Evaluating large language models (LLMs) is challenging. Traditional ground-truth-based benchmarks fail to capture the comprehensiveness and nuance of real-world queries, while LLM-as-judge benchmarks suffer from grading biases and limited query quantity. Both of them may also become contaminated over time. User-facing evaluation, such as Chatbot Arena, provides reliable signals but is costly and slow. In this work, we propose MixEval, a new paradigm for establishing efficient, gold-standard LLM evaluation by strategically mixing off-the-shelf benchmarks. It bridges (1) comprehensive and well-distributed real-world user queries and (2) efficient and fairly-graded ground-truth-based benchmarks, by matching queries mined from the web with similar queries from existing benchmarks. Based on MixEval, we further build MixEval-Hard, which offers more room for model improvement. Our benchmarks' advantages lie in (1) a 0.96 model ranking correlation with Chatbot Arena arising from the highly impartial query distribution and grading mechanism, (2) fast, cheap, and reproducible execution (6% of the time and cost of MMLU), and (3) dynamic evaluation enabled by the rapid and stable data update pipeline. We provide extensive meta-evaluation and analysis for our and existing LLM benchmarks to deepen the community's understanding of LLM evaluation and guide future research directions.

Behavior Tree (BT) planning is crucial for autonomous robot behavior control, yet its application in complex scenarios is hampered by long planning times. Pruning and heuristics are common techniques to accelerate planning, but it is difficult to design general pruning strategies and heuristic functions for BT planning problems. This paper proposes improving BT planning efficiency for everyday service robots leveraging commonsense reasoning provided by Large Language Models (LLMs), leading to model-free pre-planning action space pruning and heuristic generation. This approach takes advantage of the modularity and interpretability of BT nodes, represented by predicate logic, to enable LLMs to predict the task-relevant action predicates and objects, and even the optimal path, without an explicit action model. We propose the Heuristic Optimal Behavior Tree Expansion Algorithm (HOBTEA) with two heuristic variants and provide a formal comparison and discussion of their efficiency and optimality. We introduce a learnable and transferable commonsense library to enhance the LLM's reasoning performance without fine-tuning. The action space expansion based on the commonsense library can further increase the success rate of planning. Experiments show the theoretical bounds of commonsense pruning and heuristic, and demonstrate the actual performance of LLM learning and reasoning with the commonsense library.

To democratize large language models (LLMs) to most natural languages, it is imperative to make these models capable of understanding and generating texts in many languages, in particular low-resource ones. While recent multilingual LLMs demonstrate remarkable performance in such capabilities, these LLMs still support a limited number of human languages due to the lack of training data for low-resource languages. Moreover, these LLMs are not yet aligned with human preference for downstream tasks, which is crucial for the success of LLMs in English. In this paper, we introduce xLLaMA-100 and xBLOOM-100 (collectively xLLMs-100), which scale the multilingual capabilities of LLaMA and BLOOM to 100 languages. To do so, we construct two datasets: a multilingual instruction dataset including 100 languages, which represents the largest language coverage to date, and a cross-lingual human feedback dataset encompassing 30 languages. We perform multilingual instruction tuning on the constructed instruction data and further align the LLMs with human feedback using the DPO algorithm on our cross-lingual human feedback dataset. We evaluate the multilingual understanding and generating capabilities of xLLMs-100 on five multilingual benchmarks. Experimental results show that xLLMs-100 consistently outperforms its peers across the benchmarks by considerable margins, defining a new state-of-the-art multilingual LLM that supports 100 languages.

The advent of large language models (LLMs) has revolutionized natural language processing, enabling unprecedented capabilities in understanding and generating human-like text. However, the computational cost and convergence times associated with fine-tuning these models remain significant challenges. Low-Rank Adaptation (LoRA) has emerged as a promising method to mitigate these issues by introducing efficient fine-tuning techniques with a reduced number of trainable parameters. In this paper, we present OLoRA, an enhancement to the LoRA method that leverages orthonormal matrix initialization through QR decomposition. OLoRA significantly accelerates the convergence of LLM training while preserving the efficiency benefits of LoRA, such as the number of trainable parameters and GPU memory footprint. Our empirical evaluations demonstrate that OLoRA not only converges faster but also exhibits improved performance compared to standard LoRA across a variety of language modeling tasks. This advancement opens new avenues for more efficient and accessible fine-tuning of LLMs, potentially enabling broader adoption and innovation in natural language applications.

Head pose estimation (HPE) task requires a sophisticated understanding of 3D spatial relationships and precise numerical output of yaw, pitch, and roll Euler angles. Previous HPE studies are mainly based on Non-large language models (Non-LLMs), which rely on close-up human heads cropped from the full image as inputs and lack robustness in real-world scenario. In this paper, we present a novel framework to enhance the HPE prediction task by leveraging the visual grounding capability of CogVLM. CogVLM is a vision language model (VLM) with grounding capability of predicting object bounding boxes (BBoxes), which enables HPE training and prediction using full image information input. To integrate the HPE task into the VLM, we first cop with the catastrophic forgetting problem in large language models (LLMs) by investigating the rehearsal ratio in the data rehearsal method. Then, we propose and validate a LoRA layer-based model merging method, which keeps the integrity of parameters, to enhance the HPE performance in the framework. The results show our HPE-CogVLM achieves a 31.5% reduction in Mean Absolute Error for HPE prediction over the current Non-LLM based state-of-the-art in cross-dataset evaluation. Furthermore, we compare our LoRA layer-based model merging method with LoRA fine-tuning only and other merging methods in CogVLM. The results demonstrate our framework outperforms them in all HPE metrics.

The advent of large language models (LLMs) has dramatically advanced the state-of-the-art in numerous natural language generation tasks. For LLMs to be applied reliably, it is essential to have an accurate measure of their confidence. Currently, the most commonly used confidence score function is the likelihood of the generated sequence, which, however, conflates semantic and syntactic components. For instance, in question-answering (QA) tasks, an awkward phrasing of the correct answer might result in a lower probability prediction. Additionally, different tokens should be weighted differently depending on the context. In this work, we propose enhancing the predicted sequence probability by assigning different weights to various tokens using attention values elicited from the base LLM. By employing a validation set, we can identify the relevant attention heads, thereby significantly improving the reliability of the vanilla sequence probability confidence measure. We refer to this new score as the Contextualized Sequence Likelihood (CSL). CSL is easy to implement, fast to compute, and offers considerable potential for further improvement with task-specific prompts. Across several QA datasets and a diverse array of LLMs, CSL has demonstrated significantly higher reliability than state-of-the-art baselines in predicting generation quality, as measured by the AUROC or AUARC.

Tabular data is prevalent in many critical domains, yet it is often challenging to acquire in large quantities. This scarcity usually results in poor performance of machine learning models on such data. Data augmentation, a common strategy for performance improvement in vision and language tasks, typically underperforms for tabular data due to the lack of explicit symmetries in the input space. To overcome this challenge, we introduce TabMDA, a novel method for manifold data augmentation on tabular data. This method utilises a pre-trained in-context model, such as TabPFN, to map the data into a manifold space. TabMDA performs label-invariant transformations by encoding the data multiple times with varied contexts. This process explores the manifold of the underlying in-context models, thereby enlarging the training dataset. TabMDA is a training-free method, making it applicable to any classifier. We evaluate TabMDA on five standard classifiers and observe significant performance improvements across various tabular datasets. Our results demonstrate that TabMDA provides an effective way to leverage information from pre-trained in-context models to enhance the performance of downstream classifiers.

Diffusion Transformers have recently demonstrated unprecedented generative capabilities for various tasks. The encouraging results, however, come with the cost of slow inference, since each denoising step requires inference on a transformer model with a large scale of parameters. In this study, we make an interesting and somehow surprising observation: the computation of a large proportion of layers in the diffusion transformer, through introducing a caching mechanism, can be readily removed even without updating the model parameters. In the case of U-ViT-H/2, for example, we may remove up to 93.68% of the computation in the cache steps (46.84% for all steps), with less than 0.01 drop in FID. To achieve this, we introduce a novel scheme, named Learning-to-Cache (L2C), that learns to conduct caching in a dynamic manner for diffusion transformers. Specifically, by leveraging the identical structure of layers in transformers and the sequential nature of diffusion, we explore redundant computations between timesteps by treating each layer as the fundamental unit for caching. To address the challenge of the exponential search space in deep models for identifying layers to cache and remove, we propose a novel differentiable optimization objective. An input-invariant yet timestep-variant router is then optimized, which can finally produce a static computation graph. Experimental results show that L2C largely outperforms samplers such as DDIM and DPM-Solver, alongside prior cache-based methods at the same inference speed.

Retrieval augmented generation has revolutionized large language model (LLM) outputs by providing factual supports. Nevertheless, it struggles to capture all the necessary knowledge for complex reasoning questions. Existing retrieval methods typically divide reference documents into passages, treating them in isolation. These passages, however, are often interrelated, such as passages that are contiguous or share the same keywords. Therefore, recognizing the relatedness is crucial for enhancing the retrieval process. In this paper, we propose a novel retrieval method, called GNN-Ret, which leverages graph neural networks (GNNs) to enhance retrieval by considering the relatedness between passages. Specifically, we first construct a graph of passages by connecting passages that are structure-related and keyword-related. A graph neural network (GNN) is then leveraged to exploit the relationships between passages and improve the retrieval of supporting passages. Furthermore, we extend our method to handle multi-hop reasoning questions using a recurrent graph neural network (RGNN), named RGNN-Ret. At each step, RGNN-Ret integrates the graphs of passages from previous steps, thereby enhancing the retrieval of supporting passages. Extensive experiments on benchmark datasets demonstrate that GNN-Ret achieves higher accuracy for question answering with a single query of LLMs than strong baselines that require multiple queries, and RGNN-Ret further improves accuracy and achieves state-of-the-art performance, with up to 10.4% accuracy improvement on the 2WikiMQA dataset.

Huge memory consumption has been a major bottleneck for deploying high-throughput large language models in real-world applications. In addition to the large number of parameters, the key-value (KV) cache for the attention mechanism in the transformer architecture consumes a significant amount of memory, especially when the number of layers is large for deep language models. In this paper, we propose a novel method that only computes and caches the KVs of a small number of layers, thus significantly saving memory consumption and improving inference throughput. Our experiments on large language models show that our method achieves up to 26$\times$ higher throughput than standard transformers and competitive performance in language modeling and downstream tasks. In addition, our method is orthogonal to existing transformer memory-saving techniques, so it is straightforward to integrate them with our model, achieving further improvement in inference efficiency. Our code is available at https://github.com/whyNLP/LCKV.