Wanli Yang

Hexiang Tan, Wanli Yang, Junwei Zhang et al.

Reliable confidence is essential for trusting the outputs of LLMs, yet widely deployed post-trained LLMs (PoLLMs) typically compromise this trust with severe overconfidence. In contrast, we observe that their corresponding base LLMs often remain well-calibrated. This naturally motivates us to calibrate PoLLM confidence using the base LLM as a reference. This work proposes two ways to achieve this. A straightforward solution, BaseCal-ReEval, evaluates PoLLM's responses by feeding them into the base LLM to get average probabilities as confidence. While effective, this approach introduces additional inference overhead. To address this, we propose BaseCal-Proj, which trains a lightweight projection to map the final-layer hidden states of PoLLMs back to those of their base LLMs. These projected states are then processed by the base LLM's output layer to derive base-calibrated confidence for PoLLM's responses. Notably, BaseCal is an unsupervised, plug-and-play solution that operates without human labels or LLM modifications. Experiments across five datasets and three LLM families demonstrate the effectiveness of BaseCal, reducing Expected Calibration Error (ECE) by an average of 42.90\% compared to the best unsupervised baselines.

Linyong Zou, Liang Zhang, Xiongfei Wang et al.

An essential technique for diagnosing brain disorders is electrophysiological source imaging (ESI). While model-based optimization and deep learning methods have achieved promising results in this field, the accurate selection and refinement of features remains a central challenge for precise ESI. This paper proposes FAIR-ESI, a novel framework that adaptively refines feature importance across different views, including FFT-based spectral feature refinement, weighted temporal feature refinement, and self-attention-based patch-wise feature refinement. Extensive experiments on two simulation datasets with diverse configurations and two real-world clinical datasets validate our framework's efficacy, highlighting its potential to advance brain disorder diagnosis and offer new insights into brain function.

Wanli Yang, Hongyu Zang, Junwei Zhang et al.

Reinforcement learning (RL) has achieved remarkable success in LLM reasoning, but whether it can also improve direct recall of parametric knowledge remains an open question. We study this question in a controlled zero-shot, one-hop, closed-book QA setting with no chain-of-thought, training only on binary correctness rewards and applying fact-level train-test deduplication to ensure gains reflect improved recall rather than reasoning or memorization. Across three model families and multiple factual QA benchmarks, RL yields ~27% average relative gains, surpassing both training- and inference-time baselines alike. Mechanistically, RL primarily redistributes probability mass over existing knowledge rather than acquiring new facts, moving correct answers from the low-probability tail into reliable greedy generations. Our data-attribution study reveals that the hardest examples are the most informative: those whose answers never appear in 128 pre-RL samples (only ~18% of training data) drive ~83% of the gain, since rare correct rollouts still emerge during training and get reinforced. Together, these findings broaden the role of RL beyond reasoning, repositioning it as a tool for unlocking rather than acquiring latent parametric knowledge.

Wanli Yang, Fei Sun, Xinyu Ma et al.

Although model editing has shown promise in revising knowledge in Large Language Models (LLMs), its impact on the inherent capabilities of LLMs is often overlooked. In this work, we reveal a critical phenomenon: even a single edit can trigger model collapse, manifesting as significant performance degradation in various benchmark tasks. However, benchmarking LLMs after each edit, while necessary to prevent such collapses, is impractically time-consuming and resource-intensive. To mitigate this, we propose using perplexity as a surrogate metric, validated by extensive experiments demonstrating changes in an edited model's perplexity are strongly correlated with its downstream task performances. We further conduct an in-depth study on sequential editing, a practical setting for real-world scenarios, across various editing methods and LLMs, focusing on hard cases from our previous single edit studies. The results indicate that nearly all examined editing methods result in model collapse after only few edits. To facilitate further research, we have utilized GPT-3.5 to develop a new dataset, HardEdit, based on those hard cases. This dataset aims to establish the foundation for pioneering research in reliable model editing and the mechanisms underlying editing-induced model collapse. We hope this work can draw the community's attention to the potential risks inherent in model editing practices.

Hexiang Tan, Fei Sun, Wanli Yang et al.

While auxiliary information has become a key to enhancing Large Language Models (LLMs), relatively little is known about how LLMs merge these contexts, specifically contexts generated by LLMs and those retrieved from external sources. To investigate this, we formulate a systematic framework to identify whether LLMs' responses are attributed to either generated or retrieved contexts. To easily trace the origin of the response, we construct datasets with conflicting contexts, i.e., each question is paired with both generated and retrieved contexts, yet only one of them contains the correct answer. Our experiments reveal a significant bias in several LLMs (GPT-4/3.5 and Llama2) to favor generated contexts, even when they provide incorrect information. We further identify two key factors contributing to this bias: i) contexts generated by LLMs typically show greater similarity to the questions, increasing their likelihood of being selected; ii) the segmentation process used in retrieved contexts disrupts their completeness, thereby hindering their full utilization in LLMs. Our analysis enhances the understanding of how LLMs merge diverse contexts, offers valuable insights for advancing current LLM augmentation methods, and highlights the risk of generated misinformation for retrieval-augmented LLMs.

Wanli Yang, Fei Sun, Jiajun Tan et al.

Despite near-perfect results reported in the literature, the effectiveness of model editing in real-world applications remains unclear. To bridge this gap, we introduce QAEdit, a new benchmark aligned with widely used question answering (QA) datasets, and WILD, a task-agnostic evaluation framework designed to better reflect real-world usage of model editing. Our single editing experiments show that current editing methods perform substantially worse than previously reported (38.5% vs. 96.8%). We demonstrate that it stems from issues in the synthetic evaluation practices of prior work. Among them, the most severe is the use of teacher forcing during testing, which leaks both content and length of the ground truth, leading to overestimated performance. Furthermore, we simulate practical deployment by sequential editing, revealing that current approaches fail drastically with only 1000 edits. This work calls for a shift in model editing research toward rigorous evaluation and the development of robust, scalable methods that can reliably update knowledge in LLMs for real-world use.

Hanyang Dong, Shurong Sheng, Xiongfei Wang et al.

A Magnetoencephalography (MEG) time-series recording consists of multi-channel signals collected by superconducting sensors, with each signal's intensity reflecting magnetic field changes over time at the sensor location. Automating epileptic MEG spike detection significantly reduces manual assessment time and effort, yielding substantial clinical benefits. Existing research addresses MEG spike detection by encoding neural network inputs with signals from all channel within a time segment, followed by classification. However, these methods overlook simultaneous spiking occurred from nearby sensors. We introduce a simple yet effective paradigm that first clusters MEG channels based on their sensor's spatial position. Next, a novel convolutional input module is designed to integrate the spatial clustering and temporal changes of the signals. This module is fed into a custom MEEG-ResNet3D developed by the authors, which learns to extract relevant features and classify the input as a spike clip or not. Our method achieves an F1 score of 94.73% on a large real-world MEG dataset Sanbo-CMR collected from two centers, outperforming state-of-the-art approaches by 1.85%. Moreover, it demonstrates efficacy and stability in the Electroencephalographic (EEG) seizure detection task, yielding an improved weighted F1 score of 1.4% compared to current state-of-the-art techniques evaluated on TUSZ, whch is the largest EEG seizure dataset.

Wanli Yang, Fei Sun, Rui Tang et al.

Fine-tuning, a foundational method for adapting large language models, has long been considered ineffective for model editing. Here, we challenge this belief, arguing that the reported failure arises not from the inherent limitation of fine-tuning itself, but from adapting it to the sequential nature of the editing task, a single-pass depth-first pipeline that optimizes each sample to convergence before moving on. While intuitive, this depth-first pipeline coupled with sample-wise updating over-optimizes each edit and induces interference across edits. Our controlled experiments reveal that simply restoring fine-tuning to the standard breadth-first (i.e., epoch-based) pipeline with mini-batch optimization substantially improves its effectiveness for model editing. Moreover, fine-tuning in editing also suffers from suboptimal tuning parameter locations inherited from prior methods. Through systematic analysis of tuning locations, we derive LocFT-BF, a simple and effective localized editing method built on the restored fine-tuning framework. Extensive experiments across diverse LLMs and datasets demonstrate that LocFT-BF outperforms state-of-the-art methods by large margins. Notably, to our knowledge, it is the first to sustain 100K edits and 72B-parameter models,10 x beyond prior practice, without sacrificing general capabilities. By clarifying a long-standing misconception and introducing a principled localized tuning strategy, we advance fine-tuning from an underestimated baseline to a leading method for model editing, establishing a solid foundation for future research.

Liang Zhang, Hanyang Dong, Jia-Hong Gao et al.

Deep learning-based EEG classification is crucial for the automated detection of neurological disorders, improving diagnostic accuracy and enabling early intervention. However, the low signal-to-noise ratio of EEG signals limits model performance, making feature selection (FS) vital for optimizing representations learned by neural network encoders. Existing FS methods are seldom designed specifically for EEG diagnosis; many are architecture-dependent and lack interpretability, limiting their applicability. Moreover, most rely on single-iteration data, resulting in limited robustness to variability. To address these issues, we propose IEFS-GMB, an Information Entropy-based Feature Selection method guided by a Gradient Memory Bank. This approach constructs a dynamic memory bank storing historical gradients, computes feature importance via information entropy, and applies entropy-based weighting to select informative EEG features. Experiments on four public neurological disease datasets show that encoders enhanced with IEFS-GMB achieve accuracy improvements of 0.64% to 6.45% over baseline models. The method also outperforms four competing FS techniques and improves model interpretability, supporting its practical use in clinical settings.

Kuntao Xiao, Xiongfei Wang, Pengfei Teng et al.

The analysis of interictal epileptiform discharges (IEDs) in magnetoencephalography (MEG) or electroencephalogram (EEG) recordings represents a critical component in the diagnosis of epilepsy. However, manual analysis of these IEDs, which appear as epileptic spikes, from the large amount of MEG/EEG data is labor intensive and requires high expertise. Although automated methods have been developed to address this challenge, current approaches fail to fully emulate clinical experts' diagnostic intelligence in two key aspects: (1) their analysis on the input signals is limited to short temporal windows matching individual spike durations, missing the extended contextual patterns clinicians use to assess significance; and (2) they fail to adequately capture the dipole patterns with simultaneous positive-negative potential distributions across adjacent sensors that serve as clinicians' key diagnostic criterion for IED identification. To bridge this artificial-human intelligence gap, we propose a novel deep learning framework LV-CadeNet that integrates two key innovations: (1) a Long-View morphological feature representation that mimics expert clinicians' comprehensive assessment of both local spike characteristics and long-view contextual information, and (2) a hierarchical Encoder-Decoder NETwork that employs Convolution-Attention blocks for multi-scale spatiotemporal feature learning with progressive abstraction. Extensive evaluations confirm the superior performance of LV-CadeNet, which outperforms six state-of-the-art methods in EEG spike classification on TUEV, the largest public EEG spike dataset. Additionally, LV-CadeNet attains a significant improvement of 13.58% in balanced accuracy over the leading baseline for MEG spike detection on a clinical MEG dataset from Sanbo Brain Hospital, Capital Medical University.

Wanli Yang, Fei Sun, Jiajun Tan et al.

Despite significant progress in model editing methods, their application in real-world scenarios remains challenging as they often cause large language models (LLMs) to collapse. Among them, ROME is particularly concerning, as it could disrupt LLMs with only a single edit. In this paper, we study the root causes of such collapse. Through extensive analysis, we identify two primary factors that contribute to the collapse: i) inconsistent handling of prefixed and unprefixed keys in the parameter update equation may result in very small denominators, causing excessively large parameter updates; ii) the subject of collapse cases is usually the first token, whose unprefixed key distribution significantly differs from the prefixed key distribution in autoregressive transformers, causing the aforementioned issue to materialize. To validate our findings, we propose a simple yet effective approach: uniformly using prefixed keys during editing phase and adding prefixes during testing phase to ensure the consistency between training and testing. The experimental results show that the proposed solution can prevent model collapse while maintaining the effectiveness of the edits.