Continual learning for AI teams.
San Francisco
Joined October 2026
- Tweets55
- Following239
- Followers238
- Likes299
Pinned Tweet
Introducing Continual-ARC, a benchmark for agents that learn from experience.
If you think your harness gets better with experience, we'll rerun it on 50 new ARC-style puzzles that nobody outside unify has seen and post its share learned.
continual-arc.unify.ai/submi…
OpenClaw, Hermes Agent and Prime Agent don't meaningfully learn from experience.
On Continual-ARC, our new benchmark at @letsunifyai, none of them were able to memorize anything of value and none were able to generalize to new tasks. Training a small 9B model's weights did both.
unify retweeted
Highly recommended read.
This is a topic I have been researching and writing about recently.
The way we use agent memory in harnesses today works, but mostly for simple, one-off, and general tasks.
For more novel, complex, and domain-specific tasks, you will often find that current harness and memory systems don't generalize so well.
So the big question is: how can we encode continual learning capabilities in models and harnesses?
This is where the idea of co-evolving harnesses and models begins to make sense.
This work, along with the benchmark (Continual-ARC), provides an interesting analysis of this.
The harness is not just for steering the model and providing tools and context; it is increasingly becoming a key part of training pipelines to unlock continual learning in models. The harness continues to be important, but it can remain minimal and efficient.
OpenClaw, Hermes Agent and Prime Agent don't meaningfully learn from experience.
On Continual-ARC, our new benchmark at @letsunifyai, none of them were able to memorize anything of value and none were able to generalize to new tasks. Training a small 9B model's weights did both.
unify retweeted
OpenClaw, Hermes Agent and Prime Agent don't meaningfully learn from experience.
On Continual-ARC, our new benchmark at @letsunifyai, none of them were able to memorize anything of value and none were able to generalize to new tasks. Training a small 9B model's weights did both.
Sharing a research blog post on a classic deep learning topic - regularization! I tried a non-standard regularizer for LoRA fine-tuning of LMs: stochastic depth (drop-path), applied to the LoRA residual path instead of the model’s standard residual stream. It consistently achieves better validation loss than dropout or weight-decay baselines, across 3 datasets and 5 seeds! Worth a try if you use LoRA!
[1/9]
1/ As agentic tasks grow longer and rollouts become more expensive, the choice between critic-based PPO and critic-free GRPO becomes increasingly important.
Critics can reduce gradient variance, allowing models to learn from fewer trajectories, but they also introduce additional compute costs and potential bias from bootstrapping.
Our new research @ScaleAILabs compares PPO and GRPO under a shared actor objective across 1,200+ training runs in control tasks, short-horizon math, and long-horizon terminal agents. We studied how three factors influence this tradeoff: batch size, critic reliance, and task complexity. 🧵
Even the simplest weight edit, a deletion, can teach an LLM new skills!
Meet Deletion Tuning (#NeurIPS2026): just zeroing <0.1% of weights matches full fine-tuning's accuracy at scale. It's thousands of times smaller and avoids forgetting, a step toward continual learning. 🧵👇
1/
What if an LLM could not just edit a model's weights, but re-architect it as it trains?
We tried just that. The resulting architectures exit the known territory of transformers: the coding agents spliced code directly into the network.
On nanochat, the loss fell below ordinary training after each new code layer was introduced. On algorithmic reasoning, accuracy rose from 55% to 84%. We even let DeepSeek-V4.1-Flash act as the agent modifying itself, raising its AIME and HMMT 2026 score from 64% to 71%.
To do all this, the agents built a set of tools we’ve never seen before, which we call mechanistic intervention!
Blog: doubleai.s.gy/llm-redesign
We've posted a 156-page tech report on Beam! We have good systems and it's worth reading. It covers all phases through pre-training, mid-training, SFT, RL, and OPD and goes into the infra, data, and scientific decisions we made throughout the process
📚 reflection.ai/documents/beam…
We've released the weights for Beam!!
🤗 huggingface.co/reflection
📄 Model card: reflection.ai/beam-model-car…
🧪 Try Beam in our playground & get an API key: platform.reflection.ai/playg…
📚 AND we take open model safety seriously and have posted our safety report: reflection.ai/beam-safety-re…
The merged student beats the agentic teacher by 4.9 points on SWE-Bench Pro v1, though Table 10 routes all coding to that teacher and the student's rollouts get no reward. We wondered whether some of the lead came through the long-context lane, where the alignment teacher scores 78.0 on AA-LCR to the agentic one's 66.3.
59
Harness-S: Learning from Sibling Harnesses
A training method that uses complementary sibling harnesses to rescue failed agent rollouts and transfers those behaviors into a single policy that still works under the original harness.
Today, in the first of a series of blog posts we’re calling “Behind the Grove”, we’re offering a look into some of the technical work we do at DeepGrove.
Our first post is about our pretraining infrastructure, Grove. Grove is built around 2 backends: Grove-FSDP and DDP. We detail how we optimized both backends, culminating in an over 2x speedup over public baselines.
A production-scale pretraining run of Maple 20B-A1B (deepgrove.ai/maple-preview) on 4T tokens takes a mere 2 weeks on 64 H100s.
We view infrastructure as the backbone of our work, and as one of the many pieces that help us build toward the frontier of efficiency and performance.
To read the full post, visit deepgrove.ai/infra.
📖 The Ultimate MoE Training Handbook 📖
We spent 1/2M GPUh to optimize MoE training and wrote the most comprehensive blog about it. With 30k words and 50+ interactive figures it covers everything from profiling to large cluster training.
huggingface.co/spaces/huggin…
We've spent years learning how to quantize attention. But what about the linear recurrences (e.g. Mamba-2, GDN and KDA) replacing it?
It turns out we can rewrite their computation to reduce quantization error, while preserving the full-precision function. 🤯
Introducing LinQuant: a framework for low-precision recurrent prefill. 🧵 (1/7)
A science of post-training can be built on interpretability.
Models know when they reward hack, and we see it, yet training ignores it. We propose Mechanistic Shaping: using the model’s own concepts to write training rules.
In vision, DPO & RL, and its effects on monitoring 🧵
Ok guys so all the recent weight-space programming buzz inspired me to write a blog post on zeroth-order, weight-space methods (back to the basics), from the theory to a bit of update on our latest result.
Turns out plain weight perturbations don't just work for fine-tuning on math, they work for post-training agents too: vanilla ES takes Qwen3.5-4B from 12.1% → 21.1% on Terminal-Bench 2.1, past the released Tmax RL checkpoint (18.2%).
Better yet, no credit assignment, no critic, no masking, no messy hyperparameters, just good ol' perturb-and-update. Oh did I forget to mention that all with population size of ~30?
Here's why I think this could be huge for long-horizon agentic post-training 🧵+blog post👇
The Collaborative Intelligence Benchmark paper is out!
We tested 18 AI models as teammates on 240 tasks. The best model scores only 82.2 out of 100.
Frontier models almost never cheat. But when a teammate pushes back, they give in!
Read the paper at arxiv.org/abs/2611.02633
Looped LLMs are everywhere right now! But why are they interesting? Do they model language differently than regular LLMs?
Grown and looped transformers are better sequence models!
With @kaitlinmaile & @oswaldjoh (@Google, Paradigms of Intelligence Team) and @stefanAbauer. 🧵
Speculative decoding has been purely treated as an acceleration mechanism. We found it's also an allocator for context precision. We trained 0.06% hosted parameters for deepseek-v4.1-flash-1017 and 0.28% for GLM-5.3 and significantly improved agentic coding: alphaxiv.org/abs/2611.advers…
Where does GLM-5.3 land on DeepSWE with no draft at all? Slot 0 matches single-token decoding, 0.473 against 0.478 nats, so without a draft every token gets the precision the adversarial one gives decisions. That score would show whether the 7 points come from precision itself or from which side of a decision gets it.
as promised, here's everything.
we pretrained hobbylm from scratch, a 1.04B sparse MOE (~305M active), with our own money and compute from @jarvislabsai
here is the paper that has what worked and broke, and what we did .
thank you for lot of DMs and early support, it kept us going. this is just a start, we now hint more efficient private local models coming hereby
paper: arxiv.org/abs/2610.34582
I much prefer the concept of bringing context to the agent vs. bringing the agent to the context.
What do I mean by that?
Rather than asking an agent to discover its own context through tool calls, file exploration, and RAG queries…
Build a pipeline that understands your agent use cases and arranges for the relevant context to be materialized just in time and kept up to date.
Then have your harness match each use case to the context it needs.