Amrith Setlur
@setlur_amrith
Followers
532
Following
345
Statuses
88
Through 2024, scaling test-time compute has become key. But, what does it mean to use test-time compute effectively & efficiently + how to do it? 🤔 We wrote a blog post ✍️ with a conceptual perspective on this: 🎯Answer: meta reinforcement learning 🧵⤵️
3
22
123
Last day to submit papers to our workshop SSI-FM ICLR 2025! Both long and tiny papers are welcome.
🚨 One day left to submit your papers to our #ICLR2025 Workshop on Self-Improving Foundation Models 🚨 If you work on synthetic data generation, RL, MCTS, evolutionary search, multi-agent systems, inference-time compute or other open-ended learning methods for foundation models like LLMs, consider submitting. 📅 Deadline: 11 Feb AOE 🔗 Website:
0
2
2
RT @agarwl_: We have extended the deadline for this ICLR workshop to Feb 11. Please submit your work, even if you disagree as we prefer to…
0
6
0
Co-organized with amazing collaborators: @aviral_kumar2 @agarwl_ @FeryalMP @katie_kang_ @robertarail
0
0
0
🚨 Interested in scaling test-time compute, training LLMs with RL, synthetic data, LLM agents and more? Consider submitting your work to our ICLR 2025 workshop on self-improving foundation models w/o human supervision: Deadline Feb 7, AoE‼️
🚨 We are organizing an ICLR workshop on self-improving foundation models w/o human supervision at ICLR 2025 in Singapore! Deadline: Feb 7, AoE (submit your ICML papers!) Details: We have an amazing line up of speakers + panelists, more info coming soon.
0
0
11
This perspective was written with awesome collaborators: @QuYuxiao, Matthew Yang, @LunjunZhang, @gingsmith and @aviral_kumar2!
0
0
1
RT @goyalsachin007: Are there any unintended consequences of instruction finetuning (IFT) in LLMs ⚠️⚠️? In our new work 📢, we highlight an…
0
40
0
@seonghwan_57 Great question @seonghwan_57! To compute advantage of the prover BoK(\pi), we first train the verifier to predict Q^{pi}. As you note correctly, we use Eq.30 to transform Q^{pi} into Q^{BoK(\pi)}, and use Eq.2 to get A^{BoK(\pi)}. This way we need not retrain PAV for every K.
1
0
2
RT @arankomatsuzaki: Google presents Rewarding Progress: Scaling Automated Process Verifiers for LLM Reasoning Achieves 5 − 6× gain in sam…
0
61
0
This was work done during my internship at Google Research, with an amazing set of collaborators: @nagpalchirag, @adamjfisch, @younggeng, @jacobeisenstein, @agarwl_, Alekh Agarwal, @JonathanBerant, and @aviral_kumar2.
0
1
12