Rethinking LLM Unlearning Objectives: A Gradient Perspective and Go Beyond

Large language models (LLMs) should undergo rigorous audits to identify potential risks, such as copyright and privacy infringements. Once these risks emerge, timely updates are crucial to remove undesirable responses, ensuring legal and safe model usage. It has spurred recent research into LLM unlearning, focusing on erasing targeted undesirable knowledge without compromising the integrity of other, non-targeted responses. Existing studies have introduced various unlearning objectives to pursue LLM unlearning without necessitating complete retraining. However, each of these objectives has unique properties, and no unified framework is currently available to comprehend them thoroughly. To fill the gap, we propose a toolkit of the gradient effect (G-effect), quantifying the impacts of unlearning objectives on model performance from a gradient perspective. A notable advantage is its broad ability to detail the unlearning impacts from various aspects across instances, updating steps, and LLM layers. Accordingly, the G-effect offers new insights into identifying drawbacks of existing unlearning objectives, further motivating us to explore a series of new solutions for their mitigation and improvements. Finally, we outline promising directions that merit further studies, aiming at contributing to the community to advance this important field.

Noisy Test-Time Adaptation in Vision-Language Models

Test-time adaptation (TTA) aims to address distribution shifts between source and target data by relying solely on target data during testing. In open-world scenarios, models often encounter noisy samples, i.e., samples outside the in-distribution (ID) label space. Leveraging the zero-shot capability of pre-trained vision-language models (VLMs), this paper introduces Zero-Shot Noisy TTA (ZS-NTTA), focusing on adapting the model to target data with noisy samples during test-time in a zero-shot manner. We find existing TTA methods underperform under ZS-NTTA, often lagging behind even the frozen model. We conduct comprehensive experiments to analyze this phenomenon, revealing that the negative impact of unfiltered noisy data outweighs the benefits of clean data during model updating. Also, adapting a classifier for ID classification and noise detection hampers both sub-tasks. Built on this, we propose a framework that decouples the classifier and detector, focusing on developing an individual detector while keeping the classifier frozen. Technically, we introduce the Adaptive Noise Detector (AdaND), which utilizes the frozen model's outputs as pseudo-labels to train a noise detector. To handle clean data streams, we further inject Gaussian noise during adaptation, preventing the detector from misclassifying clean samples as noisy. Beyond the ZS-NTTA, AdaND can also improve the zero-shot out-of-distribution (ZS-OOD) detection ability of VLMs. Experiments show that AdaND outperforms in both ZS-NTTA and ZS-OOD detection. On ImageNet, AdaND achieves a notable improvement of $8.32\%$ in harmonic mean accuracy ($\text{Acc}_\text{H}$) for ZS-NTTA and $9.40\%$ in FPR95 for ZS-OOD detection, compared to SOTA methods. Importantly, AdaND is computationally efficient and comparable to the model-frozen method. The code is publicly available at: https://github.com/tmlr-group/ZS-NTTA.

Eliciting Causal Abilities in Large Language Models for Reasoning Tasks

Prompt optimization automatically refines prompting expressions, unlocking the full potential of LLMs in downstream tasks. However, current prompt optimization methods are costly to train and lack sufficient interpretability. This paper proposes enhancing LLMs' reasoning performance by eliciting their causal inference ability from prompting instructions to correct answers. Specifically, we introduce the Self-Causal Instruction Enhancement (SCIE) method, which enables LLMs to generate high-quality, low-quantity observational data, then estimates the causal effect based on these data, and ultimately generates instructions with the optimized causal effect. In SCIE, the instructions are treated as the treatment, and textual features are used to process natural language, establishing causal relationships through treatments between instructions and downstream tasks. Additionally, we propose applying Object-Relational (OR) principles, where the uncovered causal relationships are treated as the inheritable class across task objects, ensuring low-cost reusability. Extensive experiments demonstrate that our method effectively generates instructions that enhance reasoning performance with reduced training cost of prompts, leveraging interpretable textual features to provide actionable insights.

Physics Reasoner: Knowledge-Augmented Reasoning for Solving Physics Problems with Large Language Models

Physics problems constitute a significant aspect of reasoning, necessitating complicated reasoning ability and abundant physics knowledge. However, existing large language models (LLMs) frequently fail due to a lack of knowledge or incorrect knowledge application. To mitigate these issues, we propose Physics Reasoner, a knowledge-augmented framework to solve physics problems with LLMs. Specifically, the proposed framework constructs a comprehensive formula set to provide explicit physics knowledge and utilizes checklists containing detailed instructions to guide effective knowledge application. Namely, given a physics problem, Physics Reasoner solves it through three stages: problem analysis, formula retrieval, and guided reasoning. During the process, checklists are employed to enhance LLMs' self-improvement in the analysis and reasoning stages. Empirically, Physics Reasoner mitigates the issues of insufficient knowledge and incorrect application, achieving state-of-the-art performance on SciBench with an average accuracy improvement of 5.8%.

Model Inversion Attacks: A Survey of Approaches and Countermeasures

The success of deep neural networks has driven numerous research studies and applications from Euclidean to non-Euclidean data. However, there are increasing concerns about privacy leakage, as these networks rely on processing private data. Recently, a new type of privacy attack, the model inversion attacks (MIAs), aims to extract sensitive features of private data for training by abusing access to a well-trained model. The effectiveness of MIAs has been demonstrated in various domains, including images, texts, and graphs. These attacks highlight the vulnerability of neural networks and raise awareness about the risk of privacy leakage within the research community. Despite the significance, there is a lack of systematic studies that provide a comprehensive overview and deeper insights into MIAs across different domains. This survey aims to summarize up-to-date MIA methods in both attacks and defenses, highlighting their contributions and limitations, underlying modeling principles, optimization challenges, and future directions. We hope this survey bridges the gap in the literature and facilitates future research in this critical area. Besides, we are maintaining a repository to keep track of relevant research at https://github.com/AndrewZhou924/Awesome-model-inversion-attack.

Can Language Models Perform Robust Reasoning in Chain-of-thought Prompting with Noisy Rationales?

This paper investigates an under-explored challenge in large language models (LLMs): chain-of-thought prompting with noisy rationales, which include irrelevant or inaccurate reasoning thoughts within examples used for in-context learning. We construct NoRa dataset that is tailored to evaluate the robustness of reasoning in the presence of noisy rationales. Our findings on NoRa dataset reveal a prevalent vulnerability to such noise among current LLMs, with existing robust methods like self-correction and self-consistency showing limited efficacy. Notably, compared to prompting with clean rationales, base LLM drops by 1.4%-19.8% in accuracy with irrelevant thoughts and more drastically by 2.2%-40.4% with inaccurate thoughts. Addressing this challenge necessitates external supervision that should be accessible in practice. Here, we propose the method of contrastive denoising with noisy chain-of-thought (CD-CoT). It enhances LLMs' denoising-reasoning capabilities by contrasting noisy rationales with only one clean rationale, which can be the minimal requirement for denoising-purpose prompting. This method follows a principle of exploration and exploitation: (1) rephrasing and selecting rationales in the input space to achieve explicit denoising and (2) exploring diverse reasoning paths and voting on answers in the output space. Empirically, CD-CoT demonstrates an average improvement of 17.8% in accuracy over the base model and shows significantly stronger denoising capabilities than baseline methods. The source code is publicly available at: https://github.com/tmlr-group/NoisyRationales.

Mind the Gap Between Prototypes and Images in Cross-domain Finetuning

In cross-domain few-shot classification (CFC), recent works mainly focus on adapting a simple transformation head on top of a frozen pre-trained backbone with few labeled data to project embeddings into a task-specific metric space where classification can be performed by measuring similarities between image instance and prototype representations. Technically, an assumption implicitly adopted in such a framework is that the prototype and image instance embeddings share the same representation transformation. However, in this paper, we find that there naturally exists a gap, which resembles the modality gap, between the prototype and image instance embeddings extracted from the frozen pre-trained backbone, and simply applying the same transformation during the adaptation phase constrains exploring the optimal representations and shrinks the gap between prototype and image representations. To solve this problem, we propose a simple yet effective method, contrastive prototype-image adaptation (CoPA), to adapt different transformations respectively for prototypes and images similarly to CLIP by treating prototypes as text prompts. Extensive experiments on Meta-Dataset demonstrate that CoPA achieves the state-of-the-art performance more efficiently. Meanwhile, further analyses also indicate that CoPA can learn better representation clusters, enlarge the gap, and achieve minimal validation loss at the enlarged gap.

Envisioning Outlier Exposure by Large Language Models for Out-of-Distribution Detection

Detecting out-of-distribution (OOD) samples is essential when deploying machine learning models in open-world scenarios. Zero-shot OOD detection, requiring no training on in-distribution (ID) data, has been possible with the advent of vision-language models like CLIP. Existing methods build a text-based classifier with only closed-set labels. However, this largely restricts the inherent capability of CLIP to recognize samples from large and open label space. In this paper, we propose to tackle this constraint by leveraging the expert knowledge and reasoning capability of large language models (LLM) to Envision potential Outlier Exposure, termed EOE, without access to any actual OOD data. Owing to better adaptation to open-world scenarios, EOE can be generalized to different tasks, including far, near, and fine-grained OOD detection. Technically, we design (1) LLM prompts based on visual similarity to generate potential outlier class labels specialized for OOD detection, as well as (2) a new score function based on potential outlier penalty to distinguish hard OOD samples effectively. Empirically, EOE achieves state-of-the-art performance across different OOD tasks and can be effectively scaled to the ImageNet-1K dataset. The code is publicly available at: https://github.com/tmlr-group/EOE.

Less is More: One-shot Subgraph Reasoning on Large-scale Knowledge Graphs

To deduce new facts on a knowledge graph (KG), a link predictor learns from the graph structure and collects local evidence to find the answer to a given query. However, existing methods suffer from a severe scalability problem due to the utilization of the whole KG for prediction, which hinders their promise on large scale KGs and cannot be directly addressed by vanilla sampling methods. In this work, we propose the one-shot-subgraph link prediction to achieve efficient and adaptive prediction. The design principle is that, instead of directly acting on the whole KG, the prediction procedure is decoupled into two steps, i.e., (i) extracting only one subgraph according to the query and (ii) predicting on this single, query dependent subgraph. We reveal that the non-parametric and computation-efficient heuristics Personalized PageRank (PPR) can effectively identify the potential answers and supporting evidence. With efficient subgraph-based prediction, we further introduce the automated searching of the optimal configurations in both data and model spaces. Empirically, we achieve promoted efficiency and leading performances on five large-scale benchmarks. The code is publicly available at: https://github.com/tmlr-group/one-shot-subgraph.

DeepInception: Hypnotize Large Language Model to Be Jailbreaker

Despite remarkable success in various applications, large language models (LLMs) are vulnerable to adversarial jailbreaks that make the safety guardrails void. However, previous studies for jailbreaks usually resort to brute-force optimization or extrapolations of a high computation cost, which might not be practical or effective. In this paper, inspired by the Milgram experiment that individuals can harm another person if they are told to do so by an authoritative figure, we disclose a lightweight method, termed as DeepInception, which can easily hypnotize LLM to be a jailbreaker and unlock its misusing risks. Specifically, DeepInception leverages the personification ability of LLM to construct a novel nested scene to behave, which realizes an adaptive way to escape the usage control in a normal scenario and provides the possibility for further direct jailbreaks. Empirically, we conduct comprehensive experiments to show its efficacy. Our DeepInception can achieve competitive jailbreak success rates with previous counterparts and realize a continuous jailbreak in subsequent interactions, which reveals the critical weakness of self-losing on both open/closed-source LLMs like Falcon, Vicuna, Llama-2, and GPT-3.5/4/4V. Our investigation appeals that people should pay more attention to the safety aspects of LLMs and a stronger defense against their misuse risks. The code is publicly available at: https://github.com/tmlr-group/DeepInception.