What Happened in LLMs Layers when Trained for Fast vs. Slow Thinking: A Gradient Perspective

What makes a difference in the post-training of LLMs? We investigate the training patterns of different layers in large language models (LLMs), through the lens of gradient, when training with different responses and initial models. We are specifically interested in how fast vs. slow thinking affects the layer-wise gradients, given the recent popularity of training LLMs on reasoning paths such as chain-of-thoughts (CoT) and process rewards. In our study, fast thinking without CoT leads to larger gradients and larger differences of gradients across layers than slow thinking (Detailed CoT), indicating the learning stability brought by the latter. Moreover, pre-trained LLMs are less affected by the instability of fast thinking than instruction-tuned LLMs. Additionally, we study whether the gradient patterns can reflect the correctness of responses when training different LLMs using slow vs. fast thinking paths. The results show that the gradients of slow thinking can distinguish correct and irrelevant reasoning paths. As a comparison, we conduct similar gradient analyses on non-reasoning knowledge learning tasks, on which, however, trivially increasing the response length does not lead to similar behaviors of slow thinking. Our study strengthens fundamental understandings of LLM training and sheds novel insights on its efficiency and stability, which pave the way towards building a generalizable System-2 agent. Our code, data, and gradient statistics can be found in: https://github.com/MingLiiii/Layer_Gradient.

When Hindsight is Not 20/20: Testing Limits on Reflective Thinking in Large Language Models

Recent studies suggest that self-reflective prompting can significantly enhance the reasoning capabilities of Large Language Models (LLMs). However, the use of external feedback as a stop criterion raises doubts about the true extent of LLMs' ability to emulate human-like self-reflection. In this paper, we set out to clarify these capabilities under a more stringent evaluation setting in which we disallow any kind of external feedback. Our findings under this setting show a split: while self-reflection enhances performance in TruthfulQA, it adversely affects results in HotpotQA. We conduct follow-up analyses to clarify the contributing factors in these patterns, and find that the influence of self-reflection is impacted both by reliability of accuracy in models' initial responses, and by overall question difficulty: specifically, self-reflection shows the most benefit when models are less likely to be correct initially, and when overall question difficulty is higher. We also find that self-reflection reduces tendency toward majority voting. Based on our findings, we propose guidelines for decisions on when to implement self-reflection. We release the codebase for reproducing our experiments at https://github.com/yanhong-lbh/LLM-SelfReflection-Eval.

Towards Training A Chinese Large Language Model for Anesthesiology

Medical large language models (LLMs) have gained popularity recently due to their significant practical utility. However, most existing research focuses on general medicine, and there is a need for in-depth study of LLMs in specific fields like anesthesiology. To fill the gap, we introduce Hypnos, a Chinese Anesthesia model built upon existing LLMs, e.g., Llama. Hypnos' contributions have three aspects: 1) The data, such as utilizing Self-Instruct, acquired from current LLMs likely includes inaccuracies. Hypnos implements a cross-filtering strategy to improve the data quality. This strategy involves using one LLM to assess the quality of the generated data from another LLM and filtering out the data with low quality. 2) Hypnos employs a general-to-specific training strategy that starts by fine-tuning LLMs using the general medicine data and subsequently improving the fine-tuned LLMs using data specifically from Anesthesiology. The general medical data supplement the medical expertise in Anesthesiology and enhance the effectiveness of Hypnos' generation. 3) We introduce a standardized benchmark for evaluating medical LLM in Anesthesiology. Our benchmark includes both publicly available instances from the Internet and privately obtained cases from the Hospital. Hypnos outperforms other medical LLMs in anesthesiology in metrics, GPT-4, and human evaluation on the benchmark dataset.

Learning from Polar Representation: An Extreme-Adaptive Model for Long-Term Time Series Forecasting

In the hydrology field, time series forecasting is crucial for efficient water resource management, improving flood and drought control and increasing the safety and quality of life for the general population. However, predicting long-term streamflow is a complex task due to the presence of extreme events. It requires the capture of long-range dependencies and the modeling of rare but important extreme values. Existing approaches often struggle to tackle these dual challenges simultaneously. In this paper, we specifically delve into these issues and propose Distance-weighted Auto-regularized Neural network (DAN), a novel extreme-adaptive model for long-range forecasting of stremflow enhanced by polar representation learning. DAN utilizes a distance-weighted multi-loss mechanism and stackable blocks to dynamically refine indicator sequences from exogenous data, while also being able to handle uni-variate time-series by employing Gaussian Mixture probability modeling to improve robustness to severe events. We also introduce Kruskal-Wallis sampling and gate control vectors to handle imbalanced extreme data. On four real-life hydrologic streamflow datasets, we demonstrate that DAN significantly outperforms both state-of-the-art hydrologic time series prediction methods and general methods designed for long-term time series prediction.

Self-Supervised Video Transformers for Isolated Sign Language Recognition

This paper presents an in-depth analysis of various self-supervision methods for isolated sign language recognition (ISLR). We consider four recently introduced transformer-based approaches to self-supervised learning from videos, and four pre-training data regimes, and study all the combinations on the WLASL2000 dataset. Our findings reveal that MaskFeat achieves performance superior to pose-based and supervised video models, with a top-1 accuracy of 79.02% on gloss-based WLASL2000. Furthermore, we analyze these models' ability to produce representations of ASL signs using linear probing on diverse phonological features. This study underscores the value of architecture and pre-training task choices in ISLR. Specifically, our results on WLASL2000 highlight the power of masked reconstruction pre-training, and our linear probing results demonstrate the importance of hierarchical vision transformers for sign language representation.

An Extreme-Adaptive Time Series Prediction Model Based on Probability-Enhanced LSTM Neural Networks

Forecasting time series with extreme events has been a challenging and prevalent research topic, especially when the time series data are affected by complicated uncertain factors, such as is the case in hydrologic prediction. Diverse traditional and deep learning models have been applied to discover the nonlinear relationships and recognize the complex patterns in these types of data. However, existing methods usually ignore the negative influence of imbalanced data, or severe events, on model training. Moreover, methods are usually evaluated on a small number of generally well-behaved time series, which does not show their ability to generalize. To tackle these issues, we propose a novel probability-enhanced neural network model, called NEC+, which concurrently learns extreme and normal prediction functions and a way to choose among them via selective back propagation. We evaluate the proposed model on the difficult 3-day ahead hourly water level prediction task applied to 9 reservoirs in California. Experimental results demonstrate that the proposed model significantly outperforms state-of-the-art baselines and exhibits superior generalization ability on data with diverse distributions.