Refusal on harmful prompts is a key safety behaviour in instruction-tuned large language models (LLMs), yet the internal causes of this behaviour remain poorly understood. We study two public instruction tuned models-Gemma-22B-IT and LLaMA-3.1-8B-IT using sparse autoencoders (SAEs) trained on residual-stream activations. Given a harmful prompt, we search the SAE latent space for feature sets whose ablation flips the model from refusal to compliance, demonstrating causal influence and creating a jailbreak. Our search proceeds in three stages: 1. Refusal Direction - Finding a refusal mediating direction and collecting SAE features close to that direction, followed by 2. Greedy Filtering - to prune this set to obtain a minimal set and finally 3. Interaction Discovery - a factorization-machine (FM) model that captures non-linear interactions among the remaining active features and the minimal set. This pipeline yields a broad set of jailbreak-critical features, offering insight into the mechanistic basis of refusal. Moreover, we also find evidence of redundant features which remain dormant unless earlier features are suppressed. Our findings highlight the potential for fine-grained auditing and targeted intervention in safety behaviours by manipulating the interpretable latent space. Code- https://github.com/Social- AI-Studio/LLM refusal interp Extended version- https://arxiv.org/abs/2509.09708
Large language models (LLMs) have shown strong potential in automating the design of agentic workflows. However, existing methods still rely heavily on manually predefined operators, limiting generalization and scalability. To address this issue, we propose A(2)Flow, a fully automated framework for agentic workflow generation based on self-adaptive abstraction operators. A(2)Flow employs a three-stage operator extraction process: 1) Case-based Initial Operator Generation: leveraging expert demonstrations and LLM reasoning to generate case-specific operators; 2) Operator Clustering and Preliminary Abstraction: grouping similar operators across tasks to form preliminary abstractions; and 3) Deep Extraction for Abstract Execution Operators: applying long chain-of-thought prompting and multi-path reasoning to derive compact and generalizable execution operators. These operators serve as reusable building blocks for workflow construction without manual predefinition. Furthermore, we enhance node-level workflow search with an operator memory mechanism, which retains historical outputs to enrich context and improve decision-making. Experiments on general and embodied benchmarks show that A(2)Flow achieves a 2.4% and 19.3% average performance improvement and reduces resource usage by 37% over state-of-the-art baselines.
Recent years have witnessed growing scholarly interest in binary post-training quantization (PTQ) techniques for large language models (LLMs). While state-of-the-art (SOTA) binary quantization methods significantly reduce memory footprint and computational demands, they introduce additional memory overhead beyond binary weight tensors to mitigate performance degradation. Moreover, binary LLMs still suffer from substantial accuracy loss. To address these limitations, we propose MemeBQ, a novel binary PTQ framework for LLMs that reduces the memory overhead of auxiliary flag bitmaps in existing binary quantization methods. Specifically, we first design a greedy row clustering method, which leverages the similarity between the row vectors of weights to partition the weight rows into different groups. By sharing the common flag bitmap within each row group, we significantly mitigate the memory overhead associated with flag bitmaps. Besides, to improve the performance of binary LLMs, we propose a novel weight splitting method for each row group of weights, which determines the flag bitmap's values in a fine-grained way. Extensive experiments on OPT, Llama-2, and Llama-3 models demonstrate that MemeBQ reduces 50% extra memory demand while achieving comparable accuracy compared with current SOTA methods. Alternatively, MemeBQ outperforms SOTA binary quantization methods up to 7% with the same extra bits on reasoning benchmarks.
Exploration in sparse-reward tasks remains a fundamental challenge in multi-agent reinforcement learning (MARL) due to complex inter-agent interactions and the expansive exploration space. To address this issue, we propose Targeted Multi-Agent Exploration (TMAE), a novel framework that uncovers the causal relationships between the state space and the reward function, thereby reducing the exploration space and enabling more targeted exploration. Specifically, we construct a structural causal model (SCM) to model the causality between sub-state variables and sparse rewards, providing a robust analytical foundation for subsequent causal inference. Through counterfactual causal intervention, TMAE identifies the most critical subspaces for discovering rare but pivotal events while filtering out confounders. By incorporating these causal insights into the exploration process, TMAE prioritizes subspaces with stronger causal effects on sparse rewards, significantly enhancing exploration efficiency. We evaluate TMAE on a range of MARL benchmarks featuring sparse rewards, consistently demonstrating superior exploration efficiency compared to state-of-the-art methods. Furthermore, visualized causal insights derived from TMAE reveal its ability to effectively capture intricate dependencies and priorities in targeted exploration, showcasing strong alignment with prior domain knowledge.
Composed Image Retrieval (CIR) is a challenging image retrieval paradigm that enables to retrieve target images based on multimodal queries consisting of reference images and modification texts. Although substantial progress has been made in recent years, existing methods assume that all samples are correctly matched. However, in real-world scenarios, due to high triplet annotation costs, CIR datasets inevitably contain annotation errors, resulting in incorrectly matched triplets. To address this issue, the problem of Noisy Triplet Correspondence (NTC) has attracted growing attention. We argue that noise in CIR can be categorized into two types: cross-modal correspondence noise and modality-inherent noise. The former arises from mismatches across modalities, whereas the latter originates from intra-modal background interference or visual factors irrelevant to the coarse-grained modification annotations. However, modality-inherent noise is often overlooked, and research on cross-modal correspondence noise remains nascent. To tackle above issues, we propose the Invariance and discrimiNaTion-awarE Noise neTwork (INTENT), comprising two components: Visual Invariant Composition and Bi-Objective Discriminative Learning, specifically designed to handle the two-aspect noise. The former applies causal intervention on the visual side via Fast Fourier Transform (FFT) to generate intervened composed features, enforcing visual invariance and enabling the model to ignore modality-inherent noise during composition. The latter adopts collaborative optimization with both positive and negative samples, and constructs a scalable decision boundary that dynamically adjusts decisions based on the loyalty degree, enabling robust correspondence discrimination. Extensive experiments on two widely used benchmark datasets demonstrate the superiority and robustness of INTENT.
We study how the design of testing institutions, encompassing both the tests themselves and the procedures used to administer them, shapes selection outcomes in environments with multiple criteria and strategic agents. We model the testing agency as either a set of independent bureaucracies (each test administered separately) or a joint bureaucracy (where test order and personalization can be coordinated). Our mechanism design analysis shows that under a joint bureaucracy, fixed-order sequential mechanisms with stringent tests are optimal for maximizing the probability mass of qualified candidates selected. Furthermore, we demonstrate that personalizing tests through upfront communication, now increasingly feasible via AI and automation, can select all qualified candidates. Finally, we compare institutional settings and quantify the value of controlling test order, showing that the benefit depends critically on the distribution of testees and the stringency of optimal tests. Our results contribute to the design of robust, efficient, and fair testing systems in both human and AI-mediated environments.
Without manual annotations, unsupervised cross-modal hashing (UCMH) aims to achieve efficient clustering and retrieval by leveraging data interrelationships. However, the retrieval accuracy is constrained by two main aspects: 1) insufficient exploration of data relationships; 2) existing knowledge mining strategies are not well aligned with the architectural properties of multilayer perceptrons. Through summary and error analysis, the human brain is able to achieve fast learning through experience and minimal data. Inspired by this cognitive process, we propose a novel Error Notebook strategy, named ENHash, to more effectively capture similarity information between multi-modal data for fine-grained unsupervised clustering. Firstly, simulating the human process of summarizing experiences, ENHash gradually integrates the information from each batch into a global clustering representation. Secondly, drawing upon human error analysis capabilities, ENHash utilizes the summarized experiences to identify and record incorrectly predicted hash codes. Finally, by leveraging the knowledge derived from this analysis, ENHash guides the hash function to learn fine-grained patterns from the errors. To the best of our knowledge, ENHash represents the first attempt at integrating cognitively-inspired mechanisms into fine-grained UCMH optimization paradigms. We evaluate the proposed ENHash against eight state-of-the-art methods on three widely used datasets and one fine-grained cross-modal dataset. Experimental results show that ENHash achieves substantial improvements over existing approaches.
Experimental design is critical for evidence-based decision-making in healthcare, marketing, and public policy. However, designing efficient experiments across heterogeneous subgroups presents significant challenges. Existing methods often optimize for statistical power or overall sample efficiency, overlooking crucial fairness considerations across these different subgroups. To address this gap, we introduce a Fairness-Aware Contextual Track-and-Stop Design (F-CTSD) algorithm. The proposed F-CTSD algorithm provides statistical guarantees on subgroup fairness while minimizing required sample sizes. We quantify the fairness-efficiency trade-off and derive the sample complexity bound for the proposed F-CTSD algorithm under its fairness constraints. We further theoretically prove that the proposed F-CTSD algorithm consistently produces accurate treatment effect estimates even under fairness requirements, enhancing statistical reliability. Numerical experiments show that the proposed F-CTSD algorithm outperforms existing methods, achieving higher sample efficiency while reducing subgroup fairness violations by 4.95%.
Machine unlearning (MU) has emerged as a critical tool for removing sensitive or personal information from machine learning models, empowering individuals with the right to be forgotten. While MU has achieved success in classification and generative tasks, whether this technique can be effectively applied to segmentation foundation models remains uncertain. To address this issue, we propose an efficient method, Selective Concept Unlearning (SCU), to unlearn the segmentation capability of target concepts. SCU consists of several key aspects: (1) The Multi-level Forgetting Module, designed with a hierarchical three-level suppression strategy, including (i) distillation-level: Negative distillation steers model’s output distribution away from teacher’s correct outputs, erasing its learned concept recognition. (ii) attention-level: Attention suppression minimizes model’s attention to target regions. (iii) output-level: Directly erases predictions for the target by relabeling as background. (2) The Preservation Module ensures maintaining segmentation quality for non-target concepts. Additionally, we introduce a set of metrics to evaluate segmentation unlearning methods. Experiments demonstrate that SCU consistently outperforms existing baselines.
Covariance matrix estimation in high dimensions is a fundamental problem in machine learning and signal processing. A common structural assumption used to mitigate the challenges posed by high dimensionality is sparsity, which posits that most variable pairs exhibit negligible correlations. In this paper, we revisit the classical problem of positive definite sparse covariance estimation (PDSCE) introduced by Rothman (2012). Unlike many earlier approaches, this formulation incorporates a logarithmic barrier, which guarantees that the resulting covariance estimator is positive definite and thereby ensures the well-posedness of the estimation problem. However, the inclusion of the logarithmic barrier also leads to nontrivial optimization difficulties. To overcome these difficulties, we propose a dual proximal gradient method (DPGM) for solving the PDSCE problem. In contrast to existing primal-space approaches, DPGM operates directly in the dual space. This dual perspective provides several key advantages. First, DPGM significantly reduces computational costs, because positive definiteness is preserved automatically and no iterative subproblem solvers are required. Second, compared with primal optimization algorithms, DPGM offers stronger theoretical guarantees, including principled step size selection and improved iteration complexity. Extensive numerical experiments demonstrate that DPGM consistently outperforms existing methods, which confirms its effectiveness and scalability for high-dimensional sparse covariance estimation.
Few-shot graph learning remains a fundamental yet challenging problem, especially under heterophilic graph settings where connected nodes are likely to belong to different classes. In such scenarios, two key challenges arise: (1) unreliable or noisy graph structures that hinder effective message passing, and (2) semantic inconsistency: in heterophilic graphs, aggregating messages from neighbors of different classes entangles representations and introduces misleading semantics. These issues are further exacerbated by the limited labeled data inherent to few-shot learning, making it difficult to adaptively repair structure or disentangle semantics. To address these challenges, we propose DAPrompt, a Dual Alignment Prompt framework that jointly calibrates graph structure and semantic representations across the learning pipeline. In the pretraining stage, DAPrompt incorporates a graph structure learning module to denoise and repair the underlying topology, enhancing structural reliability. In the prompt tuning stage, we introduce two coordinated modules: a structure-aware prompt learner, which employs prompt tokens to repair unreliable graph structures and capture structure-level alignment, and a semantics-aligned prompt learner, which enhances the graph using target node semantics to mitigate representation noise caused by class-mismatched propagation. Extensive experiments on both node-level and graph-level few-shot benchmarks validate its effectiveness, achieving state-of-the-art performance and highlighting the value of structure-semantic dual alignment in heterophilic few-shot graph learning.
More and more organizations are relying on Machine Learning (ML) models to support internal decision-making processes. To better support such processes, it would be highly beneficial to contextualize the inductively acquired knowledge encoded in these models and enable formal reasoning over it. Despite significant progress in Neural-Symbolic AI, this specific challenge remains largely under-explored. We propose a framework that allows to integrate the knowledge induced by ML classifiers with the knowledge specified by logic-based formalisms. The framework is based on the novel notion of Hybrid Knowledge Base (HKB), consisting of two components: an ontology and a set of ML binary classifiers. As usual, the ontology provides an intensional representation of the modeled domain through logic-based axioms, while the binary classifiers implicitly encode the extensional knowledge. Specifically, a HKB associates to each concept and role mentioned in the ontology a classifier based on a set of features deemed to be relevant for the application domain, thereby virtually populating the concepts and roles with the instances and pairs of instances from the feature space. Besides the definition of the new framework, as a more technical contribution we show how to reason in this framework by studying query answering over HKBs. In particular, we investigate the computational complexity of query answering in a rich language over HKBs in which the ontology is specified in (the Description Logic counterpart of) RDFS, while the binary classifiers are represented by Multi-Layer Perceptrons.
Graph neural networks (GNNs) have demonstrated impressive performance in a broad spectrum of fields, but always suffer from the generalization problem when confronted with out-of-distribution (OOD) scenarios. Information bottleneck (IB) principle, which endeavors to learn the minimally sufficient representations for downstream tasks, has been shown to be a promising strategy in dealing with this problem. However, the IB-based methods do not inherently distinguish between causal and non-causal parts in the graph, leading to underperforming OOD generalization ability. In this paper, we develop the Graph Causal Information Bottleneck (GCIB) framework, a causal extension of the IB for graph data, which is capable of jointly compressing abundant information and capturing causal dependency from the input graph. Specifically, we endow graph IB with the ability of maintaining causal control by incorporating the underlying causal structure and introducing intervention operation. On this basis, we formulate the learning objective for GCIB and present its specific implementation. Graph representations learned by GCIB can effectively preserve causal information that fundamentally determines graph properties, resulting in outstanding OOD generalization ability. Extensive experiments on both synthetic and real-world datasets demonstrate the superiority of GCIB over state-of-the-art baselines.
Recent advancements in multi-robot navigation have explored methods that combine Large Language Models (LLMs) for tasks like scene understanding or high-level decision-making. However, these approaches face challenges with high inference latency and potential hallucinations. To address these challenges, we propose a knowledge-driven Reinforcement Learning (RL) framework, GUIDER, that utilizes an LLM in two different offline roles. First, we leverage the LLM as an offline knowledge source. Its expertise is distilled into a compact model, which is applied only when the RL agent is uncertain about its own value estimates and the model itself is confident in its prediction. Additionally, we utilize the LLM as an offline semantic engine. This process translates the LLM's high-level understanding of situational risk into a dynamic adjustment of the RL agent's behavioral style, evolving a function that optimally balances conservative and aggressive actions. We conduct extensive experiments in both terrestrial and maritime settings. Across all maritime scenarios (3–12 robots), GUIDER improves the task success rate and reduces the collision rate significantly compared to the state-of-the-art RL-based multi-robot navigation methods.
Large Language Models (LLMs) hold significant potential for enhancing healthcare applications, yet their deployment is hindered by high computational and memory demands. Model compression techniques offer solutions to reduce these demands, but their impact on medical LLMs remains underexplored. In this paper, we introduce CMedBench, the first comprehensive benchmark for evaluating compressed LLMs in medical contexts. CMedBench assesses five core dimensions: Medical Knowledge Ability, Medical Application Ability, Trustworthiness Maintenance, Compression Cross Combination, and Computational Efficiency. Through extensive empirical studies, we analyze the trade-offs between model efficiency and clinical performance across diverse models, datasets, and compression strategies. Our findings highlight critical limitations in current evaluation practices and provide a robust framework for aligning compression strategies with medical requirements. CMedBench serves as a vital resource for researchers and practitioners, guiding the development of efficient, trustworthy, and clinically effective LLMs for healthcare applications.
Modern AI services must continually adapt to newly joined domains, yet delivering high-quality customized models is hampered by label sparsity, domain shifts, and tight budgets. We formulate this challenge as the learning system expansion problem and introduce HaT, an efficient heterogeneity-aware knowledge-transfer framework. HaT first selects a small set of high-quality source models with minimal overhead, and then fuses their imperfect predictions through a sample-wise attention mixer. Later, it adaptively distills the fused knowledge into target models via a knowledge dictionary. Extensive experiments on different tasks and modalities show that HaT outperforms state-of-the-art baselines by up to 16.5% accuracy, and saves 31.1% training time and up to 93.0% traffic.
Knowledge graph construction (KGC) aims to extract valuable information from text and organize it into structured knowledge graphs (KGs). Recent methods have leveraged the strong generative capabilities of large language models (LLMs) to improve the generalization and reduce the labor costs. However, constrained by the input length of LLMs, existing methods mainly focus on extracting knowledge within individual texts and lack the capability to discover latent knowledge across texts. To fill this gap, we propose a novel method for open knowledge graph construction, termed KG-DLF. The core idea of this method is to enhance the knowledge graph construction process by discovering new facts that are consistent with the underlying contextual logic. Specifically, we first design a knowledge extractor to extract knowledge from the text. Then, a knowledge normalizer performs schema alignment on the extracted knowledge. Next, we explore a knowledge discoverer based on a clue search strategy, which leverages the logical consistency of context to mine latent facts. Finally, we design a counterfactual-based knowledge corrector, enabling the model to purify knowledge and reduce factual errors. Experimental results show that KG-DLF is capable of extracting comprehensive knowledge in open-world scenarios across three KGC benchmarks.
