Large language models are able to exploit in-context learning to access external knowledge beyond their training data through retrieval-augmentation. While promising, its inner workings remain unclear. In this work, we shed light on the mechanism of in-context retrieval augmentation for question answering by viewing a prompt as a composition of informational components. We propose an attributionbased method to identify specialized attention heads, revealing in-context heads that comprehend instructions and retrieve relevant contextual information, and parametric heads that store entities' relational knowledge. To better understand their roles, we extract function vectors and modify their attention weights to show how they can influence the answer generation process. Finally, we leverage the gained insights to trace the sources of knowledge used during inference, paving the way towards more safe and transparent language models.

Chain-of-Thought (CoT) enhances the performance of Large Language Models (LLMs) on reasoning tasks by encouraging step-by-step solutions. However, the verbosity of CoT reasoning hinders its mass deployment in efficiency-critical applications. Recently, implicit CoT approaches have emerged, which encode reasoning steps within LLM's hidden embeddings (termed "implicit reasoning") rather than explicit tokens. This approach accelerates CoT reasoning by reducing the reasoning length and bypassing some LLM components. However, existing implicit CoT methods face two significant challenges: (1) they fail to preserve the semantic alignment between the implicit reasoning (when transformed to natural language) and the ground-truth reasoning, resulting in a significant CoT performance degradation, and (2) they focus on reducing the length of the implicit reasoning; however, they neglect the considerable time cost for an LLM to generate one individual implicit reasoning token.

However, recent studies show that the alignment can be easily compromised through finetuning with only a few adversarially designed training examples.

M-7B significantly improves the performance on real-world long-context tasks (e.g., 23.5

To validate the robustness, we design extensive experiments across diverse prompts in terms of length, format, and content.

A large language model can be less helpful if it exhibits output response homogenization. But whether two responses are considered homogeneous, and whether such homogenization is problematic, both depend on the task category. For instance, in objective math tasks, we often expect no variation in the final answer but anticipate variation in the problem-solving strategy. Whereas, for creative writing tasks, we may expect variation in key narrative components (e.g. plot, genre, setting, etc), beyond the vocabulary or embedding diversity produced by temperature-sampling. Previous work addressing output homogenization often fails to conceptualize diversity in a task-dependent way. We address this gap in the literature directly by making the following contributions. (1) We present a task taxonomy comprised of eight task categories that each have distinct concepts of output homogenization. (2) We introduce task-anchored functional diversity to better evaluate output homogenization. (3) We propose a task-anchored sampling technique that increases functional diversity for task categories where homogenization is undesired, while preserving it where it is desired. (4) We challenge the perceived existence of a diversity-quality trade-off by increasing functional diversity while maintaining response quality. Overall, we demonstrate how task dependence improves the evaluation and mitigation of output homogenization.

We show that deep neural networks trained across diverse tasks exhibit remarkably similar low-dimensional parametric subspaces. We provide the first large-scale empirical evidence that demonstrates that neural networks systematically converge to shared spectral subspaces regardless of initialization, task, or domain. Through mode-wise spectral analysis of over 1100 models - including 500 Mistral-7B LoRAs, 500 Vision Transformers, and 50 LLaMA-8B models - we identify universal subspaces capturing majority variance in just a few principal directions. By applying spectral decomposition techniques to the weight matrices of various architectures trained on a wide range of tasks and datasets, we identify sparse, joint subspaces that are consistently exploited, within shared architectures across diverse tasks and datasets. Our findings offer new insights into the intrinsic organization of information within deep networks and raise important questions about the possibility of discovering these universal subspaces without the need for extensive data and computational resources. Furthermore, this inherent structure has significant implications for model reusability, multi-task learning, model merging, and the development of training and inference-efficient algorithms, potentially reducing the carbon footprint of large-scale neural models.

Standard human preference-based alignment methods, such as Reinforcement Learning from Human Feedback (RLHF), are a cornerstone for aligning large language models (LLMs) with human values. However, these methods typically assume that preference data is clean and that all labels are equally reliable. In practice, large-scale preference datasets contain substantial noise due to annotator mistakes, inconsistent instructions, varying expertise, and even adversarial or low-effort feedback. This mismatch between recorded labels and ground-truth preferences can misguide training and degrade model performance. To address this issue, we introduce Robust Enhanced Policy Optimization (RE-PO), which uses an expectation-maximization procedure to infer the posterior correctness of each label and then adaptively reweight data points in the training loss to mitigate label noise. We further generalize this idea by establishing a theoretical link between arbitrary preference losses and their underlying probabilistic models, enabling a systematic transformation of existing alignment algorithms into robust counterparts and elevating RE-PO from a single method to a general framework for robust preference alignment. Theoretically, we prove that, under a perfectly calibrated model, RE-PO recovers the true noise level of the dataset. Empirically, we show that RE-PO consistently improves four state-of-the-art alignment methods (DPO, IPO, SimPO, and CPO); when applied to Mistral and Llama 3 models, the RE-PO-enhanced variants increase AlpacaEval 2 win rates by up to 7.0 percent over their respective baselines.

Trained on various human-authored corpora, Large Language Models (LLMs) have demonstrated a certain capability of reflecting specific human-like traits (e.g., personality or values) by prompting, benefiting applications like personalized LLMs and social simulations. However, existing methods suffer from the superficial elicitation problem: LLMs can only be steered to mimic shallow and unstable stylistic patterns, failing to embody the desired traits precisely and consistently across diverse tasks like humans. To address this challenge, we propose IROTE, a novel in-context method for stable and transferable trait elicitation. Drawing on psychological theories suggesting that traits are formed through identity-related reflection, our method automatically generates and optimizes a textual self-reflection within prompts, which comprises self-perceived experience, to stimulate LLMs' trait-driven behavior. The optimization is performed by iteratively maximizing an information-theoretic objective that enhances the connections between LLMs' behavior and the target trait, while reducing noisy redundancy in reflection without any fine-tuning, leading to evocative and compact trait reflection. Extensive experiments across three human trait systems manifest that one single IROTE-generated self-reflection can induce LLMs' stable impersonation of the target trait across diverse downstream tasks beyond simple questionnaire answering, consistently outperforming existing strong baselines.