Yuhui Li

Yuhui Li, Fangyun Wei, Jinjing Zhao et al.

Large language models (LLMs) often demonstrate inconsistencies with human preferences. Previous research typically gathered human preference data and then aligned the pre-trained models using reinforcement learning or instruction tuning, a.k.a. the finetuning step. In contrast, aligning frozen LLMs without requiring alignment data is more appealing. This work explores the potential of the latter setting. We discover that by integrating self-evaluation and rewind mechanisms, unaligned LLMs can directly produce responses consistent with human preferences via self-boosting. We introduce a novel inference method, Rewindable Auto-regressive INference (RAIN), that allows pre-trained LLMs to evaluate their own generation and use the evaluation results to guide rewind and generation for AI safety. Notably, RAIN operates without the need of extra data for model alignment and abstains from any training, gradient computation, or parameter updates. Experimental results evaluated by GPT-4 and humans demonstrate the effectiveness of RAIN: on the HH dataset, RAIN improves the harmlessness rate of LLaMA 30B from 82% of vanilla inference to 97%, while maintaining the helpfulness rate. On the TruthfulQA dataset, RAIN improves the truthfulness of the already-well-aligned LLaMA-2-chat 13B model by 5%.

Yuhui Li, Zejia Wu, Chao Zhang et al.

We study the problem of out-of-distribution (o.o.d.) generalization where spurious correlations of attributes vary across training and test domains. This is known as the problem of correlation shift and has posed concerns on the reliability of machine learning. In this work, we introduce the concepts of direct and indirect effects from causal inference to the domain generalization problem. We argue that models that learn direct effects minimize the worst-case risk across correlation-shifted domains. To eliminate the indirect effects, our algorithm consists of two stages: in the first stage, we learn an indirect-effect representation by minimizing the prediction error of domain labels using the representation and the class labels; in the second stage, we remove the indirect effects learned in the first stage by matching each data with another data of similar indirect-effect representation but of different class labels in the training and validation phase. Our approach is shown to be compatible with existing methods and improve the generalization performance of them on correlation-shifted datasets. Experiments on 5 correlation-shifted datasets and the DomainBed benchmark verify the effectiveness of our approach.

Qiankun Zuo, Yanfei Zhu, Libin Lu et al.

Brain network analysis is a useful approach to studying human brain disorders because it can distinguish patients from healthy people by detecting abnormal connections. Due to the complementary information from multiple modal neuroimages, multimodal fusion technology has a lot of potential for improving prediction performance. However, effective fusion of multimodal medical images to achieve complementarity is still a challenging problem. In this paper, a novel hierarchical structural-functional connectivity fusing (HSCF) model is proposed to construct brain structural-functional connectivity matrices and predict abnormal brain connections based on functional magnetic resonance imaging (fMRI) and diffusion tensor imaging (DTI). Specifically, the prior knowledge is incorporated into the separators for disentangling each modality of information by the graph convolutional networks (GCN). And a disentangled cosine distance loss is devised to ensure the disentanglement's effectiveness. Moreover, the hierarchical representation fusion module is designed to effectively maximize the combination of relevant and effective features between modalities, which makes the generated structural-functional connectivity more robust and discriminative in the cognitive disease analysis. Results from a wide range of tests performed on the public Alzheimer's Disease Neuroimaging Initiative (ADNI) database show that the proposed model performs better than competing approaches in terms of classification evaluation. In general, the proposed HSCF model is a promising model for generating brain structural-functional connectivities and identifying abnormal brain connections as cognitive disease progresses.

Yuhui Li, Fangyun Wei, Chao Zhang et al.

Inference with modern Large Language Models (LLMs) is expensive and time-consuming, and speculative sampling has proven to be an effective solution. Most speculative sampling methods such as EAGLE use a static draft tree, implicitly assuming that the acceptance rate of draft tokens depends only on their position. Interestingly, we found that the acceptance rate of draft tokens is also context-dependent. In this paper, building upon EAGLE, we propose EAGLE-2, which introduces a new technique of context-aware dynamic draft tree into drafting modeling. This improvement leverages the fact that the draft model of EAGLE is well-calibrated: the confidence scores from the draft model approximate acceptance rates with small errors. We conducted extensive evaluations on three series of LLMs and six tasks, with EAGLE-2 achieving speedup ratios 3.05x-4.26x, which is 20%-40% faster than EAGLE-1. EAGLE-2 also ensures that the distribution of the generated text remains unchanged, making it a lossless acceleration algorithm.

Yuhui Li, Fangyun Wei, Chao Zhang et al.

Autoregressive decoding makes the inference of Large Language Models (LLMs) time-consuming. In this paper, we reconsider speculative sampling and derive two key observations. Firstly, autoregression at the feature (second-to-top-layer) level is more straightforward than at the token level. Secondly, the inherent uncertainty in feature (second-to-top-layer) level autoregression constrains its performance. Based on these insights, we introduce EAGLE (Extrapolation Algorithm for Greater Language-model Efficiency), a simple yet highly efficient speculative sampling framework. By incorporating a token sequence advanced by one time step, EAGLE effectively resolves the uncertainty, enabling precise second-to-top-layer feature prediction with minimal overhead. We conducted comprehensive evaluations of EAGLE, including all models from the Vicuna and LLaMA2-Chat series, the MoE model Mixtral 8x7B Instruct, and tasks in dialogue, code generation, mathematical reasoning, and instruction following. For LLaMA2-Chat 70B, EAGLE achieved a latency speedup ratio of 2.7x-3.5x, doubled throughput, while maintaining the distribution of the generated text.