In-Context Learning (ICL) has emerged as a pivotal capability of auto-regressive large language models, yet it is hindered by a notable sensitivity to the ordering of context examples regardless of their mutual independence. To address this issue, recent studies have introduced several variant algorithms of ICL that achieve permutation invariance. However, many of these do not exhibit comparable performance with the standard auto-regressive ICL algorithm. In this work, we identify two crucial elements in the design of an invariant ICL algorithm: information non-leakage and context interdependence, which are not simultaneously achieved by any of the existing methods. These investigations lead us to the proposed Invariant ICL (InvICL), a methodology designed to achieve invariance in ICL while ensuring the two properties. Empirically, our findings reveal that InvICL surpasses previous models, both invariant and non-invariant, in most benchmark datasets, showcasing superior generalization capabilities across varying input lengths. Code is available at https://github.com/PKU-ML/InvICL.
The standardization of clinical data elements (CDEs) aims to ensure consistent and comprehensive patient information across various healthcare systems. Existing methods often falter when standardizing CDEs of varying representation and complex structure, impeding data integration and interoperability in clinical research. We introduce CDE-Mapper, an innovative framework that leverages Retrieval-Augmented Generation approach combined with Large Language Models to automate the linking of CDEs to controlled vocabularies. Our modular approach features query decomposition to manage varying levels of CDEs complexity, integrates expert-defined rules within prompt engineering, and employs in-context learning alongside multiple retriever components to resolve terminological ambiguities. In addition, we propose a knowledge reservoir validated by a human-in-loop approach, achieving accurate concept linking for future applications while minimizing computational costs. For four diverse datasets, CDE-Mapper achieved an average of 7.2\% higher accuracy improvement compared to baseline methods. This work highlights the potential of advanced language models in improving data harmonization and significantly advancing capabilities in clinical decision support systems and research.
Large-scale models trained on extensive datasets have become the standard due to their strong generalizability across diverse tasks. In-context learning (ICL), widely used in natural language processing, leverages these models by providing task-specific prompts without modifying their parameters. This paradigm is increasingly being adapted for computer vision, where models receive an input-output image pair, known as an in-context pair, alongside a query image to illustrate the desired output. However, the success of visual ICL largely hinges on the quality of these prompts. To address this, we propose Enhanced Instruct Me More (E-InMeMo), a novel approach that incorporates learnable perturbations into in-context pairs to optimize prompting. Through extensive experiments on standard vision tasks, E-InMeMo demonstrates superior performance over existing state-of-the-art methods. Notably, it improves mIoU scores by 7.99 for foreground segmentation and by 17.04 for single object detection when compared to the baseline without learnable prompts. These results highlight E-InMeMo as a lightweight yet effective strategy for enhancing visual ICL. Code is publicly available at: https://github.com/Jackieam/E-InMeMo
One of the primary aspirations in reinforcement learning research is developing general-purpose agents capable of rapidly adapting to and mastering novel tasks. While RL gaming agents have mastered many Atari games, they remain slow and costly to train for each game. In this work, we demonstrate that latest reasoning LLMs with out-of-domain RL post-training can play a challenging Atari game called Frogger under a zero-shot setting. We then investigate the effect of in-context learning and the amount of reasoning effort on LLM performance. Lastly, we demonstrate a way to bootstrap traditional RL method with LLM demonstrations, which significantly improves their performance and sample efficiency. Our implementation is open sourced at https://github.com/AlienKevin/frogger.
Language models have emerged as powerful predictors of the viability of biological sequences. During training these models learn the rules of the grammar obeyed by sequences of amino acids or nucleotides. Once trained, these models can take a sequence as input and produce a likelihood score as an output; a higher likelihood implies adherence to the learned grammar and correlates with experimental fitness measurements. Here we show that in-context learning can distort the relationship between fitness and likelihood scores of sequences. This phenomenon most prominently manifests as anomalously high likelihood scores for sequences that contain repeated motifs. We use protein language models with different architectures trained on the masked language modeling objective for our experiments, and find transformer-based models to be particularly vulnerable to this effect. This behavior is mediated by a look-up operation where the model seeks the identity of the masked position by using the other copy of the repeated motif as a reference. This retrieval behavior can override the model's learned priors. This phenomenon persists for imperfectly repeated sequences, and extends to other kinds of biologically relevant features such as reversed complement motifs in RNA sequences that fold into hairpin structures.
User intent classification is an important task in information retrieval. Previously, user intents were classified manually and automatically; the latter helped to avoid hand labelling of large datasets. Recent studies explored whether LLMs can reliably determine user intent. However, researchers have recognized the limitations of using generative LLMs for classification tasks. In this study, we empirically compare user intent classification into informational, navigational, and transactional categories, using weak supervision and LLMs. Specifically, we evaluate LLaMA-3.1-8B-Instruct and LLaMA-3.1-70B-Instruct for in-context learning and LLaMA-3.1-8B-Instruct for fine-tuning, comparing their performance to an established baseline classifier trained using weak supervision (ORCAS-I). Our results indicate that while LLMs outperform weak supervision in recall, they continue to struggle with precision, which shows the need for improved methods to balance both metrics effectively.
LLMs are increasingly explored for bundle generation, thanks to their reasoning capabilities and knowledge. However, deploying large-scale LLMs introduces significant efficiency challenges, primarily high computational costs during fine-tuning and inference due to their massive parameterization. Knowledge distillation (KD) offers a promising solution, transferring expertise from large teacher models to compact student models. This study systematically investigates knowledge distillation approaches for bundle generation, aiming to minimize computational demands while preserving performance. We explore three critical research questions: (1) how does the format of KD impact bundle generation performance? (2) to what extent does the quantity of distilled knowledge influence performance? and (3) how do different ways of utilizing the distilled knowledge affect performance? We propose a comprehensive KD framework that (i) progressively extracts knowledge (patterns, rules, deep thoughts); (ii) captures varying quantities of distilled knowledge through different strategies; and (iii) exploits complementary LLM adaptation techniques (in-context learning, supervised fine-tuning, combination) to leverage distilled knowledge in small student models for domain-specific adaptation and enhanced efficiency. Extensive experiments provide valuable insights into how knowledge format, quantity, and utilization methodologies collectively shape LLM-based bundle generation performance, exhibiting KD's significant potential for more efficient yet effective LLM-based bundle generation.
In-Context Learning (ICL) has gained prominence due to its ability to perform tasks without requiring extensive training data and its robustness to noisy labels. A typical ICL workflow involves selecting localized examples relevant to a given input using sparse or dense embedding-based similarity functions. However, relying solely on similarity-based selection may introduce topical biases in the retrieved contexts, potentially leading to suboptimal downstream performance. We posit that reranking the retrieved context to enhance topical diversity can improve downstream task performance. To achieve this, we leverage maximum marginal relevance (MMR) which balances topical similarity with inter-example diversity. Our experimental results demonstrate that diversifying the selected examples leads to consistent improvements in downstream performance across various context sizes and similarity functions. The implementation of our approach is made available at https://github.com/janak11111/Diverse-ICL.
Transformers have become a standard architecture in machine learning, demonstrating strong in-context learning (ICL) abilities that allow them to learn from the prompt at inference time. However, uncertainty quantification for ICL remains an open challenge, particularly in noisy regression tasks. This paper investigates whether ICL can be leveraged for distribution-free uncertainty estimation, proposing a method based on conformal prediction to construct prediction intervals with guaranteed coverage. While traditional conformal methods are computationally expensive due to repeated model fitting, we exploit ICL to efficiently generate confidence intervals in a single forward pass. Our empirical analysis compares this approach against ridge regression-based conformal methods, showing that conformal prediction with in-context learning (CP with ICL) achieves robust and scalable uncertainty estimates. Additionally, we evaluate its performance under distribution shifts and establish scaling laws to guide model training. These findings bridge ICL and conformal prediction, providing a theoretically grounded and new framework for uncertainty quantification in transformer-based models.
In-context learning (ICL) has emerged as a powerful capability of large language models (LLMs), enabling them to perform new tasks based on a few provided examples without explicit fine-tuning. Despite their impressive adaptability, these models remain vulnerable to subtle adversarial perturbations and exhibit unpredictable behavior when faced with linguistic variations. Inspired by software testing principles, we introduce a software testing-inspired framework, called MMT4NL, for evaluating the trustworthiness of in-context learning by utilizing adversarial perturbations and software testing techniques. It includes diverse evaluation aspects of linguistic capabilities for testing the ICL capabilities of LLMs. MMT4NL is built around the idea of crafting metamorphic adversarial examples from a test set in order to quantify and pinpoint bugs in the designed prompts of ICL. Our philosophy is to treat any LLM as software and validate its functionalities just like testing the software. Finally, we demonstrate applications of MMT4NL on the sentiment analysis and question-answering tasks. Our experiments could reveal various linguistic bugs in state-of-the-art LLMs.
In-context learning (ICL) allows a language model to improve its problem-solving capability when provided with suitable information in context. Since the choice of in-context information can be determined based on the problem itself, in-context learning is analogous to human learning from teachers in a classroom. Recent works (Didolkar et al., 2024a; 2024b) show that ICL performance can be improved by leveraging a frontier large language model's (LLM) ability to predict required skills to solve a problem, popularly referred to as an LLM's metacognition, and using the recommended skills to construct necessary in-context examples. While this skill-based strategy boosts ICL performance in larger models, its gains on small language models (SLMs) have been minimal, highlighting a performance gap in ICL capabilities. We investigate this gap and show that skill-based prompting can hurt SLM performance on easy questions by introducing unnecessary information, akin to cognitive overload. To address this, we introduce AdaptMI, an adaptive approach to selecting skill-based in-context Math Instructions for SLMs. Inspired by cognitive load theory from human pedagogy, our method only introduces skill-based examples when the model performs poorly. We further propose AdaptMI+, which adds examples targeted to the specific skills missing from the model's responses. On 5-shot evaluations across popular math benchmarks and five SLMs (1B--7B; Qwen, Llama), AdaptMI+ improves accuracy by up to 6% over naive skill-based strategies.
We present an embodied robotic system with an LLM-driven agent-orchestration architecture for autonomous household object management. The system integrates memory-augmented task planning, enabling robots to execute high-level user commands while tracking past actions. It employs three specialized agents: a routing agent, a task planning agent, and a knowledge base agent, each powered by task-specific LLMs. By leveraging in-context learning, our system avoids the need for explicit model training. RAG enables the system to retrieve context from past interactions, enhancing long-term object tracking. A combination of Grounded SAM and LLaMa3.2-Vision provides robust object detection, facilitating semantic scene understanding for task planning. Evaluation across three household scenarios demonstrates high task planning accuracy and an improvement in memory recall due to RAG. Specifically, Qwen2.5 yields best performance for specialized agents, while LLaMA3.1 excels in routing tasks. The source code is available at: https://github.com/marc1198/chat-hsr.
Domain-adaptive continual pretraining (DAPT) is a state-of-the-art technique that further trains a language model (LM) on its pretraining task, e.g., language masking. Although popular, it requires a significant corpus of domain-related data, which is difficult to obtain for specific domains in languages other than English, such as the process industry in the German language. This paper introduces an efficient approach called ICL-augmented pretraining or ICL-APT that leverages in-context learning (ICL) and k-nearest neighbors (kNN) to augment target data with domain-related and in-domain texts, significantly reducing GPU time while maintaining strong model performance. Our results show that this approach performs better than traditional DAPT by 3.5 points of the average IR metrics (e.g., mAP, MRR, and nDCG) and requires almost 4 times less computing time, providing a cost-effective solution for industries with limited computational capacity. The findings highlight the broader applicability of this framework to other low-resource industries, making NLP-based solutions more accessible and feasible in production environments.
Large Language Models (LLMs) have demonstrated remarkable capabilities in In-Context Learning (ICL). However, the fixed position length constraints in pre-trained models limit the number of demonstration examples. Recent efforts to extend context suffer from attention dispersion as the number of demonstrations increases. In this paper, we introduce Mitigating Attention Dispersion in large-scale ICL (MateICL) that enables LLMs to maintain effective self-attention as the context size grows. We first split the context into multiple windows, each filled to the model's context capacity, which are processed separately. Then, we introduce an additional layer to recalibrate the attention weights, prioritizing the query tokens as the number of demonstrations increases. Our empirical results show that MateICL can effectively leverage larger contexts to improve ICL performance. Compared to retrieval-based baselines, MateICL consistently achieves better performance without requiring an externally trained retrieval model. Despite recent advances in inference strategies (e.g., 32k token contexts), our results demonstrate that MateICL remains beneficial in computationally resource-constrained settings. The code is publicly available at https://github.com/amurtadha/MateICL.
Large language models exhibit exciting capabilities, yet can show surprisingly narrow generalization from finetuning. E.g. they can fail to generalize to simple reversals of relations they are trained on, or fail to make simple logical deductions based on trained information. These failures to generalize from fine-tuning can hinder practical application of these models. On the other hand, language models' in-context learning shows different inductive biases, and can generalize better in some cases. Here, we explore these differences in generalization between in-context- and fine-tuning-based learning. To do so, we constructed several novel datasets to evaluate and improve models' abilities to generalize from finetuning data. The datasets are designed to create clean tests of generalization, by isolating the knowledge in the dataset from that in pretraining. We expose pretrained large models to controlled subsets of the information in these datasets -- either in context, or through fine-tuning -- and evaluate their performance on test sets that require various types of generalization. We find overall that in data-matched settings, in-context learning can generalize more flexibly than fine-tuning (though we also find some qualifications of prior findings, such as cases when fine-tuning can generalize to reversals embedded in a larger structure of knowledge). We build on these findings to propose a method to enable improved generalization from fine-tuning: adding in-context inferences to finetuning data. We show that this method improves generalization across various splits of our datasets and other benchmarks. Our results have implications for understanding the inductive biases of different modes of learning in language models, and practically improving their performance.
In this paper, we address the task of multimodal-to-speech generation, which aims to synthesize high-quality speech from multiple input modalities: text, video, and reference audio. This task has gained increasing attention due to its wide range of applications, such as film production, dubbing, and virtual avatars. Despite recent progress, existing methods still suffer from limitations in speech intelligibility, audio-video synchronization, speech naturalness, and voice similarity to the reference speaker. To address these challenges, we propose AlignDiT, a multimodal Aligned Diffusion Transformer that generates accurate, synchronized, and natural-sounding speech from aligned multimodal inputs. Built upon the in-context learning capability of the DiT architecture, AlignDiT explores three effective strategies to align multimodal representations. Furthermore, we introduce a novel multimodal classifier-free guidance mechanism that allows the model to adaptively balance information from each modality during speech synthesis. Extensive experiments demonstrate that AlignDiT significantly outperforms existing methods across multiple benchmarks in terms of quality, synchronization, and speaker similarity. Moreover, AlignDiT exhibits strong generalization capability across various multimodal tasks, such as video-to-speech synthesis and visual forced alignment, consistently achieving state-of-the-art performance. The demo page is available at https://mm.kaist.ac.kr/projects/AlignDiT .
Large Language Models (LLMs) have shown promise in clinical decision support, yet their application to triage remains underexplored. We systematically investigate the capabilities of LLMs in emergency department triage through two key dimensions: (1) robustness to distribution shifts and missing data, and (2) counterfactual analysis of intersectional biases across sex and race. We assess multiple LLM-based approaches, ranging from continued pre-training to in-context learning, as well as machine learning approaches. Our results indicate that LLMs exhibit superior robustness, and we investigate the key factors contributing to the promising LLM-based approaches. Furthermore, in this setting, we identify gaps in LLM preferences that emerge in particular intersections of sex and race. LLMs generally exhibit sex-based differences, but they are most pronounced in certain racial groups. These findings suggest that LLMs encode demographic preferences that may emerge in specific clinical contexts or particular combinations of characteristics.
Long-term action anticipation (LTA) aims to predict future actions over an extended period. Previous approaches primarily focus on learning exclusively from video data but lack prior knowledge. Recent researches leverage large language models (LLMs) by utilizing text-based inputs which suffer severe information loss. To tackle these limitations single-modality methods face, we propose a novel Intention-Conditioned Vision-Language (ICVL) model in this study that fully leverages the rich semantic information of visual data and the powerful reasoning capabilities of LLMs. Considering intention as a high-level concept guiding the evolution of actions, we first propose to employ a vision-language model (VLM) to infer behavioral intentions as comprehensive textual features directly from video inputs. The inferred intentions are then fused with visual features through a multi-modality fusion strategy, resulting in intention-enhanced visual representations. These enhanced visual representations, along with textual prompts, are fed into LLM for future action anticipation. Furthermore, we propose an effective example selection strategy jointly considers visual and textual similarities, providing more relevant and informative examples for in-context learning. Extensive experiments with state-of-the-art performance on Ego4D, EPIC-Kitchens-55, and EGTEA GAZE+ datasets fully demonstrate the effectiveness and superiority of the proposed method.
Large language models (LLMs) have shown to be valuable tools for tackling process mining tasks. Existing studies report on their capability to support various data-driven process analyses and even, to some extent, that they are able to reason about how processes work. This reasoning ability suggests that there is potential for LLMs to tackle semantics-aware process mining tasks, which are tasks that rely on an understanding of the meaning of activities and their relationships. Examples of these include process discovery, where the meaning of activities can indicate their dependency, whereas in anomaly detection the meaning can be used to recognize process behavior that is abnormal. In this paper, we systematically explore the capabilities of LLMs for such tasks. Unlike prior work, which largely evaluates LLMs in their default state, we investigate their utility through both in-context learning and supervised fine-tuning. Concretely, we define five process mining tasks requiring semantic understanding and provide extensive benchmarking datasets for evaluation. Our experiments reveal that while LLMs struggle with challenging process mining tasks when used out of the box or with minimal in-context examples, they achieve strong performance when fine-tuned for these tasks across a broad range of process types and industries.
Recent improvements in large language model (LLM) performance on academic benchmarks, such as MATH and GSM8K, have emboldened their use as standalone tutors and as simulations of human learning. However, these new applications require more than evaluations of final solution generation. We introduce TutorGym to evaluate these applications more directly. TutorGym is a standard interface for testing artificial intelligence (AI) agents within existing intelligent tutoring systems (ITS) that have been tested and refined in classroom studies, including Cognitive Tutors (CTAT), Apprentice Tutors, and OATutors. TutorGym is more than a simple problem-solution benchmark, it situates AI agents within the interactive interfaces of existing ITSs. At each step of problem-solving, AI agents are asked what they would do as a tutor or as a learner. As tutors, AI agents are prompted to provide tutoring support -- such as generating examples, hints, and step-level correctness feedback -- which can be evaluated directly against the adaptive step-by-step support provided by existing ITSs. As students, agents directly learn from ITS instruction, and their mistakes and learning trajectories can be compared to student data. TutorGym establishes a common framework for training and evaluating diverse AI agents, including LLMs, computational models of learning, and reinforcement learning agents, within a growing suite of learning environments. Currently, TutorGym includes 223 different tutor domains. In an initial evaluation, we find that current LLMs are poor at tutoring -- none did better than chance at labeling incorrect actions, and next-step actions were correct only ~52-70% of the time -- but they could produce remarkably human-like learning curves when trained as students with in-context learning.