The growing volume of academic papers has made it increasingly difficult for researchers to efficiently extract key information. While large language models (LLMs) based agents are capable of automating question answering (QA) workflows for scientific papers, there still lacks a comprehensive and realistic benchmark to evaluate their capabilities. Moreover, training an interactive agent for this specific task is hindered by the shortage of high-quality interaction trajectories. In this work, we propose AirQA, a human-annotated comprehensive paper QA dataset in the field of artificial intelligence (AI), with 13,948 papers and 1,246 questions, that encompasses multi-task, multi-modal and instance-level evaluation. Furthermore, we propose ExTrActor, an automated framework for instruction data synthesis. With three LLM-based agents, ExTrActor can perform example generation and trajectory collection without human intervention. Evaluations of multiple open-source and proprietary models show that most models underperform on AirQA, demonstrating the quality of our dataset. Extensive experiments confirm that ExTrActor consistently improves the multi-turn tool-use capability of small models, enabling them to achieve performance comparable to larger ones.

Traditional Human-Swarm Interaction (HSI) methods often lack intuitive real-time adaptive interfaces, making decision making slower and increasing cognitive load while limiting command flexibility. To solve this, we present SwarmChat, a context-aware, multimodal interaction system powered by Large Language Models (LLMs). SwarmChat enables users to issue natural language commands to robotic swarms using multiple modalities, such as text, voice, or teleoperation. The system integrates four LLM-based modules: Context Generator, Intent Recognition, Task Planner, and Modality Selector. These modules collaboratively generate context from keywords, detect user intent, adapt commands based on real-time robot state, and suggest optimal communication modalities. Its three-layer architecture offers a dynamic interface with both fixed and customizable command options, supporting flexible control while optimizing cognitive effort. The preliminary evaluation also shows that the SwarmChat's LLM modules provide accurate context interpretation, relevant intent recognition, and effective command delivery, achieving high user satisfaction.

We present our submission to Task 3 (Discourse Relation Classification) of the DISRPT 2025 shared task. Task 3 introduces a unified set of 17 discourse relation labels across 39 corpora in 16 languages and six discourse frameworks, posing significant multilingual and cross-formalism challenges. We first benchmark the task by fine-tuning multilingual BERT-based models (mBERT, XLM-RoBERTa-Base, and XLM-RoBERTa-Large) with two argument-ordering strategies and progressive unfreezing ratios to establish strong baselines. We then evaluate prompt-based large language models (namely Claude Opus 4.0) in zero-shot and few-shot settings to understand how LLMs respond to the newly proposed unified labels. Finally, we introduce HiDAC, a Hierarchical Dual-Adapter Contrastive learning model. Results show that while larger transformer models achieve higher accuracy, the improvements are modest, and that unfreezing the top 75% of encoder layers yields performance comparable to full fine-tuning while training far fewer parameters. Prompt-based models lag significantly behind fine-tuned transformers, and HiDAC achieves the highest overall accuracy (67.5%) while remaining more parameter-efficient than full fine-tuning.

We introduce seqBench, a parametrized benchmark for probing sequential reasoning limits in Large Language Models (LLMs) through precise, multi-dimensional control over several key complexity dimensions. seqBench allows systematic variation of (1) the logical depth, defined as the number of sequential actions required to solve the task; (2) the number of backtracking steps along the optimal path, quantifying how often the agent must revisit prior states to satisfy deferred preconditions (e.g., retrieving a key after encountering a locked door); and (3) the noise ratio, defined as the ratio between supporting and distracting facts about the environment. Our evaluations on state-of-the-art LLMs reveal a universal failure pattern: accuracy collapses exponentially beyond a model-specific logical depth. Unlike existing benchmarks, seqBench's fine-grained control facilitates targeted analyses of these reasoning failures, illuminating universal scaling laws and statistical limits, as detailed in this paper alongside its generation methodology and evaluation metrics. We find that even top-performing models systematically fail on seqBench's structured reasoning tasks despite minimal search complexity, underscoring key limitations in their commonsense reasoning capabilities. Designed for future evolution to keep pace with advancing models, the seqBench datasets are publicly released to spur deeper scientific inquiry into LLM reasoning, aiming to establish a clearer understanding of their true potential and current boundaries for robust real-world application.

Guardrails are critical for the safe deployment of Large Language Models (LLMs)-powered software. Unlike traditional rule-based systems with limited, predefined input-output spaces that inherently constrain unsafe behavior, LLMs enable open-ended, intelligent interactions--opening the door to jailbreak attacks through user inputs. Guardrails serve as a protective layer, filtering unsafe prompts before they reach the LLM. However, prior research shows that jailbreak attacks can still succeed over 70% of the time, even against advanced models like GPT-4o. While guardrails such as LlamaGuard report up to 95% accuracy, our preliminary analysis shows their performance can drop sharply--to as low as 12%--when confronted with unseen attacks. This highlights a growing software engineering challenge: how to build a post-deployment guardrail that adapts dynamically to emerging threats? To address this, we propose AdaptiveGuard, an adaptive guardrail that detects novel jailbreak attacks as out-of-distribution (OOD) inputs and learns to defend against them through a continual learning framework. Through empirical evaluation, AdaptiveGuard achieves 96% OOD detection accuracy, adapts to new attacks in just two update steps, and retains over 85% F1-score on in-distribution data post-adaptation, outperforming other baselines. These results demonstrate that AdaptiveGuard is a guardrail capable of evolving in response to emerging jailbreak strategies post deployment. We release our AdaptiveGuard and studied datasets at https://github.com/awsm-research/AdaptiveGuard to support further research.

Determining whether another system, biological or artificial, possesses phenomenal consciousness has long been a central challenge in consciousness studies. This attribution problem has become especially pressing with the rise of large language models and other advanced AI systems, where debates about "AI consciousness" implicitly rely on some criterion for deciding whether a given system is conscious. In this paper, we propose a substrate-independent, logically rigorous, and counterfeit-resistant sufficiency criterion for phenomenal consciousness. We argue that any machine satisfying this criterion should be regarded as conscious with at least the same level of confidence with which we attribute consciousness to other humans. Building on this criterion, we develop a formal framework and specify a set of operational principles that guide the design of systems capable of meeting the sufficiency condition. We further argue that machines engineered according to this framework can, in principle, realize phenomenal consciousness. As an initial validation, we show that humans themselves can be viewed as machines that satisfy this framework and its principles. If correct, this proposal carries significant implications for philosophy, cognitive science, and artificial intelligence. It offers an explanation for why certain qualia, such as the experience of red, are in principle irreducible to physical description, while simultaneously providing a general reinterpretation of human information processing. Moreover, it suggests a path toward a new paradigm of AI beyond current statistics-based approaches, potentially guiding the construction of genuinely human-like AI.

Unlike in language or vision, one of the fundamental challenges in robot learning is the lack of access to vast data resources. We can further break down the problem into (1) data sparsity from the angle of data representation and (2) data scarcity from the angle of data quantity. In this thesis, I will discuss selected works on two domains: (1) tactile sensing and (2) rehabilitation robots, which are exemplars of data sparsity and scarcity, respectively. Tactile sensing is an essential modality for robotics, but tactile data are often sparse, and for each interaction with the physical world, tactile sensors can only obtain information about the local area of contact. I will discuss my work on learning vision-free tactile-only exploration and manipulation policies through model-free reinforcement learning to make efficient use of sparse tactile information. On the other hand, rehabilitation robots are an example of data scarcity to the extreme due to the significant challenge of collecting biosignals from disabled-bodied subjects at scale for training. I will discuss my work in collaboration with the medical school and clinicians on intent inferral for stroke survivors, where a hand orthosis developed in our lab collects a set of biosignals from the patient and uses them to infer the activity that the patient intends to perform, so the orthosis can provide the right type of physical assistance at the right moment. My work develops machine learning algorithms that enable intent inferral with minimal data, including semi-supervised, meta-learning, and generative AI methods.

A variety of recent methods guide large language model outputs via the inference-time addition of steering vectors to residual-stream or attention-head representations. In contrast, we propose to inject steering vectors directly into the query and value representation spaces within attention heads. We provide evidence that a greater portion of these spaces exhibit high linear discriminability of concepts --a key property motivating the use of steering vectors-- than attention head outputs. We analytically characterize the effect of our method, which we term DISentangled COmmunication (DISCO) Steering, on attention head outputs. Our analysis reveals that DISCO disentangles a strong but underutilized baseline, steering attention inputs, which implicitly modifies queries and values in a rigid manner. In contrast, DISCO's direct modulation of these components enables more granular control. We find that DISCO achieves superior performance over a number of steering vector baselines across multiple datasets on LLaMA 3.1 8B and Gemma 2 9B, with steering efficacy scoring up to 19.1% higher than the runner-up. Our results support the conclusion that the query and value spaces are powerful building blocks for steering vector methods.

Quantifying identity fusion -- the psychological merging of self with another entity or abstract target (e.g., a religious group, political party, ideology, value, brand, belief, etc.) -- is vital for understanding a wide range of group-based human behaviors. We introduce the Cognitive Linguistic Identity Fusion Score (CLIFS), a novel metric that integrates cognitive linguistics with large language models (LLMs), which builds on implicit metaphor detection. Unlike traditional pictorial and verbal scales, which require controlled surveys or direct field contact, CLIFS delivers fully automated, scalable assessments while maintaining strong alignment with the established verbal measure. In benchmarks, CLIFS outperforms both existing automated approaches and human annotation. As a proof of concept, we apply CLIFS to violence risk assessment to demonstrate that it can improve violence risk assessment by more than 240%. Building on our identification of a new NLP task and early success, we underscore the need to develop larger, more diverse datasets that encompass additional fusion-target domains and cultural backgrounds to enhance generalizability and further advance this emerging area. CLIFS models and code are public at https://github.com/DevinW-sudo/CLIFS.

Aspect-based sentiment analysis (ABSA) in natural language processing enables organizations to understand customer opinions on specific product aspects. While deep learning models are widely used for English ABSA, their application in Arabic is limited due to the scarcity of labeled data. Researchers have attempted to tackle this issue by using pre-trained contextualized language models such as BERT. However, these models are often based on fact-based data, which can introduce bias in domain-specific tasks like ABSA. To our knowledge, no studies have applied adaptive pre-training with Arabic contextualized models for ABSA. This research proposes a novel approach using domain-adaptive pre-training for aspect-sentiment classification (ASC) and opinion target expression (OTE) extraction. We examine fine-tuning strategies - feature extraction, full fine-tuning, and adapter-based methods - to enhance performance and efficiency, utilizing multiple adaptation corpora and contextualized models. Our results show that in-domain adaptive pre-training yields modest improvements. Adapter-based fine-tuning is a computationally efficient method that achieves competitive results. However, error analyses reveal issues with model predictions and dataset labeling. In ASC, common problems include incorrect sentiment labeling, misinterpretation of contrastive markers, positivity bias for early terms, and challenges with conflicting opinions and subword tokenization. For OTE, issues involve mislabeling targets, confusion over syntactic roles, difficulty with multi-word expressions, and reliance on shallow heuristics. These findings underscore the need for syntax- and semantics-aware models, such as graph convolutional networks, to more effectively capture long-distance relations and complex aspect-based opinion alignments.