
PhD Candidate in AI @imperialcollege Member @ic_arl @safe_trusted_ai RL | Continual learning | Robust AI systems
London, UK
Joined July 2021
- Tweets5
- Following97
- Followers11
- Likes30
Most RL methods can adapt. Much fewer can adapt without forgetting safety.
In our paper SafeAdapt (arxiv.org/abs/2604.09452), we study how to perform downstream policy updates while preserving its safety on earlier tasks. To be presented at @AAMASConference.
#RL #ContinualRL
Maksim Anisimov retweeted
Introducing the Environments Hub
RL environments are the key bottleneck to the next wave of AI progress, but big labs are locking them down
We built a community platform for crowdsourcing open environments, so anyone can contribute to open-source AGI
It is insane how much thinking can improve an LLM's coding accuracy.
openai.com/index/introducing…
48
In my experience, o3 model makes much more mistakes than a few months ago. It outputs tables with unfilled cells, unparseable LaTeX code and wrong math formulas.
Have you noticed the same? I wonder if this be explained by model collapse, reward hacking or something else.
43
GSPO is a new RL method by @Alibaba_Qwen for LLM fine-tuning that outperforms GRPO's training stability. The key is to use importance ratio on a sequence, NOT a token level.
Nice to see researchers using first principles to motivate their idea.
Paper: arxiv.org/abs/2507.18071
1
93


