Long-form text generation remains a significant challenge for large language models (LLMs), particularly in maintaining coherence, ensuring logical consistency, and preserving text quality as sequence length increases. To address these limitations, we propose SuperWriter-Agent, an agent-based framework designed to enhance the quality and consistency of long-form text generation. SuperWriter-Agent introduces explicit structured thinking-through planning and refinement stages into the generation pipeline, guiding the model to follow a more deliberate and cognitively grounded process akin to that of a professional writer. Based on this framework, we construct a supervised fine-tuning dataset to train a 7B SuperWriter-LM. We further develop a hierarchical Direct Preference Optimization (DPO) procedure that uses Monte Carlo Tree Search (MCTS) to propagate final quality assessments and optimize each generation step accordingly. Empirical results across diverse benchmarks demonstrate that SuperWriter-LM achieves state-of-the-art performance, surpassing even larger-scale baseline models in both automatic evaluation and human evaluation. Furthermore, comprehensive ablation studies demonstrate the effectiveness of hierarchical DPO and underscore the value of incorporating structured thinking steps to improve the quality of long-form text generation.
Recent advances in Large Language Models (LLMs) have enabled strong performance in long-form writing, yet existing supervised fine-tuning (SFT) approaches suffer from limitations such as data saturation and restricted learning capacity bounded by teacher signals. In this work, we present Writing-RL: an Adaptive Curriculum Reinforcement Learning framework to advance long-form writing capabilities beyond SFT. The framework consists of three key components: Margin-aware Data Selection strategy that prioritizes samples with high learning potential, Pairwise Comparison Reward mechanism that provides discriminative learning signals in the absence of verifiable rewards, and Dynamic Reference Scheduling approach, which plays a particularly critical role by adaptively adjusting task difficulty based on evolving model performance. Experiments on 7B-scale writer models show that our RL framework largely improves long-form writing performance over strong SFT baselines. Furthermore, we observe that models trained with long-output RL generalize surprisingly well to long-input reasoning tasks, potentially offering a promising perspective for rethinking long-context training.
We introduce Frankentexts, a new type of long-form narratives produced by LLMs under the extreme constraint that most tokens (e.g., 90%) must be copied verbatim from human writings. This task presents a challenging test of controllable generation, requiring models to satisfy a writing prompt, integrate disparate text fragments, and still produce a coherent narrative. To generate Frankentexts, we instruct the model to produce a draft by selecting and combining human-written passages, then iteratively revise the draft while maintaining a user-specified copy ratio. We evaluate the resulting Frankentexts along three axes: writing quality, instruction adherence, and detectability. Gemini-2.5-Pro performs surprisingly well on this task: 81% of its Frankentexts are coherent and 100% relevant to the prompt. Notably, up to 59% of these outputs are misclassified as human-written by detectors like Pangram, revealing limitations in AI text detectors. Human annotators can sometimes identify Frankentexts through their abrupt tone shifts and inconsistent grammar between segments, especially in longer generations. Beyond presenting a challenging generation task, Frankentexts invite discussion on building effective detectors for this new grey zone of authorship, provide training data for mixed authorship detection, and serve as a sandbox for studying human-AI co-writing processes.
Large language models (LLMs) are widely used for long-form text generation. However, factual errors in the responses would undermine their reliability. Despite growing attention to LLM factuality, the effect of response length on factuality remains underexplored. In this work, we systematically investigate this relationship by first introducing an automatic and bi-level long-form factuality evaluation framework, which achieves high agreement with human annotations while being cost-effective. Using this framework, we conduct controlled experiments and find that longer responses exhibit lower factual precision, confirming the presence of length bias. To explain this phenomenon, we empirically examine three hypotheses: error propagation, long context, and facts exhaustion. Our results reveal that facts exhaustion, where the model gradually exhausts more reliable knowledge, is the primary cause of factual degradation, rather than the other two hypotheses.
While Large Language Models (LLMs) can generate fluent and convincing responses, they are not necessarily correct. This is especially apparent in the popular decompose-then-verify factuality evaluation pipeline, where LLMs evaluate generations by decomposing the generations into individual, valid claims. Factuality evaluation is especially important for medical answers, since incorrect medical information could seriously harm the patient. However, existing factuality systems are a poor match for the medical domain, as they are typically only evaluated on objective, entity-centric, formulaic texts such as biographies and historical topics. This differs from condition-dependent, conversational, hypothetical, sentence-structure diverse, and subjective medical answers, which makes decomposition into valid facts challenging. We propose MedScore, a new approach to decomposing medical answers into condition-aware valid facts. Our method extracts up to three times more valid facts than existing methods, reducing hallucination and vague references, and retaining condition-dependency in facts. The resulting factuality score significantly varies by decomposition method, verification corpus, and used backbone LLM, highlighting the importance of customizing each step for reliable factuality evaluation.
Metrics like FactScore and VeriScore that evaluate long-form factuality operate by decomposing an input response into atomic claims and then individually verifying each claim. While effective and interpretable, these methods incur numerous LLM calls and can take upwards of 100 seconds to evaluate a single response, limiting their practicality in large-scale evaluation and training scenarios. To address this, we propose VeriFastScore, which leverages synthetic data to fine-tune Llama3.1 8B for simultaneously extracting and verifying all verifiable claims within a given text based on evidence from Google Search. We show that this task cannot be solved via few-shot prompting with closed LLMs due to its complexity: the model receives ~4K tokens of evidence on average and needs to concurrently decompose claims, judge their verifiability, and verify them against noisy evidence. However, our fine-tuned VeriFastScore model demonstrates strong correlation with the original VeriScore pipeline at both the example level (r=0.80) and system level (r=0.94) while achieving an overall speedup of 6.6x (9.9x excluding evidence retrieval) over VeriScore. To facilitate future factuality research, we publicly release our VeriFastScore model and synthetic datasets.
Teaching large language models (LLMs) to be faithful in the provided context is crucial for building reliable information-seeking systems. Therefore, we propose a systematic framework, CANOE, to improve the faithfulness of LLMs in both short-form and long-form generation tasks without human annotations. Specifically, we first synthesize short-form question-answering (QA) data with four diverse tasks to construct high-quality and easily verifiable training data without human annotation. Also, we propose Dual-GRPO, a rule-based reinforcement learning method that includes three tailored rule-based rewards derived from synthesized short-form QA data, while simultaneously optimizing both short-form and long-form response generation. Notably, Dual-GRPO eliminates the need to manually label preference data to train reward models and avoids over-optimizing short-form generation when relying only on the synthesized short-form QA data. Experimental results show that CANOE greatly improves the faithfulness of LLMs across 11 different downstream tasks, even outperforming the most advanced LLMs, e.g., GPT-4o and OpenAI o1.
Large Language Models (LLMs) frequently produce factoid hallucinations - plausible yet incorrect answers. A common mitigation strategy is model alignment, which improves factual accuracy by training on curated factual and non-factual pairs. However, this approach often relies on a stronger model (e.g., GPT-4) or an external knowledge base to assess factual correctness, which may not always be accessible. To address this, we propose Atomic Consistency Preference Optimization (ACPO), a self-supervised preference-tuning method that enhances factual accuracy without external supervision. ACPO leverages atomic consistency signals, i.e., the agreement of individual facts across multiple stochastic responses, to identify high- and low-quality data pairs for model alignment. By eliminating the need for costly GPT calls, ACPO provides a scalable and efficient approach to improving factoid question-answering. Despite being self-supervised, empirical results demonstrate that ACPO outperforms FactAlign, a strong supervised alignment baseline, by 1.95 points on the LongFact and BioGen datasets, highlighting its effectiveness in enhancing factual reliability without relying on external models or knowledge bases.
We employ a unified framework for computing the information capacity of biological signaling systems using Fisher Information. By deriving closed-form or easily computable information capacity formulas, we quantify how well different signaling models, including binomial, multinomial, Poisson, Gaussian, and Gamma distributions, can discriminate among input signals. These expressions clarify how key features such as signal range, noise scaling, pathway length, and receivers' diversity shape the theoretical limits of sensing. In particular, we show how signal-to-noise ratio and fold-change sensitivity arise naturally within the Fisher formalism, and how signal degradation in cascades imposes linear information loss. Our results provide intuitive, analytically grounded tools to benchmark and guide the analysis of real signaling systems, without requiring computationally expensive mutual information estimation. While motivated by cellular communication, the framework generalizes to any system where noisy input-output relationships constrain transmission fidelity, including synthetic biology, sensor networks, and engineered communication channels.
Despite recent advances in video understanding, the capabilities of Large Video Language Models (LVLMs) to perform video-based causal reasoning remains underexplored, largely due to the absence of relevant and dedicated benchmarks for evaluating causal reasoning in visually grounded and goal-driven settings. To fill this gap, we introduce a novel benchmark named Video-based long-form Causal Reasoning (VCRBench). We create VCRBench using procedural videos of simple everyday activities, where the steps are deliberately shuffled with each clip capturing a key causal event, to test whether LVLMs can identify, reason about, and correctly sequence the events needed to accomplish a specific goal. Moreover, the benchmark is carefully designed to prevent LVLMs from exploiting linguistic shortcuts, as seen in multiple-choice or binary QA formats, while also avoiding the challenges associated with evaluating open-ended QA. Our evaluation of state-of-the-art LVLMs on VCRBench suggests that these models struggle with video-based long-form causal reasoning, primarily due to their difficulty in modeling long-range causal dependencies directly from visual observations. As a simple step toward enabling such capabilities, we propose Recognition-Reasoning Decomposition (RRD), a modular approach that breaks video-based causal reasoning into two sub-tasks of video recognition and causal reasoning. Our experiments on VCRBench show that RRD significantly boosts accuracy on VCRBench, with gains of up to 25.2%. Finally, our thorough analysis reveals interesting insights, for instance, that LVLMs primarily rely on language knowledge for complex video-based long-form causal reasoning tasks.
\Ac{LFQA} aims to generate lengthy answers to complex questions. This scenario presents great flexibility as well as significant challenges for evaluation. Most evaluations rely on deterministic metrics that depend on string or n-gram matching, while the reliability of large language model-based evaluations for long-form answers remains relatively unexplored. We address this gap by conducting an in-depth study of long-form answer evaluation with the following research questions: (i) To what extent do existing automatic evaluation metrics serve as a substitute for human evaluations? (ii) What are the limitations of existing evaluation metrics compared to human evaluations? (iii) How can the effectiveness and robustness of existing evaluation methods be improved? We collect 5,236 factoid and non-factoid long-form answers generated by different large language models and conduct a human evaluation on 2,079 of them, focusing on correctness and informativeness. Subsequently, we investigated the performance of automatic evaluation metrics by evaluating these answers, analyzing the consistency between these metrics and human evaluations. We find that the style, length of the answers, and the category of questions can bias the automatic evaluation metrics. However, fine-grained evaluation helps mitigate this issue on some metrics. Our findings have important implications for the use of large language models for evaluating long-form question answering. All code and datasets are available at https://github.com/bugtig6351/lfqa_evaluation.
Kernel mean embeddings -- integrals of a kernel with respect to a probability distribution -- are essential in Bayesian quadrature, but also widely used in other computational tools for numerical integration or for statistical inference based on the maximum mean discrepancy. These methods often require, or are enhanced by, the availability of a closed-form expression for the kernel mean embedding. However, deriving such expressions can be challenging, limiting the applicability of kernel-based techniques when practitioners do not have access to a closed-form embedding. This paper addresses this limitation by providing a comprehensive dictionary of known kernel mean embeddings, along with practical tools for deriving new embeddings from known ones. We also provide a Python library that includes minimal implementations of the embeddings.
Artificial Kuramoto oscillatory neurons were recently introduced as an alternative to threshold units. Empirical evidence suggests that oscillatory units outperform threshold units in several tasks including unsupervised object discovery and certain reasoning problems. The proposed coupling mechanism for these oscillatory neurons is heterogeneous, combining a generalized Kuramoto equation with standard coupling methods used for threshold units. In this research note, we present a theoretical framework that clearly distinguishes oscillatory neurons from threshold units and establishes a coupling mechanism between them. We argue that, from a biological standpoint, oscillatory and threshold units realise distinct aspects of neural coding: roughly, threshold units model intensity of neuron firing, while oscillatory units facilitate information exchange by frequency modulation. To derive interaction between these two types of units, we constrain their dynamics by focusing on dynamical systems that admit Lyapunov functions. For threshold units, this leads to Hopfield associative memory model, and for oscillatory units it yields a specific form of generalized Kuramoto model. The resulting dynamical systems can be naturally coupled to form a Hopfield-Kuramoto associative memory model, which also admits a Lyapunov function. Various forms of coupling are possible. Notably, oscillatory neurons can be employed to implement a low-rank correction to the weight matrix of a Hopfield network. This correction can be viewed either as a form of Hebbian learning or as a popular LoRA method used for fine-tuning of large language models. We demonstrate the practical realization of this particular coupling through illustrative toy experiments.
Process reward models (PRMs) have proven effective for test-time scaling of Large Language Models (LLMs) on challenging reasoning tasks. However, reward hacking issues with PRMs limit their successful application in reinforcement fine-tuning. In this paper, we identify the main cause of PRM-induced reward hacking: the canonical summation-form credit assignment in reinforcement learning (RL), which defines the value as cumulative gamma-decayed future rewards, easily induces LLMs to hack steps with high rewards. To address this, we propose PURE: Process sUpervised Reinforcement lEarning. The key innovation of PURE is a min-form credit assignment that formulates the value function as the minimum of future rewards. This method significantly alleviates reward hacking by limiting the value function range and distributing advantages more reasonably. Through extensive experiments on 3 base models, we show that PRM-based approaches enabling min-form credit assignment achieve comparable reasoning performance to verifiable reward-based methods within only 30% steps. In contrast, the canonical sum-form credit assignment collapses training even at the beginning! Additionally, when we supplement PRM-based fine-tuning with just 10% verifiable rewards, we further alleviate reward hacking and produce the best fine-tuned model based on Qwen2.5-Math-7B in our experiments, achieving 82.5% accuracy on AMC23 and 53.3% average accuracy across 5 benchmarks. Moreover, we summarize the observed reward hacking cases and analyze the causes of training collapse. Code and models are available at https://github.com/CJReinforce/PURE.
This paper presents a review for the NTIRE 2025 Challenge on Short-form UGC Video Quality Assessment and Enhancement. The challenge comprises two tracks: (i) Efficient Video Quality Assessment (KVQ), and (ii) Diffusion-based Image Super-Resolution (KwaiSR). Track 1 aims to advance the development of lightweight and efficient video quality assessment (VQA) models, with an emphasis on eliminating reliance on model ensembles, redundant weights, and other computationally expensive components in the previous IQA/VQA competitions. Track 2 introduces a new short-form UGC dataset tailored for single image super-resolution, i.e., the KwaiSR dataset. It consists of 1,800 synthetically generated S-UGC image pairs and 1,900 real-world S-UGC images, which are split into training, validation, and test sets using a ratio of 8:1:1. The primary objective of the challenge is to drive research that benefits the user experience of short-form UGC platforms such as Kwai and TikTok. This challenge attracted 266 participants and received 18 valid final submissions with corresponding fact sheets, significantly contributing to the progress of short-form UGC VQA and image superresolution. The project is publicly available at https://github.com/lixinustc/KVQE- ChallengeCVPR-NTIRE2025.
This paper presents a Bayesian estimation framework for Wiener models, focusing on learning nonlinear output functions under known linear state dynamics. We derive a closed-form optimal affine estimator for the unknown parameters, characterized by the so-called "dynamic basis statistics" (DBS). Several features of the proposed estimator are studied, including Bayesian unbiasedness, closed-form posterior statistics, error monotonicity in trajectory length, and consistency condition (also known as persistent excitation). In the special case of Fourier basis functions, we demonstrate that the closed-form description is computationally available, as the Fourier DBS enjoys explicit expressions. Furthermore, we identify an inherent inconsistency in the Fourier bases for single-trajectory measurements, regardless of the input excitation. Leveraging the closed-form estimation error, we develop an active learning algorithm synthesizing input signals to minimize estimation error. Numerical experiments validate the efficacy of our approach, showing significant improvements over traditional regularized least-squares methods.
Long-form generation is crucial for a wide range of practical applications, typically categorized into short-to-long and long-to-long generation. While short-to-long generations have received considerable attention, generating long texts from extremely long resources remains relatively underexplored. The primary challenge in long-to-long generation lies in effectively integrating and analyzing relevant information from extensive inputs, which remains difficult for current large language models (LLMs). In this paper, we propose LLM$\times$MapReduce-V2, a novel test-time scaling strategy designed to enhance the ability of LLMs to process extremely long inputs. Drawing inspiration from convolutional neural networks, which iteratively integrate local features into higher-level global representations, LLM$\times$MapReduce-V2 utilizes stacked convolutional scaling layers to progressively expand the understanding of input materials. Both quantitative and qualitative experimental results demonstrate that our approach substantially enhances the ability of LLMs to process long inputs and generate coherent, informative long-form articles, outperforming several representative baselines. Both LLM$\times$MapReduce-V2 and SurveyEval are publicly available at https://github.com/thunlp/LLMxMapReduce .
Efficient understanding of long-form videos remains a significant challenge in computer vision. In this work, we revisit temporal search paradigms for long-form video understanding, studying a fundamental issue pertaining to all state-of-the-art (SOTA) long-context vision-language models (VLMs). In particular, our contributions are two-fold: First, we formulate temporal search as a Long Video Haystack problem, i.e., finding a minimal set of relevant frames (typically one to five) among tens of thousands of frames from real-world long videos given specific queries. To validate our formulation, we create LV-Haystack, the first benchmark containing 3,874 human-annotated instances with fine-grained evaluation metrics for assessing keyframe search quality and computational efficiency. Experimental results on LV-Haystack highlight a significant research gap in temporal search capabilities, with SOTA keyframe selection methods achieving only 2.1% temporal F1 score on the LVBench subset. Next, inspired by visual search in images, we re-think temporal searching and propose a lightweight keyframe searching framework, T*, which casts the expensive temporal search as a spatial search problem. T* leverages superior visual localization capabilities typically used in images and introduces an adaptive zooming-in mechanism that operates across both temporal and spatial dimensions. Our extensive experiments show that when integrated with existing methods, T* significantly improves SOTA long-form video understanding performance. Specifically, under an inference budget of 32 frames, T* improves GPT-4o's performance from 50.5% to 53.1% and LLaVA-OneVision-72B's performance from 56.5% to 62.4% on LongVideoBench XL subset. Our PyTorch code, benchmark dataset and models are included in the Supplementary material.
Generating high-quality stories spanning thousands of tokens requires competency across a variety of skills, from tracking plot and character arcs to keeping a consistent and engaging style. Due to the difficulty of sourcing labeled datasets and precise quality measurements, most work using large language models (LLMs) for long-form story generation uses combinations of hand-designed prompting techniques to elicit author-like behavior. This is a manual process that is highly dependent on the specific story-generation task. Motivated by the recent success of applying RL with Verifiable Rewards to domains like math and coding, we propose a general story-generation task (Next-Chapter Prediction) and a reward formulation (Verified Rewards via Completion Likelihood Improvement) that allows us to use an unlabeled book dataset as a learning signal for reasoning. We learn to reason over a story's condensed information and generate a detailed plan for the next chapter. Our reasoning is evaluated via the chapters it helps a story-generator create, and compared against non-trained and supervised finetuning (SFT) baselines. Pairwise human judgments reveal the chapters our learned reasoning produces are preferred across almost all metrics, and the effect is more pronounced in Scifi and Fantasy genres.
As AI models tackle increasingly complex problems, ensuring reliable human oversight becomes more challenging due to the difficulty of verifying solutions. Approaches to scaling AI supervision include debate, in which two agents engage in structured dialogue to help a judge evaluate claims; critique, in which models identify potential flaws in proposed solutions; and prover-verifier games, in which a capable 'prover' model generates solutions that must be verifiable by a less capable 'verifier'. Evaluations of the scalability of these and similar approaches to difficult problems benefit from datasets that include (1) long-form expert-verified correct solutions and (2) long-form flawed solutions with annotations highlighting specific errors, but few are available. To address this gap, we present FindTheFlaws, a group of five diverse datasets spanning medicine, mathematics, science, coding, and the Lojban language. Each dataset contains questions and long-form solutions with expert annotations validating their correctness or identifying specific error(s) in the reasoning. We evaluate frontier models' critiquing capabilities and observe a range of performance that can be leveraged for scalable oversight experiments: models performing more poorly on particular datasets can serve as judges/verifiers for more capable models. Additionally, for some task/dataset combinations, expert baselines exceed even top model performance, making them more beneficial for scalable oversight experiments.