Elucidating the rationale behind neural models' outputs has been challenging in the machine learning field, which is indeed applicable in this age of large language models (LLMs) and in-context learning (ICL). When it comes to estimating input attributions (IA), ICL poses a new issue of interpreting which example in the prompt, consisting of a set of examples, contributed to identifying the task/rule to be solved. To this end, in this paper, we introduce synthetic diagnostic tasks inspired by the poverty of the stimulus design in inductive reasoning; here, most in-context examples are ambiguous w.r.t. their underlying rule, and one critical example disambiguates the task demonstrated. The question is whether conventional IA methods can identify such an example in interpreting the inductive reasoning process in ICL. Our experiments provide several practical findings; for example, a certain simple IA method works the best, and the larger the model, the generally harder it is to interpret the ICL with gradient-based IA methods.
Science and engineering problems fall in the category of complex conceptual problems that require specific conceptual information (CI) like math/logic -related know-how, process information, or engineering guidelines to solve them. Large Language Models (LLMs) are promising agents to solve such complex conceptual problems due to their implications in advancing engineering and science tasks like assisted problem-solving. But vanilla LLMs, trained on open-world data, lack the necessary CI. In this work, we specifically explore shallow customization methods (SCMs) of LLMs for solving complex conceptual problems. We propose two novel SCM algorithms for LLM, to augment LLMs with CI and enable LLMs to solve complex conceptual problems: Conceptual In-Context Learning (C-ICL) and Chain of Concepts (CoC). The problem tackled in this paper is generation of proprietary data models in the engineering/industry domain based on conceptual information in data modelling guidelines. We evaluate our algorithms on varied sizes of the OpenAI LLMs against four evaluation metrics related to syntactic and semantic correctness, time and cost incurred. The proposed algorithms perform better than currently popular LLM SCMs like In-context Learning (ICL) and Chain of Thoughts (CoT). It was observed that as compared to CoT, response correctness increased by 30.6% and 29.88% for the new SCMs C-ICL and CoC respectively. Qualitative analysis suggests that the proposed new SCMs activate emergent capabilities in LLMs, previously unobserved in the existing SCMs. They make problem-solving processes more transparent and reduce hallucinations and the tendency of model responses to copy examples from prompts (parroting).
Existing continual learning (CL) methods mainly rely on fine-tuning or adapting large language models (LLMs). They still suffer from catastrophic forgetting (CF). Little work has been done to exploit in-context learning (ICL) to leverage the extensive knowledge within LLMs for CL without updating any parameters. However, incrementally learning each new task in ICL necessitates adding training examples from each class of the task to the prompt, which hampers scalability as the prompt length increases. This issue not only leads to excessively long prompts that exceed the input token limit of the underlying LLM but also degrades the model's performance due to the overextended context. To address this, we introduce InCA, a novel approach that integrates an external continual learner (ECL) with ICL to enable scalable CL without CF. The ECL is built incrementally to pre-select a small subset of likely classes for each test instance. By restricting the ICL prompt to only these selected classes, InCA prevents prompt lengths from becoming excessively long, while maintaining high performance. Experimental results demonstrate that InCA significantly outperforms existing CL baselines, achieving substantial performance gains.
Recently developed large language models (LLMs) have presented promising new avenues to address data scarcity in low-resource scenarios. In few-shot aspect-based sentiment analysis (ABSA), previous efforts have explored data augmentation techniques, which prompt LLMs to generate new samples by modifying existing ones. However, these methods fail to produce adequately diverse data, impairing their effectiveness. Besides, some studies apply in-context learning for ABSA by using specific instructions and a few selected examples as prompts. Though promising, LLMs often yield labels that deviate from task requirements. To overcome these limitations, we propose DS$^2$-ABSA, a dual-stream data synthesis framework targeted for few-shot ABSA. It leverages LLMs to synthesize data from two complementary perspectives: \textit{key-point-driven} and \textit{instance-driven}, which effectively generate diverse and high-quality ABSA samples in low-resource settings. Furthermore, a \textit{label refinement} module is integrated to improve the synthetic labels. Extensive experiments demonstrate that DS$^2$-ABSA significantly outperforms previous few-shot ABSA solutions and other LLM-oriented data generation methods.
Foundation models have vast potential to enable diverse AI applications. The powerful yet incomplete nature of these models has spurred a wide range of mechanisms to augment them with capabilities such as in-context learning, information retrieval, and code interpreting. We propose Vieira, a declarative framework that unifies these mechanisms in a general solution for programming with foundation models. Vieira follows a probabilistic relational paradigm and treats foundation models as stateless functions with relational inputs and outputs. It supports neuro-symbolic applications by enabling the seamless combination of such models with logic programs, as well as complex, multi-modal applications by streamlining the composition of diverse sub-models. We implement Vieira by extending the Scallop compiler with a foreign interface that supports foundation models as plugins. We implement plugins for 12 foundation models including GPT, CLIP, and SAM. We evaluate Vieira on 9 challenging tasks that span language, vision, and structured and vector databases. Our evaluation shows that programs in Vieira are concise, can incorporate modern foundation models, and have comparable or better accuracy than competitive baselines.
This paper introduces LongBench v2, a benchmark designed to assess the ability of LLMs to handle long-context problems requiring deep understanding and reasoning across real-world multitasks. LongBench v2 consists of 503 challenging multiple-choice questions, with contexts ranging from 8k to 2M words, across six major task categories: single-document QA, multi-document QA, long in-context learning, long-dialogue history understanding, code repository understanding, and long structured data understanding. To ensure the breadth and the practicality, we collect data from nearly 100 highly educated individuals with diverse professional backgrounds. We employ both automated and manual review processes to maintain high quality and difficulty, resulting in human experts achieving only 53.7% accuracy under a 15-minute time constraint. Our evaluation reveals that the best-performing model, when directly answers the questions, achieves only 50.1% accuracy. In contrast, the o1-preview model, which includes longer reasoning, achieves 57.7%, surpassing the human baseline by 4%. These results highlight the importance of enhanced reasoning ability and scaling inference-time compute to tackle the long-context challenges in LongBench v2. The project is available at https://longbench2.github.io.
Large language models (LLMs) demonstrate an impressive ability to utilise information within the context of their input sequences to appropriately respond to data unseen by the LLM during its training procedure. This ability is known as in-context learning (ICL). Humans and non-human animals demonstrate similar abilities, however their neural architectures differ substantially from LLMs. Despite this, a critical component within LLMs, the attention mechanism, resembles modern associative memory models, widely used in and influenced by the computational neuroscience community to model biological memory systems. Using this connection, we introduce an associative memory model capable of performing ICL. We use this as inspiration for a novel residual stream architecture which allows information to directly flow between attention heads. We test this architecture during training within a two-layer Transformer and show its ICL abilities manifest more quickly than without this modification. We then apply our architecture in small language models with 8 million parameters, focusing on attention head values, with results also indicating improved ICL performance at this larger and more naturalistic scale.
We applied few-shot in-context learning on the OPT-1.3B model for the natural language inference task and employed knowledge distillation to internalize the context information, reducing model parameter from 1.3B to 125M and achieving a size reduction from 2.5GB to 0.25GB. Compared to using in-context learning alone on similarly sized models, this context distillation approach achieved a nearly 50% improvement in out-of-domain accuracy, demonstrating superior knowledge transfer capabilities over prompt-based methods. Furthermore, this approach reduced memory consumption by up to 60% while delivering a 20% improvement in out-of-domain accuracy compared to conventional pattern-based fine-tuning.
The security of Large Language Models (LLMs) has become an important research topic since the emergence of ChatGPT. Though there have been various effective methods to defend against jailbreak attacks, prefilling attacks remain an unsolved and popular threat against open-sourced LLMs. In-Context Learning (ICL) offers a computationally efficient defense against various jailbreak attacks, yet no effective ICL methods have been developed to counter prefilling attacks. In this paper, we: (1) show that ICL can effectively defend against prefilling jailbreak attacks by employing adversative sentence structures within demonstrations; (2) characterize the effectiveness of this defense through the lens of model size, number of demonstrations, over-defense, integration with other jailbreak attacks, and the presence of safety alignment. Given the experimental results and our analysis, we conclude that there is no free lunch for defending against prefilling jailbreak attacks with ICL. On the one hand, current safety alignment methods fail to mitigate prefilling jailbreak attacks, but adversative structures within ICL demonstrations provide robust defense across various model sizes and complex jailbreak attacks. On the other hand, LLMs exhibit similar over-defensiveness when utilizing ICL demonstrations with adversative structures, and this behavior appears to be independent of model size.
Long context understanding remains challenging for large language models due to their limited context windows. This paper introduces Long Input Fine-Tuning (LIFT) for long context modeling, a novel framework that enhances LLM performance on long-context tasks by adapting model parameters to the context at test time. LIFT enables efficient processing of lengthy inputs without the computational burden of offline long-context adaptation, and can improve the long-context capabilities of arbitrary short-context models. The framework is further enhanced by integrating in-context learning and pre-LIFT supervised fine-tuning. The combination of in-context learning and LIFT enables short-context models like Llama 3 to handle arbitrarily long contexts and consistently improves their performance on popular long-context benchmarks like LooGLE and LongBench. We also provide a comprehensive analysis of the strengths and limitations of LIFT on long context understanding, offering valuable directions for future research.
Annotation of medical images, such as MRI and CT scans, is crucial for evaluating treatment efficacy and planning radiotherapy. However, the extensive workload of medical professionals limits their ability to annotate large image datasets, posing a bottleneck for AI applications in medical imaging. To address this, we propose In-context Cascade Segmentation (ICS), a novel method that minimizes annotation requirements while achieving high segmentation accuracy for sequential medical images. ICS builds on the UniverSeg framework, which performs few-shot segmentation using support images without additional training. By iteratively adding the inference results of each slice to the support set, ICS propagates information forward and backward through the sequence, ensuring inter-slice consistency. We evaluate the proposed method on the HVSMR dataset, which includes segmentation tasks for eight cardiac regions. Experimental results demonstrate that ICS significantly improves segmentation performance in complex anatomical regions, particularly in maintaining boundary consistency across slices, compared to baseline methods. The study also highlights the impact of the number and position of initial support slices on segmentation accuracy. ICS offers a promising solution for reducing annotation burdens while delivering robust segmentation results, paving the way for its broader adoption in clinical and research applications.
Large language models (LLMs) excel in various tasks but remain vulnerable to jailbreak attacks, where adversaries manipulate prompts to generate harmful outputs. Examining jailbreak prompts helps uncover the shortcomings of LLMs. However, current jailbreak methods and the target model's defenses are engaged in an independent and adversarial process, resulting in the need for frequent attack iterations and redesigning attacks for different models. To address these gaps, we propose a Reverse Embedded Defense Attack (REDA) mechanism that disguises the attack intention as the "defense". intention against harmful content. Specifically, REDA starts from the target response, guiding the model to embed harmful content within its defensive measures, thereby relegating harmful content to a secondary role and making the model believe it is performing a defensive task. The attacking model considers that it is guiding the target model to deal with harmful content, while the target model thinks it is performing a defensive task, creating an illusion of cooperation between the two. Additionally, to enhance the model's confidence and guidance in "defensive" intentions, we adopt in-context learning (ICL) with a small number of attack examples and construct a corresponding dataset of attack examples. Extensive evaluations demonstrate that the REDA method enables cross-model attacks without the need to redesign attack strategies for different models, enables successful jailbreak in one iteration, and outperforms existing methods on both open-source and closed-source models.
Recently, large language models (LLMs) have significantly improved the performance of text-to-SQL systems. Nevertheless, many state-of-the-art (SOTA) approaches have overlooked the critical aspect of system robustness. Our experiments reveal that while LLM-driven methods excel on standard datasets, their accuracy is notably compromised when faced with adversarial perturbations. To address this challenge, we propose a robust text-to-SQL solution, called Solid-SQL, designed to integrate with various LLMs. We focus on the pre-processing stage, training a robust schema-linking model enhanced by LLM-based data augmentation. Additionally, we design a two-round, structural similarity-based example retrieval strategy for in-context learning. Our method achieves SOTA SQL execution accuracy levels of 82.1% and 58.9% on the general Spider and Bird benchmarks, respectively. Furthermore, experimental results show that Solid-SQL delivers an average improvement of 11.6% compared to baselines on the perturbed Spider-Syn, Spider-Realistic, and Dr. Spider benchmarks.
Multimodal Large Language Models (MLLMs) have significantly advanced visual tasks by integrating visual representations into large language models (LLMs). The textual modality, inherited from LLMs, equips MLLMs with abilities like instruction following and in-context learning. In contrast, the visual modality enhances performance in downstream tasks by leveraging rich semantic content, spatial information, and grounding capabilities. These intrinsic modalities work synergistically across various visual tasks. Our research initially reveals a persistent imbalance between these modalities, with text often dominating output generation during visual instruction tuning. This imbalance occurs when using both full fine-tuning and parameter-efficient fine-tuning (PEFT) methods. We then found that re-balancing these modalities can significantly reduce the number of trainable parameters required, inspiring a direction for further optimizing visual instruction tuning. We introduce Modality Linear Representation-Steering (MoReS) to achieve the goal. MoReS effectively re-balances the intrinsic modalities throughout the model, where the key idea is to steer visual representations through linear transformations in the visual subspace across each model layer. To validate our solution, we composed LLaVA Steering, a suite of models integrated with the proposed MoReS method. Evaluation results show that the composed LLaVA Steering models require, on average, 500 times fewer trainable parameters than LoRA needs while still achieving comparable performance across three visual benchmarks and eight visual question-answering tasks. Last, we present the LLaVA Steering Factory, an in-house developed platform that enables researchers to quickly customize various MLLMs with component-based architecture for seamlessly integrating state-of-the-art models, and evaluate their intrinsic modality imbalance.
Humans distill complex experiences into fundamental abstractions that enable rapid learning and adaptation. Similarly, autoregressive transformers exhibit adaptive learning through in-context learning (ICL), which begs the question of how. In this paper, we propose concept encoding-decoding mechanism to explain ICL by studying how transformers form and use internal abstractions in their representations. On synthetic ICL tasks, we analyze the training dynamics of a small transformer and report the coupled emergence of concept encoding and decoding. As the model learns to encode different latent concepts (e.g., ``Finding the first noun in a sentence.") into distinct, separable representations, it concureently builds conditional decoding algorithms and improve its ICL performance. We validate the existence of this mechanism across pretrained models of varying scales (Gemma-2 2B/9B/27B, Llama-3.1 8B/70B). Further, through mechanistic interventions and controlled finetuning, we demonstrate that the quality of concept encoding is causally related and predictive of ICL performance. Our empirical insights shed light into better understanding the success and failure modes of large language models via their representations.
Geometry mathematics problems pose significant challenges for large language models (LLMs) because they involve visual elements and spatial reasoning. Current methods primarily rely on symbolic character awareness to address these problems. Considering geometry problem solving is a relatively nascent field with limited suitable datasets and currently almost no work on solid geometry problem solving, we collect a geometry question-answer dataset by sourcing geometric data from Chinese high school education websites, referred to as GeoMath. It contains solid geometry questions and answers with accurate reasoning steps as compensation for existing plane geometry datasets. Additionally, we propose a Large Multi-modal Model (LMM) framework named Geo-LLaVA, which incorporates retrieval augmentation with supervised fine-tuning (SFT) in the training stage, called meta-training, and employs in-context learning (ICL) during inference to improve performance. Our fine-tuned model with ICL attains the state-of-the-art performance of 65.25% and 42.36% on selected questions of the GeoQA dataset and GeoMath dataset respectively with proper inference steps. Notably, our model initially endows the ability to solve solid geometry problems and supports the generation of reasonable solid geometry picture descriptions and problem-solving steps. Our research sets the stage for further exploration of LLMs in multi-modal math problem-solving, particularly in geometry math problems.
Detecting music entities such as song titles or artist names is a useful application to help use cases like processing music search queries or analyzing music consumption on the web. Recent approaches incorporate smaller language models (SLMs) like BERT and achieve high results. However, further research indicates a high influence of entity exposure during pre-training on the performance of the models. With the advent of large language models (LLMs), these outperform SLMs in a variety of downstream tasks. However, researchers are still divided if this is applicable to tasks like entity detection in texts due to issues like hallucination. In this paper, we provide a novel dataset of user-generated metadata and conduct a benchmark and a robustness study using recent LLMs with in-context-learning (ICL). Our results indicate that LLMs in the ICL setting yield higher performance than SLMs. We further uncover the large impact of entity exposure on the best performing LLM in our study.
Yes! In the present-day documenting and preserving endangered languages, the application of Large Language Models (LLMs) presents a promising approach. This paper explores how LLMs, particularly through in-context learning, can assist in generating grammatical information for low-resource languages with limited amount of data. We takes Moklen as a case study to evaluate the efficacy of LLMs in producing coherent grammatical rules and lexical entries using only bilingual dictionaries and parallel sentences of the unknown language without building the model from scratch. Our methodology involves organising the existing linguistic data and prompting to efficiently enable to generate formal XLE grammar. Our results demonstrate that LLMs can successfully capture key grammatical structures and lexical information, although challenges such as the potential for English grammatical biases remain. This study highlights the potential of LLMs to enhance language documentation efforts, providing a cost-effective solution for generating linguistic data and contributing to the preservation of endangered languages.
In-context learning (ICL) enables large language models (LLMs) to adapt to new tasks without fine-tuning by leveraging contextual information provided within a prompt. However, ICL relies not only on contextual clues but also on the global knowledge acquired during pretraining for the next token prediction. Analyzing this process has been challenging due to the complex computational circuitry of LLMs. This paper investigates the balance between in-context information and pretrained bigram knowledge in token prediction, focusing on the induction head mechanism, a key component in ICL. Leveraging the fact that a two-layer transformer can implement the induction head mechanism with associative memories, we theoretically analyze the logits when a two-layer transformer is given prompts generated by a bigram model. In the experiments, we design specific prompts to evaluate whether the outputs of a two-layer transformer align with the theoretical results.
Radiology report generation (RRG) models typically focus on individual exams, often overlooking the integration of historical visual or textual data, which is crucial for patient follow-ups. Traditional methods usually struggle with long sequence dependencies when incorporating historical information, but large language models (LLMs) excel at in-context learning, making them well-suited for analyzing longitudinal medical data. In light of this, we propose a novel Historical-Constrained Large Language Models (HC-LLM) framework for RRG, empowering LLMs with longitudinal report generation capabilities by constraining the consistency and differences between longitudinal images and their corresponding reports. Specifically, our approach extracts both time-shared and time-specific features from longitudinal chest X-rays and diagnostic reports to capture disease progression. Then, we ensure consistent representation by applying intra-modality similarity constraints and aligning various features across modalities with multimodal contrastive and structural constraints. These combined constraints effectively guide the LLMs in generating diagnostic reports that accurately reflect the progression of the disease, achieving state-of-the-art results on the Longitudinal-MIMIC dataset. Notably, our approach performs well even without historical data during testing and can be easily adapted to other multimodal large models, enhancing its versatility.