Inner Thinking Transformer: Leveraging Dynamic Depth Scaling to Foster Adaptive Internal Thinking

Large language models (LLMs) face inherent performance bottlenecks under parameter constraints, particularly in processing critical tokens that demand complex reasoning. Empirical analysis reveals challenging tokens induce abrupt gradient spikes across layers, exposing architectural stress points in standard Transformers. Building on this insight, we propose Inner Thinking Transformer (ITT), which reimagines layer computations as implicit thinking steps. ITT dynamically allocates computation through Adaptive Token Routing, iteratively refines representations via Residual Thinking Connections, and distinguishes reasoning phases using Thinking Step Encoding. ITT enables deeper processing of critical tokens without parameter expansion. Evaluations across 162M-466M parameter models show ITT achieves 96.5\% performance of a 466M Transformer using only 162M parameters, reduces training data by 43.2\%, and outperforms Transformer/Loop variants in 11 benchmarks. By enabling elastic computation allocation during inference, ITT balances performance and efficiency through architecture-aware optimization of implicit thinking pathways.

BeamLoRA: Beam-Constraint Low-Rank Adaptation

Due to the demand for efficient fine-tuning of large language models, Low-Rank Adaptation (LoRA) has been widely adopted as one of the most effective parameter-efficient fine-tuning methods. Nevertheless, while LoRA improves efficiency, there remains room for improvement in accuracy. Herein, we adopt a novel perspective to assess the characteristics of LoRA ranks. The results reveal that different ranks within the LoRA modules not only exhibit varying levels of importance but also evolve dynamically throughout the fine-tuning process, which may limit the performance of LoRA. Based on these findings, we propose BeamLoRA, which conceptualizes each LoRA module as a beam where each rank naturally corresponds to a potential sub-solution, and the fine-tuning process becomes a search for the optimal sub-solution combination. BeamLoRA dynamically eliminates underperforming sub-solutions while expanding the parameter space for promising ones, enhancing performance with a fixed rank. Extensive experiments across three base models and 12 datasets spanning math reasoning, code generation, and commonsense reasoning demonstrate that BeamLoRA consistently enhances the performance of LoRA, surpassing the other baseline methods.

OmniRL: In-Context Reinforcement Learning by Large-Scale Meta-Training in Randomized Worlds

We introduce OmniRL, a highly generalizable in-context reinforcement learning (ICRL) model that is meta-trained on hundreds of thousands of diverse tasks. These tasks are procedurally generated by randomizing state transitions and rewards within Markov Decision Processes. To facilitate this extensive meta-training, we propose two key innovations: 1. An efficient data synthesis pipeline for ICRL, which leverages the interaction histories of diverse behavior policies; and 2. A novel modeling framework that integrates both imitation learning and reinforcement learning (RL) within the context, by incorporating prior knowledge. For the first time, we demonstrate that in-context learning (ICL) alone, without any gradient-based fine-tuning, can successfully tackle unseen Gymnasium tasks through imitation learning, online RL, or offline RL. Additionally, we show that achieving generalized ICRL capabilities-unlike task identification-oriented few-shot learning-critically depends on long trajectories generated by variant tasks and diverse behavior policies. By emphasizing the potential of ICL and departing from pre-training focused on acquiring specific skills, we further underscore the significance of meta-training aimed at cultivating the ability of ICL itself.

Curiosity-Driven Reinforcement Learning from Human Feedback

Reinforcement learning from human feedback (RLHF) has proven effective in aligning large language models (LLMs) with human preferences, but often at the cost of reduced output diversity. This trade-off between diversity and alignment quality remains a significant challenge. Drawing inspiration from curiosity-driven exploration in reinforcement learning, we introduce curiosity-driven RLHF (CD-RLHF), a framework that incorporates intrinsic rewards for novel states, alongside traditional sparse extrinsic rewards, to optimize both output diversity and alignment quality. We demonstrate the effectiveness of CD-RLHF through extensive experiments on a range of tasks, including text summarization and instruction following. Our approach achieves significant gains in diversity on multiple diversity-oriented metrics while maintaining alignment with human preferences comparable to standard RLHF. We make our code publicly available at https://github.com/ernie-research/CD-RLHF.

K-space Diffusion Model Based MR Reconstruction Method for Simultaneous Multislice Imaging

Simultaneous Multi-Slice(SMS) is a magnetic resonance imaging (MRI) technique which excites several slices concurrently using multiband radiofrequency pulses to reduce scanning time. However, due to its variable data structure and difficulty in acquisition, it is challenging to integrate SMS data as training data into deep learning frameworks.This study proposed a novel k-space diffusion model of SMS reconstruction that does not utilize SMS data for training. Instead, it incorporates Slice GRAPPA during the sampling process to reconstruct SMS data from different acquisition modes.Our results demonstrated that this method outperforms traditional SMS reconstruction methods and can achieve higher acceleration factors without in-plane aliasing.

Mixture of Hidden-Dimensions Transformer

Transformer models encounter challenges in scaling hidden dimensions efficiently, as uniformly increasing them inflates computational and memory costs while failing to emphasize the most relevant features for each token. For further understanding, we study hidden dimension sparsity and observe that trained Transformers utilize only a small fraction of token dimensions, revealing an "activation flow" pattern. Notably, there are shared sub-dimensions with sustained activation across multiple consecutive tokens and specialized sub-dimensions uniquely activated for each token. To better model token-relevant sub-dimensions, we propose MoHD (Mixture of Hidden Dimensions), a sparse conditional activation architecture. Particularly, MoHD employs shared sub-dimensions for common token features and a routing mechanism to dynamically activate specialized sub-dimensions. To mitigate potential information loss from sparsity, we design activation scaling and group fusion mechanisms to preserve activation flow. In this way, MoHD expands hidden dimensions with negligible increases in computation or parameters, efficient training and inference while maintaining performance. Evaluations across 10 NLP tasks show that MoHD surpasses Vanilla Transformers in parameter efficiency and task performance. It achieves 1.7% higher performance with 50% fewer activation parameters and 3.7% higher performance with a 3x parameter expansion at constant activation cost. MOHD offers a new perspective for scaling the model, showcasing the potential of hidden dimension sparsity to boost efficiency

MONOPOLY: Learning to Price Public Facilities for Revaluing Private Properties with Large-Scale Urban Data

The value assessment of private properties is an attractive but challenging task which is widely concerned by a majority of people around the world. A prolonged topic among us is ``\textit{how much is my house worth?}''. To answer this question, most experienced agencies would like to price a property given the factors of its attributes as well as the demographics and the public facilities around it. However, no one knows the exact prices of these factors, especially the values of public facilities which may help assess private properties. In this paper, we introduce our newly launched project ``Monopoly'' (named after a classic board game) in which we propose a distributed approach for revaluing private properties by learning to price public facilities (such as hospitals etc.) with the large-scale urban data we have accumulated via Baidu Maps. To be specific, our method organizes many points of interest (POIs) into an undirected weighted graph and formulates multiple factors including the virtual prices of surrounding public facilities as adaptive variables to parallelly estimate the housing prices we know. Then the prices of both public facilities and private properties can be iteratively updated according to the loss of prediction until convergence. We have conducted extensive experiments with the large-scale urban data of several metropolises in China. Results show that our approach outperforms several mainstream methods with significant margins. Further insights from more in-depth discussions demonstrate that the ``Monopoly'' is an innovative application in the interdisciplinary field of business intelligence and urban computing, and it will be beneficial to tens of millions of our users for investments and to the governments for urban planning as well as taxation.

DuMapper: Towards Automatic Verification of Large-Scale POIs with Street Views at Baidu Maps

With the increased popularity of mobile devices, Web mapping services have become an indispensable tool in our daily lives. To provide user-satisfied services, such as location searches, the point of interest (POI) database is the fundamental infrastructure, as it archives multimodal information on billions of geographic locations closely related to people's lives, such as a shop or a bank. Therefore, verifying the correctness of a large-scale POI database is vital. To achieve this goal, many industrial companies adopt volunteered geographic information (VGI) platforms that enable thousands of crowdworkers and expert mappers to verify POIs seamlessly; but to do so, they have to spend millions of dollars every year. To save the tremendous labor costs, we devised DuMapper, an automatic system for large-scale POI verification with the multimodal street-view data at Baidu Maps. DuMapper takes the signboard image and the coordinates of a real-world place as input to generate a low-dimensional vector, which can be leveraged by ANN algorithms to conduct a more accurate search through billions of archived POIs in the database for verification within milliseconds. It can significantly increase the throughput of POI verification by $50$ times. DuMapper has already been deployed in production since \DuMPOnline, which dramatically improves the productivity and efficiency of POI verification at Baidu Maps. As of December 31, 2021, it has enacted over $405$ million iterations of POI verification within a 3.5-year period, representing an approximate workload of $800$ high-performance expert mappers.

Quantum Neural Network for Accelerated Magnetic Resonance Imaging

Magnetic resonance image reconstruction starting from undersampled k-space data requires the recovery of many potential nonlinear features, which is very difficult for algorithms to recover these features. In recent years, the development of quantum computing has discovered that quantum convolution can improve network accuracy, possibly due to potential quantum advantages. This article proposes a hybrid neural network containing quantum and classical networks for fast magnetic resonance imaging, and conducts experiments on a quantum computer simulation system. The experimental results indicate that the hybrid network has achieved excellent reconstruction results, and also confirm the feasibility of applying hybrid quantum-classical neural networks into the image reconstruction of rapid magnetic resonance imaging.

Optimized Magnetic Resonance Fingerprinting Using Ziv-Zakai Bound

Magnetic Resonance Fingerprinting (MRF) has emerged as a promising quantitative imaging technique within the field of Magnetic Resonance Imaging (MRI), offers comprehensive insights into tissue properties by simultaneously acquiring multiple tissue parameter maps in a single acquisition. Sequence optimization is crucial for improving the accuracy and efficiency of MRF. In this work, a novel framework for MRF sequence optimization is proposed based on the Ziv-Zakai bound (ZZB). Unlike the Cram\'er-Rao bound (CRB), which aims to enhance the quality of a single fingerprint signal with deterministic parameters, ZZB provides insights into evaluating the minimum mismatch probability for pairs of fingerprint signals within the specified parameter range in MRF. Specifically, the explicit ZZB is derived to establish a lower bound for the discrimination error in the fingerprint signal matching process within MRF. This bound illuminates the intrinsic limitations of MRF sequences, thereby fostering a deeper understanding of existing sequence performance. Subsequently, an optimal experiment design problem based on ZZB was formulated to ascertain the optimal scheme of acquisition parameters, maximizing discrimination power of MRF between different tissue types. Preliminary numerical experiments show that the optimized ZZB scheme outperforms both the conventional and CRB schemes in terms of the reconstruction accuracy of multiple parameter maps.