Objectives: We aim to dynamically retrieve informative demonstrations, enhancing in-context learning in multimodal large language models (MLLMs) for disease classification. Methods: We propose a Retrieval-Augmented In-Context Learning (RAICL) framework, which integrates retrieval-augmented generation (RAG) and in-context learning (ICL) to adaptively select demonstrations with similar disease patterns, enabling more effective ICL in MLLMs. Specifically, RAICL examines embeddings from diverse encoders, including ResNet, BERT, BioBERT, and ClinicalBERT, to retrieve appropriate demonstrations, and constructs conversational prompts optimized for ICL. We evaluated the framework on two real-world multi-modal datasets (TCGA and IU Chest X-ray), assessing its performance across multiple MLLMs (Qwen, Llava, Gemma), embedding strategies, similarity metrics, and varying numbers of demonstrations. Results: RAICL consistently improved classification performance. Accuracy increased from 0.7854 to 0.8368 on TCGA and from 0.7924 to 0.8658 on IU Chest X-ray. Multi-modal inputs outperformed single-modal ones, with text-only inputs being stronger than images alone. The richness of information embedded in each modality will determine which embedding model can be used to get better results. Few-shot experiments showed that increasing the number of retrieved examples further enhanced performance. Across different similarity metrics, Euclidean distance achieved the highest accuracy while cosine similarity yielded better macro-F1 scores. RAICL demonstrated consistent improvements across various MLLMs, confirming its robustness and versatility. Conclusions: RAICL provides an efficient and scalable approach to enhance in-context learning in MLLMs for multimodal disease classification.
The notion of language modeling has gradually shifted in recent years from a distribution over finite-length strings to general-purpose prediction models for textual inputs and outputs, following appropriate alignment phases. This paper analyzes the distinction between distribution estimation and response prediction in the context of LLMs, and their often conflicting goals. We examine the training phases of LLMs, which include pretraining, in-context learning, and preference tuning, and also the common use cases for their output probabilities, which include completion probabilities and explicit probabilities as output. We argue that the different settings lead to three distinct intended output distributions. We demonstrate that NLP works often assume that these distributions should be similar, which leads to misinterpretations of their experimental findings. Our work sets firmer formal foundations for the interpretation of LLMs, which will inform ongoing work on the interpretation and use of LLMs' induced distributions.
Graph In-Context Learning, with the ability to adapt pre-trained graph models to novel and diverse downstream graphs without updating any parameters, has gained much attention in the community. The key to graph in-context learning is to perform downstream graphs conditioned on chosen prompt examples. Existing methods randomly select subgraphs or edges as prompts, leading to noisy graph prompts and inferior model performance. Additionally, due to the gap between pre-training and testing graphs, when the number of classes in the testing graphs is much greater than that in the training, the in-context learning ability will also significantly deteriorate. To tackle the aforementioned challenges, we develop a multi-stage adaptive prompt optimization method GraphPrompter, which optimizes the entire process of generating, selecting, and using graph prompts for better in-context learning capabilities. Firstly, Prompt Generator introduces a reconstruction layer to highlight the most informative edges and reduce irrelevant noise for graph prompt construction. Furthermore, in the selection stage, Prompt Selector employs the $k$-nearest neighbors algorithm and pre-trained selection layers to dynamically choose appropriate samples and minimize the influence of irrelevant prompts. Finally, we leverage a Prompt Augmenter with a cache replacement strategy to enhance the generalization capability of the pre-trained model on new datasets. Extensive experiments show that GraphPrompter effectively enhances the in-context learning ability of graph models. On average across all the settings, our approach surpasses the state-of-the-art baselines by over 8%. Our code is released at https://github.com/karin0018/GraphPrompter.
Advances in deep learning and generative modeling have driven interest in data-driven molecule discovery pipelines, whereby machine learning (ML) models are used to filter and design novel molecules without requiring prohibitively expensive first-principles simulations. Although the discovery of novel molecules that extend the boundaries of known chemistry requires accurate out-of-distribution (OOD) predictions, ML models often struggle to generalize OOD. Furthermore, there are currently no systematic benchmarks for molecular OOD prediction tasks. We present BOOM, $\boldsymbol{b}$enchmarks for $\boldsymbol{o}$ut-$\boldsymbol{o}$f-distribution $\boldsymbol{m}$olecular property predictions -- a benchmark study of property-based out-of-distribution models for common molecular property prediction models. We evaluate more than 140 combinations of models and property prediction tasks to benchmark deep learning models on their OOD performance. Overall, we do not find any existing models that achieve strong OOD generalization across all tasks: even the top performing model exhibited an average OOD error 3x larger than in-distribution. We find that deep learning models with high inductive bias can perform well on OOD tasks with simple, specific properties. Although chemical foundation models with transfer and in-context learning offer a promising solution for limited training data scenarios, we find that current foundation models do not show strong OOD extrapolation capabilities. We perform extensive ablation experiments to highlight how OOD performance is impacted by data generation, pre-training, hyperparameter optimization, model architecture, and molecular representation. We propose that developing ML models with strong OOD generalization is a new frontier challenge in chemical ML model development. This open-source benchmark will be made available on Github.
Humans and intelligent animals can effortlessly internalize new information ("news") and accurately extract the implications for performing downstream tasks. While large language models (LLMs) can achieve this through in-context learning (ICL) when the news is explicitly given as context, fine-tuning remains challenging for the models to consolidate learning in weights. In this paper, we introduce $\textit{New News}$, a dataset composed of hypothetical yet plausible news spanning multiple domains (mathematics, coding, discoveries, leaderboards, events), accompanied by downstream evaluation questions whose correct answers critically depend on understanding and internalizing the news. We first demonstrate a substantial gap between naive fine-tuning and in-context learning (FT-ICL gap) on our news dataset. To address this gap, we explore a suite of self-play data generation protocols -- paraphrases, implications and Self-QAs -- designed to distill the knowledge from the model with context into the weights of the model without the context, which we term $\textit{System-2 Fine-tuning}$ (Sys2-FT). We systematically evaluate ICL and Sys2-FT performance across data domains and model scales with the Qwen 2.5 family of models. Our results demonstrate that the self-QA protocol of Sys2-FT significantly improves models' in-weight learning of the news. Furthermore, we discover the $\textit{contexual shadowing effect}$, where training with the news $\textit{in context}$ followed by its rephrases or QAs degrade learning of the news. Finally, we show preliminary evidence of an emerging scaling law of Sys2-FT.
In-context learning (ICL)-the ability of transformer-based models to perform new tasks from examples provided at inference time-has emerged as a hallmark of modern language models. While recent works have investigated the mechanisms underlying ICL, its feasibility under formal privacy constraints remains largely unexplored. In this paper, we propose a differentially private pretraining algorithm for linear attention heads and present the first theoretical analysis of the privacy-accuracy trade-off for ICL in linear regression. Our results characterize the fundamental tension between optimization and privacy-induced noise, formally capturing behaviors observed in private training via iterative methods. Additionally, we show that our method is robust to adversarial perturbations of training prompts, unlike standard ridge regression. All theoretical findings are supported by extensive simulations across diverse settings.
Legal practice requires careful adherence to procedural rules. In the United States, few are more complex than those found in The Bluebook: A Uniform System of Citation. Compliance with this system's 500+ pages of byzantine formatting instructions is the raison d'etre of thousands of student law review editors and the bete noire of lawyers everywhere. To evaluate whether large language models (LLMs) are able to adhere to the procedures of such a complicated system, we construct an original dataset of 866 Bluebook tasks and test flagship LLMs from OpenAI, Anthropic, Google, Meta, and DeepSeek. We show (1) that these models produce fully compliant Bluebook citations only 69%-74% of the time and (2) that in-context learning on the Bluebook's underlying system of rules raises accuracy only to 77%. These results caution against using off-the-shelf LLMs to automate aspects of the law where fidelity to procedure is paramount.
Visual In-Context Learning (VICL) enables adaptively solving vision tasks by leveraging pixel demonstrations, mimicking human-like task completion through analogy. Prompt selection is critical in VICL, but current methods assume the existence of a single "ideal" prompt in a pool of candidates, which in practice may not hold true. Multiple suitable prompts may exist, but individually they often fall short, leading to difficulties in selection and the exclusion of useful context. To address this, we propose a new perspective: prompt condensation. Rather than relying on a single prompt, candidate prompts collaborate to efficiently integrate informative contexts without sacrificing resolution. We devise Condenser, a lightweight external plugin that compresses relevant fine-grained context across multiple prompts. Optimized end-to-end with the backbone, Condenser ensures accurate integration of contextual cues. Experiments demonstrate Condenser outperforms state-of-the-arts across benchmark tasks, showing superior context compression, scalability with more prompts, and enhanced computational efficiency compared to ensemble methods, positioning it as a highly competitive solution for VICL. Code is open-sourced at https://github.com/gimpong/CVPR25-Condenser.
In NLP, fine-tuning LLMs is effective for various applications but requires high-quality annotated data. However, manual annotation of data is labor-intensive, time-consuming, and costly. Therefore, LLMs are increasingly used to automate the process, often employing in-context learning (ICL) in which some examples related to the task are given in the prompt for better performance. However, manually selecting context examples can lead to inefficiencies and suboptimal model performance. This paper presents comprehensive experiments comparing several LLMs, considering different embedding models, across various datasets for the Named Entity Recognition (NER) task. The evaluation encompasses models with approximately $7$B and $70$B parameters, including both proprietary and non-proprietary models. Furthermore, leveraging the success of Retrieval-Augmented Generation (RAG), it also considers a method that addresses the limitations of ICL by automatically retrieving contextual examples, thereby enhancing performance. The results highlight the importance of selecting the appropriate LLM and embedding model, understanding the trade-offs between LLM sizes and desired performance, and the necessity to direct research efforts towards more challenging datasets.
Unmanned Aerial Vehicles (UAVs) are increasingly being used in various private and commercial applications, e.g. traffic control, package delivery, and Search and Rescue (SAR) operations. Machine Learning (ML) methods used in UAV-assisted Sensor Networks (UASNETs) and especially in Deep Reinforcement Learning (DRL) face challenges such as complex and lengthy model training, gaps between simulation and reality, and low sample efficiency, which conflict with the urgency of emergencies such as SAR operations. This paper proposes In-Context Learning (ICL)-based Data Collection Scheduling (ICLDC) scheme, as an alternative to DRL in emergencies. The UAV collects and transmits logged sensory data, to an LLM, to generate a task description in natural language, from which it obtains a data collection schedule to be executed by the UAV. The system continuously adapts by adding feedback to task descriptions and utilizing feedback for future decisions. This method is tested against jailbreaking attacks, where task description is manipulated to undermine network performance, highlighting the vulnerability of LLMs to such attacks. The proposed ICLDC outperforms the Maximum Channel Gain by reducing cumulative packet loss by approximately 56\%. ICLDC presents a promising direction for intelligent scheduling and control in UAV-assisted data collection.
To develop trustworthy Vision-Language Models (VLMs), it is essential to address adversarial robustness and hallucination mitigation, both of which impact factual accuracy in high-stakes applications such as defense and healthcare. Existing methods primarily focus on either adversarial defense or hallucination post-hoc correction, leaving a gap in unified robustness strategies. We introduce \textbf{Hydra}, an adaptive agentic framework that enhances plug-in VLMs through iterative reasoning, structured critiques, and cross-model verification, improving both resilience to adversarial perturbations and intrinsic model errors. Hydra employs an Action-Critique Loop, where it retrieves and critiques visual information, leveraging Chain-of-Thought (CoT) and In-Context Learning (ICL) techniques to refine outputs dynamically. Unlike static post-hoc correction methods, Hydra adapts to both adversarial manipulations and intrinsic model errors, making it robust to malicious perturbations and hallucination-related inaccuracies. We evaluate Hydra on four VLMs, three hallucination benchmarks, two adversarial attack strategies, and two adversarial defense methods, assessing performance on both clean and adversarial inputs. Results show that Hydra surpasses plug-in VLMs and state-of-the-art (SOTA) dehallucination methods, even without explicit adversarial defenses, demonstrating enhanced robustness and factual consistency. By bridging adversarial resistance and hallucination mitigation, Hydra provides a scalable, training-free solution for improving the reliability of VLMs in real-world applications.
Automated diagnosis of eczema using images acquired from digital camera can enable individuals to self-monitor their recovery. The process entails first segmenting out the eczema region from the image and then measuring the severity of eczema in the segmented region. The state-of-the-art methods for automated eczema diagnosis rely on deep neural networks such as convolutional neural network (CNN) and have shown impressive performance in accurately measuring the severity of eczema. However, these methods require massive volume of annotated data to train which can be hard to obtain. In this paper, we propose a self-supervised learning framework for automated eczema diagnosis under limited training data regime. Our framework consists of two stages: i) Segmentation, where we use an in-context learning based algorithm called SegGPT for few-shot segmentation of eczema region from the image; ii) Feature extraction and classification, where we extract DINO features from the segmented regions and feed it to a multi-layered perceptron (MLP) for 4-class classification of eczema severity. When evaluated on a dataset of annotated "in-the-wild" eczema images, we show that our method outperforms (Weighted F1: 0.67 $\pm$ 0.01) the state-of-the-art deep learning methods such as finetuned Resnet-18 (Weighted F1: 0.44 $\pm$ 0.16) and Vision Transformer (Weighted F1: 0.40 $\pm$ 0.22). Our results show that self-supervised learning can be a viable solution for automated skin diagnosis where labeled data is scarce.
In-context learning (ICL) enables Large Vision-Language Models (LVLMs) to adapt to new tasks without parameter updates, using a few demonstrations from a large support set. However, selecting informative demonstrations leads to high computational and memory costs. While some methods explore selecting a small and representative coreset in the text classification, evaluating all support set samples remains costly, and discarded samples lead to unnecessary information loss. These methods may also be less effective for image classification due to differences in feature spaces. Given these limitations, we propose Key-based Coreset Optimization (KeCO), a novel framework that leverages untapped data to construct a compact and informative coreset. We introduce visual features as keys within the coreset, which serve as the anchor for identifying samples to be updated through different selection strategies. By leveraging untapped samples from the support set, we update the keys of selected coreset samples, enabling the randomly initialized coreset to evolve into a more informative coreset under low computational cost. Through extensive experiments on coarse-grained and fine-grained image classification benchmarks, we demonstrate that KeCO effectively enhances ICL performance for image classification task, achieving an average improvement of more than 20\%. Notably, we evaluate KeCO under a simulated online scenario, and the strong performance in this scenario highlights the practical value of our framework for resource-constrained real-world scenarios.
In information retrieval (IR), providing appropriate clarifications to better understand users' information needs is crucial for building a proactive search-oriented dialogue system. Due to the strong in-context learning ability of large language models (LLMs), recent studies investigate prompting methods to generate clarifications using few-shot or Chain of Thought (CoT) prompts. However, vanilla CoT prompting does not distinguish the characteristics of different information needs, making it difficult to understand how LLMs resolve ambiguities in user queries. In this work, we focus on the concept of ambiguity for clarification, seeking to model and integrate ambiguities in the clarification process. To this end, we comprehensively study the impact of prompting schemes based on reasoning and ambiguity for clarification. The idea is to enhance the reasoning abilities of LLMs by limiting CoT to predict first ambiguity types that can be interpreted as instructions to clarify, then correspondingly generate clarifications. We name this new prompting scheme Ambiguity Type-Chain of Thought (AT-CoT). Experiments are conducted on various datasets containing human-annotated clarifying questions to compare AT-CoT with multiple baselines. We also perform user simulations to implicitly measure the quality of generated clarifications under various IR scenarios.
Instruct models, obtained from various instruction tuning or post-training steps, are commonly deemed superior and more usable than their base counterpart. While the model gains instruction following ability, instruction tuning may lead to forgetting the knowledge from pre-training or it may encourage the model being overly conversational or verbose. This, in turn, can lead to degradation of in-context few-shot learning performance. In this work, we study the performance trajectory between base and instruct models by scaling down the strength of instruction-tuning via the partial adaption method. We show that, across several model families and model sizes, reducing the strength of instruction-tuning results in material improvement on a few-shot in-context learning benchmark covering a variety of classic natural language tasks. This comes at the cost of losing some degree of instruction following ability as measured by AlpacaEval. Our study shines light on the potential trade-off between in-context learning and instruction following abilities that is worth considering in practice.
WebShell attacks, in which malicious scripts are injected into web servers, are a major cybersecurity threat. Traditional machine learning and deep learning methods are hampered by issues such as the need for extensive training data, catastrophic forgetting, and poor generalization. Recently, Large Language Models (LLMs) have gained attention for code-related tasks, but their potential in WebShell detection remains underexplored. In this paper, we make two major contributions: (1) a comprehensive evaluation of seven LLMs, including GPT-4, LLaMA 3.1 70B, and Qwen 2.5 variants, benchmarked against traditional sequence- and graph-based methods using a dataset of 26.59K PHP scripts, and (2) the Behavioral Function-Aware Detection (BFAD) framework, designed to address the specific challenges of applying LLMs to this domain. Our framework integrates three components: a Critical Function Filter that isolates malicious PHP function calls, a Context-Aware Code Extraction strategy that captures the most behaviorally indicative code segments, and Weighted Behavioral Function Profiling (WBFP) that enhances in-context learning by prioritizing the most relevant demonstrations based on discriminative function-level profiles. Our results show that larger LLMs achieve near-perfect precision but lower recall, while smaller models exhibit the opposite trade-off. However, all models lag behind previous State-Of-The-Art (SOTA) methods. With BFAD, the performance of all LLMs improved, with an average F1 score increase of 13.82%. Larger models such as GPT-4, LLaMA 3.1 70B, and Qwen 2.5 14B outperform SOTA methods, while smaller models such as Qwen 2.5 3B achieve performance competitive with traditional methods. This work is the first to explore the feasibility and limitations of LLMs for WebShell detection, and provides solutions to address the challenges in this task.
Hate speech detection is a crucial area of research in natural language processing, essential for ensuring online community safety. However, detecting implicit hate speech, where harmful intent is conveyed in subtle or indirect ways, remains a major challenge. Unlike explicit hate speech, implicit expressions often depend on context, cultural subtleties, and hidden biases, making them more challenging to identify consistently. Additionally, the interpretation of such speech is influenced by external knowledge and demographic biases, resulting in varied detection results across different language models. Furthermore, Large Language Models often show heightened sensitivity to toxic language and references to vulnerable groups, which can lead to misclassifications. This over-sensitivity results in false positives (incorrectly identifying harmless statements as hateful) and false negatives (failing to detect genuinely harmful content). Addressing these issues requires methods that not only improve detection precision but also reduce model biases and enhance robustness. To address these challenges, we propose a novel method, which utilizes in-context learning without requiring model fine-tuning. By adaptively retrieving demonstrations that focus on similar groups or those with the highest similarity scores, our approach enhances contextual comprehension. Experimental results show that our method outperforms current state-of-the-art techniques. Implementation details and code are available at TBD.
Sparse autoencoders (SAEs) are a popular tool for interpreting large language model activations, but their utility in addressing open questions in interpretability remains unclear. In this work, we demonstrate their effectiveness by using SAEs to deepen our understanding of the mechanism behind in-context learning (ICL). We identify abstract SAE features that (i) encode the model's knowledge of which task to execute and (ii) whose latent vectors causally induce the task zero-shot. This aligns with prior work showing that ICL is mediated by task vectors. We further demonstrate that these task vectors are well approximated by a sparse sum of SAE latents, including these task-execution features. To explore the ICL mechanism, we adapt the sparse feature circuits methodology of Marks et al. (2024) to work for the much larger Gemma-1 2B model, with 30 times as many parameters, and to the more complex task of ICL. Through circuit finding, we discover task-detecting features with corresponding SAE latents that activate earlier in the prompt, that detect when tasks have been performed. They are causally linked with task-execution features through the attention and MLP sublayers.
Chain-of-Thought (CoT) prompting has emerged as a powerful technique to improve in-context learning (ICL) in large language models (LLMs) by breaking complex reasoning into intermediate steps. However, the ability of CoT to generalize under distribution shift remains poorly understood. In this work, we extend a latent-variable framework for CoT prompting and study its behavior on two prototypical out-of-distribution (OOD) scenarios: (i) the latent variables for CoT steps are permuted into novel combinations, and (ii) the latent variables uniformly scaled by a factor. Our experiments demonstrate that CoT inference generalizes effectively to OOD samples whose latent variables closely resemble those seen during training, but its performance degrades as this similarity decreases. These findings provide foundational insights into the strengths and limitations of CoT prompting under OOD conditions and suggest directions for developing more resilient reasoning strategies in future LLMs.
Aspect-based summarization aims to generate summaries tailored to specific aspects, addressing the resource constraints and limited generalizability of traditional summarization approaches. Recently, large language models have shown promise in this task without the need for training. However, they rely excessively on prompt engineering and face token limits and hallucination challenges, especially with in-context learning. To address these challenges, in this paper, we propose a novel framework for aspect-based summarization: Self-Aspect Retrieval Enhanced Summary Generation. Rather than relying solely on in-context learning, given an aspect, we employ an embedding-driven retrieval mechanism to identify its relevant text segments. This approach extracts the pertinent content while avoiding unnecessary details, thereby mitigating the challenge of token limits. Moreover, our framework optimizes token usage by deleting unrelated parts of the text and ensuring that the model generates output strictly based on the given aspect. With extensive experiments on benchmark datasets, we demonstrate that our framework not only achieves superior performance but also effectively mitigates the token limitation problem.