A comprehensive index of artificial intelligence and machine-learning research with AI-generated summaries, citation metrics, and direct links to papers and code.
Unknown
Proposes conformal-calibrated rewards with procedural regeneration to mitigate benchmark contamination in scientific RLVR.
Unknown
Relax introduces latent exploration for RLVR to improve reasoning in large reasoning models.
Unknown
Proposes native reasoning models that train language models to reason on unverifiable data, moving beyond SFT and RLVR.
Vatsal Venkatkrishna, Nico Daheim, Iryna Gurevych
Introduces parameter-space exploration for LLM RL via Perturbed Parameter Policy Optimization (3PO), improving downstream performance over GRPO at similar FLOPs.
Yifan Ding, Xincheng Wei, Yoshua Y. Li, et al.
SAF introduces a stable advantage fusion framework for on-policy distillation, resolving magnitude and temporal mismatches to improve RLVR+OPD training.
Jonas Hubotter, Frederike Lubeck, L. Behric, et al.
SDPO converts tokenized feedback from verifiable environments into a dense learning signal via self-distillation, improving sample efficiency and accuracy over RLVR baselines.
Adithya Bhaskar, Xi Ye, Danqi Chen
RLMT extends RLVR to open-ended chat tasks by training LMs to think before responding, outperforming standard RLHF across multiple benchmarks.
Unknown
RLVR improves low-k pass@k efficiency but restricts high-k reasoning capacity by reducing exploration of successful paths already present in base models.
Unknown
Trains an LLM with RL using a verifiable reward on just one example, matching performance of larger datasets in math reasoning.
Unknown
A large language model built on Mistral Small, enhanced with SFT, RLVR, and inference optimization for Indian languages and reasoning tasks.
Unknown
Tulu V3 presents a fully open post-training recipe for Llama 3.1 models, achieving state-of-the-art performance via SFT, DPO, and RLVR.
Unknown
Proposes post-training world models via reinforcement learning to improve their generality and task performance.