Key-value (KV) caching plays an essential role in accelerating decoding for transformer-based autoregressive large language models (LLMs). However, the amount of memory required to store the KV cache can become prohibitive at long sequence lengths and large batch sizes. Since the invention of the transformer, two of the most effective interventions discovered for reducing the size of the KV cache have been Multi-Query Attention (MQA) and its generalization, Grouped-Query Attention (GQA). MQA and GQA both modify the design of the attention block so that multiple query heads can share a single key/value head, reducing the number of distinct key/value heads by a large factor while only minimally degrading accuracy. In this paper, we show that it is possible to take Multi-Query Attention a step further by also sharing key and value heads between adjacent layers, yielding a new attention design we call Cross-Layer Attention (CLA). With CLA, we find that it is possible to reduce the size of the KV cache by another $2\times$ while maintaining nearly the same accuracy as unmodified MQA. In experiments training 1Band 3B-parameter models from scratch, we demonstrate that CLA provides a Pareto improvement over the memory/accuracy tradeoffs which are possible with traditional MQA, potentially enabling future models to operate at longer sequence lengths and larger batch sizes than would otherwise be possible.

Large language models (LLMs) have been widely deployed across numerous fields. Reinforcement Learning from Human Feedback (RLHF) leverages reward models (RMs) as proxies for human preferences to align LLM behaviors with human values, making the accuracy, reliability, and interpretability of RMs critical for effective alignment. However, traditional RMs lack interpretability, offer limited insight into the reasoning behind reward assignments, and are inflexible toward user preference shifts. While recent multidimensional RMs aim for improved interpretability, they often fail to provide feature-level attribution and require costly annotations. To overcome these limitations, we introduce the Sparse Autoencoder-enhanced Reward Model (SARM), a novel architecture that integrates a pretrained Sparse Autoencoder (SAE) into a reward model. SARM maps the hidden activations of LLM-based RM into an interpretable, sparse, and monosemantic feature space, from which a scalar head aggregates feature activations to produce transparent and conceptually meaningful reward scores. Empirical evaluations demonstrate that SARM facilitates direct feature-level attribution of reward assignments, allows dynamic adjustment to preference shifts, and achieves superior alignment performance compared to conventional reward models. Our code is available at https://github.com/schrieffer-z/sarm.

Large language models (LLMs) with billions of parameters demonstrate impressive performance.

We thank all the reviewers for the constructive comments. We address the major concerns below. Reproducibility: 1) learning to draft details; 2) feature details; 3) discussions on the computing resources used. The search tree is updated based on four steps of MCTS. The learning rate is set to 0.001 with Adam.

This paper presents a Deep Reinforcement Learning (DRL) system for Xiangqi (Chinese Chess) that integrates neural networks with Monte Carlo Tree Search (MCTS) to enable strategic self-play and self-improvement. Addressing the underexplored complexity of Xiangqi--including its unique board layout, piece movement constraints, and victory conditions--our approach combines policy-value networks with MCTS to simulate move consequences and refine decision-making.By overcoming challenges such as Xiangqi's high branching factor and asymmetrical piece dynamics, our work advances AI capabilities in culturally significant strategy games while providing insights for adapting DRL-MCTS frameworks to domain-specific rule systems.

Key-value (KV) caching plays an essential role in accelerating decoding for transformer-based autoregressive large language models (LLMs). However, the amount of memory required to store the KV cache can become prohibitive at long sequence lengths and large batch sizes. Since the invention of the transformer, two of the most effective interventions discovered for reducing the size of the KV cache have been Multi-Query Attention (MQA) and its generalization, Grouped-Query Attention (GQA). MQA and GQA both modify the design of the attention block so that multiple query heads can share a single key/value head, reducing the number of distinct key/value heads by a large factor while only minimally degrading accuracy. In this paper, we show that it is possible to take Multi-Query Attention a step further by also sharing key and value heads between adjacent layers, yielding a new attention design we call Cross-Layer Attention (CLA).

Games have played a pivotal role in advancing artificial intelligence, with AI agents using sophisticated techniques to compete. Despite the success of neural network based game AIs, their performance often requires significant computational resources. In this paper, we present Rapfi, an efficient Gomoku agent that outperforms CNN-based agents in limited computation environments. Rapfi leverages a compact neural network with a pattern-based codebook distilled from CNNs, and an incremental update scheme that minimizes computation when input changes are minor. This new network uses computation that is orders of magnitude less to reach a similar accuracy of much larger neural networks such as Resnet. Thanks to our incremental update scheme, depth-first search methods such as the α-β search can be significantly accelerated. With a carefully tuned evaluation and search, Rapfi reached strength surpassing Katagomo, the strongest open-source Gomoku AI based on AlphaZero's algorithm, under limited computational resources where accelerators like GPUs are absent. Rapfi ranked first among 520 Gomoku agents on Botzone and won the championship in GomoCup 2024. Artificial intelligence in board games like Go, Chess, and Shogi has progressed rapidly with the advent of deep neural networks.

Reward models (RM) play a critical role in aligning generations of large language models (LLM) to human expectations. However, prevailing RMs fail to capture the stochasticity within human preferences and cannot effectively evaluate the reliability of reward predictions. To address these issues, we propose Uncertain-aware RM (URM) and Uncertain-aware RM Ensemble (URME) to incorporate and manage uncertainty in reward modeling. URM can model the distribution of disentangled attributes within human preferences, while URME quantifies uncertainty through discrepancies in the ensemble, thereby identifying potential lack of knowledge during reward evaluation. Experiment results indicate that the proposed URM achieves state-of-the-art performance compared to models with the same size, demonstrating the effectiveness of modeling uncertainty within human preferences. Furthermore, empirical results show that through uncertainty quantification, URM and URME can identify unreliable predictions to improve the quality of reward evaluations.

Chess has long been regarded as a proving ground for advancing artificial intelligence (AI). The importance of strategic planning and rational decisionmaking in strong chess play make the game an ideal domain in which to test systems aiming to master these human-like cognitive capabilities. Traditional chess engines employ a specialized tree-search algorithm paired with a handcrafted evaluation function. Most modern chess engines follow this recipe but use an efficiently updated neural network (NNUE) rather than a handcrafted evaluation function. The AlphaZero engine [13] introduced a different recipe based on Monte Carlo Tree Search (MCTS) and deep neural networks.