're launching Project Beacon to make safety a default for open intelligence. We're partnering with @GoodfireAI to bring inference and in-line monitoring together for the first

7 18 124 25K 33

We are hosting our first research conference. At All-Gather we'll be hearing from legends of the field including @polynoamial @MillionInt @ctnzr @cHHillee and many more. Plus a bunch of chalk talks and fire sides of work happening at the frontier of inference and training. Apply to join! baseten.co/all-gather/

Baseten @baseten

a week ago

Progress that moves the frontier is almost never made alone. We’re thrilled to announce All-Gather, our first-ever research conference that will bring together the brightest minds in research, founders, and engineers to join us for a day of debate and exchange. Come hear Noam

23 19 217 67K 99

Progress that moves the frontier is almost never made alone. We’re thrilled to announce All-Gather, our first-ever research conference that will bring together the brightest minds in research, founders, and engineers to join us for a day of debate and exchange. Come hear Noam Brown, Sarah Guo, Horace He, Bryan Catanzaro, Jerry Tworek, Sabri Eyuboglu, Tuhin Srivastava, and many more debate pre-training, post-training, infrastructure, model performance, and emerging fields. Join us on November 18th in San Francisco. Research has been part of our DNA from our early days of optimizing open-source models through our recent launch of Base Labs. We’ve grown with the open ecosystem, and we’re excited to keep building alongside it. Space is limited, but we’d love you to join us. Apply here: baseten.co/all-gather/

Corrected link for paper 2/3: The Optimization Landscape of Learning Compacted Context Models arxiv.org/abs/2610.05885

We’re excited to announce that 3/3 papers from our research team were accepted at NeurIPS workshops this year! Two papers will be featured in the continual learning workshop, and another one in the grounded computer vision workshop. More on the papers below 🧵

We are very excited to present this work at NeurIPS in Sydney from December 6-12. Base Labs will also have a large booth and presence at the conference; we are actively recruiting and are excited to talk to any aspirational researchers in the field. Please stop by if you are attending!

3/3: We study continual learning in the iterative regime, writing a series of invented facts into model weights sequentially with different methods (SFT, LoRA, OPSD). After 20 such sequential writes, we find that training with varied examples (e.g. rephrasing of the same fact) maintains far better performance across a variety of questions than just repeating the same fact. We also find that this forgetting does not necessarily imply erasure, with older facts maintaining a presence in log-probabilities. Paper: arxiv.org/pdf/2607.11020

This remains an early toy exploration and we do generally believe there is much extractable structure in more general RL optimization dynamics. We are excited to share more on this soon. As a part of our commitment to doing our work in the open, we provide all code and checkpoints within the article: labs.baseten.co/articles/the-u…

Each week we share ongoing work from across the Base Labs team. We call these releases 'rollouts'. This week we share our progress in understanding the low-dimensional structure of RL training and whether recent discoveries about rank-1 extrapolation of early updates really hold. We find that performance gains largely come from easier problems and fixing doom loops, while harder environments are less linearly extrapolatable.