Sparse Autoencoders (SAEs) are a cornerstone of mechanistic interpretability. Existing training methods inherit the Block Training paradigm from LLM pre-training, which introduces destructive gradient noise in instruct models due to attention leakage from unrelated contexts. Using GSNR analysis, we theoretically characterize this issue and propose Finetuning-aligned Sequential Training (FAST), a sequential training paradigm specifically designed for instruct models. FAST aligns SAE training with the data distribution and activation patterns of instruct models, substantially improving both reconstruction fidelity and feature interpretability. Experimental results show that FAST achieves higher GSNR, a significantly lower log-scaled MSE of 0.6468 compared to the baseline's 5.1985, and a near-zero Delta Loss (-0.51% to 0.37%). Moreover, on Llama-3.2-3B-it, FAST produces 21.1% high-quality features, substantially outperforming baseline methods that achieve 7.0% and 10.2%. We further find that intervening on special token activations through SAEs can improve generation quality, revealing new opportunities for fine-grained control. Our codes are available as open source at https://github.com/Geaming2002/FAST.

Given the audio-visual clip of the speaker, facial reaction generation aims to predict the listener's facial reactions. The challenge lies in capturing the relevance between video and audio while balancing appropriateness, realism, and diversity. While prior works have mostly focused on uni-modal inputs or simplified reaction mappings, recent approaches such as PerFRDiff have explored multi-modal inputs and the one-to-many nature of appropriate reaction mappings. In this work, we propose the Facial Reaction Diffusion (ReactDiff) framework that uniquely integrates a Multi-Modality Transformer with conditional diffusion in the latent space for enhanced reaction generation. Unlike existing methods, ReactDiff leverages intra- and inter-class attention for fine-grained multi-modal interaction, while the latent diffusion process between the encoder and decoder enables diverse yet contextually appropriate outputs. Experimental results demonstrate that ReactDiff significantly outperforms existing approaches, achieving a facial reaction correlation of 0.26 and diversity score of 0.094 while maintaining competitive realism. The code is open-sourced at github.

Deep brain stimulation (DBS) has been suggested to be a safe and effective therapeutic option for drug addiction. Although sustained abstinence is better predicted by at least 5 years of drug-free duration, few studies have followed DBS-treated addicted individuals for more than 5 years. Twenty patients with treatment-resistant heroin addiction were enrolled in this prospective, single-center, open-label pilot study. Patients who received DBS of the nucleus accumbens (NAc) with or without the anterior limb of the internal capsule (ALIC) were followed up for over 5 years, and continuous abstinence, heroin craving (HC), psychometric instrument scores and late positive potential (LPP) amplitudes served as the outcome parameters. Twelve patients maintained sustained abstinence for five years after DBS treatment, and no severe complications or adverse events occurred. A linear mixed-effects model revealed a significant main effect of postsurgical time and relapse status on HC visual analog scale (HC-VAS), SF-36, SCL-90, HAMD, and Y-BOCS scores. Preoperative marital status, HAMD cognitive and SCL-90 psychoticism subscores significantly differed between abstinent and relapsed patients, and persistent abstinence was correlated with moderate acute (VAS 5–8) and delayed (VAS 4–6) psychopsychiatric responses to stimulation. ERP analysis revealed a significant decrease in the drug-pleasant LPP amplitude (400–1000 ms) from baseline to 2–3 years after surgery. The present study suggested that DBS of the NAc/ALIC is effective for reducing heroin craving and preventing relapse in the long term and may reverse motivational attentional bias from drug-related stimuli to pleasant stimuli for heroin-addicted individuals.

In recent years, with the rapid advancements in Natural Language Processing (NLP) technologies, large models have become widespread. Traditional reinforcement learning algorithms have also started experimenting with language models to optimize training. However, they still fundamentally rely on the Markov Decision Process (MDP) for reinforcement learning, and do not fully exploit the advantages of language models for dealing with long sequences of problems. The Decision Transformer (DT) introduced in 2021 is the initial effort to completely transform the reinforcement learning problem into a challenge within the NLP domain. It attempts to use text generation techniques to create reinforcement learning trajectories, addressing the issue of finding optimal trajectories. However, the article places the training trajectory data of reinforcement learning directly into a basic language model for training. Its aim is to predict the entire trajectory, encompassing state and reward information. This approach deviates from the reinforcement learning training objective of finding the optimal action. Furthermore, it generates redundant information in the output, impacting the final training effectiveness of the agent. This paper proposes a more reasonable network model structure, the Action-Translator Transformer (ATT), to predict only the next action of the agent. This makes the language model more interpretable for the reinforcement learning problem. We test our model in simulated gaming scenarios and compare it with current mainstream methods in the offline reinforcement learning field. Based on the presented experimental results, our model demonstrates superior performance. We hope that introducing this model will inspire new ideas and solutions for combining language models and reinforcement learning, providing fresh perspectives for offline reinforcement learning research.

It is posited that cognitive and affective dysfunction in patients with major depression disorder (MDD) may be caused by dysfunctional signal propagation in the brain. By leveraging dynamic causal modeling, we investigated large-scale directed signal propagation (effective connectivity) among distributed large-scale brain networks with 43 MDD patients and 56 healthy controls. The results revealed the existence of two mutual inhibitory systems: the anterior default mode network, auditory network, sensorimotor network, salience network and visual networks formed an “emotional” brain, while the posterior default mode network, central executive networks, cerebellum and dorsal attention network formed a “rational brain”. These two networks exhibited excitatory intra-system connectivity and inhibitory inter-system connectivity. Patients were characterized by potentiated intra-system connections within the “emotional/sensory brain”, as well as over-inhibition of the “rational brain” by the “emotional/sensory brain”. The hierarchical architecture of the large-scale effective connectivity networks was then analyzed using a PageRank algorithm which revealed a shift of the controlling role of the “rational brain” to the “emotional/sensory brain” in the patients. These findings inform basic organization of distributed large-scale brain networks and furnish a better characterization of the neural mechanisms of depression, which may facilitate effective treatment.

The nucleus accumbens (NAc) and the anterior limb of internal capsule (ALIC) are effective targets for treating addiction using deep brain stimulation (DBS). However, there have been no reports on the electrophysiological characteristics of addiction nuclei at the single-cell level in humans. This study aimed to investigate the electrical activity characteristics of neurons in the NAc and ALIC using microelectrode recording (MER) during DBS surgery in patients with addiction, and six patients with addiction were included (five with heroin addiction and one with alcohol addiction). The microelectrode recording trajectories were reconstructed and recording sites at different depths were determined by merging the pre- and post-operative images in the FrameLink system. The results showed that among the 256 neurons, 204 (80 %) were burst neurons. NAc neurons accounted for the majority (57 %), and the mean firing rate (MFR) was the highest (1.94 Hz). ALIC neurons accounted for the least (14 %), and MFR was the lowest (0.44 Hz). MFR increased after entering the NAc and decreased after entering the ALIC. In the patients with addiction treated using DBS, the single-cell level electrophysiological characteristics of the different nuclei were found to be distinct along the surgical trajectory.

目的 基于功能MRI(fMRI)动态低频振幅(dALFF)观察孤独症谱系障碍(ASD)患儿脑活动特征.方法 于国际神经影像数据共享倡议组织公开数据集ABIDE Ⅱ中选取21例ASD患儿(ASD组)和37名健康儿童(对照组),对其脑fMRI数据进行预处理,计算dALFF值,观察组间脑活动差异.结果 相比对照组,ASD组左侧海马旁回dALFF值降低(AlphaSim校正P＜0.05,簇大小＞56)、左侧枕上回dALFF值升高(AlphaSim校正P＜0.05,簇大小＞56).结论 ASD患儿左侧海马旁回脑活动减弱而左侧枕上回脑活动增强.

To evaluate the potential effect of radiofrequency ablation and deep brain stimulation in patients with treatment-refractory Tourette syndrome (TS), this study enrolled thirteen patients with TS who were admitted to our hospital between August 2002 and September 2018. Four patients received a single- or multi-target radiofrequency ablation after local, potentiated, or general anesthesia; eight patients underwent deep brain stimulation (DBS) surgery; and one patient underwent both ablation and DBS surgery. The severity of tics and obsessive compulsive disorder symptoms and the quality of life were evaluated using the Yale Global Tic Severity Scale (YGTSS), Yale–Brown Obsessive Compulsive Scale (YBOCS), and Gilles de la Tourette Syndrome Quality of Life scale (GTS-QOL), respectively, before surgery, one month after surgery, and at the final follow-up after surgery, which was conducted in December 2018. A paired-sample t test and a multiple linear regression analysis were performed to analyze the data. All patients underwent the operation successfully without any severe complications. Overall, the YGTSS total scores at one month post-surgery (44.1 ± 22.3) and at the final visit (35.1 ± 23.7) were significantly decreased compared with those at baseline (75.1 ± 6.2; both p < 0.05). Additionally, the YBOCS scores at one month post-surgery (16.5 ± 10.1) and at the final visit (12.0 ± 9.5) were significantly decreased compared with those at baseline (22.5 ± 13.1; both p < 0.05). Furthermore, the GTS-QOL scores at one month post-surgery (44.0 ± 12.8) and at the final visit (31.0 ± 17.8) were significantly decreased compared with those at baseline (58.4 ± 14.2; both p < 0.05). Results from a multiple linear regression analysis revealed that the improvement in the YGTSS total score was independently associated with the improvement in the GTS-QOL score at one month post-surgery (standardized β = 0.716, p = 0.023) and at the final visit (standardized β = 1.064, p = 0.000). Conversely, changes in YBOCS scores did not correlate with changes in GTS-QOL scores (p > 0.05). Our results demonstrate that tics, psychiatric symptoms, and the quality of life in patients with intractable TS may be relieved by stereotactic ablation surgery and deep brain stimulation. Furthermore, it appears that the improvement in tics contributes more to the post-operative quality of life of patients than does the improvement in obsessive compulsive symptoms.

Two different but interacting neural systems exist in the human brain: the task positive networks and task negative networks. One of the most important task positive networks is the central executive network (CEN), while the task negative network generally refers to the default mode network (DMN), which usually demonstrates task-induced deactivation. Although previous studies have clearly shown the association of both the CEN and DMN with major depressive disorder (MDD), how the causal interactions between these two networks change in depressed patients remains unclear. In the current study, 99 subjects (43 patients with MDD and 56 healthy controls) were recruited with their resting-state fMRI data collected. After data preprocessing, spectral dynamic causal modeling (spDCM) was used to investigate the causal interactions within and between the DMN and CEN. Group commonalities and differences in causal interaction patterns within and between the CEN and DMN in patients and controls were assessed by a parametric empirical Bayes (PEB) model. Both subject groups demonstrated significant effective connectivity between regions of the CEN and DMN. In particular, we detected inhibitory influences from the CEN to the DMN with node-level PEB analyses, which may help to explain the anticorrelations between these two networks consistently reported in previous studies. Compared with healthy controls, patients with MDD showed increased effective connectivity within the CEN and decreased connectivity from regions of the CEN to DMN, suggesting impaired control of the DMN by the CEN in these patients. These findings might provide new insights into the neural substrates of MDD.

Objective: The symptom-related neurobiology characteristic of schizophrenia in the brain from a network perspective is still poorly understood, leading to a lack of potential biologically-based markers and difficulty identifying therapeutic targets. We aim to test the dysregulated cross-network interactions among the Salience Network (SN), Central Executive Network (CEN) and Default Mode Network (DMN) and how they contributed to different symptoms in schizophrenia patients. Methods: We examined network interactions among the SN, CEN and DMN in 76 patients with schizophrenia vs. 80 well-matched controls using dynamic causal modeling (DCM). We further analyzed the relation between network dynamics and Positive and Negative Syndrome Scale (PANSS). Results: We observed that the DMN, CEN and SN across healthy controls and schizophrenia patients showed several similarities within or between-network pattern in the resting state. Comparing schizophrenia to controls, SN-centered cross-network interactions were most significantly reduced. Crucially, the strength of connections from CEN subnetwork 1 to DMN subnetwork 1 was positively correlated with the Positive Score of PANSS. The connection from the DMN subnetwork 2 to CEN subnetwork 2 was negatively correlated with the Negative Score of PANSS. Conclusions: Our study provides strong evidence for the dysregulation among SN, CEN and DMN in a triplenetwork perspective in schizophrenia. The connection between DMN and CEN could be clinically-relevant neurobiological signature of schizophrenia symptoms. Our study indicated that the description of brain triple network hypothesis could be a novel and possible bio-marker for schizophrenia.

Matrix decomposition is an important technology in collaborative filtering recommendation algorithm. Based on SVD, some improvements, e.g., RSVD algorithm with offset term and regular term, and SVD++ algorithm with explicit and implicit feedback, was proposed. Considering the influence of time factor on the recommendation process, we discussed a time factor to the algorithms, which were associated with the users, the projects and the interaction offset. Based on the different time decay characteristics of video and users, an algorithm with exponential correction called timeSVD++EXP is proposed. The results show that the proposed algorithm can improve the prediction accuracy of the model when root mean square error is used as the evaluation index. And the prediction accuracy of the model is further increased by considering the time factor and exponential correction.

Rationale: Neuroadaptations in the medial prefrontal cortex (mPFC) and Nucleus Accumbens (NAc) play a role in the disruption of control-reward circuits in opioid addiction. Small Conductance Calcium-Activated Potassium (SK) channels in the mPFC have been implicated in neuronal excitability changes during morphine withdrawal. However, the mechanism that modulates SK channels during withdrawal is still unknown. Methods: Rats were exposed for one week to daily morphine injections (10 mg·kg-1 s.c.) followed by conditional place preference (CPP) assessment. One week after withdrawal, electrophysiological, morphological and molecular biological methods were applied to investigate the effects of morphine on SK channels in mPFC, including infralimbic (IL), prelimbic (PrL) cortices and NAc (core and shell). We verified the hypothesis that Rac1, a member of Rho family of small GTPases, implicated in SK channel regulation, modulate SK channel neuroadaptations during opiate withdrawal. Results: One week after morphine withdrawal, the neuronal excitability of layer 5 pyramidal neurons in IL was decreased, but not in PrL. Whereas, the excitability was increased in NAc-shell, but not in NAc-core. In mPFC, the expression of the SK3 subunit was enhanced after one-week of withdrawal compared to controls. In the IL, Rac1 signaling was increased during withdrawal, and the Rac1 inhibitor NSC23766 disrupted SK current, which increased neuronal firing. Suppression of Rac1 inhibited morphine-induced CPP and expression of SK channels in IL. Conclusions: These findings highlight the potential value of SK channels and the upstream molecule Rac1, which may throw light on the therapeutic mechanism of neuromodulation treatment for opioid dependence.

Executive function is a complex involving multiple advanced brain functions like planning, working memory, mental flexibility and psychomotor. Previous researches indicated that executive function may be impaired after acute or chronic high-altitude exposure, while the underlying neurobiological mechanism has not been totally clarified. In the present study, based on 69 young healthy volunteers immigrating to high-altitude, Stroop test was utilized to identify the potential impairment of executive function after two-year high-altitude exposure while resting-state functional MRI (rs-fMRI) technology was employed to observe the alteration of resting-state networks. Stroop test indicated that the subjects experienced significantly lower accuracies and prolonged responding time after two-year exposure. Resting-state network analysis displayed a significantly decreased degree of co-activation within the left/right frontoparietal network, sensorimotor network, and auditory network after exposure. In the frontoparietal network, decreased co-activation intensity was found in left angular gyrus, while in the right frontoparietal network, decreased co-activation intensity was found in left precentral gyrus and postcentral gyrus. Similarly, as for sensorimotor and auditory network, left middle frontal gyrus and left superior temporal gyrus was identified to have decreased co-activation, respectively. Moreover, the responding delays in ST (part II) were negatively correlated with the signal intensity alteration of the right frontoparietal network. All these evidences indicated that the high-altitude exposure induced alteration in above resting state networks may be the functional basis of executive control impairment.

MRI technology has been widely used in brain connectivity analysis to explore the neurobiological mechanism of ASD. Analysis of brain structural connectivity found atypical growth and development in patients with ASD, excessive growth in infancy, and excessive decline in functions from adolescence to middle age. Functional connectivity analysis suggests that default mode network (DMN) may be related to theory of mind (ToM) and self-referential processing, abnormal salience network (SN) connections can lead to social communication disorders, and abnormal reward network (RN) may be related to repetitive and restricted behaviors. Effective connectivity analysis found that information transmission between brain regions was abnormal in ASD patients, which was associated with social dysfunction. This article introduces the latest development of MRI research, and discusses the differences between ASD patients and normal human brain connections, so as to provide theoretical reference for further investigations of the neural activity mechanism of ASD patients.

目的 分析静脉压迫所导致三叉神经痛(TN)的术前影像学评估和显微血管减压术中责任静脉的处理.方法 回顾分析空军军医大学唐都医院神经外科自2017年12月至2018年12月收治的15例存在静脉压迫的TN患者的临床资料,分析术前影像学评估及术中处理.结果 所有患者均通过术前3D-TOF及3D-FIESTA检查,判断责任血管为单纯静脉型或动脉及静脉共同作用型,其中单纯静脉型11例,动脉及静脉共同作用型4例.通过术中分析,符合术前判断.所有患者静脉均得到有效保留,手术疗效满意.结论 通过术前影像学评估能有效预判TN的责任血管,从而指导术中操作.针对静脉压迫型TN,术中应尽量保护静脉,避免因损伤静脉所导致的术后并发症.

目的 探讨不同性别的重度抑郁症(MDD)患者在杏仁核脑区的脑动态功能连接差异.方法 将30例女性MDD患者、13例男性MDD患者、26例女性正常被试、30例男性正常被试分为4组进行rs-fMRI分析.取大脑的杏仁核脑区作为ROI,对各个被试的动态脑功能连接进行计算,使用K均值聚类方法对被试的动态脑功能连接图进行聚类.结果 根据K均值聚类结果,结合3种最优聚类评价指标,得到4组被试的最优聚类类别数,分别是MDD男性患者14类,女性患者7类,正常男性被试2类,正常女性被试3类.结论 MDD患者的脑区活动比正常被试紊乱.女性MDD患者的脑区活动比男性MDD患者的脑区活动紊乱.

目的:基于磁共振成像探究大脑镜像同伦脑区的功能连接强度在自闭症(autism spectrum disorder,ASD)儿童与正常儿童间的差异性,寻找潜在的ASD神经生物学机制.方法:募集83例年龄在2～6岁的ASD患儿(ASD组)和70例相同年龄的典型发育(typically developing,TD)儿童(TD组).对2组分别进行睡眠状态下功能磁共振成像扫描,分析体素镜像同伦连接(voxel-mirrored homotopic connectivity,VMHC)强度.使用单样本t检验分别得到2组的VMHC强度图谱,最后通过双样本t检验方法研究2组间大脑半球间连接的差异.结果:经错误发现率(false disco-very rate,FDR)控制校正后,与TD组相比,ASD组在颞上回、楔前叶、内侧额叶、尾状核的VMHC强度显著降低(P<0.05).结论:ASD患儿在静息状态下大脑两半球间功能连接降低,交互沟通能力下降.该研究为理解ASD的神经生物学机制提供了新视角.

Objective To evaluate the long-term efficacy of deep brain stimulation (DBS) for the treatment of primary generalized dystonia (PGD) and to analyze the factors influencing the clinical outcome.Methods We retrospectively analyzed the data of 21 patients with PGD who underwent DBS from December 2004 to April 2013 at Department of Neurosurgery,Tangdu Hospital,Air Force Medical University.DBS targeting the subthalamic nucleus (STN-DBS) was applied in 11 patients,and DBS targeting the globus pallidus internus (GPi-DBS) was performed in 10 patients.The severity of dystonia was assessed using the Burke-Fahn-Marsden Dystonia Rating Scale (BFMDRS).Multiple linear regression analysis was used to explore the factors affecting clinical outcomes.Results The follow-up period ranged from 5 years to 11 years,and the median time was 6.5 years.The average improvement rate of BFMDRS movement scores was 65.7％ ± 13.9％ and the median improvement rate of disability scores was 54.5 ％ [IQR (interquartile range):17.0％] in the 21 patients with PGD at the last follow-up.The mean/median improvement rates of movement score and disability score in STN-DBS group were 64.4％ ± 15.8％ and 57.1％ (IQR:20.2％) respectively,and those in GPi-DBS group were 67.1％ ± 12.2％ and 52.8％ (IQR:18.0％) respectively.There was no significant difference in the average improvement rate of BFMDRS movement score or disability score between STN-DBS and GPi-DBS group (both P ＞ 0.05).Multiple linear regression analysis showed that only the course of disease had a significant effect on clinical outcome,which was negatively correlated with the results (t =-7.082,P ＜ 0.001).The remaining variables including gender,age of onset,age of surgery,DBS target and preoperative movement score had no effect on clinical efficacy (all P ＞ 0.05).Conclusions DBS treatment of PGD is effective.Both GPi and STN can be used as alternative therapeutic targets and their efficacies seem similar.Patients with shorter disease duration may achieve better clinical outcomes.

Objective To explore the effects of abnormal effective connectivity within the default mode network (DMN)in relapsed patients with major depressive disorder.Methods Resting-state functional magnetic resonance imaging (rs-fMRI)data were collected from 21 patients of first-episode depression and 16 patients with relapsed depression,and 37 matched healthy controls.The effective connectivity within the DMN was investigated with spectral dynamic causal modeling (spDCM)method.Results spDCM analysis showed that the effective connections from left parietal cortex(LPC)to right parietal cortex(RPC)and medial frontal cortex(mPFC)were significantly decreased,while the connection from posterior cingulate cortex(PCC)to mPFC was increased in patients of first episode depression compared to healthy controls.Furthermore,the connectivity between mPFC and LPC were enhanced in patients with recurrent depression compared withhealthy control subjects,as well as PCC.Meanwhile,the connectivity between mPFC and PCC was enhanced in patients with relapse depression compared with patients of first episode depression.Conclusion Both first-episode and relapsed patients demonstrated abnormal effective connectivity of LPC,implicating that abnormal LPC connectivity may be associated with the neural substrates of depression.In contrast,the patients with relapsed depression showed aberrant connectivity with the mPFC,suggesting that abnormal effective connectivity of the mPFC may play an important role in the relapse of depression.

Background Deep brain stimulation (DBS) is currently used to treat addiction, with the nucleus accumbens (NAc) as one promising target. The anterior limb of the internal capsule (ALIC) is also a potential target, as it carries fiber tracts connecting the mesocorticolimbic circuits that are crucially involved in several psychiatric disorders, including addiction. Stimulating the NAc and ALIC simultaneously may have a synergistic effect against addiction. Methods Eight patients with a long history of heroin use and multiple relapses, despite optimal conventional treatments, were enrolled. Customized electrodes were implanted through the ALIC into the NAc, and deep brain stimulation (DBS) treatment began two weeks after surgery. The patients were followed for at least 24 months. The duration of drug-free time, severity of drug cravings, psychometric evaluations, and PET studies of glucose metabolism before and after DBS were conducted. All adverse events were recorded. Results With DBS, five patients were abstinent for more than three years, two relapsed after abstaining for six months, and one was lost of follow-up at three months. The degree of cravings for drug use after DBS was reduced if the patients remained abstinent (p < 0.001). Simultaneous DBS of the NAc and ALIC also improved the quality of life, alleviated psychiatric symptoms, and increased glucose metabolism in addiction-related brain regions. Moreover, stimulation-related adverse events were few and reversible. Conclusions Simultaneous DBS of the NAc and ALIC appears to be safe, with few side effects, and may prevent long-term heroin relapse after detoxification in certain patients. (This trial was registered at ClinicalTrials.gov, NCT01274988).