We investigate the mechanisms underlying a range of list-processing tasks in LLMs, and we find that LLMs have learned to encode a compact, causal representation of a general filtering operation that mirrors the generic "filter" function of functional programming. Using causal mediation analysis on a diverse set of list-processing tasks, we find that a small number of attention heads, which we dub filter heads, encode a compact representation of the filtering predicate in their query states at certain tokens. We demonstrate that this predicate representation is general and portable: it can be extracted and reapplied to execute the same filtering operation on different collections, presented in different formats, languages, or even in tasks. However, we also identify situations where transformer LMs can exploit a different strategy for filtering: eagerly evaluating if an item satisfies the predicate and storing this intermediate result as a flag directly in the item representations. Our results reveal that transformer LMs can develop human-interpretable implementations of abstract computational operations that generalize in ways that are surprisingly similar to strategies used in traditional functional programming patterns.

Prompt tuning has emerged as a lightweight strategy for adapting foundation models to downstream tasks, particularly for resource-constrained systems. As pre-trained prompts become valuable assets, combining multiple source prompts offers a promising approach to enhance generalization for new tasks by leveraging complementary knowledge. However, naive aggregation often overlooks different source prompts have different contribution potential to the target task. To address this, we propose HGPrompt, a dynamic framework that learns optimal ensemble weights. These weights are optimized by jointly maximizing an information-theoretic metric for transferability and minimizing gradient conflicts via a novel regularization strategy. Specifically, we propose a differentiable prompt transferability metric to captures the dis-criminability of prompt-induced features on the target task. Meanwhile, HGPrompt match the gradient variances with respect to different source prompts based on Hessian and Fisher Information, ensuring stable and coherent knowledge transfer while suppressing gradient conflicts among them. Extensive experiments on the large-scale VT AB benchmark demonstrate the state-of-the-art performance of HGPrompt, validating its effectiveness in learning an optimal ensemble for effective multi-source prompt transfer.

Modern NLP systems increasingly rely on parameter-efficient adaptation methods to customize large pre-trained language models (PLMs) for new tasks without updating all model parameters. Among these, prompt tuning has become popular for its simplicity and low memory footprint: it learns a small set of continuous prompt embeddings while keeping the PLM frozen [5]. This allows adaptation to many tasks with minimal added parameters for both language models [2] and vision-language models [4]. Despite this efficiency, most existing prompt tuning methods are designed for single-task learning, where prompts are optimized independently for each task without sharing knowledge across them. This design is limiting in multi-task settings, where tasks often share semantic structure, labels, or data domains that can be exploited for transfer. Learning isolated prompts from scratch for each task misses opportunities for cross-task knowledge sharing--especially problematic in few-shot scenarios where data is scarce. Moreover, with only a single prompt vector per task, it becomes difficult to represent complex task relationships or adapt to varying label spaces, risking both underfitting of shared patterns and overfitting of task-specific noise. These challenges motivate the need for a more modular and transferable approach to prompt tuning that can share generalizable representations across tasks while preserving task-specific specialization in a parameter-efficient manner.

With the rapid advancement of large language models , code generation has become a key benchmark for evaluating LLM capabilities. However, existing benchmarks face two major challenges: (1) the escalating cost of constructing high-quality test suites and reference solutions, and (2) the increasing risk of data contamination, which undermines the reliability of benchmark-based evaluations. In this paper, we propose BIS, a prompt-centric evaluation framework that enables ground-truth-free prediction of LLM performance on code generation tasks. Rather than executing generated code, BIS estimates performance metrics by analyzing the prompt distribution alone. Built on importance sampling theory and implemented using Importance Weighted Autoencoders, our method reweights samples from existing annotated benchmarks to estimate performance on new, unseen benchmarks. To stabilize the estimation, we introduce weight truncation strategies and compute marginal expectations across the fitted distributions. BIS serves as a complementary tool that supports benchmark development and validation under constrained resources, offering actionable and quick feedback for prompt selection and contamination assessment. We conduct extensive experiments involving 8,000 evaluation points across 4 CodeLlama models and 9 diverse benchmarks. Our framework achieves an average absolute prediction error of 1.1% for code correctness scores, with best- and worst-case errors of 0.3% and 1.9%, respectively. It also generalizes well to other metrics, attaining average absolute errors of 2.15% for pass@1. These results demonstrate the reliability and broad applicability of BIS, which can significantly reduce the cost and effort of benchmarking LLMs in code-related tasks.

Just as humans display language patterns influenced by their native tongue when speaking new languages, LLMs often default to English-centric responses even when generating in other languages. Nevertheless, we observe that local cultural information persists within the models and can be readily activated for cultural customization. We first demonstrate that explicitly providing cultural context in prompts significantly improves the models' ability to generate culturally localized responses. We term the disparity in model performance with versus without explicit cultural context the explicit-implicit localization gap, indicating that while cultural knowledge exists within LLMs, it may not naturally surface in multilingual interactions if cultural context is not explicitly provided. Despite the explicit prompting benefit, however, the answers reduce in diversity and tend toward stereotypes. Second, we identify an explicit cultural customization vector, conserved across all non-English languages we explore, which enables LLMs to be steered from the synthetic English cultural world-model toward each non-English cultural world. Steered responses retain the diversity of implicit prompting and reduce stereotypes to dramatically improve the potential for customization. We discuss the implications of explicit cultural customization for understanding the conservation of alternative cultural world models within LLMs, and their controllable utility for translation, cultural customization, and the possibility of making the explicit implicit through soft control for expanded LLM function and appeal.

Large language models (LLMs) face significant challenges when balancing multiple high-level objectives, such as generating coherent, relevant, and high-quality responses while maintaining efficient task adaptation across diverse tasks. To address these challenges, we introduce Transfer-Prompting, a novel two-stage framework designed to enhance cross-task adaptation in prompt generation. The framework comprises two key components: (1) source prompt construction, which refines the original prompts on source task datasets to generate source prompts with enhanced generalization ability, and (2) target prompt generation, which enhances cross-task adaptation of target prompts by fine-tuning a set of high-scored source prompts on task-specific datasets. In each optimization cycle, a reference LLM generates candidate prompts based on historical prompt-score pairs and task descriptions in our designed reference prompt. These candidate prompts are refined iteratively, while a scorer LLM evaluates their effectiveness using the multi-dimensional metrics designed in the objective prompts evaluator-a novel contribution in this work that provides a holistic evaluation of prompt quality and task performance. This feedback loop facilitates continuous refinement, optimizing both prompt quality and task-specific outcomes. We validate Transfer-Prompting through extensive experiments across 25 LLMs, including 7 foundational models and 18 specialized models, evaluated on 9 diverse datasets. The results demonstrate that Transfer-Prompting significantly improves task-specific performance, highlighting its potential for enhancing cross-task adaptation in LLMs. The code is available at https://github.com/llm172/Transfer-Prompting.

A central question in multilingual language modeling is whether large language models (LLMs) develop a universal concept representation, disentangled from specific languages. In this paper, we address this question by analyzing latent representations (latents) during a word translation task in transformer-based LLMs. We strategically extract latents from a source translation prompt and insert them into the forward pass on a target translation prompt. By doing so, we find that the output language is encoded in the latent at an earlier layer than the concept to be translated. Building on this insight, we conduct two key experiments. First, we demonstrate that we can change the concept without changing the language and vice versa through activation patching alone. Second, we show that patching with the mean over latents across different languages does not impair and instead improves the models' performance in translating the concept. Our results provide evidence for the existence of language-agnostic concept representations within the investigated models.

In recent years, multi-task prompt tuning has garnered considerable attention for its inherent modularity and potential to enhance parameter-efficient transfer learning across diverse tasks. This paper aims to analyze and improve the performance of multiple tasks by facilitating the transfer of knowledge between their corresponding prompts in a multi-task setting. Our proposed approach decomposes the prompt for each target task into a combination of shared prompts (source prompts) and a task-specific prompt (private prompt). During training, the source prompts undergo fine-tuning and are integrated with the private prompt to drive the target prompt for each task. We present and compare multiple methods for combining source prompts to construct the target prompt, analyzing the roles of both source and private prompts within each method. We investigate their contributions to task performance and offer flexible, adjustable configurations based on these insights to optimize performance. Our empirical findings clearly showcase improvements in accuracy and robustness compared to the conventional practice of prompt tuning and related works. Notably, our results substantially outperform other methods in the field in few-shot settings, demonstrating superior performance in various tasks across GLUE benchmark, among other tasks. This achievement is attained with a significantly reduced amount of training data, making our method a promising one for few-shot settings.

Definition bias is a negative phenomenon that can mislead models. Definition bias in information extraction appears not only across datasets from different domains but also within datasets sharing the same domain. We identify two types of definition bias in IE: bias among information extraction datasets and bias between information extraction datasets and instruction tuning datasets. To systematically investigate definition bias, we conduct three probing experiments to quantitatively analyze it and discover the limitations of unified information extraction and large language models in solving definition bias. To mitigate definition bias in information extraction, we propose a multi-stage framework consisting of definition bias measurement, bias-aware fine-tuning, and task-specific bias mitigation. Experimental results demonstrate the effectiveness of our framework in addressing definition bias. Resources of this paper can be found at https://github.com/EZ-hwh/definition-bias

Inspecting the information encoded in hidden representations of large language models (LLMs) can explain models' behavior and verify their alignment with human values. Given the capabilities of LLMs in generating human-understandable text, we propose leveraging the model itself to explain its internal representations in natural language. We introduce a framework called Patchscopes and show how it can be used to answer a wide range of questions about an LLM's computation. We show that prior interpretability methods based on projecting representations into the vocabulary space and intervening on the LLM computation can be viewed as instances of this framework. Moreover, several of their shortcomings such as failure in inspecting early layers or lack of expressivity can be mitigated by Patchscopes. Beyond unifying prior inspection techniques, Patchscopes also opens up new possibilities such as using a more capable model to explain the representations of a smaller model, and unlocks new applications such as self-correction in multi-hop reasoning.