Matthew Macfarlane
@MattVMacfarlane
Followers
696
Following
3K
Media
9
Statuses
171
PhD Candidate @AmlabUvA. Working on Reinforcement Learning , Search, Reasoning and Adaptive Computation. Prev intern @Microsoft. Views Are Not My Own.
Amsterdam
Joined September 2024
Check out our work on the Latent Program Network for inductive program synthesis! . It is a new architecture for latent program search that enables efficient test-time adaptation without the need for parameter fine-tuning.
Introducing Latent Program Network (LPN), a new architecture for inductive program synthesis that builds in test-time adaption by learning a latent space that can be used for search 🔎.Inspired by @arcprize 🧩, we designed LPN to tackle out-of-distribution reasoning tasks!
0
2
23
Some great work from @AmirhoseinRj and @levilelis on neural policies vs programmatic policies for OOD generalization. I'm looking forward to discussing such topics further at the Workshop on Programmatic Representations for Agent Learning @icmlconf, which Levi is co-organising.
Previous work has shown that programmatic policies—computer programs written in a domain-specific language—generalize to out-of-distribution problems more easily than neural policies. Is this really the case? 🧵
0
1
8
Thrilled to see our NeurIPS 2024 paper, Sequential Monte Carlo Policy Optimisation (, featured in Kevin's Reinforcement Learning: A Comprehensive Overview, which additionally recognises SMC as a competitive, scalable online planner. A fantastic modern.
I'm happy to announce that v2 of my RL tutorial is now online. I added a new chapter on multi-agent RL, and improved the sections on 'RL as inference' and 'RL+LLMs' (although latter is still WIP), fixed some typos, etc.
0
9
66
Checkout @ClementBonnet16 discussing searching latent program spaces tommorow on @MLStreetTalk !.
We spoke with @ClementBonnet16 at NeurIPS about his extremely innovative approach to the @arcprize using a form of test time inference where you search a latent space of a VAE before making an optimal prediction. @fchollet was so impressed, he hired Clem shortly after! 😃 -
0
2
25
3/ Searching Latent Program Spaces @ClementBonnet16 ( similarly uses a meta-induction network, but our insight was that gradient descent can be performed in the compact task embedding space, as opposed to the parameter space.
0
0
12
2/ Neural Program Meta-Induction explores transductively mapping input-output examples of a program and a single input to predict the output (named Meta). Essentially identical to ideas applied to ARC-AGI, they then performed gradient-based fine-tuning of model parameters (named
1
0
6
1/ An interesting outcome of the @arcprize 2024 was the high performance of test-time fine-tuning. It is common in the field of AI to emphasize novelty over historical connections, so I'd like to highlight an early application of this exact idea, applied to program synthesis from
2
7
68
Epic Xmas haul ⛄ looking forward to getting stuck in #activeinference @NickJChater @MITCoCoSci @ClementBonnet16
1
0
8
RT @CohereForAI: Are you interested in policy + people + AI? Interested in program synthesis? How about reinforcement learning?🧐 . December….
0
5
0
If you are interested in scalable search based 🔎 policy improvement operators come chat with me at our poster on SPO tommorow!.
Excited to share our latest work on Sequential Monte Carlo Policy Optimisation (SPO)🔥— a scalable, search-based RL algorithm leveraging SMC as a policy improvement operator for both continuous and discrete environments!.📍 Catch us tomorrow at #NeurIPS2024 (poster #94776) from
1
3
13
Great to have Searching Latent Program Spaces ( recognised with a 🥉3rd place paper award in the @arcprize! It was a pleasure working on this with @ClementBonnet16 . Looking forward to continuing to develop these ideas even further.
Today we're announcing the winners of ARC Prize 2024. We're also publishing an extensive technical report on what we learned from the competition (link in the next tweet). The state-of-the-art went from 33% to 55.5%, the largest single-year increase we've seen since 2020. The.
1
5
30
I'm also attending the System-2 Reasoning workshop (Sun, 15 Dec, 8:55 a.m.).@fchollet will discuss recent approaches in the last iteration of the ARC challenge @arcprize. Happy to discuss my paper Searching Latent Program Spaces @ClementBonnet16 , submitted to the competition.
0
2
7
Catch the poster for SPO in West Ballroom A-D (#6600) on Wed, 11 Dec, 11 a.m. PST – 2 p.m. PST.
1
0
4
Excited to share I’ll be at #NeurIPS2024 🔥I will be presenting SPO: Sequential Monte Carlo Policy Optimisation: 🔎. If you'd like to discuss this or topics like search, test-time adaptation, or program synthesis reach out and we can catch up in person.
1
1
10
RT @jparkerholder: Introducing 🧞Genie 2 🧞 - our most capable large-scale foundation world model, which can generate a diverse array of cons….
0
477
0
RT @arcprize: Paper submissions (due today by 11:59pm UTC) are flowing in!. We're excited to share these intriguing approaches from top uni….
0
9
0
RT @NatashaEve4: Introducing BenchAgents: a framework for automated benchmark creation, using multiple LLM agents that interact with each o….
0
9
0