A comprehensive index of artificial intelligence and machine-learning research with AI-generated summaries, citation metrics, and direct links to papers and code.
Gokul Swamy, Sanjiban Choudhury, Wen Sun, et al.
This paper investigates why two-stage RL fine-tuning outperforms offline MLE, finding that RL's value lies in searching over a reduced policy space defined by simple verifiers.
Corby Rosset, Pratyush Sharma, Andrew Zhao, et al.
The Universal Verifier improves computer use agent evaluation by combining rubric design, process/outcome rewards, cascading-error-free scoring, and divide-and-conquer context management, achieving human-level agreement and near-zero false positives.
Xingyu Tan, Xiaoyang Wang, Qing Liu, et al.
SkillZip is a contract-preserving graph compression framework that compresses skill libraries into reversible macros, achieving 3.46x compression with high dependency preservation and verifier reachability.
Omatharv Bharat Vaidya, Connor Thomas Jerzak, Zayne Rea Sprague, et al.
CALVER is a training-free symbolic verifier that scores LLM reasoning traces against Pearl's causal criteria, outperforming voting methods in multi-valid-answer causal reasoning tasks.
J. Kirchner, Yining Chen, Harri Edwards, et al.
Proposes legibility training via a Prover-Verifier Game to make LLM chain-of-thought reasoning easier for humans to verify, improving trust in model outputs.
Aayush Karan, Yilun Du
Proposes a simple iterative sampling algorithm that elicits reasoning from base LLMs at inference time, matching or outperforming RL post-training on single-shot tasks without additional training or verifiers.
Unknown
V-STaR iteratively improves LLM reasoning by training a DPO verifier on correct and incorrect solutions, boosting generator and verifier performance.
Unknown
OmniMath introduces a comprehensive Olympiad-level math benchmark with 4428 problems across 33 sub-domains and 10 difficulty levels, using GPT-4o and an open-source verifier OmniJudge for rigorous evaluation.
Unknown
Nover introduces verifier-free reinforcement learning for language models, eliminating the need for large verifier models and their computational costs.
Unknown
This paper proposes Double-Helix Co-Training, a method that iteratively improves a computer-use agent generator and an LLM-based verifier to provide accurate reward signals for reinforcement learning.