Language Models can Self-Improve at State-Value Estimation for Better Search

Very excited about this new work by Ethan Mendes on self-imprving state value estimation for more efficient search without labels or rewards.

LLMs use fixed strategies for all questions. This is inefficient for complex reasoning. This paper introduces self-taught lookahead (STL). It improves value estimation in language models by learning from state transitions without ground truth rewards. 📌 Self-taught lookahead

🚨o3-mini vastly outperforms DeepSeek-R1 on an unseen probabilistic reasoning task! Introducing k-anonymity estimation: a novel task to assess privacy risks in sensitive texts Unlike conventional math and logical reasoning, this is difficult for both humans and AI models. 1/7

Want to learn about Llama's pre-training? Mike Lewis will be giving a Keynote at NAACL 2025 in Albuquerque, NM on May 1. 2025.naacl.org NAACL HLT 2025

Wondering what review scores you need to get accepted at ACL? Maybe this data from NAACL 2025 can help: gist.github.com/aritter/8b65a9…

Excited to share our paper on uncovering prompting patterns in Frontier LLMs using automated meta-analysis at ACL 2025 #ACL2025

📢We conduct a systematic study to demystify the synergy between SFT and RL for reasoning models. The result? We trained a 7B model - AceReason-Nemotron-1.1, significantly improved from version 1.0 on math and coding benchmarks. ✅AIME2025 (math): 53.6% -> 64.8% ✅LiveCodeBench

❤️🌎 Introducing CARE: Multilingual Multicultural Human Preference Learning 3490 culturally relevant prompts + 31.7k Human/AI-written responses rated by multilingual speakers 💡 Key insights: - Even a small amount of cultural data improves popular LLMs consistently. - Deepseek-v3

🎉 Excited to see that our paper on cost-efficient data annotation for LLMs won an SAC Highlight Award! 🔗 Check out @mohit_rag18's work here: aclanthology.org/2025.acl-long.…

Unfortunately, I had to miss out on attending in person in Vienna, but glad to see the recognition. We need more research on understanding data and posttraining of LLMs. Always a pleasure working with Alan and Junmo Kang

What if LLMs can forecast their own scores on unseen benchmarks from just a task description? We are the first to study text description→performance prediction, giving practitioners an early read on outcomes so they can plan what to build—before paying full price 💸

New work led by Jungsoo Park with a really surprising result: LLMs can predict their performance on a benchmark using only a natural language description of the task