Klotski: Efficient Mixture-of-Expert Inference via Expert-Aware Multi-Batch Pipeline

Mixture of Experts (MoE), with its distinctive sparse structure, enables the scaling of language models up to trillions of parameters without significantly increasing computational costs. However, the substantial parameter size presents a challenge for inference, as the expansion in GPU memory cannot keep pace with the growth in parameters. Although offloading techniques utilise memory from the CPU and disk and parallelise the I/O and computation for efficiency, the computation for each expert in MoE models is often less than the I/O, resulting in numerous bubbles in the pipeline. Therefore, we propose Klotski, an efficient MoE inference engine that significantly reduces pipeline bubbles through a novel expert-aware multi-batch pipeline paradigm. The proposed paradigm uses batch processing to extend the computation time of the current layer to overlap with the loading time of the next layer. Although this idea has been effectively applied to dense models, more batches may activate more experts in the MoE, leading to longer loading times and more bubbles. Thus, unlike traditional approaches, we balance computation and I/O time and minimise bubbles by orchestrating their inference orders based on their heterogeneous computation and I/O requirements and activation patterns under different batch numbers. Moreover, to adapt to different hardware environments and models, we design a constraint-sensitive I/O-compute planner and a correlation-aware expert prefetcher for a schedule that minimises pipeline bubbles. Experimental results demonstrate that Klotski achieves a superior throughput-latency trade-off compared to state-of-the-art techniques, with throughput improvements of up to 85.12x.

Leveraging Chain of Thought towards Empathetic Spoken Dialogue without Corresponding Question-Answering Data

Empathetic dialogue is crucial for natural human-computer interaction, allowing the dialogue system to respond in a more personalized and emotionally aware manner, improving user satisfaction and engagement. The emergence of large language models (LLMs) has revolutionized dialogue generation by harnessing their powerful capabilities and shown its potential in multimodal domains. Many studies have integrated speech with text-based LLMs to take speech question as input and output text response. However, the lack of spoken question-answering datasets that include speech style information to supervised fine-tuning (SFT) limits the performance of these systems. As a result, while these systems excel at understanding speech content, they often struggle to generate empathetic responses. In response, we propose a novel approach that circumvents the need for question-answering data, called Listen, Perceive, and Express (LPE). Our method employs a two-stage training process, initially guiding the LLM to listen the content and perceive the emotional aspects of speech. Subsequently, we utilize Chain-of-Thought (CoT) prompting to unlock the model's potential for expressing empathetic responses based on listened spoken content and perceived emotional cues. We employ experiments to prove the effectiveness of proposed method. To our knowledge, this is the first attempt to leverage CoT for speech-based dialogue.

MiniMax-01: Scaling Foundation Models with Lightning Attention

We introduce MiniMax-01 series, including MiniMax-Text-01 and MiniMax-VL-01, which are comparable to top-tier models while offering superior capabilities in processing longer contexts. The core lies in lightning attention and its efficient scaling. To maximize computational capacity, we integrate it with Mixture of Experts (MoE), creating a model with 32 experts and 456 billion total parameters, of which 45.9 billion are activated for each token. We develop an optimized parallel strategy and highly efficient computation-communication overlap techniques for MoE and lightning attention. This approach enables us to conduct efficient training and inference on models with hundreds of billions of parameters across contexts spanning millions of tokens. The context window of MiniMax-Text-01 can reach up to 1 million tokens during training and extrapolate to 4 million tokens during inference at an affordable cost. Our vision-language model, MiniMax-VL-01 is built through continued training with 512 billion vision-language tokens. Experiments on both standard and in-house benchmarks show that our models match the performance of state-of-the-art models like GPT-4o and Claude-3.5-Sonnet while offering 20-32 times longer context window. We publicly release MiniMax-01 at https://github.com/MiniMax-AI.

Continuous Knowledge-Preserving Decomposition for Few-Shot Continual Learning

Few-shot class-incremental learning (FSCIL) involves learning new classes from limited data while retaining prior knowledge, and often results in catastrophic forgetting. Existing methods either freeze backbone networks to preserve knowledge, which limits adaptability, or rely on additional modules or prompts, introducing inference overhead. To this end, we propose Continuous Knowledge-Preserving Decomposition for FSCIL (CKPD-FSCIL), a framework that decomposes a model's weights into two parts: one that compacts existing knowledge (knowledge-sensitive components) and another that carries redundant capacity to accommodate new abilities (redundant-capacity components). The decomposition is guided by a covariance matrix from replay samples, ensuring principal components align with classification abilities. During adaptation, we freeze the knowledge-sensitive components and only adapt the redundant-capacity components, fostering plasticity while minimizing interference without changing the architecture or increasing overhead. Additionally, CKPD introduces an adaptive layer selection strategy to identify layers with redundant capacity, dynamically allocating adapters. Experiments on multiple benchmarks show that CKPD-FSCIL outperforms state-of-the-art methods.

Correcting Large Language Model Behavior via Influence Function

Recent advancements in AI alignment techniques have significantly improved the alignment of large language models (LLMs) with static human preferences. However, the dynamic nature of human preferences can render some prior training data outdated or even erroneous, ultimately causing LLMs to deviate from contemporary human preferences and societal norms. Existing methodologies, whether they involve the curation of new data for continual alignment or the manual correction of outdated data for re-alignment, demand costly human resources. To address this challenge, we propose a novel approach, Large Language Model Behavior Correction with Influence Function Recall and Post-Training (LANCET), which requires no human involvement. LANCET consists of two phases: (1) using influence functions to identify the training data that significantly impact undesirable model outputs, and (2) applying an Influence function-driven Bregman Optimization (IBO) technique to adjust the model's behavior based on these influence distributions. Our experiments demonstrate that LANCET effectively and efficiently correct inappropriate behaviors of LLMs. Furthermore, LANCET can outperform methods that rely on collecting human preferences, and it enhances the interpretability of learning human preferences within LLMs.

GraphLoRA: Empowering LLMs Fine-Tuning via Graph Collaboration of MoE

Low-Rank Adaptation (LoRA) is a parameter-efficient fine-tuning method that has been widely adopted in various downstream applications of LLMs. Together with the Mixture-of-Expert (MoE) technique, fine-tuning approaches have shown remarkable improvements in model capability. However, the coordination of multiple experts in existing studies solely relies on the weights assigned by the simple router function. Lack of communication and collaboration among experts exacerbate the instability of LLMs due to the imbalance load problem of MoE. To address this issue, we propose a novel MoE graph-based LLM fine-tuning framework GraphLoRA, in which a graph router function is designed to capture the collaboration signals among experts by graph neural networks (GNNs). GraphLoRA enables all experts to understand input knowledge and share information from neighbor experts by aggregating operations. Besides, to enhance each expert's capability and their collaborations, we design two novel coordination strategies: the Poisson distribution-based distinction strategy and the Normal distribution-based load balance strategy. Extensive experiments on four real-world datasets demonstrate the effectiveness of our GraphLoRA in parameter-efficient fine-tuning of LLMs, showing the benefits of facilitating collaborations of multiple experts in the graph router of GraphLoRA.

Centaur: Bridging the Impossible Trinity of Privacy, Efficiency, and Performance in Privacy-Preserving Transformer Inference

As pre-trained models, like Transformers, are increasingly deployed on cloud platforms for inference services, the privacy concerns surrounding model parameters and inference data are becoming more acute. Current Privacy-Preserving Transformer Inference (PPTI) frameworks struggle with the "impossible trinity" of privacy, efficiency, and performance. For instance, Secure Multi-Party Computation (SMPC)-based solutions offer strong privacy guarantees but come with significant inference overhead and performance trade-offs. On the other hand, PPTI frameworks that use random permutations achieve inference efficiency close to that of plaintext and maintain accurate results but require exposing some model parameters and intermediate results, thereby risking substantial privacy breaches. Addressing this "impossible trinity" with a single technique proves challenging. To overcome this challenge, we propose Centaur, a novel hybrid PPTI framework. Unlike existing methods, Centaur protects model parameters with random permutations and inference data with SMPC, leveraging the structure of Transformer models. By designing a series of efficient privacy-preserving algorithms, Centaur leverages the strengths of both techniques to achieve a better balance between privacy, efficiency, and performance in PPTI. We comprehensively evaluate the effectiveness of Centaur on various types of Transformer models and datasets. Experimental results demonstrate that the privacy protection capabilities offered by Centaur can withstand various existing model inversion attack methods. In terms of performance and efficiency, Centaur not only maintains the same performance as plaintext inference but also improves inference speed by $5.0-30.4$ times.

SpecFuse: Ensembling Large Language Models via Next-Segment Prediction

Ensembles of generative large language models (LLMs) can integrate the strengths of different LLMs to compensate for the limitations of individual models. However, recent work has focused on training an additional fusion model to combine complete responses from multiple LLMs, failing to tap into their collaborative potential to generate higher-quality responses. Moreover, as the additional fusion model is trained on a specialized dataset, these methods struggle with generalizing to open-domain queries from online users. In this paper, we propose SpecFuse, a novel ensemble framework that outputs the fused result by iteratively producing the next segment through collaboration among LLMs. This is achieved through cyclic execution of its inference and verification components. In each round, the inference component invokes each base LLM to generate candidate segments in parallel, and the verify component calls these LLMs again to predict the ranking of the segments. The top-ranked segment is then broadcast to all LLMs, encouraging them to generate higher-quality segments in the next round. This approach also allows the base LLMs to be plug-and-play, without any training or adaptation, avoiding generalization limitations. Furthermore, to conserve computational resources, we propose a model exit mechanism that dynamically excludes models exhibiting poor performance in previous rounds during each query response. In this way, it effectively reduces the number of model calls while maintaining overall performance.

Self-Generated Critiques Boost Reward Modeling for Language Models

Reward modeling is crucial for aligning large language models (LLMs) with human preferences, especially in reinforcement learning from human feedback (RLHF). However, current reward models mainly produce scalar scores and struggle to incorporate critiques in a natural language format. We hypothesize that predicting both critiques and the scalar reward would improve reward modeling ability. Motivated by this, we propose Critic-RM, a framework that improves reward models using self-generated critiques without extra supervision. Critic-RM employs a two-stage process: generating and filtering high-quality critiques, followed by joint fine-tuning on reward prediction and critique generation. Experiments across benchmarks show that Critic-RM improves reward modeling accuracy by 3.7%-7.3% compared to standard reward models and LLM judges, demonstrating strong performance and data efficiency. Additional studies further validate the effectiveness of generated critiques in rectifying flawed reasoning steps with 2.5%-3.2% gains in improving reasoning accuracy.

Instruct or Interact? Exploring and Eliciting LLMs' Capability in Code Snippet Adaptation Through Prompt Engineering

Code snippet adaptation is a fundamental activity in the software development process. Unlike code generation, code snippet adaptation is not a "free creation", which requires developers to tailor a given code snippet in order to fit specific requirements and the code context. Recently, large language models (LLMs) have confirmed their effectiveness in the code generation task with promising results. However, their performance on adaptation, a reuse-oriented and context-dependent code change prediction task, is still unclear. To bridge this gap, we conduct an empirical study to investigate the performance and issues of LLMs on the adaptation task. We first evaluate the adaptation performances of three popular LLMs and compare them to the code generation task. Our result indicates that their adaptation ability is weaker than generation, with a nearly 15% decrease on pass@1 and more context-related errors. By manually inspecting 200 cases, we further investigate the causes of LLMs' sub-optimal performance, which can be classified into three categories, i.e., Unclear Requirement, Requirement Misalignment and Context Misapplication. Based on the above empirical research, we propose an interactive prompting approach to eliciting LLMs' adaptation ability. Experimental result reveals that our approach greatly improve LLMs' adaptation performance. The best-performing Human-LLM interaction successfully solves 159 out of the 202 identified defects and improves the pass@1 and pass@5 by over 40% compared to the initial instruction-based prompt. Considering human efforts, we suggest multi-agent interaction as a trade-off, which can achieve comparable performance with excellent generalization ability. We deem that our approach could provide methodological assistance for autonomous code snippet reuse and adaptation with LLMs.