The recent successes of large language models (LLMs) have led to a surge of theoretical research into language generation. A recent line of work proposes an abstract view, called language generation in the limit, where generation is seen as a game between an adversary and an algorithm: the adversary generates strings from an unknown language $K$, chosen from a countable collection of candidate languages, and after seeing a finite set of these strings, the algorithm must generate new strings from $K$ that it has not seen before. This formalism highlights a key tension: the trade-off between validity (the algorithm should only produce strings from the language) and breadth (it should be able to produce many strings from the language). This trade-off is central in applied language generation as well, where it appears as a balance between hallucination (generating invalid utterances) and mode collapse (generating only a restricted set of outputs). Despite its importance, this trade-off has been challenging to study quantitatively. We develop ways to quantify this trade-off by formalizing breadth using measures of density. Existing algorithms for language generation in the limit produce output sets that can have zero density in the true language, and this important failure of breadth might seem unavoidable. We show, however, that such a failure is not necessary: we provide an algorithm for language generation in the limit whose outputs have strictly positive density in $K$. We also study the internal representations built by these algorithms, specifically the sequence of hypothesized candidate languages they consider, and show that achieving the strongest form of breadth may require oscillating indefinitely between high- and low-density representations. Our analysis introduces a novel topology on language families, with notions of convergence and limit points playing a key role.
Zero-shot referring image segmentation aims to locate and segment the target region based on a referring expression, with the primary challenge of aligning and matching semantics across visual and textual modalities without training. Previous works address this challenge by utilizing Vision-Language Models and mask proposal networks for region-text matching. However, this paradigm may lead to incorrect target localization due to the inherent ambiguity and diversity of free-form referring expressions. To alleviate this issue, we present LGD (Leveraging Generative Descriptions), a framework that utilizes the advanced language generation capabilities of Multi-Modal Large Language Models to enhance region-text matching performance in Vision-Language Models. Specifically, we first design two kinds of prompts, the attribute prompt and the surrounding prompt, to guide the Multi-Modal Large Language Models in generating descriptions related to the crucial attributes of the referent object and the details of surrounding objects, referred to as attribute description and surrounding description, respectively. Secondly, three visual-text matching scores are introduced to evaluate the similarity between instance-level visual features and textual features, which determines the mask most associated with the referring expression. The proposed method achieves new state-of-the-art performance on three public datasets RefCOCO, RefCOCO+ and RefCOCOg, with maximum improvements of 9.97% in oIoU and 11.29% in mIoU compared to previous methods.
Retrieval-augmented generation (RAG) has become a widely adopted paradigm for enabling knowledge-grounded large language models (LLMs). However, standard RAG pipelines often fail to ensure that model reasoning remains consistent with the evidence retrieved, leading to factual inconsistencies or unsupported conclusions. In this work, we reinterpret RAG as Retrieval-Augmented Reasoning and identify a central but underexplored problem: \textit{Reasoning Misalignment}-the divergence between an LLM's internal reasoning trajectory and the evidential constraints provided by retrieval. To address this issue, we propose \textsc{AlignRAG}, a novel iterative framework grounded in Critique-Driven Alignment (CDA). At the heart of \textsc{AlignRAG} lies a \textit{contrastive critique synthesis} mechanism that generates retrieval-sensitive critiques while mitigating self-bias. This mechanism trains a dedicated retrieval-augmented \textit{Critic Language Model (CLM)} using labeled critiques that distinguish between evidence-aligned and misaligned reasoning. Alignment signals for supervision are obtained through self-supervised or externally guided labeling strategies. The resulting CLM is explicitly optimized for evidence sensitivity, enabling it to detect and revise reasoning errors during inference without relying solely on self-generated feedback. Empirical evaluations show that our 8B-parameter CLM improves performance over the Self-Refine baseline by 12.1\% on out-of-domain tasks and outperforms a standard 72B-parameter CLM by 2.2\%, while remaining compatible with existing RAG architectures as a plug-and-play module. Overall, AlignRAG offers a principled solution for aligning model reasoning with retrieved evidence, substantially improving the factual reliability and robustness of RAG systems.
Research on evaluating and analyzing large language models (LLMs) has been extensive for resource-rich languages such as English, yet their performance in languages such as Persian has received considerably less attention. This paper introduces FarsEval-PKBETS benchmark, a subset of FarsEval project for evaluating large language models in Persian. This benchmark consists of 4000 questions and answers in various formats, including multiple choice, short answer and descriptive responses. It covers a wide range of domains and tasks,including medicine, law, religion, Persian language, encyclopedic knowledge, human preferences, social knowledge, ethics and bias, text generation, and respecting others' rights. This bechmark incorporates linguistics, cultural, and local considerations relevant to the Persian language and Iran. To ensure the questions are challenging for current LLMs, three models -- Llama3-70B, PersianMind, and Dorna -- were evaluated using this benchmark. Their average accuracy was below 50%, meaning they provided fully correct answers to fewer than half of the questions. These results indicate that current language models are still far from being able to solve this benchmark
Dense retrieval is a crucial task in Information Retrieval (IR) and is the foundation for downstream tasks such as re-ranking. Recently, large language models (LLMs) have shown compelling semantic understanding capabilities and are appealing to researchers studying dense retrieval. LLMs, as decoder-style generative models, are competent at language generation while falling short on modeling global information due to the lack of attention to tokens afterward. Inspired by the classical word-based language modeling approach for IR, i.e., the query likelihood (QL) model, we seek to sufficiently utilize LLMs' generative ability by QL maximization. However, instead of ranking documents with QL estimation, we introduce an auxiliary task of QL maximization to yield a better backbone for contrastively learning a discriminative retriever. We name our model as LLM-QL. To condense global document semantics to a single vector during QL modeling, LLM-QL has two major components, Attention Stop (AS) and Input Corruption (IC). AS stops the attention of predictive tokens to previous tokens until the ending token of the document. IC masks a portion of tokens in the input documents during prediction. Experiments on MSMARCO show that LLM-QL can achieve significantly better performance than other LLM-based retrievers and using QL estimated by LLM-QL for ranking outperforms word-based QL by a large margin.
Increasing test-time computation has emerged as a promising direction for improving language model performance, particularly in scenarios where model finetuning is impractical or impossible due to computational constraints or private model weights. However, existing test-time search methods using a reward model (RM) often degrade in quality as compute scales, due to the over-optimization of what are inherently imperfect reward proxies. We introduce QAlign, a new test-time alignment approach. As we scale test-time compute, QAlign converges to sampling from the optimal aligned distribution for each individual prompt. By adopting recent advances in Markov chain Monte Carlo for text generation, our method enables better-aligned outputs without modifying the underlying model or even requiring logit access. We demonstrate the effectiveness of QAlign on mathematical reasoning benchmarks (GSM8K and GSM-Symbolic) using a task-specific RM, showing consistent improvements over existing test-time compute methods like best-of-n and majority voting. Furthermore, when applied with more realistic RMs trained on the Tulu 3 preference dataset, QAlign outperforms direct preference optimization (DPO), best-of-n, majority voting, and weighted majority voting on a diverse range of datasets (GSM8K, MATH500, IFEval, MMLU-Redux, and TruthfulQA). A practical solution to aligning language models at test time using additional computation without degradation, our approach expands the limits of the capability that can be obtained from off-the-shelf language models without further training.
We present FourierNAT, a novel non-autoregressive Transformer (NAT) architecture that employs Fourier-based mixing in the decoder to generate output sequences in parallel. While traditional NAT approaches often face challenges with capturing global dependencies, our method leverages a discrete Fourier transform to mix token embeddings across the entire sequence dimension, coupled with learned frequency-domain gating. This allows the model to efficiently propagate context without explicit autoregressive steps. Empirically, FourierNAT achieves competitive results against leading NAT baselines on standard benchmarks like WMT machine translation and CNN/DailyMail summarization, providing significant speed advantages over autoregressive Transformers. We further demonstrate that learned frequency-domain parameters allow the model to adaptively focus on long-range or short-range dependencies, partially mitigating the well-known coherence gaps in one-pass NAT generation. Overall, FourierNAT highlights the potential of integrating spectral-domain operations to accelerate and improve parallel text generation. This approach can potentially provide great computational and time savings in inference tasks LLMs.
The rapid development of Large Language Models (LLMs) has intensified concerns about content traceability and potential misuse. Existing watermarking schemes for sampled text often face trade-offs between maintaining text quality and ensuring robust detection against various attacks. To address these issues, we propose a novel watermarking scheme that improves both detectability and text quality by introducing a cumulative watermark entropy threshold. Our approach is compatible with and generalizes existing sampling functions, enhancing adaptability. Experimental results across multiple LLMs show that our scheme significantly outperforms existing methods, achieving over 80\% improvements on widely-used datasets, e.g., MATH and GSM8K, while maintaining high detection accuracy.
Personalization with retrieval-augmented generation (RAG) often fails to capture fine-grained features of authors, making it hard to identify their unique traits. To enrich the RAG context, we propose providing Large Language Models (LLMs) with author-specific features, such as average sentiment polarity and frequently used words, in addition to past samples from the author's profile. We introduce a new feature called Contrastive Examples: documents from other authors are retrieved to help LLM identify what makes an author's style unique in comparison to others. Our experiments show that adding a couple of sentences about the named entities, dependency patterns, and words a person uses frequently significantly improves personalized text generation. Combining features with contrastive examples boosts the performance further, achieving a relative 15% improvement over baseline RAG while outperforming the benchmarks. Our results show the value of fine-grained features for better personalization, while opening a new research dimension for including contrastive examples as a complement with RAG. We release our code publicly.
Generating novel and creative scientific hypotheses is a cornerstone in achieving Artificial General Intelligence. Large language and reasoning models have the potential to aid in the systematic creation, selection, and validation of scientifically informed hypotheses. However, current foundation models often struggle to produce scientific ideas that are both novel and feasible. One reason is the lack of a dedicated dataset that frames Scientific Hypothesis Generation (SHG) as a Natural Language Generation (NLG) task. In this paper, we introduce HypoGen, the first dataset of approximately 5500 structured problem-hypothesis pairs extracted from top-tier computer science conferences structured with a Bit-Flip-Spark schema, where the Bit is the conventional assumption, the Spark is the key insight or conceptual leap, and the Flip is the resulting counterproposal. HypoGen uniquely integrates an explicit Chain-of-Reasoning component that reflects the intellectual process from Bit to Flip. We demonstrate that framing hypothesis generation as conditional language modelling, with the model fine-tuned on Bit-Flip-Spark and the Chain-of-Reasoning (and where, at inference, we only provide the Bit), leads to improvements in the overall quality of the hypotheses. Our evaluation employs automated metrics and LLM judge rankings for overall quality assessment. We show that by fine-tuning on our HypoGen dataset we improve the novelty, feasibility, and overall quality of the generated hypotheses. The HypoGen dataset is publicly available at huggingface.co/datasets/UniverseTBD/hypogen-dr1.
Recent advances in Large Language Models have demonstrated their capabilities across a variety of tasks. However, automatically extracting implicit knowledge from natural language remains a significant challenge, as machines lack active experience with the physical world. Given this scenario, semantic knowledge graphs can serve as conceptual spaces that guide the automated text generation reasoning process to achieve more efficient and explainable results. In this paper, we apply a logic-augmented generation (LAG) framework that leverages the explicit representation of a text through a semantic knowledge graph and applies it in combination with prompt heuristics to elicit implicit analogical connections. This method generates extended knowledge graph triples representing implicit meaning, enabling systems to reason on unlabeled multimodal data regardless of the domain. We validate our work through three metaphor detection and understanding tasks across four datasets, as they require deep analogical reasoning capabilities. The results show that this integrated approach surpasses current baselines, performs better than humans in understanding visual metaphors, and enables more explainable reasoning processes, though still has inherent limitations in metaphor understanding, especially for domain-specific metaphors. Furthermore, we propose a thorough error analysis, discussing issues with metaphorical annotations and current evaluation methods.
Recent advances in reasoning-focused large language models (LLMs) mark a shift from general LLMs toward models designed for complex decision-making, a crucial aspect in medicine. However, their performance in specialized domains like ophthalmology remains underexplored. This study comprehensively evaluated and compared the accuracy and reasoning capabilities of four newly developed reasoning-focused LLMs, namely DeepSeek-R1, OpenAI o1, o3-mini, and Gemini 2.0 Flash-Thinking. Each model was assessed using 5,888 multiple-choice ophthalmology exam questions from the MedMCQA dataset in zero-shot setting. Quantitative evaluation included accuracy, Macro-F1, and five text-generation metrics (ROUGE-L, METEOR, BERTScore, BARTScore, and AlignScore), computed against ground-truth reasonings. Average inference time was recorded for a subset of 100 randomly selected questions. Additionally, two board-certified ophthalmologists qualitatively assessed clarity, completeness, and reasoning structure of responses to differential diagnosis questions.O1 (0.902) and DeepSeek-R1 (0.888) achieved the highest accuracy, with o1 also leading in Macro-F1 (0.900). The performance of models across the text-generation metrics varied: O3-mini excelled in ROUGE-L (0.151), o1 in METEOR (0.232), DeepSeek-R1 and o3-mini tied for BERTScore (0.673), DeepSeek-R1 (-4.105) and Gemini 2.0 Flash-Thinking (-4.127) performed best in BARTScore, while o3-mini (0.181) and o1 (0.176) led AlignScore. Inference time across the models varied, with DeepSeek-R1 being slowest (40.4 seconds) and Gemini 2.0 Flash-Thinking fastest (6.7 seconds). Qualitative evaluation revealed that DeepSeek-R1 and Gemini 2.0 Flash-Thinking tended to provide detailed and comprehensive intermediate reasoning, whereas o1 and o3-mini displayed concise and summarized justifications.
Vision-language models (VLMs) show promise for autonomous driving but often lack transparent reasoning capabilities that are critical for safety. We investigate whether explicitly modeling reasoning during fine-tuning enhances VLM performance on driving decision tasks. Using GPT-4o, we generate structured reasoning chains for driving scenarios from the DriveLM benchmark with category-specific prompting strategies. We compare reasoning-based fine-tuning, answer-only fine-tuning, and baseline instruction-tuned models across multiple small VLM families (Llama 3.2, Llava 1.5, and Qwen 2.5VL). Our results demonstrate that reasoning-based fine-tuning consistently outperforms alternatives, with Llama3.2-11B-reason achieving the highest performance. Models fine-tuned with reasoning show substantial improvements in accuracy and text generation quality, suggesting explicit reasoning enhances internal representations for driving decisions. These findings highlight the importance of transparent decision processes in safety-critical domains and offer a promising direction for developing more interpretable autonomous driving systems.
An agent's intention often remains hidden behind the black-box nature of embodied policies. Communication using natural language statements that describe the next action can provide transparency towards the agent's behavior. We aim to insert transparent behavior directly into the learning process, by transforming the problem of policy learning into a language generation problem and combining it with traditional autoregressive modelling. The resulting model produces transparent natural language statements followed by tokens representing the specific actions to solve long-horizon tasks in the Language-Table environment. Following previous work, the model is able to learn to produce a policy represented by special discretized tokens in an autoregressive manner. We place special emphasis on investigating the relationship between predicting actions and producing high-quality language for a transparent agent. We find that in many cases both the quality of the action trajectory and the transparent statement increase when they are generated simultaneously.
The landscape of image generation has rapidly evolved, from early GAN-based approaches to diffusion models and, most recently, to unified generative architectures that seek to bridge understanding and generation tasks. Recent advances, especially the GPT-4o, have demonstrated the feasibility of high-fidelity multimodal generation, their architectural design remains mysterious and unpublished. This prompts the question of whether image and text generation have already been successfully integrated into a unified framework for those methods. In this work, we conduct an empirical study of GPT-4o's image generation capabilities, benchmarking it against leading open-source and commercial models. Our evaluation covers four main categories, including text-to-image, image-to-image, image-to-3D, and image-to-X generation, with more than 20 tasks. Our analysis highlights the strengths and limitations of GPT-4o under various settings, and situates it within the broader evolution of generative modeling. Through this investigation, we identify promising directions for future unified generative models, emphasizing the role of architectural design and data scaling. For a high-definition version of the PDF, please refer to the link on GitHub: \href{https://github.com/Ephemeral182/Empirical-Study-of-GPT-4o-Image-Gen}{https://github.com/Ephemeral182/Empirical-Study-of-GPT-4o-Image-Gen}.
For high-quality texts, single-score metrics seldom provide actionable feedback. In contrast, span annotation - pointing out issues in the text by annotating their spans - can guide improvements and provide insights. Until recently, span annotation was limited to human annotators or fine-tuned encoder models. In this study, we automate span annotation with large language models (LLMs). We compare expert or skilled crowdworker annotators with open and proprietary LLMs on three tasks: data-to-text generation evaluation, machine translation evaluation, and propaganda detection in human-written texts. In our experiments, we show that LLMs as span annotators are straightforward to implement and notably more cost-efficient than human annotators. The LLMs achieve moderate agreement with skilled human annotators, in some scenarios comparable to the average agreement among the annotators themselves. Qualitative analysis shows that reasoning models outperform their instruction-tuned counterparts and provide more valid explanations for annotations. We release the dataset of more than 40k model and human annotations for further research.
Reasoning-enabled large language models (LLMs) have recently demonstrated impressive performance in complex logical and mathematical tasks, yet their effectiveness in evaluating natural language generation remains unexplored. This study systematically compares reasoning-based LLMs (DeepSeek-R1 and OpenAI o3) with their non-reasoning counterparts across machine translation (MT) and text summarization (TS) evaluation tasks. We evaluate eight models across three architectural categories, including state-of-the-art reasoning models, their distilled variants (ranging from 8B to 70B parameters), and equivalent conventional, non-reasoning LLMs. Our experiments on WMT23 and SummEval benchmarks reveal that the benefits of reasoning capabilities are highly model and task-dependent: while OpenAI o3-mini models show consistent performance improvements with increased reasoning intensity, DeepSeek-R1 underperforms compared to its non-reasoning variant, with exception to certain aspects of TS evaluation. Correlation analysis demonstrates that increased reasoning token usage positively correlates with evaluation quality in o3-mini models. Furthermore, our results show that distillation of reasoning capabilities maintains reasonable performance in medium-sized models (32B) but degrades substantially in smaller variants (8B). This work provides the first comprehensive assessment of reasoning LLMs for NLG evaluation and offers insights into their practical use.
We present MedHal, a novel large-scale dataset specifically designed to evaluate if models can detect hallucinations in medical texts. Current hallucination detection methods face significant limitations when applied to specialized domains like medicine, where they can have disastrous consequences. Existing medical datasets are either too small, containing only a few hundred samples, or focus on a single task like Question Answering or Natural Language Inference. MedHal addresses these gaps by: (1) incorporating diverse medical text sources and tasks; (2) providing a substantial volume of annotated samples suitable for training medical hallucination detection models; and (3) including explanations for factual inconsistencies to guide model learning. We demonstrate MedHal's utility by training and evaluating a baseline medical hallucination detection model, showing improvements over general-purpose hallucination detection approaches. This resource enables more efficient evaluation of medical text generation systems while reducing reliance on costly expert review, potentially accelerating the development of medical AI research.
Advancements in emotion aware language processing increasingly shape vital NLP applications ranging from conversational AI and affective computing to computational psychology and creative content generation. Existing emotion datasets either lack emotional granularity or fail to capture necessary stylistic diversity, limiting the advancement of effective emotion conditioned text generation systems. Seeking to bridge this crucial gap between granularity and style diversity, this paper introduces a novel systematically constructed dataset named ELSA Emotion and Language Style Alignment Dataset leveraging fine grained emotion taxonomies adapted from existing sources such as dair ai emotion dataset and GoEmotions taxonomy. This dataset comprises multiple emotionally nuanced variations of original sentences regenerated across distinct contextual styles such as conversational, formal, poetic, and narrative, using advanced Large Language Models LLMs. Rigorous computational evaluation using metrics such as perplexity, embedding variance, readability, lexical diversity, and semantic coherence measures validates the datasets emotional authenticity, linguistic fluency, and textual diversity. Comprehensive metric analyses affirm its potential to support deeper explorations into emotion conditioned style adaptive text generation. By enabling precision tuned emotionally nuanced language modeling, our dataset creates fertile ground for research on fine grained emotional control, prompt driven explanation, interpretability, and style adaptive expressive language generation with LLMs.
Recently, the personalization of Large Language Models (LLMs) to generate content that aligns with individual user preferences has garnered widespread attention. Personalized Retrieval-Augmented Generation (RAG), which retrieves relevant documents from the user's history to reflect their preferences and enhance LLM generation, is one commonly used approach for personalization. However, existing personalized RAG methods do not consider that the histories of similar users can also assist in personalized generation for the current user, meaning that collaborative information between users can also benefit personalized generation. Inspired by the application of collaborative filtering in recommender systems, we propose a method called CFRAG, which adapts Collaborative Filtering to RAG for personalized text generation. However, this presents two challenges: (1)~how to incorporate collaborative information without explicit user similarity labels? (2)~how to retrieve documents that support personalized LLM generation? For Challenge 1, we use contrastive learning to train user embeddings to retrieve similar users and introduce collaborative information. For Challenge 2, we design a personalized retriever and reranker to retrieve the top-$k$ documents from these users' histories. We take into account the user's preference during retrieval and reranking. Then we leverage feedback from the LLM to fine-tune the personalized retriever and reranker, enabling them to retrieve documents that meet the personalized generation needs of the LLM. Experimental results on the Language Model Personalization (LaMP) benchmark validate the effectiveness of CFRAG. Further analysis confirms the importance of incorporating collaborative information.