Language models can retain dangerous knowledge and skills even after extensive safety fine-tuning, posing both misuse and misalignment risks. Recent studies show that even specialized unlearning methods can be easily reversed. To address this, we systematically evaluate many existing and novel components of unlearning methods and identify ones crucial for irreversible unlearning. We introduce Disruption Masking, a technique in which we only allow updating weights, where the signs of the unlearning gradient and the retaining gradient are the same. This ensures all updates are non-disruptive. Additionally, we identify the need for normalizing the unlearning gradients, and also confirm the usefulness of meta-learning. We combine these insights into MUDMAN (Meta-Unlearning with Disruption Masking and Normalization) and validate its effectiveness at preventing the recovery of dangerous capabilities. MUDMAN outperforms the prior TAR method by 40\%, setting a new state-of-the-art for robust unlearning.
This paper presents a Technology Assisted Review (TAR) stopping approach based on Reinforcement Learning (RL). Previous such approaches offered limited control over stopping behaviour, such as fixing the target recall and tradeoff between preferring to maximise recall or cost. These limitations are overcome by introducing a novel RL environment, GRLStop, that allows a single model to be applied to multiple target recalls, balances the recall/cost tradeoff and integrates a classifier. Experiments were carried out on six benchmark datasets (CLEF e-Health datasets 2017-9, TREC Total Recall, TREC Legal and Reuters RCV1) at multiple target recall levels. Results showed that the proposed approach to be effective compared to multiple baselines in addition to offering greater flexibility.
Harmful fine-tuning attack poses a serious threat to the online fine-tuning service. Vaccine, a recent alignment-stage defense, applies uniform perturbation to all layers of embedding to make the model robust to the simulated embedding drift. However, applying layer-wise uniform perturbation may lead to excess perturbations for some particular safety-irrelevant layers, resulting in defense performance degradation and unnecessary memory consumption. To address this limitation, we propose Targeted Vaccine (T-Vaccine), a memory-efficient safety alignment method that applies perturbation to only selected layers of the model. T-Vaccine follows two core steps: First, it uses gradient norm as a statistical metric to identify the safety-critical layers. Second, instead of applying uniform perturbation across all layers, T-Vaccine only applies perturbation to the safety-critical layers while keeping other layers frozen during training. Results show that T-Vaccine outperforms Vaccine in terms of both defense effectiveness and resource efficiency. Comparison with other defense baselines, e.g., RepNoise and TAR also demonstrate the superiority of T-Vaccine. Notably, T-Vaccine is the first defense that can address harmful fine-tuning issues for a 7B pre-trained models trained on consumer GPUs with limited memory (e.g., RTX 4090). Our code is available at https://github.com/Lslland/T-Vaccine.
Rapid advances in the capabilities of large language models (LLMs) have raised widespread concerns regarding their potential for malicious use. Open-weight LLMs present unique challenges, as existing safeguards lack robustness to tampering attacks that modify model weights. For example, recent works have demonstrated that refusal and unlearning safeguards can be trivially removed with a few steps of fine-tuning. These vulnerabilities necessitate new approaches for enabling the safe release of open-weight LLMs. We develop a method, called TAR, for building tamper-resistant safeguards into open-weight LLMs such that adversaries cannot remove the safeguards even after thousands of steps of fine-tuning. In extensive evaluations and red teaming analyses, we find that our method greatly improves tamper-resistance while preserving benign capabilities. Our results demonstrate that tamper-resistance is a tractable problem, opening up a promising new avenue to improve the safety and security of open-weight LLMs.
The goal of screening prioritisation in systematic reviews is to identify relevant documents with high recall and rank them in early positions for review. This saves reviewing effort if paired with a stopping criterion, and speeds up review completion if performed alongside downstream tasks. Recent studies have shown that neural models have good potential on this task, but their time-consuming fine-tuning and inference discourage their widespread use for screening prioritisation. In this paper, we propose an alternative approach that still relies on neural models, but leverages dense representations and relevance feedback to enhance screening prioritisation, without the need for costly model fine-tuning and inference. This method exploits continuous relevance feedback from reviewers during document screening to efficiently update the dense query representation, which is then applied to rank the remaining documents to be screened. We evaluate this approach across the CLEF TAR datasets for this task. Results suggest that the investigated dense query-driven approach is more efficient than directly using neural models and shows promising effectiveness compared to previous methods developed on the considered datasets. Our code is available at https://github.com/ielab/dense-screening-feedback.
Advanced diffusion models (DMs) perform impressively in image super-resolution (SR), but the high memory and computational costs hinder their deployment. Binarization, an ultra-compression algorithm, offers the potential for effectively accelerating DMs. Nonetheless, due to the model structure and the multi-step iterative attribute of DMs, existing binarization methods result in significant performance degradation. In this paper, we introduce a novel binarized diffusion model, BI-DiffSR, for image SR. First, for the model structure, we design a UNet architecture optimized for binarization. We propose the consistent-pixel-downsample (CP-Down) and consistent-pixel-upsample (CP-Up) to maintain dimension consistent and facilitate the full-precision information transfer. Meanwhile, we design the channel-shuffle-fusion (CS-Fusion) to enhance feature fusion in skip connection. Second, for the activation difference across timestep, we design the timestep-aware redistribution (TaR) and activation function (TaA). The TaR and TaA dynamically adjust the distribution of activations based on different timesteps, improving the flexibility and representation alability of the binarized module. Comprehensive experiments demonstrate that our BI-DiffSR outperforms existing binarization methods. Code is released at: https://github.com/zhengchen1999/BI-DiffSR.
Object tracking is a fundamental task in computer vision with broad practical applications across various domains, including traffic monitoring, robotics, and autonomous vehicle tracking. In this project, we aim to develop a sophisticated aerial vehicle system known as Track Anything Rapter (TAR), designed to detect, segment, and track objects of interest based on user-provided multimodal queries, such as text, images, and clicks. TAR utilizes cutting-edge pre-trained models like DINO, CLIP, and SAM to estimate the relative pose of the queried object. The tracking problem is approached as a Visual Servoing task, enabling the UAV to consistently focus on the object through advanced motion planning and control algorithms. We showcase how the integration of these foundational models with a custom high-level control algorithm results in a highly stable and precise tracking system deployed on a custom-built PX4 Autopilot-enabled Voxl2 M500 drone. To validate the tracking algorithm's performance, we compare it against Vicon-based ground truth. Additionally, we evaluate the reliability of the foundational models in aiding tracking in scenarios involving occlusions. Finally, we test and validate the model's ability to work seamlessly with multiple modalities, such as click, bounding box, and image templates.
High Recall Retrieval (HRR), such as eDiscovery and medical systematic review, is a search problem that optimizes the cost of retrieving most relevant documents in a given collection. Iterative approaches, such as iterative relevance feedback and uncertainty sampling, are shown to be effective under various operational scenarios. Despite neural models demonstrating success in other text-related tasks, linear models such as logistic regression, in general, are still more effective and efficient in HRR since the model is trained and retrieves documents from the same fixed collection. In this work, we leverage SPLADE, an efficient retrieval model that transforms documents into contextualized sparse vectors, for HRR. Our approach combines the best of both worlds, leveraging both the contextualization from pretrained language models and the efficiency of linear models. It reduces 10% and 18% of the review cost in two HRR evaluation collections under a one-phase review workflow with a target recall of 80%. The experiment is implemented with TARexp and is available at https://github.com/eugene-yang/LSR-for-TAR.
We present RLStop, a novel Technology Assisted Review (TAR) stopping rule based on reinforcement learning that helps minimise the number of documents that need to be manually reviewed within TAR applications. RLStop is trained on example rankings using a reward function to identify the optimal point to stop examining documents. Experiments at a range of target recall levels on multiple benchmark datasets (CLEF e-Health, TREC Total Recall, and Reuters RCV1) demonstrated that RLStop substantially reduces the workload required to screen a document collection for relevance. RLStop outperforms a wide range of alternative approaches, achieving performance close to the maximum possible for the task under some circumstances.
Screening documents is a tedious and time-consuming aspect of high-recall retrieval tasks, such as compiling a systematic literature review, where the goal is to identify all relevant documents for a topic. To help streamline this process, many Technology-Assisted Review (TAR) methods leverage active learning techniques to reduce the number of documents requiring review. BERT-based models have shown high effectiveness in text classification, leading to interest in their potential use in TAR workflows. In this paper, we investigate recent work that examined the impact of further pre-training epochs on the effectiveness and efficiency of a BERT-based active learning pipeline. We first report that we could replicate the original experiments on two specific TAR datasets, confirming some of the findings: importantly, that further pre-training is critical to high effectiveness, but requires attention in terms of selecting the correct training epoch. We then investigate the generalisability of the pipeline on a different TAR task, that of medical systematic reviews. In this context, we show that there is no need for further pre-training if a domain-specific BERT backbone is used within the active learning pipeline. This finding provides practical implications for using the studied active learning pipeline within domain-specific TAR tasks.
Technology Assisted Review (TAR) stopping rules aim to reduce the cost of manually assessing documents for relevance by minimising the number of documents that need to be examined to ensure a desired level of recall. This paper extends an effective stopping rule using information derived from a text classifier that can be trained without the need for any additional annotation. Experiments on multiple data sets (CLEF e-Health, TREC Total Recall, TREC Legal and RCV1) showed that the proposed approach consistently improves performance and outperforms several alternative methods.
Technology Assisted Review (TAR), which aims to reduce the effort required to screen collections of documents for relevance, is used to develop systematic reviews of medical evidence and identify documents that must be disclosed in response to legal proceedings. Stopping methods are algorithms which determine when to stop screening documents during the TAR process, helping to ensure that workload is minimised while still achieving a high level of recall. This paper proposes a novel stopping method based on point processes, which are statistical models that can be used to represent the occurrence of random events. The approach uses rate functions to model the occurrence of relevant documents in the ranking and compares four candidates, including one that has not previously been used for this purpose (hyperbolic). Evaluation is carried out using standard datasets (CLEF e-Health, TREC Total Recall, TREC Legal), and this work is the first to explore stopping method robustness by reporting performance on a range of rankings of varying effectiveness. Results show that the proposed method achieves the desired level of recall without requiring an excessive number of documents to be examined in the majority of cases and also compares well against multiple alternative approaches.
Most existing MRI reconstruction methods perform tar-geted reconstruction of the entire MR image without tak-ing specific tissue regions into consideration. This may fail to emphasize the reconstruction accuracy on im-portant tissues for diagnosis. In this study, leveraging a combination of the properties of k-space data and the diffusion process, our novel scheme focuses on mining the multi-frequency prior with different strategies to pre-serve fine texture details in the reconstructed image. In addition, a diffusion process can converge more quickly if its target distribution closely resembles the noise distri-bution in the process. This can be accomplished through various high-frequency prior extractors. The finding further solidifies the effectiveness of the score-based gen-erative model. On top of all the advantages, our method improves the accuracy of MRI reconstruction and accel-erates sampling process. Experimental results verify that the proposed method successfully obtains more accurate reconstruction and outperforms state-of-the-art methods.
This foundational research provides additional support for using the Fuzzy ARTMAP neural network as a classification algorithm in the TAR domain. While research opportunities exist to improve recall performance and explanation, the robust recall results from this study and the proof-of-concept demonstration of If-Then rules for tf-idf vectorization strongly substantiate that a Fuzzy ARTMAP-based TAR system is a potentially viable explainable alternative to "black box" TAR systems.
Threshold Autoregressive (TAR) models have been widely used by statisticians for non-linear time series forecasting during the past few decades, due to their simplicity and mathematical properties. On the other hand, in the forecasting community, general-purpose tree-based regression algorithms (forests, gradient-boosting) have become popular recently due to their ease of use and accuracy. In this paper, we explore the close connections between TAR models and regression trees. These enable us to use the rich methodology from the literature on TAR models to define a hierarchical TAR model as a regression tree that trains globally across series, which we call SETAR-Tree. In contrast to the general-purpose tree-based models that do not primarily focus on forecasting, and calculate averages at the leaf nodes, we introduce a new forecasting-specific tree algorithm that trains global Pooled Regression (PR) models in the leaves allowing the models to learn cross-series information and also uses some time-series-specific splitting and stopping procedures. The depth of the tree is controlled by conducting a statistical linearity test commonly employed in TAR models, as well as measuring the error reduction percentage at each node split. Thus, the proposed tree model requires minimal external hyperparameter tuning and provides competitive results under its default configuration. We also use this tree algorithm to develop a forest where the forecasts provided by a collection of diverse SETAR-Trees are combined during the forecasting process. In our evaluation on eight publicly available datasets, the proposed tree and forest models are able to achieve significantly higher accuracy than a set of state-of-the-art tree-based algorithms and forecasting benchmarks across four evaluation metrics.
The lack of high fidelity and publicly available longitudinal children face datasets is one of the main limiting factors in the development of face recognition systems for children. In this work, we introduce the Young Face Aging (YFA) dataset for analyzing the performance of face recognition systems over short age-gaps in children. We expand previous work by comparing YFA with several publicly available cross-age adult datasets to quantify the effects of short age-gap in adults and children. Our analysis confirms a statistically significant and matcher independent decaying relationship between the match scores of ArcFace-Focal, MagFace, and Facenet matchers and the age-gap between the gallery and probe images in children, even at the short age-gap of 6 months. However, our result indicates that the low verification performance reported in previous work might be due to the intra-class structure of the matcher and the lower quality of the samples. Our experiment using YFA and a state-of-the-art, quality-aware face matcher (MagFace) indicates 98.3% and 94.9% TAR at 0.1% FAR over 6 and 36 Months age-gaps, respectively, suggesting that face recognition may be feasible for children for age-gaps of up to three years.
Technology-assisted review (TAR) is an important industrial application of information retrieval (IR) and machine learning (ML). While a small TAR research community exists, the complexity of TAR software and workflows is a major barrier to entry. Drawing on past open source TAR efforts, as well as design patterns from the IR and ML open source software, we present an open source Python framework for conducting experiments on TAR algorithms. Key characteristics of this framework are declarative representations of workflows and experiment plans, the ability for components to play variable numbers of workflow roles, and state maintenance and restart capabilities. Users can draw on reference implementations of standard TAR algorithms while incorporating novel components to explore their research interests. The framework is available at https://github.com/eugene-yang/tarexp.
Representation knowledge distillation aims at transferring rich information from one model to another. Common approaches for representation distillation mainly focus on the direct minimization of distance metrics between the models' embedding vectors. Such direct methods may be limited in transferring high-order dependencies embedded in the representation vectors, or in handling the capacity gap between the teacher and student models. Moreover, in standard knowledge distillation, the teacher is trained without awareness of the student's characteristics and capacity. In this paper, we explore two mechanisms for enhancing representation distillation using classifier sharing between the teacher and student. We first investigate a simple scheme where the teacher's classifier is connected to the student backbone, acting as an additional classification head. Then, we propose a student-aware mechanism that asks to tailor the teacher model to a student with limited capacity by training the teacher with a temporary student's head. We analyze and compare these two mechanisms and show their effectiveness on various datasets and tasks, including image classification, fine-grained classification, and face verification. In particular, we achieve state-of-the-art results for face verification on the IJB-C dataset for a MobileFaceNet model: TAR@(FAR=1e-5)=93.7\%. Code is available at https://github.com/Alibaba-MIIL/HeadSharingKD.
It is a well-known approach for fringe groups and organizations to use euphemisms -- ordinary-sounding and innocent-looking words with a secret meaning -- to conceal what they are discussing. For instance, drug dealers often use "pot" for marijuana and "avocado" for heroin. From a social media content moderation perspective, though recent advances in NLP have enabled the automatic detection of such single-word euphemisms, no existing work is capable of automatically detecting multi-word euphemisms, such as "blue dream" (marijuana) and "black tar" (heroin). Our paper tackles the problem of euphemistic phrase detection without human effort for the first time, as far as we are aware. We first perform phrase mining on a raw text corpus (e.g., social media posts) to extract quality phrases. Then, we utilize word embedding similarities to select a set of euphemistic phrase candidates. Finally, we rank those candidates by a masked language model -- SpanBERT. Compared to strong baselines, we report 20-50% higher detection accuracies using our algorithm for detecting euphemistic phrases.
Content moderation (removing or limiting the distribution of posts based on their contents) is one tool social networks use to fight problems such as harassment and disinformation. Manually screening all content is usually impractical given the scale of social media data, and the need for nuanced human interpretations makes fully automated approaches infeasible. We consider content moderation from the perspective of technology-assisted review (TAR): a human-in-the-loop active learning approach developed for high recall retrieval problems in civil litigation and other fields. We show how TAR workflows, and a TAR cost model, can be adapted to the content moderation problem. We then demonstrate on two publicly available content moderation data sets that a TAR workflow can reduce moderation costs by 20% to 55% across a variety of conditions.