iiinova — where your digital self keeps evolving

Reposted by Aditi Mavalankar

Abhinav Moudgil @amoudgl.bsky.social · 31/03/2026

Introducing Celo2: Towards Learned Optimization Free Lunch We show that learned optimizers can generalize to practical tasks like GPT-3 1.3B pretraining and several out-of-distribution vision/RL tasks from limited meta-training (~4.5 GPU hours)! 🧵

celo2 language model pretraining results

121

Reposted by Aditi Mavalankar

Tom Schaul @schaul.bsky.social · 16/03/2026

DeepMind's RL team is hiring a research scientist: if you're passionate about RL, come work with us! And if you know people who might be interested, please share: job-boards.greenhouse.io/deepmind/job...

job-boards.greenhouse.io

Research Scientist, Reinforcement Learning

London, UK

12914

Reposted by Aditi Mavalankar

Tom Schaul @schaul.bsky.social · 27/08/2025

Where do some of Reinforcement Learning's great thinkers stand today? Find out! Keynotes of the RL Conference are online: www.youtube.com/playlist?lis... Wanting vs liking, Agent factories, Theoretical limit of LLMs, Pluralist value, RL teachers, Knowledge flywheels (guess who talked about which!)

17624

Aditi Mavalankar @aditimavalankar.bsky.social · 13/07/2025

On my way to #ICML2025 to present our algorithm that strongly scales with inference compute, in both performance and sample diversity! 🚀 Reach out if you’d like to chat more!

082

Reposted by Aditi Mavalankar

Abhinav Moudgil @amoudgl.bsky.social · 15/06/2025

New side project! assayer: A simple Python-RQ based tool to automatically monitor and evaluate ML model checkpoints offline during training.

141

Reposted by Aditi Mavalankar

Tom Schaul @schaul.bsky.social · 22/05/2025

Ever thought of joining DeepMind's RL team? We're recruiting for a research engineering role in London: job-boards.greenhouse.io/deepmind/job... Please spread the word!

job-boards.greenhouse.io

Research Engineer, Reinforcement Learning

London, UK

1288

Reposted by Aditi Mavalankar

Tom Schaul @schaul.bsky.social · 20/03/2025

When faced with a challenge (like debugging) it helps to think back to examples of how you've overcome challenges in the past. Same for LLMs! The method we introduce in this paper is efficient because examples are chosen for their complementarity, leading to much steeper inference-time scaling! 🧪

0195

Aditi Mavalankar @aditimavalankar.bsky.social · 17/03/2025

This was a really fun collaboration with my brilliant collaborators Hassan Mansoor, Zita Marinho, Masha Samsikova, and @schaul.bsky.social!

010

Aditi Mavalankar @aditimavalankar.bsky.social · 17/03/2025

In addition to this, AuPair has been shown to work better across CodeForces difficulty levels and preserve coverage of problem categories from the training data distribution (see paper for more details).

110

Aditi Mavalankar @aditimavalankar.bsky.social · 17/03/2025

4) the responses produced by the model have high diversity for the more performant models.

110

Aditi Mavalankar @aditimavalankar.bsky.social · 17/03/2025

3) our approach exhibits strong scaling with inference-time compute, and even after 100+ LLM calls, we do not see plateauing in the scaling curve;

110

Aditi Mavalankar @aditimavalankar.bsky.social · 17/03/2025

2) we observe strong generalisation across datasets and models, implying that the process of curating these examples can be performed once and the benefits in performance can be reaped multiple times;

110

Aditi Mavalankar @aditimavalankar.bsky.social · 17/03/2025

Injecting different examples into the prompt has several benefits: 1) we see significant gains in performance compared to best-of-N and self-repair baselines on multiple model families: Gemini, Gemma, and GPT;

110

Aditi Mavalankar @aditimavalankar.bsky.social · 17/03/2025

Fun fact: the title “AuPair” has multiple interpretations: at a higher level, it guides LLMs to better behaviour with a predefined set of examples; it is also a conjunction of Au, the chemical symbol for gold, and pair, i.e. golden pairs!

110

Aditi Mavalankar @aditimavalankar.bsky.social · 17/03/2025

For the coding domain, a golden example pair, or AuPair, contains the problem description, an incorrect guess, and a fix that improves the solution.

110

Aditi Mavalankar @aditimavalankar.bsky.social · 17/03/2025

Our submodular approach yields a fixed ordered set of complementary and useful AuPairs. For a budget of N LLM calls, the model is given N different prompts to answer the same question, where each prompt contains a different golden example.

110

Aditi Mavalankar @aditimavalankar.bsky.social · 17/03/2025

The key idea underlying our approach is simple: our approach curates a fixed set of golden examples (AuPairs) that are provided as 1-shot in-context examples during inference. We show that using AuPairs significantly improves code repair performance and scales well with inference compute!

110

Reposted by Aditi Mavalankar

Tom Schaul @schaul.bsky.social · 28/11/2024

Are there limits to what you can learn in a closed system? Do we need human feedback in training? Is scale all we need? Should we play language games? What even is "recursive self-improvement"? Thoughts about this and more here: arxiv.org/abs/2411.16905

arxiv.org

Boundless Socratic Learning with Language Games

An agent trained within a closed system can master any desired capability, as long as the following three conditions hold: (a) it receives sufficiently informative and aligned feedback, (b) its covera...

710919