Van Seijen et al. [2017] propose to split a state into different sub-states, each with a sub-reward obtained bytraining ageneral valuefunction, andlearnmultiple valuefunctions withsub-rewards. The architecture is rather limited due to requiring prior knowledge of how to split into sub-states.

Reward decomposition, which aims to decompose the full reward into multiple sub-rewards, has been proven beneficial for improving sample efficiency in reinforcement learning. Existing works on discovering reward decomposition are mostly policy dependent, which constrains diverse or disentangled behavior between different policies induced by different sub-rewards. In this work, we propose a set of novel reward decomposition principles by constraining uniqueness and compactness of different state features/representations relevant to different sub-rewards. Our principles encourage sub-rewards with minimal relevant features, while maintaining the uniqueness of each sub-reward. We derive a deep learning algorithm based on our principle, and term our method as RD$^2$, since we learn reward decomposition and representation decomposition jointly. RD$^2$ is evaluated on a toy case, where we have the true reward structure, and some Atari environments where reward structure exists but is unknown to the agent to demonstrate the effectiveness of RD$^2$ against existing reward decomposition methods.

With the rapid evolution of large language models (LLMs), the demand for automated, high-performance system kernels has emerged as a key enabler for accelerating development and deployment. We introduce TritonRL, a domain-specialized LLM for Triton kernel generation, trained with a novel training framework that enables robust and automated kernel synthesis. Unlike general-purpose programming languages, Triton kernel generation faces unique challenges due to data scarcity and incomplete evaluation criteria, vulnerable to reward hacking. Our approach addresses these challenges end-to-end by distilling Triton-specific knowledge through supervised fine-tuning on curated datasets, and further improving code quality via reinforcement learning (RL) with robust, verifiable rewards and hierarchical reward assignment. Our RL framework robustly detects reward hacking and guides both reasoning traces and code tokens through fine-grained verification and hierarchical reward decomposition, enabling the model to generate high-quality Triton kernels that can truly replace existing modules. With robust and fine-grained evaluation, our experiments on KernelBench demonstrate that TritonRL achieves state-of-the-art correctness and speedup, surpassing all other Triton-specific models and underscoring the effectiveness of our RL-based training paradigm.

Neural Information Processing Systems http://nips.cc/

We propose a large language model based reward decomposition framework for aligning dialogue agents using only a single session-level feedback signal. We leverage the reasoning capabilities of a frozen, pretrained large language model (LLM) to infer fine-grained local implicit rewards by decomposing global, session-level feedback. Our first text-only variant prompts the LLM to perform reward decomposition using only the dialogue transcript. The second multimodal variant incorporates additional behavioral cues, such as pitch, gaze, and facial affect, expressed as natural language descriptions. These inferred turn-level rewards are distilled into a lightweight reward model, which we utilize for RL-based fine-tuning for dialogue generation. We evaluate both text-only and multimodal variants against state-of-the-art reward decomposition methods and demonstrate notable improvements in human evaluations of conversation quality, suggesting that LLMs are strong reward decomposers that obviate the need for manual reward shaping and granular human feedback.

The use of reward functions to structure AI learning and decision making is core to the current reinforcement learning paradigm; however, without careful design of reward functions, agents can learn to solve problems in ways that may be considered ``undesirable" or ``unethical. Without thorough understanding of the incentives a reward function creates, it can be difficult to impose principled yet general control mechanisms over its behavior. In this paper, we study methods for constructing guardrails for AI agents that use reward functions to learn decision making. We introduce a novel approach, which we call strategy masking, to explicitly learn and then suppress undesirable AI agent behavior. We apply our method to study lying in AI agents and show that strategy masking can effectively modify agent behavior by suppressing, or actively penalizing, the reward dimension for lying such that agents act more honestly while not compromising their ability to perform effectively.