Large language models (LLMs) are trained on extensive datasets that encapsulate substantial world knowledge. However, their outputs often include confidently stated inaccuracies. Earlier works suggest that LLMs encode truthfulness as a distinct linear feature, termed the "truth direction", which can classify truthfulness reliably. We address several open questions about the truth direction: (i) whether LLMs universally exhibit consistent truth directions; (ii) whether sophisticated probing techniques are necessary to identify truth directions; and (iii) how the truth direction generalizes across diverse contexts. Our findings reveal that not all LLMs exhibit consistent truth directions, with stronger representations observed in more capable models, particularly in the context of logical negation. Additionally, we demonstrate that truthfulness probes trained on declarative atomic statements can generalize effectively to logical transformations, question-answering tasks, in-context learning, and external knowledge sources. Finally, we explore the practical application of truthfulness probes in selective question-answering, illustrating their potential to improve user trust in LLM outputs. These results advance our understanding of truth directions and provide new insights into the internal representations of LLM beliefs. Our code is public at https://github.com/colored-dye/truthfulness_probe_generalization
Forecasting over Temporal Knowledge Graphs (TKGs) which predicts future facts based on historical ones has received much attention. Recent studies have introduced Large Language Models (LLMs) for this task to enhance the models' generalization abilities. However, these models perform forecasting via simultaneously learning two kinds of entangled knowledge in the TKG: (1) general patterns, i.e., invariant temporal structures shared across different scenarios; and (2) scenario information, i.e., factual knowledge engaged in specific scenario, such as entities and relations. As a result, the learning processes of these two kinds of knowledge may interfere with each other, which potentially impact the generalization abilities of the models. To enhance the generalization ability of LLMs on this task, in this paper, we propose a General-to-Specific learning framework (G2S) that disentangles the learning processes of the above two kinds of knowledge. In the general learning stage, we mask the scenario information in different TKGs and convert it into anonymous temporal structures. After training on these structures, the model is able to capture the general patterns across different TKGs. In the specific learning stage, we inject the scenario information into the structures via either in-context learning or fine-tuning modes. Experimental results show that G2S effectively improves the generalization abilities of LLMs.
In-Context Learning (ICL) enhances the performance of large language models (LLMs) with demonstrations. However, obtaining these demonstrations primarily relies on manual effort. In most real-world scenarios, users are often unwilling or unable to provide such demonstrations. Inspired by the human analogy, we explore a new ICL paradigm CrossICL to study how to utilize existing source task demonstrations in the ICL for target tasks, thereby obtaining reliable guidance without any additional manual effort. To explore this, we first design a two-stage alignment strategy to mitigate the interference caused by gaps across tasks, as the foundation for our experimental exploration. Based on it, we conduct comprehensive exploration of CrossICL, with 875 NLP tasks from the Super-NI benchmark and six types of LLMs, including GPT-4o. Experimental results demonstrate the effectiveness of CrossICL and provide valuable insights on questions like the criteria for selecting cross-task demonstrations, as well as the types of task-gap-induced interference in CrossICL.
In-context learning (ICL) enables Large Language Models (LLMs) to adapt to new tasks using few examples, with task vectors - specific hidden state activations - hypothesized to encode task information. Existing studies are limited by small-scale benchmarks, restricting comprehensive analysis. We introduce QuiteAFew, a novel dataset of 3,096 diverse few-shot tasks, each with 30 input-output pairs derived from the Alpaca dataset. Experiments with Llama-3-8B on QuiteAFew reveal: (1) task vector performance peaks at an intermediate layer (e.g., 15th), (2) effectiveness varies significantly by task type, and (3) complex tasks rely on multiple, subtask-specific vectors rather than a single vector, suggesting distributed task knowledge representation.
Despite their central role in the success of foundational models and large-scale language modeling, the theoretical foundations governing the operation of Transformers remain only partially understood. Contemporary research has largely focused on their representational capacity for language comprehension and their prowess in in-context learning, frequently under idealized assumptions such as linearized attention mechanisms. Initially conceived to model sequence-to-sequence transformations, a fundamental and unresolved question is whether Transformers can robustly perform functional regression over sequences of input tokens. This question assumes heightened importance given the inherently non-Euclidean geometry underlying real-world data distributions. In this work, we establish that Transformers equipped with softmax-based nonlinear attention are uniformly consistent when tasked with executing Ordinary Least Squares (OLS) regression, provided both the inputs and outputs are embedded in hyperbolic space. We derive deterministic upper bounds on the empirical error which, in the asymptotic regime, decay at a provable rate of $\mathcal{O}(t^{-1/2d})$, where $t$ denotes the number of input tokens and $d$ the embedding dimensionality. Notably, our analysis subsumes the Euclidean setting as a special case, recovering analogous convergence guarantees parameterized by the intrinsic dimensionality of the data manifold. These theoretical insights are corroborated through empirical evaluations on real-world datasets involving both continuous and categorical response variables.
Large language models (LLMs) have shown impressive performance on downstream tasks through in-context learning (ICL), which heavily relies on the demonstrations selected from annotated datasets. However, these datasets often exhibit long-tailed class distributions in real-world scenarios, leading to biased demonstration selection. In this work, we show that such class imbalances significantly degrade the ICL performance across various tasks, regardless of selection methods. Moreover, classical rebalancing methods, which focus solely on class weights, yield poor performance due to neglecting condition bias--skewed feature distributions within classes. To address this, we propose Reweighting with Conditional Bias (dubbed RCB), a simple and complementary approach to enhance ICL performance under class imbalance. In particular, RCB estimates conditional bias using a balanced subset and re-weights demonstration scores based on both class weight and conditional bias. In effect, RCB prevents over-selection from dominant classes while preserving the efficacy of current selection methods. Extensive experiments on common benchmarks demonstrate the effectiveness of our method, improving the average accuracy of current selection methods by up to 5.42%.
We examine limits of in-context learning (ICL) in transformer models trained from scratch, focusing on function approximation tasks as a controlled setting to uncover fundamental behaviors. While we show empirically that transformer models can generalize, approximating unseen classes of polynomial (non linear) functions, they cannot generalize beyond certain values. We provide both empirical and mathematical arguments explaining that these limitations stem from architectural components, namely layer normalization and the attention scoring function, softmax. Together, our findings reveal structural constraints on ICL that are often masked in more complex NLP tasks but that need to be understood to improve robustness and interpretability in transformer-based models.
In-Context Learning (ICL) allows Large Language Models (LLMs) to adapt to new tasks with just a few examples, but their predictions often suffer from systematic biases, leading to unstable performances in classification. While calibration techniques are proposed to mitigate these biases, we show that, in the logit space, many of these methods are equivalent to merely shifting the LLM's decision boundary without having the ability to alter its orientation. This proves inadequate when biases cause the LLM to be severely misdirected. To address these limitations and provide a unifying framework, we propose Supervised Calibration (SC), a loss-minimization based framework which learns an optimal, per-class affine transformation of the LLM's predictive probabilities in the logit space without requiring external data beyond the context. By using a more expressive functional class, SC not only subsumes many existing calibration methods in ICL as special cases, but also enables the ability to alter and even completely reverse the orientation of the LLM's decision boundary. Furthermore, SC's loss-based nature facilitates the seamless integration of two purpose-built regularization techniques: context-invariance and directional trust-region. The former is designed to tackle the instability issue in ICL, while the latter controls the degree of calibration. Finally, SC delivers state-of-the-art performance over calibration baselines in the 4-shot, 8-shot, and 16-shot settings across all nine datasets for Mistral-7B-Instruct-v0.3, LLaMA-2-7B-chat, and Qwen2-7B-Instruct.
Detecting toxic content using language models is important but challenging. While large language models (LLMs) have demonstrated strong performance in understanding Chinese, recent studies show that simple character substitutions in toxic Chinese text can easily confuse the state-of-the-art (SOTA) LLMs. In this paper, we highlight the multimodal nature of Chinese language as a key challenge for deploying LLMs in toxic Chinese detection. First, we propose a taxonomy of 3 perturbation strategies and 8 specific approaches in toxic Chinese content. Then, we curate a dataset based on this taxonomy, and benchmark 9 SOTA LLMs (from both the US and China) to assess if they can detect perturbed toxic Chinese text. Additionally, we explore cost-effective enhancement solutions like in-context learning (ICL) and supervised fine-tuning (SFT). Our results reveal two important findings. (1) LLMs are less capable of detecting perturbed multimodal Chinese toxic contents. (2) ICL or SFT with a small number of perturbed examples may cause the LLMs "overcorrect'': misidentify many normal Chinese contents as toxic.
Currently, zero-shot voice conversion systems are capable of synthesizing the voice of unseen speakers. However, most existing approaches struggle to accurately replicate the speaking style of the source speaker or mimic the distinctive speaking style of the target speaker, thereby limiting the controllability of voice conversion. In this work, we propose Discl-VC, a novel voice conversion framework that disentangles content and prosody information from self-supervised speech representations and synthesizes the target speaker's voice through in-context learning with a flow matching transformer. To enable precise control over the prosody of generated speech, we introduce a mask generative transformer that predicts discrete prosody tokens in a non-autoregressive manner based on prompts. Experimental results demonstrate the superior performance of Discl-VC in zero-shot voice conversion and its remarkable accuracy in prosody control for synthesized speech.
In-context learning (ICL) enables large language models (LLMs) to perform new tasks using only a few demonstrations. In Named Entity Recognition (NER), demonstrations are typically selected based on semantic similarity to the test instance, ignoring training labels and resulting in suboptimal performance. We introduce DEER, a new method that leverages training labels through token-level statistics to improve ICL performance. DEER first enhances example selection with a label-guided, token-based retriever that prioritizes tokens most informative for entity recognition. It then prompts the LLM to revisit error-prone tokens, which are also identified using label statistics, and make targeted corrections. Evaluated on five NER datasets using four different LLMs, DEER consistently outperforms existing ICL methods and approaches the performance of supervised fine-tuning. Further analysis shows its effectiveness on both seen and unseen entities and its robustness in low-resource settings.
Factual knowledge extraction aims to explicitly extract knowledge parameterized in pre-trained language models for application in downstream tasks. While prior work has been investigating the impact of supervised fine-tuning data on the factuality of large language models (LLMs), its mechanism remains poorly understood. We revisit this impact through systematic experiments, with a particular focus on the factuality gap that arises when fine-tuning on known versus unknown knowledge. Our findings show that this gap can be mitigated at the inference stage, either under out-of-distribution (OOD) settings or by using appropriate in-context learning (ICL) prompts (i.e., few-shot learning and Chain of Thought (CoT)). We prove this phenomenon theoretically from the perspective of knowledge graphs, showing that the test-time prompt may diminish or even overshadow the impact of fine-tuning data and play a dominant role in knowledge extraction. Ultimately, our results shed light on the interaction between finetuning data and test-time prompt, demonstrating that ICL can effectively compensate for shortcomings in fine-tuning data, and highlighting the need to reconsider the use of ICL prompting as a means to evaluate the effectiveness of fine-tuning data selection methods.
In this position paper we raise critical awareness of a realistic view of LLM capabilities that eschews extreme alternative views that LLMs are either "stochastic parrots" or in possession of "emergent" advanced reasoning capabilities, which, due to their unpredictable emergence, constitute an existential threat. Our middle-ground view is that LLMs extrapolate from priors from their training data, and that a mechanism akin to in-context learning enables the targeting of the appropriate information from which to extrapolate. We call this "context-directed extrapolation." Under this view, substantiated though existing literature, while reasoning capabilities go well beyond stochastic parroting, such capabilities are predictable, controllable, not indicative of advanced reasoning akin to high-level cognitive capabilities in humans, and not infinitely scalable with additional training. As a result, fears of uncontrollable emergence of agency are allayed, while research advances are appropriately refocused on the processes of context-directed extrapolation and how this interacts with training data to produce valuable capabilities in LLMs. Future work can therefore explore alternative augmenting techniques that do not rely on inherent advanced reasoning in LLMs.
Using Large Language Models (LLMs) to generate training data can potentially be a preferable way to improve zero or few-shot NLP tasks. However, many problems remain to be investigated for this direction. For the task of Relation Extraction (RE), we find that samples generated by directly prompting LLMs may easily have high structural similarities with each other. They tend to use a limited variety of phrasing while expressing the relation between a pair of entities. Therefore, in this paper, we study how to effectively improve the diversity of the training samples generated with LLMs for RE, while also maintaining their correctness. We first try to make the LLMs produce dissimilar samples by directly giving instructions in In-Context Learning (ICL) prompts. Then, we propose an approach to fine-tune LLMs for diversity training sample generation through Direct Preference Optimization (DPO). Our experiments on commonly used RE datasets show that both attempts can improve the quality of the generated training data. We also find that comparing with directly performing RE with an LLM, training a non-LLM RE model with its generated samples may lead to better performance.
In-Context Learning (ICL) technique based on Large Language Models (LLMs) has gained prominence in Named Entity Recognition (NER) tasks for its lower computing resource consumption, less manual labeling overhead, and stronger generalizability. Nevertheless, most ICL-based NER methods depend on large-parameter LLMs: the open-source models demand substantial computational resources for deployment and inference, while the closed-source ones incur high API costs, raise data-privacy concerns, and hinder community collaboration. To address this question, we propose an Ensemble Learning Method for Named Entity Recognition (EL4NER), which aims at aggregating the ICL outputs of multiple open-source, small-parameter LLMs to enhance overall performance in NER tasks at less deployment and inference cost. Specifically, our method comprises three key components. First, we design a task decomposition-based pipeline that facilitates deep, multi-stage ensemble learning. Second, we introduce a novel span-level sentence similarity algorithm to establish an ICL demonstration retrieval mechanism better suited for NER tasks. Third, we incorporate a self-validation mechanism to mitigate the noise introduced during the ensemble process. We evaluated EL4NER on multiple widely adopted NER datasets from diverse domains. Our experimental results indicate that EL4NER surpasses most closed-source, large-parameter LLM-based methods at a lower parameter cost and even attains state-of-the-art (SOTA) performance among ICL-based methods on certain datasets. These results show the parameter efficiency of EL4NER and underscore the feasibility of employing open-source, small-parameter LLMs within the ICL paradigm for NER tasks.
In-context learning, the ability of large language models to perform tasks using only examples provided in the prompt, has recently been adapted for time series forecasting. This paradigm enables zero-shot prediction, where past values serve as context for forecasting future values, making powerful forecasting tools accessible to non-experts and increasing the performance when training data are scarce. Most existing zero-shot forecasting approaches rely on transformer architectures, which, despite their success in language, often fall short of expectations in time series forecasting, where recurrent models like LSTMs frequently have the edge. Conversely, while LSTMs are well-suited for time series modeling due to their state-tracking capabilities, they lack strong in-context learning abilities. We introduce TiRex that closes this gap by leveraging xLSTM, an enhanced LSTM with competitive in-context learning skills. Unlike transformers, state-space models, or parallelizable RNNs such as RWKV, TiRex retains state-tracking, a critical property for long-horizon forecasting. To further facilitate its state-tracking ability, we propose a training-time masking strategy called CPM. TiRex sets a new state of the art in zero-shot time series forecasting on the HuggingFace benchmarks GiftEval and Chronos-ZS, outperforming significantly larger models including TabPFN-TS (Prior Labs), Chronos Bolt (Amazon), TimesFM (Google), and Moirai (Salesforce) across both short- and long-term forecasts.
Bargaining, a critical aspect of real-world interactions, presents challenges for large language models (LLMs) due to limitations in strategic depth and adaptation to complex human factors. Existing benchmarks often fail to capture this real-world complexity. To address this and enhance LLM capabilities in realistic bargaining, we introduce a comprehensive framework centered on utility-based feedback. Our contributions are threefold: (1) BargainArena, a novel benchmark dataset with six intricate scenarios (e.g., deceptive practices, monopolies) to facilitate diverse strategy modeling; (2) human-aligned, economically-grounded evaluation metrics inspired by utility theory, incorporating agent utility and negotiation power, which implicitly reflect and promote opponent-aware reasoning (OAR); and (3) a structured feedback mechanism enabling LLMs to iteratively refine their bargaining strategies. This mechanism can positively collaborate with in-context learning (ICL) prompts, including those explicitly designed to foster OAR. Experimental results show that LLMs often exhibit negotiation strategies misaligned with human preferences, and that our structured feedback mechanism significantly improves their performance, yielding deeper strategic and opponent-aware reasoning.
A restaurant dinner or a hotel stay may lead to memorable experiences when guests encounter unexpected aspects that also match their interests. For example, an origami-making station in the waiting area of a restaurant may be both surprising and enjoyable for a customer who is passionate about paper crafts. Similarly, an exhibit of 18th century harpsichords would be atypical for a hotel lobby and likely pique the interest of a guest who has a passion for Baroque music. Motivated by this insight, in this paper we introduce the new task of engineering serendipity through recommendations of items with atypical aspects. We describe an LLM-based system pipeline that extracts atypical aspects from item reviews, then estimates and aggregates their user-specific utility in a measure of serendipity potential that is used to rerank a list of items recommended to the user. To facilitate system development and evaluation, we introduce a dataset of Yelp reviews that are manually annotated with atypical aspects and a dataset of artificially generated user profiles, together with crowdsourced annotations of user-aspect utility values. Furthermore, we introduce a custom procedure for dynamic selection of in-context learning examples, which is shown to improve LLM-based judgments of atypicality and utility. Experimental evaluations show that serendipity-based rankings generated by the system are highly correlated with ground truth rankings for which serendipity scores are computed from manual annotations of atypical aspects and their user-dependent utility. Overall, we hope that the new recommendation task and the associated system presented in this paper catalyze further research into recommendation approaches that go beyond accuracy in their pursuit of enhanced user satisfaction. The datasets and the code are made publicly available at https://github.com/ramituncc49er/ATARS .
Recent years have witnessed extensive exploration of Large Language Models (LLMs) on the field of Recommender Systems (RS). There are currently two commonly used strategies to enable LLMs to have recommendation capabilities: 1) The "Guidance-Only" strategy uses in-context learning to exploit and amplify the inherent semantic understanding and item recommendation capabilities of LLMs; 2) The "Tuning-Only" strategy uses supervised fine-tuning (SFT) to fine-tune LLMs with the aim of fitting them to real recommendation data. However, neither of these strategies can effectively bridge the gap between the knowledge space of LLMs and recommendation, and their performance do not meet our expectations. To better enable LLMs to learn recommendation knowledge, we combine the advantages of the above two strategies and proposed a novel "Guidance+Tuning" method called Self-Optimized Fine-Tuning (SOFT), which adopts the idea of curriculum learning. It first employs self-distillation to construct an auxiliary easy-to-learn but meaningful dataset from a fine-tuned LLM. Then it further utilizes a self-adaptive curriculum scheduler to enable LLMs to gradually learn from simpler data (self-distilled data) to more challenging data (real RS data). Extensive experiments demonstrate that SOFT significantly enhances the recommendation accuracy (37.59\% on average) of LLM-based methods. The code is available via https://anonymous.4open.science/r/Self-Optimized-Fine-Tuning-264E
Large Language Models (LLMs) and Knowledge Graphs (KGs) offer a promising approach to robust and explainable Question Answering (QA). While LLMs excel at natural language understanding, they suffer from knowledge gaps and hallucinations. KGs provide structured knowledge but lack natural language interaction. Ideally, an AI system should be both robust to missing facts as well as easy to communicate with. This paper proposes such a system that integrates LLMs and KGs without requiring training, ensuring adaptability across different KGs with minimal human effort. The resulting approach can be classified as a specific form of a Retrieval Augmented Generation (RAG) with a KG, thus, it is dubbed Knowledge Graph-extended Retrieval Augmented Generation (KG-RAG). It includes a question decomposition module to enhance multi-hop information retrieval and answer explainability. Using In-Context Learning (ICL) and Chain-of-Thought (CoT) prompting, it generates explicit reasoning chains processed separately to improve truthfulness. Experiments on the MetaQA benchmark show increased accuracy for multi-hop questions, though with a slight trade-off in single-hop performance compared to LLM with KG baselines. These findings demonstrate KG-RAG's potential to improve transparency in QA by bridging unstructured language understanding with structured knowledge retrieval.