Papers6

#reward hacking

Reinforcement World Model Learning for LLM-based Agents

Large language models are great at words, but they struggle to predict what will happen after they act in a changing world.

#Reinforcement World Model Learning#world modeling#LLM agents

Pushing the Boundaries of Natural Reasoning: Interleaved Bonus from Formal-Logic Verification

Intermediate

Chuxue Cao, Jinluan Yang et al.Jan 30arXiv

Large language models sometimes reach the right answer for the wrong reasons, which is risky and confusing.

#formal logic verification#interleaved verification#neuro-symbolic reasoning

Latent Adversarial Regularization for Offline Preference Optimization

Intermediate

Enyi Jiang, Yibo Jacky Zhang et al.Jan 29arXiv

This paper introduces GANPO, a new way to train language models from human preferences by guiding the model using its hidden thoughts (latent space) instead of just its visible words (token space).

#GANPO#latent space regularization#offline preference optimization

DenseGRPO: From Sparse to Dense Reward for Flow Matching Model Alignment

Intermediate

Haoyou Deng, Keyu Yan et al.Jan 28arXiv

DenseGRPO teaches image models using lots of small, timely rewards instead of one final score at the end.

#DenseGRPO#flow matching#GRPO

BAPO: Boundary-Aware Policy Optimization for Reliable Agentic Search

Intermediate

Shiyu Liu, Yongjing Yin et al.Jan 16arXiv

RL-trained search agents often sound confident even when they don’t know, which can mislead people.

#agentic search#reinforcement learning#boundary awareness

GARDO: Reinforcing Diffusion Models without Reward Hacking

Intermediate

Haoran He, Yuxiao Ye et al.Dec 30arXiv

GARDO is a new way to fine-tune text-to-image diffusion models with reinforcement learning without getting tricked by bad reward signals.

#GARDO#reward hacking#gated KL regularization