Evaluating the open-ended text generation of large language models (LLMs) is challenging because of the lack of a clear ground truth and the high cost of human or LLM-based assessments. We propose a novel benchmark that evaluates LLMs using n-gram statistics and rules, without relying on human judgement or LLM-as-a-judge approaches. Using 50 question and reference answer sets, we introduce three new metrics based on n-grams and rules: Fluency, Truthfulness, and Helpfulness. Our benchmark strongly correlates with GPT-4o-based evaluations while requiring significantly fewer computational resources, demonstrating its effectiveness as a scalable alternative for assessing LLMs' open-ended generation capabilities.
Grammatical error correction (GEC) aims to correct grammatical, spelling, and semantic errors in natural language text. With the growing of large language models (LLMs), direct text generation has gradually become the focus of the GEC methods, and few-shot in-context learning presents a cost-effective solution. However, selecting effective in-context examples remains challenging, as the similarity between input texts does not necessarily correspond to similar grammatical error patterns. In this paper, we propose a novel retrieval method based on natural language grammatical error explanations (GEE) to address this issue. Our method retrieves suitable few-shot demonstrations by matching the GEE of the test input with that of pre-constructed database samples, where explanations for erroneous samples are generated by LLMs. We conducted multilingual GEC few-shot experiments on both major open-source and closed-source LLMs. Experiments across five languages show that our method outperforms existing semantic and BM25-based retrieval techniques, without requiring additional training or language adaptation. This also suggests that matching error patterns is key to selecting examples.
As large language models (LLMs) advance, their ability to perform in-context learning and few-shot language generation has improved significantly. This has spurred using LLMs to produce high-quality synthetic data to enhance the performance of smaller models like online retrievers or weak LLMs. However, LLM-generated synthetic data often differs from the real data in key language attributes (e.g., styles, tones, content proportions, etc.). As a result, mixing these synthetic data directly with real data may distort the original data distribution, potentially hindering performance improvements. To solve this, we introduce SynAlign: a synthetic data generation and filtering framework based on key attribute distribution matching. Before generation, SynAlign employs an uncertainty tracker surrogated by the Gaussian Process model to iteratively select data clusters distinct from selected ones as demonstrations for new data synthesis, facilitating the efficient exploration diversity of the real data. Then, a latent attribute reasoning method is employed: the LLM summarizes linguistic attributes of demonstrations and then synthesizes new data based on them. This approach facilitates synthesizing diverse data with linguistic attributes that appear in real data.After generation, the Maximum Mean Discrepancy is used as the objective function to learn the sampling weight of each synthetic data, ensuring distribution matching with the real data. Our experiments on multiple text prediction tasks show significant performance improvements. We also conducted an online A/B test on an online retriever to demonstrate SynAlign's effectiveness.
The rise of Large Language Models (LLMs) in software engineering, particularly in code generation, has garnered significant attention. However, assessing the quality of AI-generated code remains a challenge due to the inherent complexity of programming tasks and the lack of robust evaluation metrics that align well with human judgment. Traditional token-based metrics such as BLEU and ROUGE, while commonly used in natural language processing, exhibit weak correlations with human assessments in code intelligence and verification tasks. Furthermore, these metrics are primarily research focused and are not designed for seamless integration into the software development lifecycle, limiting their practical utility for developers seeking to improve code quality and security. AI-assisted coding has been shown to be more beneficial for senior developers, as they possess the expertise to critically evaluate the generated code for correctness, completeness, and compliance. In contrast, junior developers may struggle to identify hallucinations, missing functionality, or incorrect logic in AI-generated code. To bridge this gap, This paper introduces a novel scoring mechanism called the SBC score, which is based on a reverse generation technique that leverages the natural language generation capabilities of LLMs. Unlike direct code analysis, our approach reconstructs system requirements from AI-generated code and compares them with the original specifications to quantify accuracy. The SBC score combines semantic similarity, BLEU, and completeness analysis, providing actionable insights to developers by highlighting missing features and hallucinations. Our code and datasets are available on GitHub
Hallucinated facts in large language models (LLMs) have recently been shown to obey a statistical lower bound determined by the monofact rate (related to the classical Good-Turing missing mass estimator) minus model miscalibration (Kalai & Vempala, 2024). We present the first empirical investigation of this three-way relationship in classical n-gram models and fine-tuned encoder-decoder Transformers. By generating training data from Pareto distributions with varying shape parameters, we systematically control the monofact rates and establish its positive relationship with hallucination. To bridge theory and practice, we derive an empirical analog of the hallucination bound by replacing the population miscalibration term (Section 2.1) with an empirical bin-wise KL divergence and confirm its practical viability. We then introduce selective upweighting -- a simple yet effective technique that strategically repeats as little as 5% of training examples -- to deliberately inject miscalibration into the model. This intervention reduces hallucination by up to 40%, challenging universal deduplication policies. Our experiments reveal a critical trade-off: selective upweighting maintains pre-injection levels of accuracy while substantially reducing hallucination, whereas standard training gradually improves accuracy but fails to address persistently high hallucination, indicating an inherent tension in optimization objectives.
Speculative decoding has been widely used to accelerate auto-regressive (AR) text generation. However, its effectiveness for visual AR models remains limited due to token selection ambiguity, where multiple tokens share similarly low probabilities and thus reduce acceptance rates. Recently, relaxed speculative decoding with dynamic tree drafting was proposed to mitigate this ambiguity, demonstrating promising results in accelerating visual AR models. However, we observe that token selection ambiguity still negatively affects dynamic tree drafting, resulting in shallow draft trees and limited acceleration. To overcome this issue, we introduce LANTERN++, a refined framework that integrates static tree drafting with a tailored relaxed acceptance condition, allowing drafts to be selected independently of low-confidence predictions. This enables the acceptance of deeper sequences, improving decoding efficiency while preserving image quality. Extensive experiments on state-of-the-art visual AR models demonstrate that LANTERN++ significantly accelerates inference, achieving up to $\mathbf{\times 2.56}$ speedup over standard AR decoding while maintaining high image quality. The code is publicly available at https://github.com/jadohu/LANTERN.
To understand the complexity of sequence classification tasks, Hahn et al. (2021) proposed sensitivity as the number of disjoint subsets of the input sequence that can each be individually changed to change the output. Though effective, calculating sensitivity at scale using this framework is costly because of exponential time complexity. Therefore, we introduce a Sensitivity-based Multi-Armed Bandit framework (SMAB), which provides a scalable approach for calculating word-level local (sentence-level) and global (aggregated) sensitivities concerning an underlying text classifier for any dataset. We establish the effectiveness of our approach through various applications. We perform a case study on CHECKLIST generated sentiment analysis dataset where we show that our algorithm indeed captures intuitively high and low-sensitive words. Through experiments on multiple tasks and languages, we show that sensitivity can serve as a proxy for accuracy in the absence of gold data. Lastly, we show that guiding perturbation prompts using sensitivity values in adversarial example generation improves attack success rate by 15.58%, whereas using sensitivity as an additional reward in adversarial paraphrase generation gives a 12.00% improvement over SOTA approaches. Warning: Contains potentially offensive content.
Equipping large language models (LLMs) with latent-space memory has attracted increasing attention as they can extend the context window of existing language models. However, retaining information from the distant past remains a challenge. For example, MemoryLLM (Wang et al., 2024a), as a representative work with latent-space memory, compresses past information into hidden states across all layers, forming a memory pool of 1B parameters. While effective for sequence lengths up to 16k tokens, it struggles to retain knowledge beyond 20k tokens. In this work, we address this limitation by introducing M+, a memory-augmented model based on MemoryLLM that significantly enhances long-term information retention. M+ integrates a long-term memory mechanism with a co-trained retriever, dynamically retrieving relevant information during text generation. We evaluate M+ on diverse benchmarks, including long-context understanding and knowledge retention tasks. Experimental results show that M+ significantly outperforms MemoryLLM and recent strong baselines, extending knowledge retention from under 20k to over 160k tokens with similar GPU memory overhead.
Effective linguistic choices that attract potential customers play crucial roles in advertising success. This study aims to explore the linguistic features of ad texts that influence human preferences. Although the creation of attractive ad texts is an active area of research, progress in understanding the specific linguistic features that affect attractiveness is hindered by several obstacles. First, human preferences are complex and influenced by multiple factors, including their content, such as brand names, and their linguistic styles, making analysis challenging. Second, publicly available ad text datasets that include human preferences are lacking, such as ad performance metrics and human feedback, which reflect people's interests. To address these problems, we present AdParaphrase, a paraphrase dataset that contains human preferences for pairs of ad texts that are semantically equivalent but differ in terms of wording and style. This dataset allows for preference analysis that focuses on the differences in linguistic features. Our analysis revealed that ad texts preferred by human judges have higher fluency, longer length, more nouns, and use of bracket symbols. Furthermore, we demonstrate that an ad text-generation model that considers these findings significantly improves the attractiveness of a given text. The dataset is publicly available at: https://github.com/CyberAgentAILab/AdParaphrase.
Existing methods fail to effectively steer Large Language Models (LLMs) between textual reasoning and code generation, leaving symbolic computing capabilities underutilized. We introduce CodeSteer, an effective method for guiding LLM code/text generation. We construct a comprehensive benchmark SymBench comprising 37 symbolic tasks with adjustable complexity and also synthesize datasets of 12k multi-turn guidance/generation trajectories and 5.5k guidance comparison pairs. We fine-tune the Llama-3-8B model with a newly designed multi-turn supervised fine-tuning (SFT) and direct preference optimization (DPO). The resulting model, CodeSteerLLM, augmented with the proposed symbolic and self-answer checkers, effectively guides the code/text generation of larger models. Augmenting GPT-4o with CodeSteer raises its average performance score from 53.3 to 86.4, even outperforming the existing best LLM OpenAI o1 (82.7), o1-preview (74.8), and DeepSeek R1 (76.8) across all 37 tasks (28 seen, 9 unseen). Trained for GPT-4o, CodeSteer demonstrates superior generalizability, providing an average 41.8 performance boost on Claude, Mistral, and GPT-3.5. CodeSteer-guided LLMs fully harness symbolic computing to maintain strong performance on highly complex tasks. Models, Datasets, and Codes are available at https://github.com/yongchao98/CodeSteer-v1.0 and https://huggingface.co/yongchao98.
State-of-the-art visual generation models, such as Diffusion Models (DMs) and Vision Auto-Regressive Models (VARs), produce highly realistic images. While prior work has successfully mitigated Not Safe For Work (NSFW) content in the visual domain, we identify a novel threat: the generation of NSFW text embedded within images. This includes offensive language, such as insults, racial slurs, and sexually explicit terms, posing significant risks to users. We show that all state-of-the-art DMs (e.g., SD3, Flux, DeepFloyd IF) and VARs (e.g., Infinity) are vulnerable to this issue. Through extensive experiments, we demonstrate that existing mitigation techniques, effective for visual content, fail to prevent harmful text generation while substantially degrading benign text generation. As an initial step toward addressing this threat, we explore safety fine-tuning of the text encoder underlying major DM architectures using a customized dataset. Thereby, we suppress NSFW generation while preserving overall image and text generation quality. Finally, to advance research in this area, we introduce ToxicBench, an open-source benchmark for evaluating NSFW text generation in images. ToxicBench provides a curated dataset of harmful prompts, new metrics, and an evaluation pipeline assessing both NSFW-ness and generation quality. Our benchmark aims to guide future efforts in mitigating NSFW text generation in text-to-image models and is available at https://github.com/sprintml/ToxicBench
Retrieval, re-ranking, and retrieval-augmented generation (RAG) are critical components of modern applications in information retrieval, question answering, or knowledge-based text generation. However, existing solutions are often fragmented, lacking a unified framework that easily integrates these essential processes. The absence of a standardized implementation, coupled with the complexity of retrieval and re-ranking workflows, makes it challenging for researchers to compare and evaluate different approaches in a consistent environment. While existing toolkits such as Rerankers and RankLLM provide general-purpose reranking pipelines, they often lack the flexibility required for fine-grained experimentation and benchmarking. In response to these challenges, we introduce Rankify, a powerful and modular open-source toolkit designed to unify retrieval, re-ranking, and RAG within a cohesive framework. Rankify supports a wide range of retrieval techniques, including dense and sparse retrievers, while incorporating state-of-the-art re-ranking models to enhance retrieval quality. Additionally, Rankify includes a collection of pre-retrieved datasets to facilitate benchmarking, available at Huggingface (https://huggingface.co/datasets/abdoelsayed/reranking-datasets-light). To encourage adoption and ease of integration, we provide comprehensive documentation (http://rankify.readthedocs.io/), an open-source implementation on GitHub (https://github.com/DataScienceUIBK/rankify), and a PyPI package for easy installation (https://pypi.org/project/rankify/). As a unified and lightweight framework, Rankify allows researchers and practitioners to advance retrieval and re-ranking methodologies while ensuring consistency, scalability, and ease of use.
Which functions can be used as activations in deep neural networks? This article explores families of functions based on orthonormal bases, including the Hermite polynomial basis and the Fourier trigonometric basis, as well as a basis resulting from the tropicalization of a polynomial basis. Our study shows that, through simple variance-preserving initialization and without additional clamping mechanisms, these activations can successfully be used to train deep models, such as GPT-2 for next-token prediction on OpenWebText and ConvNeXt for image classification on ImageNet. Our work addresses the issue of exploding and vanishing activations and gradients, particularly prevalent with polynomial activations, and opens the door for improving the efficiency of large-scale learning tasks. Furthermore, our approach provides insight into the structure of neural networks, revealing that networks with polynomial activations can be interpreted as multivariate polynomial mappings. Finally, using Hermite interpolation, we show that our activations can closely approximate classical ones in pre-trained models by matching both the function and its derivative, making them especially useful for fine-tuning tasks. These activations are available in the torchortho library, which can be accessed via: https://github.com/K-H-Ismail/torchortho.
Large Language Models (LLMs) have revolutionized Natural Language Processing (NLP) and Artificial Intelligence (AI), unlocking unprecedented capabilities. This rapid advancement has spurred research into various aspects of LLMs, their text generation & reasoning capability, and potential misuse, fueling the necessity for robust detection methods. While numerous prior research has focused on detecting LLM-generated text (AI text) and thus checkmating them, our study investigates a relatively unexplored territory: portraying the nuanced distinctions between human and AI texts across text segments. Whether LLMs struggle with or excel at incorporating linguistic ingenuity across different text segments carries substantial implications for determining their potential as effective creative assistants to humans. Through an analogy with the structure of chess games-comprising opening, middle, and end games-we analyze text segments (introduction, body, and conclusion) to determine where the most significant distinctions between human and AI texts exist. While AI texts can approximate the body segment better due to its increased length, a closer examination reveals a pronounced disparity, highlighting the importance of this segment in AI text detection. Additionally, human texts exhibit higher cross-segment differences compared to AI texts. Overall, our research can shed light on the intricacies of human-AI text distinctions, offering novel insights for text detection and understanding.
Human evaluation is the gold-standard for evaluating text generation models. It is also expensive, and to fit budgetary constraints, a random subset of the test data is often chosen in practice. The randomly selected data may not accurately represent test performance, making this approach economically inefficient for model comparison. Thus, in this work, we develop a suite of selectors to get the most informative datapoints for human evaluation while taking the evaluation costs into account. We show that selectors based on variance in automated metric scores, diversity in model outputs, or Item Response Theory outperform random selection. We further develop an approach to distill these selectors to the scenario where the model outputs are not yet available. In particular, we introduce source-based estimators, which predict item usefulness for human evaluation just based on the source texts. We demonstrate the efficacy of our selectors in two common NLG tasks, machine translation and summarization, and show that up to only ~50% of the test data is needed to produce the same evaluation result as the entire data. Our implementations are published in the subset2evaluate package.
Aligning Large Language Models (LLMs) with human values and away from undesirable behaviors (such as hallucination) has become increasingly important. Recently, steering LLMs towards a desired behavior via activation editing has emerged as an effective method to mitigate harmful generations at inference-time. Activation editing modifies LLM representations by preserving information from positive demonstrations (e.g., truthful) and minimising information from negative demonstrations (e.g., hallucinations). When these demonstrations come from a private dataset, the aligned LLM may leak private information contained in those private samples. In this work, we present the first study of aligning LLM behavior with private datasets. Our work proposes the \textit{\underline{P}rivate \underline{S}teering for LLM \underline{A}lignment (PSA)} algorithm to edit LLM activations with differential privacy (DP) guarantees. We conduct extensive experiments on seven different benchmarks with open-source LLMs of different sizes (0.5B to 7B) and model families (LlaMa, Qwen, Mistral and Gemma). Our results show that PSA achieves DP guarantees for LLM alignment with minimal loss in performance, including alignment metrics, open-ended text generation quality, and general-purpose reasoning. We also develop the first Membership Inference Attack (MIA) for evaluating and auditing the empirical privacy for the problem of LLM steering via activation editing. Our attack is tailored for activation editing and relies solely on the generated texts without their associated probabilities. Our experiments support the theoretical guarantees by showing improved guarantees for our \textit{PSA} algorithm compared to several existing non-private techniques.
The rise of large language models (LLMs) has led to more diverse and higher-quality machine-generated text. However, their high expressive power makes it difficult to control outputs based on specific business instructions. In response, benchmarks focusing on the controllability of LLMs have been developed, but several issues remain: (1) They primarily cover major languages like English and Chinese, neglecting low-resource languages like Japanese; (2) Current benchmarks employ task-specific evaluation metrics, lacking a unified framework for selecting models based on controllability across different use cases. To address these challenges, this research introduces LCTG Bench, the first Japanese benchmark for evaluating the controllability of LLMs. LCTG Bench provides a unified framework for assessing control performance, enabling users to select the most suitable model for their use cases based on controllability. By evaluating nine diverse Japanese-specific and multilingual LLMs like GPT-4, we highlight the current state and challenges of controllability in Japanese LLMs and reveal the significant gap between multilingual models and Japanese-specific models.
Evaluating personalized text generated by large language models (LLMs) is challenging, as only the LLM user, i.e., prompt author, can reliably assess the output, but re-engaging the same individuals across studies is infeasible. This paper addresses the challenge of evaluating personalized text generation by introducing ExPerT, an explainable reference-based evaluation framework. ExPerT leverages an LLM to extract atomic aspects and their evidence from the generated and reference texts, match the aspects, and evaluate their alignment based on content and writing style -- two key attributes in personalized text generation. Additionally, ExPerT generates detailed, fine-grained explanations for every step of the evaluation process, enhancing transparency and interpretability. Our experiments demonstrate that ExPerT achieves a 7.2% relative improvement in alignment with human judgments compared to the state-of-the-art text generation evaluation methods. Furthermore, human evaluators rated the usability of ExPerT's explanations at 4.7 out of 5, highlighting its effectiveness in making evaluation decisions more interpretable.
Large language models (LLMs) have demonstrated immense potential across various tasks. However, research for exploring and improving the capabilities of LLMs in interpreting graph structures remains limited. To address this gap, we conduct a comprehensive evaluation of prompting current open-source LLMs on graph-to-text generation tasks. Although we explored the optimal prompting strategies and proposed a novel and effective diversity-difficulty-based few-shot sample selection method, we found that the improvements from tuning-free approaches were incremental, as LLMs struggle with planning on complex graphs, particularly those with a larger number of triplets. To further improve LLMs in planning with graph sequences and grounding in truth, we introduce a new graph-to-text dataset, PlanGTG, annotated with two sub-tasks: reordering and attribution. Through extensive automatic and human evaluations, we demonstrate significant improvements in the quality of generated text from both few-shot learning and fine-tuning perspectives using the PlanGTG dataset. Our study paves the way for new research directions in graph-to-text generation. PlanGTG datasets can be found in https://github.com/probe2/kg_text.
We present Pix2Cap-COCO, the first panoptic pixel-level caption dataset designed to advance fine-grained visual understanding. To achieve this, we carefully design an automated annotation pipeline that prompts GPT-4V to generate pixel-aligned, instance-specific captions for individual objects within images, enabling models to learn more granular relationships between objects and their contexts. This approach results in 167,254 detailed captions, with an average of 22.94 words per caption. Building on Pix2Cap-COCO, we introduce a novel task, panoptic segmentation-captioning, which challenges models to recognize instances in an image and provide detailed descriptions for each simultaneously. To benchmark this task, we design a robust baseline based on X-Decoder. The experimental results demonstrate that Pix2Cap-COCO is a particularly challenging dataset, as it requires models to excel in both fine-grained visual understanding and detailed language generation. Furthermore, we leverage Pix2Cap-COCO for Supervised Fine-Tuning (SFT) on large multimodal models (LMMs) to enhance their performance. For example, training with Pix2Cap-COCO significantly improves the performance of GPT4RoI, yielding gains in CIDEr +1.4%, ROUGE +0.4%, and SPICE +0.5% on Visual Genome dataset, and strengthens its region understanding ability on the ViP-BENCH, with an overall improvement of +5.1%, including notable increases in recognition accuracy +11.2% and language generation quality +22.2%.