←
Base Labs63 posts
Base Labs avatar✓

Base Labs ✓

@baselabs

A research lab by @baseten working to advance and democratize open-source intelligence.

◷ Joined July 2026▧ 25 Photos and videos
Base Labs avatar
Base Labs@baselabs·Nov 14

For this week’s rollout, we continue our exploration into current on-policy learning methods and the various attempts to distill expert demonstrations or knowledge in an on-policy manner. We dive into gradient alignment rewards, where we reward on-policy rollouts by how much the likelihood gradient aligns with that of an expert trajectory. Here, we provide some mathematical motivation for the reward, show that this reward is still efficiently computable (contrary to prior work suggestions), and demonstrate that even though the resulting RL algorithm is standard on-policy RL, gradient alignment rewards still behave like SFT: overfitting to specifics of the expert like wording.

Phil Howes avatar
⟳ Base Labs retweeted
Phil Howes@saltyph·Nov 13

we built an inference engine for deepseek v4.1-flash from scratch. the core engine is just rust + cuda. no pytorch, no sglang, no dsls, etc. ~1740 TPS decode on 4xB300, 2060TPS on 8xB300. 1M token prefill in ~12s. it's up on github

Michael Feil avatar
⟳ Base Labs retweeted
Michael Feil@feilsystem·Nov 9

Baseten runs 9B and 27B - our secrets: - we trained the models to only need a single inference pass per state, and not one prefill-step per question (e.g. perplexity flash v1.2 can be 30x more expensive) - we politely asked codex to make the inference faster, make no mistakes. Worked quite well, acoording to AA. artificialanalysis.ai/articl…

Post attachment
Base Labs avatar
Base Labs@baselabs·Nov 7

OPSD and related variants have been popular areas of post-training research in 2026. Whilst the original papers were commendable work with respect to on-policy learning and the dense feedback they provided, we should treat the bias that a hinted teacher injects as incongruent with long-horizon continual learning. This week we unpack some of the literature, alongside our own attempted fixes and observations to understand why this is the case.

Post attachment
Base Labs avatar
Base Labs@baselabs·Oct 30

We found it fun to dig into the literature (including of some recent ICLR submissions...) to write this up. Overall, we find that learned value functions remain hard to justify computationally over pure group reward-based methods, although they can provide some uplift in cheaper, targeted value prediction tasks.

Base Labs avatar
Base Labs@baselabs·Oct 30

We explore a hybrid PPO/GRPO model, where the value function instead learns a residual from the held-out group estimated value, with the hope that this is a cheaper-to-compute proxy to the true value function. Indeed, this theoretically gives comparable variance reduction to full learned value functions, and we find this can be cheaply computed by only evaluating a handful of later layers.

Base Labs avatar
Base Labs@baselabs·Oct 30

We can actually quantify the Pareto frontier of this tradeoff, based on the amount of variance reduction in the full advantage estimation using this learned value function. No value function can be truly perfect, so this threshold gets harder and harder to cross computationally as more accuracy is required.

Base Labs avatar
Base Labs@baselabs·Oct 30

For this week's Rollout, we dig deeper into a resurging question: is it computationally worth it to learn a value function in RL, or is it better to use that compute on more rollouts?

Post attachment
Base Labs avatar
Base Labs@baselabs·Oct 29

We're releasing joey-1, our smallest and fastest decision model and the best 2B model in its class. We trained joey-1 using the same data mixture and approach used to train elephant-1 and mammoth-1 and are, as always, open sourcing the weights on Hugging Face. On @baseten Model APIs, we're serving joey-1 at $0.01 per million tokens, a tenth of the mammoth-1 price.

Post attachment
Base Labs avatar
Base Labs@baselabs·Oct 23

If we wish to train frontier open-source models, we need to aggregate data as well as compute. This aggregation relies on high quality RL environments. Building good RL environments is a science, and at @baseten our Applied Research team has been developing them hand in hand with customers. Today we're partnering with 12 vetted environment companies to accelerate this work. Very soon, we'll open-source a set of frontier environments for everyone to use and train on, and show you all our early plans for building this out.

Post attachment