Diego Calanzone
@diegocalanzone
Followers
203
Following
4K
Statuses
555
« artificia docuit fames » 📖 deep learning, reasoning. 🧪 drug design @Mila_Quebec 🏛️ AI grad @UniTrento https://t.co/1Y4kDPdp4B
127.0.0.1
Joined April 2015
🥳 "Logically Consistent Language Models via Neuro-Symbolic Integration" just accepted at #ICLR2025! We focus on instilling logical rules in LLMs with an efficient loss, leading to higher factuality & (self) consistency. How? 🧵
2
5
25
@soldni @kylelostat @mechanicaldirk @natolambert @allen_ai it is so cool that you can hold the model in your hands with this
0
0
1
RT @uavster: Apple gets it. Robots are going to be everywhere, but they won’t look like robots. Check out their new paper ELEGNT. I believ…
0
887
0
RT @jxmnop: humanity's real alignment crisis is in recommendation systems, not language models ironic to see hordes of people streaming in…
0
33
0
RT @proceduralia: What if you could just describe in words to an AI agent the skills it should learn? The future of AI is human-AI collabo…
0
2
0
RT @norabelrose: We find that the probability of sampling a network at random— or local volume for short— decreases exponentially as the ne…
0
1
0
NLP & Vision researchers: here’s a bridge to RL!
✨New blog post✨: my attempt as a vision researcher at finally understanding RLHF -- a deep dive into PPO & DeepSeek's GRPO! No hot take, I promise.
0
0
1
my current fear: what if we could just plug our initialisation model to a search algorithm with good signal
Obligatory "actually my lab invented test-time-compute" post. In "Stay on topic with Classifier-Free Guidance," we show that CFG enables a model to expend twice as much compute at inference time and match the performance of a model twice as large.
0
0
0
RT @jxmnop: most important thing we learned from R1? that there’s no secret revolutionary technique that’s only known by openAI. no magic…
0
109
0
RT @chelseabfinn: Disappointed with your ICLR paper being rejected? Ten years ago today, Sergey and I finished training some of the first…
0
171
0
Finally, LOgically COnsistent (LoCo) LLaMas can outperform solver-based baselines and SFT! I thank @tetraduzione and @looselycorrect for the guidance in realizing this project, get in touch or come to chat in Singapore!
0
2
6
This is hacking. Don’t confuse it with AI.
0
0
1
@y0b1byte but it’s much better to generate plots with notebooks as the data is loaded in the memory once…best practices?
0
0
0