Retrieval-Augmented Generation (RAG) significantly enhances Large Language Models (LLMs) by providing access to external knowledge. However, current research primarily focuses on retrieval quality, often overlooking the critical ”integration bottleneck”: even when relevant documents are retrieved, LLMs frequently fail to utilize them effectively due to conflicts with their internal parametric knowledge. In this paper, we argue that implicitly resolving this conflict in a single generation pass is suboptimal. We introduce GuarantRAG, a framework that explicitly decouples reasoning from evidence integration. First, we generate an ”Inner-Answer” based solely on parametric knowledge to capture the model’s reasoning flow. Second, to guarantee faithful evidence extraction, we generate a ”Refer-Answer” using a novel Contrastive DPO objective. This objective treats the parametric Inner-Answer as a negative constraint and the retrieved documents as positive ground truth, forcing the model to suppress internal hallucinations in favor of external evidence during this phase. Finally, rather than naive concatenation or using the DPO trained model directly, we propose a joint decoding mechanism that dynamically fuses the logical coherence of the Inner-Answer with the factual precision of the Refer-Answer at the token level. Experiments on five QA benchmarks demonstrate that GuarantRAG improves accuracy by up to 12.1% and reduces hallucinations by 16.3% compared to standard and dynamic RAG baselines.
论文检索
输入标题、作者或关键词，从 1,386 篇学术成果中精准定位
UniversalRAG: Retrieval-Augmented Generation over Corpora of Diverse Modalities and Granularities
PDF ↗Retrieval-Augmented Generation (RAG) has shown substantial promise in improving factual accuracy by grounding model responses with external knowledge relevant to queries. However, most existing approaches are limited to a text-only corpus, and while recent efforts have extended RAG to other modalities such as images and videos, they typically operate over a single modality-specific corpus. In contrast, real-world queries vary widely in the type of knowledge they require, which a single type of knowledge source cannot address. To address this, we introduce UniversalRAG, an any-to-any RAG framework designed to retrieve and integrate knowledge from heterogeneous sources with diverse modalities and granularities. Specifically, motivated by the observation that forcing all modalities into a unified representation space derived from a single aggregated corpus causes a modality gap, where the retrieval tends to favor items from the same modality as the query, we propose modality-aware routing, which dynamically identifies the most appropriate modality-specific corpus and performs targeted retrieval within it, and further justify its effectiveness with a theoretical analysis. Moreover, beyond modality, we organize each modality into multiple granularity levels, enabling fine-tuned retrieval tailored to the complexity and scope of the query. We validate UniversalRAG on 10 benchmarks of multiple modalities, showing its superiority over various modality-specific and unified baselines.
Chunks as Arms: Multi-Armed Bandit-Guided Sampling for Long-Context LLM Preference Optimization
PDF ↗Long-context modeling is critical for a wide range of real-world tasks, including long-context question answering, summarization, and complex reasoning tasks. Recent studies have explored fine-tuning Large Language Models (LLMs) with synthetic data to enhance their long-context capabilities. However, the effectiveness of such approaches is often limited by the low diversity and factual inconsistencies in the generated data. To address these challenges, we propose LongMab, a novel framework that leverages a Multi-Armed Bandit (MAB) rollout strategy to identify the most informative chunks from the given long context for sampling high-quality and diverse responses and constructing preference data pairs for Direct Preference Optimization (DPO) training. Specifically, we treat context chunks as arms of MAB, select chunks based on their expected reward scores to input into LLMs to generate responses, and iteratively update these scores based on reward feedback. Both exploration and exploitation during the rollout process enable the LLM to focus on the most relevant context segments, thereby generating and collecting high-quality and diverse responses. Experimental results on both Llama and Qwen show the effectiveness of LongMab by achieving more than a 4% improvement on long-context reasoning benchmarks. All data and code will be released on https://github.com/NEUIR/LongMab-PO.
Multi-hop claim verification is inherently challenging, requiring multi-step reasoning to construct verification chains while iteratively searching for information to uncover hidden bridging facts. This process is fundamentally interleaved, as effective reasoning relies on dynamically retrieved evidence, while effective search demands reasoning to refine queries based on partial information. To achieve this, we propose Hierarchical Agent Reasoning and Information Search (HARIS), explicitly modeling the coordinated process of reasoning-driven searching and search-informed reasoning. HARIS consists of a high-level reasoning agent that focuses on constructing the main verification chain, generating factual questions when more information is needed, and a low-level search agent that iteratively retrieves more information, refining its search based on intermediate findings. This design allows each agent to specialize in its respective task, enhancing verification accuracy and interpretability. HARIS is trained using reinforcement learning with outcome-based rewards. Experimental results on the EX-FEVER and HOVER benchmarks demonstrate that HARIS achieves strong performance, greatly advancing multi-hop claim verification.
The ability to control LLMs’ emulated emotional states and personality traits is an essential step in enabling rich, human-centered interactions in socially interactive settings. We introduce PsySET, a Psychologically-informed benchmark to evaluate LLM Steering Effectiveness and Trustworthiness across the emotion and personality domains. Our study spans four models from different LLM families paired with various steering strategies, including prompting, fine-tuning, and representation engineering. Our results indicate that prompting is consistently effective but limited in intensity control, whereas vector injections achieve finer controllability while slightly reducing output quality. Moreover, we explore the trustworthiness of steered LLMs by assessing safety, truthfulness, fairness, and ethics, highlighting potential side effects and behavioral shifts. Notably, we observe idiosyncratic effects; for instance, even a positive emotion like joy can degrade robustness to adversarial factuality, lower privacy awareness, and increase preferential bias. Meanwhile, anger predictably elevates toxicity yet strengthens leakage resistance. Our framework establishes the first holistic evaluation of emotion and personality steering, offering insights into its interpretability and reliability for socially interactive applications.
As Large Language Models (LLMs) become increasingly used for question-answering (QA), relying on static, pre-annotated references for evaluation poses significant challenges in cost, scalability, and completeness. Meanwhile, using LLMs themselves as evaluators without external grounding remains unreliable for objective tasks, as they systematically over-accept incorrect answers, fabricate supporting rationales, and degrade sharply on questions that fall outside their training data. We propose Search-AuGmented Evaluation (SAGE), a framework to assess LLM outputs without fixed ground-truth answers. Unlike conventional metrics that compare to static references or depend solely on LLM-as-a-judge knowledge, SAGE acts as an agent that actively retrieves and synthesizes external evidence. It iteratively generates web queries, collects information, summarizes findings, and refines subsequent searches through reflection. By reducing dependence on static reference-driven evaluation protocols, SAGE offers a scalable and adaptive alternative for evaluating the factuality of LLMs. Experimental results on multiple free-form QA benchmarks show that SAGE achieves substantial to perfect agreement with human evaluations.
Natural Language Inference (NLI) models have been used in various ways to improve the factuality of LLM outputs. This is typically done by applying an NLI model to judge whether the model output is entailed from the supposed evidence, triggering some corrective actions, such as beam reranking at inference time or RL rewards during training. While NLI models are trained to detect factual inconsistencies over complete sentences, decisions in the common autoregressive generation architecture are made for each evolving text prefix, during decoding. Addressing this setting, we generalize the entailment detection task to apply over arbitrary text prefixes, and suggest its utility for improving generation faithfulness. Providing suitable evaluation and training datasets for this task, we train MiniTruePrefixes, a novel specialized model that better detects factual inconsistencies over text prefixes, outperforming comparable baseline NLI models by 5-14 F1 points in prefix-level entailment. We further demonstrate that integrating MiniTruePrefixes into a controlled decoding framework substantially improves factual consistency in abstractive summarization. When guided by MiniTruePrefixes, LLaMA-3.2-3B-Instruct matches the faithfulness and runtime of the 8B model from the same model family, while using only half the memory.
Large language models (LLMs) integrated with retrieval-augmented generation (RAG) have become a dominant framework for building intelligent assistants. In real-world applications such as ChatGPT with web search, the retrieved document often comes from diverse, potentially unreliable sources and may contain inconsistent claims. Unlike traditional search engines that rely on users to manually compare information, LLM-based systems typically feed all retrieved content into the model’s context, requiring LLMs to autonomously identify, differentiate, and reason over conflicting viewpoints. Unlike mainstream LLM evaluation tasks like math and code generation that are primarily focused on reasoning with factual context, question-answering with multi-source references requires fundamentally different capabilities to identify and reason over knowledge contradictions. In this paper, we introduce ConfRAG, a benchmark for evaluating LLMs’ reasoning capability over real-world conflicting documents retrieved from the web. It consists of 1,814 real-world questions, each paired with an average of 9.58 retrieved paragraphs from heterogeneous online sources. A total of 57.2% of the questions exhibit explicit contradictions. We further propose three structured evaluation tasks, answer clustering, answer coverage, and reason coverage, to quantify a model’s ability to organize and explain contradictory content. Experiments with state-of-the-art models such as GPT-4.1 and Claude-3-7-Sonnet reveal substantial performance gaps, highlighting the need for more targeted research in contradiction-aware question answering. To the best of our knowledge, ConfRAG is the first benchmark specifically designed to evaluate contradiction-aware reasoning on real-world long web documents.
The Plain Writing Act in the United States requires government documents to be written in clear and simple language. However, existing summarization systems struggle to address diverse linguistic and cognitive barriers among general readers. We propose NRLB (No Reader Left Behind), a unified multi-agent framework for plain language summarization that simulates three representative reader groups: elementary school students, non-native speakers, and readers with attention deficits. NRLB integrates template-based planning with an iterative feedback loop guided by simulated readers and domain expert revision to address comprehension barriers such as unknown terms, missing contexts, and confusing sentences. Evaluations across multiple datasets demonstrate consistent improvements in both readability and factuality. Human evaluation further supports these findings, with annotator preference rates ranging from 55% to 76%, highlighting NRLB’s ability to generate summaries that are both faithful to the source and accessible to a wide range of readers.
Applications · Chemistry, Physics, and Earth Sciences
Climate change research increasingly requires AI systems that reason across text, dynamic visual content, and scientific figures, yet existing climate QA benchmarks are small, mostly textual, and cover a narrow range of models. We introduce MMClima, a large-scale multimodal climate question answering framework with over 104k expert-validated question–answer pairs spanning articles, video transcriptions, and figures across five core climate science domains. MMClima is constructed via automated claim extraction and QA synthesis with human-in-the-loop validation to ensure both scale and reliability. Using MMClima, we benchmark state-of-the-art multimodal language models on tasks requiring factual recall, visual interpretation, and cross-modal synthesis. We additionally fine-tune on the textual split to produce mmclima-70b-txt, a domain-adapted baseline that outperforms strong open- and closed-source models on textual QA. We release the dataset, evaluation pipeline, fine-tuned model weights, and data creation framework to support standardized multimodal evaluation for climate science.
Deep Learning · Large Language Models
In this work, we reveal that Large Language Models (LLMs) possess intrinsic behavioral plasticity—akin to chameleons adapting their coloration to environmental cues—that can be *exposed* through token-conditional generation and *stabilized* via reinforcement learning. Specifically, by conditioning generation on carefully selected token prefixes sampled from responses exhibiting desired behaviors, LLMs seamlessly adapt their behavioral modes at inference time (e.g., switching from step-by-step reasoning to direct answering) without retraining. Based on this insight, we propose **To**ken-**Co**nditioned **R**einforcement **L**earning (**ToCoRL**), a principled framework that leverages RL to internalize this chameleon-like plasticity, transforming transient inference-time adaptations into stable and learnable behavioral patterns. ToCoRL guides exploration with token-conditional generation and keep enhancing exploitation, enabling emergence of appropriate behaviors. Extensive experiments show that ToCoRL enables precise behavioral control without capability degradation. Notably, we show that large reasoning models, while performing strongly on complex mathematics, can be effectively adapted to excel at factual question answering, which was a capability previously hindered by their step-by-step reasoning patterns.
Deep Learning · Large Language Models
Large Language Models (LLMs) have inherent limitations of faithfulness and factuality, commonly referred to as hallucinations. Several benchmarks have been developed that provide a test bed for factuality evaluation within the context of English-centric datasets, while relying on supplementary informative context like web links or text passages but ignoring the available structured factual resources. To this end, Knowledge Graphs (KGs) have been identified as a useful aid for hallucination mitigation, as they provide a structured way to represent the facts about entities and their relations with minimal linguistic overhead. We bridge the lack of KG paths and multilinguality for factual language modeling within the existing hallucination evaluation benchmarks and propose a KG-based multilingual, multihop benchmark called MultiHal framed for generative text evaluation. As part of our data collection pipeline, we mined 140k KG-paths from open-domain KGs, from which we pruned noisy KG-paths, curating a high-quality subset of 25.9k. Our baseline evaluation shows an absolute scale improvement by approximately 0.12 to 0.36 points for the semantic similarity score, 0.16 to 0.36 for NLI entailment and 0.29 to 0.42 for hallucination detection in KG-RAG over vanilla QA across multiple languages and multiple models, demonstrating the potential of KG integration. We anticipate MultiHal will foster future research towards several graph-based hallucination mitigation and fact-checking tasks.
Deep Learning · Large Language Models
Large Language Models (LLMs) show strong reasoning abilities, yet their reliability is hindered by hallucinations, where fluent reasoning becomes factually or logically incorrect. Most existing uncertainty-based detectors rely on sequence-level averaging, which ignores the step-wise dynamics of reasoning and often misclassifies hard-but-correct or easy-but-wrong samples. We propose a dynamic perspective that models reasoning as a trajectory on a latent \emph{Evidence Manifold}, where each step is supported by local evidence. Hallucinations are characterized as \emph{Evidence Drops}, i.e., sudden declines in local evidence support that indicate topological deviations from this manifold. Based on this insight, we design a training-free and model-agnostic detector that identifies hallucinations via the worst-case Evidence Drop and enables step-level error localization. Experiments on GSM8K, MATH, and ProcessBench show consistent improvements over sequence-level uncertainty baselines in selective accuracy and risk–coverage trade-offs.
Deep Learning · Large Language Models
Multimodal large language models (MLLMs) are increasingly used for real-world tasks involving multi-step reasoning and long-form generation, where reliability requires grounding model outputs in heterogeneous input sources and verifying individual factual claims. However, existing multimodal grounding benchmarks and evaluation methods focus on simplified, observation-based scenarios or limited modalities and fail to assess attributions in complex multimodal reasoning. We introduce MURGAT (Multimodal Reasoning with Grounded Attribution), a benchmark for evaluating fact-level multimodal attribution in settings that require reasoning beyond direct observation. Given inputs spanning video, audio, and other modalities, MURGAT requires models to generate answers with explicit reasoning and precise citations, where each citation specifies both modality and temporal segment. To enable reliable assessment, we introduce an automatic evaluation framework that strongly correlates with human judgments (r = 0.84). Benchmarking against human and automated scores reveals that even strong MLLMs frequently hallucinate citations despite correct reasoning. Moreover, we see a key trade-off: increasing reasoning depth or enforcing structured grounding often degrades accuracy, highlighting a significant gap between internal reasoning and verifiable attribution
Applications · Computer Vision
LiDAR point cloud anomaly detection is critical for autonomous system safety, yet most existing methods rely only on visible measurements, overlooking occlusion as a structured consequence of the LiDAR sensing process. We argue that anomalies are characterized not only by what is observed, but also by the spatial voids they create, which alter occlusion patterns and volumetric visibility. We propose Counterfactual Occlusion-Visibility Anomaly Learning (COVAL), a framework that intervenes on volumetric visibility during training. Using physics-conformed synthetic anomaly construction, COVAL generates paired factual and counterfactual observations with identical scene geometry but different occlusion patterns. Then, we introduce two complementary objectives: Visibility-Variant Counterfactual Reconstruction, which models occlusion-induced missing regions, and Visibility-Invariant Counterfactual Consistency, which enforces stable representations across visibility changes. Together, these objectives isolate anomaly-induced structural missingness and in turn refine representation of normal scenes, thus improving anomaly sensitivity at test time. Experiments on standard LiDAR anomaly segmentation benchmarks show that COVAL achieves state-of-the-art performance.
Modern post-trained language models are increasingly capable, but remain prone to extrinsic hallucinations. We target the utility degradation issue that prior hallucination-reduction methods often struggle to avoid, and propose online RL with Binary Retrieval-Augmented Reward (Binary RAR) to reduce hallucinations while preserving general capabilities. Binary RAR assigns a reward of 1 if a response contains no factual contradictions with retrieved evidence, and 0 otherwise. We theoretically show that this method reduces the probability of error-containing responses while preserving the distribution of error-free responses. This helps preserve the model’s capabilities, whereas other methods often degrade them. We evaluate Binary RAR on multiple widely used models. On Qwen3-8B, it reduces long-form hallucination rates by 39.3\% and short-form hallucination rates by 54.4\%, outperforming supervised learning and preference optimization baselines. Our error analysis shows that continuous factuality rewards (e.g., VeriScore) can be exploited via reward hacking by producing fewer or more generic claims, whereas Binary RAR is more robust and better preserves general capabilities, including instruction following, math, and coding.
Current evaluation paradigms for generative models rely heavily on retrieval-based metrics such as exact match accuracy, creating a bottleneck particularly in domains requiring scientific discovery and creative reasoning. These metrics penalize any deviation from the training distribution, treating all non-factual outputs as errors. This position paper argues that rigidly minimizing these deviations induces a form of epistemic mode collapse that suppresses the stochastic exploration required for innovation. We propose the Higher-Dimensional Cognitive Hypothesis (HDCH), positing that valuable hallucinations represent geodesic traversals in a high-dimensional latent space that appear as errors only when projected onto the lower-dimensional manifold of established knowledge. We introduce a formal distinction between Type I (factually inconsistent noise) and Type II (factually novel but structurally coherent) exploratory hypotheses based on information geometry. Through experiments, we demonstrate that maximizing discovery requires calibrated instability, peaking at a critical thermodynamic phase transition. Furthermore, we advocate for an evaluation framework that optimizes an Exploratory Signal-to-Noise Ratio (ESNR), balancing the novelty of outputs against their structural plausibility. We conclude that evolving evaluation from validating static retrieval to incentivizing calibrated latent exploration is essential to unlock the full, discovery-oriented potential of generative AI.
Deep Learning · Large Language Models
Large language model (LLM) agents increasingly rely on external tools such as search engines to solve complex, multi-step problems, yet their rollouts are structurally heterogeneous: variations in tool-call number, placement, and outcomes induce distinct behaviors and reward distributions. As a result, policy gradient methods with a single global baseline suffer from *cross-stratum bias*, an "apples-to-oranges" comparison that distorts credit assignment and impedes exploration. To address this issue, we propose *Stratified GRPO*. Its core component, *Stratified Advantage Normalization* (SAN), partitions trajectories into homogeneous strata based on structural properties and computes advantages locally within each stratum, ensuring comparisons only among true peers. We show that SAN eliminates cross-stratum bias, yields conditionally unbiased unit-variance estimates within strata, and preserves the global unbiasedness and unit-variance properties of standard normalization, resulting in a more reliable learning signal. To improve robustness in finite-sample regimes, we further linearly blend SAN with the global estimator. Experiments on factual QA and deep-research agent benchmarks demonstrate that Stratified GRPO consistently outperforms GRPO by up to 12.6 points, achieving higher training rewards, improved training stability, and more effective search policies. These results establish stratification as a principled remedy for structural heterogeneity in RL for LLM search agents.
Current multimodal benchmarks often conflate reasoning with domain-specific knowledge, making it difficult to isolate and evaluate general reasoning abilities in non-expert settings. To address this, we introduce VisualPuzzles, a benchmark that targets visual reasoning while deliberately minimizing reliance on specialized knowledge. VisualPuzzles consists of diverse questions spanning five categories: algorithmic, analogical, deductive, inductive, and spatial reasoning. One major source of our questions is manually translated logical reasoning questions from the Chinese Civil Service Examination. Experiments show that VisualPuzzles requires significantly less intensive domain-specific knowledge and more complex reasoning compared to benchmarks like MMMU, enabling us to better evaluate genuine multimodal reasoning. Evaluations show that state-of-the-art multimodal large language models consistently lag behind human performance on VisualPuzzles, and that strong performance on knowledge-intensive benchmarks does not necessarily translate to success on reasoning-focused, knowledge-light tasks. Additionally, reasoning enhancements such as scaling up inference compute (with "thinking" modes) yield inconsistent gains across models and task types, and we observe no clear correlation between model size and performance. We also found that models exhibit different reasoning and answering patterns on VisualPuzzles compared to benchmarks with heavier emphasis on knowledge. VisualPuzzles offers a clearer lens through which to evaluate reasoning capabilities beyond factual recall and domain knowledge.
COFT: Counterfactual–Conformal Decoding for Fair Chain‑of‑Thought Reasoning in Large Language Models
Social Aspects · Fairness
Large language models (LLMs) can reveal and amplify societal biases during chain-of-thought (CoT) generation. We present COFT (Chain of Fair Thought), a training-free decoding method that applies token-level fairness control at decode time, with distribution-free marginal validity guarantees (under exchangeability) for any frozen causal language model. COFT operates in three stages. First, it creates a masked counterfactual prompt by replacing sensitive spans with neutral tokens. Second, it compares the factual and masked logit distributions through lightweight logit fusion to attenuate attribute-driven biases. Third, it uses dual-branch split-conformal calibration to certify per-step candidate token sets at a user-chosen risk level. We evaluate COFT across six models and multiple bias benchmarks. Our method reduces standard bias metrics by 30–55\% (median 38\%) while preserving task utility and language quality. Reasoning accuracies remain unchanged within run-to-run noise margins. The computational overhead is modest, equivalent to one additional cached forward pass (<=11%). COFT offers a clear, auditable path to safer CoT generation with significant bias reduction, negligible utility loss, and no requirement for retraining, auxiliary classifiers, or weight access.