As large language model (LLM) based systems take on high-stakes roles in real-world decision-making, they must reconcile competing instructions from multiple sources within a single prompt context. Enforcing an instruction hierarchy, where higher-level directives override lower-priority requests, is critical to the reliability and controllability of LLMs. In this work, we reframe instruction hierarchy resolution as a reasoning task. The model must first "think" about the relationship between a given user prompt and higher-priority instructions before generating a response. To enable this capability, we construct VerIH, a training dataset of constraint-following tasks with verifiable answers, comprising aligned and conflicting system–user instructions. We show that lightweight reinforcement learning with VerIH effectively transfers general reasoning capabilities of models to instruction prioritization. Our method leads to consistent improvements across multiple model families and scales on instruction following and instruction hierarchy benchmarks, achieving ~20% absolute improvement in conflict setups. Our method also generalizes to safety-critical scenarios beyond the training distribution, exhibiting increased robustness against jailbreak and prompt injection, reducing absolute attack success rates by up to 20%. Our results establish reasoning over instruction hierarchies as a practical mechanism for improving AI reliability, where targeted updates to system prompts produce predictable, controllable, and robust changes in model behavior.
论文检索
输入标题、作者或关键词，从 12,319 篇学术成果中精准定位
Despite the widespread adoption of text embedding models, selecting the optimal model for a specific target corpus remains challenging due to the lack of task-specific labels. While task-agnostic evaluation offers a promising solution by relying on unlabeled data, existing approaches based on kernel estimators or Gaussian mixtures fail to model high-dimensional distributions effectively, resulting in unstable rankings. To address this limitation, we propose FLARE (Flow-based Label-free Assessment of Representation Embeddings), which employs normalizing flows to estimate information sufficiency in high-dimensional spaces. By learning invertible transformations, flows enable exact density estimation while mitigating the instability inherent in distance-based methods. We provide theoretical guarantees showing that our estimation error depends on the data’s intrinsic structure rather than its raw dimensionality. Experiments across 11 datasets demonstrate that FLARE achieves a strong Spearman’s ρ (up to 0.90) with supervised benchmarks, remaining robust even for high-dimensional embeddings (d ≥ 3,584).
Parametric Knowledge is Not All You Need: Toward Honest Large Language Models via Retrieval of Pretraining Data
PDF ↗Large language models (LLMs) are highly capable of answering questions, but they are often unaware of their own knowledge boundary, i.e., knowing what they know and what they don’t know. As a result, they can generate factually incorrect responses on topics they do not have enough knowledge of, commonly known as hallucination. Rather than hallucinating, a language model should be more honest and respond with "I don’t know" when it does not have enough knowledge about a topic. Many methods have been proposed to improve LLM honesty, but their evaluations lack robustness, as they do not take into account the knowledge that the LLM has ingested during its pretraining. In this paper, we propose a more robust evaluation benchmark dataset for LLM honesty by utilizing Pythia, a truly open LLM with publicly available pretraining data. In addition, we also propose a novel method for harnessing the pretraining data to build a more honest LLM.
Generating spoken dialogue is inherently more complex than monologue text-to-speech (TTS), as it demands both realistic turn-taking and the maintenance of distinct speaker timbres. While existing autoregressive (AR) models have made progress, they often suffer from high inference latency and stability issues. To overcome these limitations, we propose ZipVoice-Dialog, a non-autoregressive (NAR) zero-shot spoken dialogue generation model based on flow-matching. Observing that applying vanilla flow-matching to dialogue generation leads to poor speech intelligibility and turn-taking precision, we introduce two simple yet effective methods to adapt flow-matching architectures for dialogue generation: (1) a curriculum learning strategy to ensure robust speech-text alignment, and (2) speaker-turn embeddings to govern precise speaker turn-taking. Additionally, we introduce dedicated strategies to support stereo dialogue generation.Recognizing the lack of training datasets in this field, we curate and release OpenDialog, the first large-scale (6.8k hours) open-source spoken dialogue dataset derived from in-the-wild speech data. Moreover, for fair and rigorous evaluations, we established a benchmark to comprehensively evaluate dialogue generation models. Experiments demonstrate the effectiveness of the proposed methods and dataset, showing that ZipVoice-Dialog achieves superior performance in inference speed, intelligibility, speaker turn-taking accuracy, and speaker similarity. Our code, model checkpoints, and the OpenDialog dataset are publicly available.
KnowDR-REC: Auditing Knowledge-Conditioned Visual Grounding in Referring Expression Comprehension
PDF ↗While Multimodal Large Language Models (MLLMs) have demonstrated the capacity for multi-modal reasoning, current Referring Expression Comprehension (REC) benchmarks lag behind, predominantly relying on intra-image cues and neglecting the integration of external world knowledge, which significantly impedes the evolution of REC towards real-world applications. This limitation obscures a model’s true capability to conduct textual reasoning (entity resolution), resolve spatial location (visual grounding), and verify reference validity (hallucination rejection). To address this, we introduce KnowDR-REC, a targeted audit benchmark comprising 1,042 positive triplets derived from real-world knowledge, along with rigorously matched negative samples. Unlike traditional datasets, we implement a controllable counterfactual evaluation mechanism that subjects textual expressions to single-factor perturbations (entity, relation, or time) to test sensitivity to fine-grained factual changes. Extensive evaluation of 18 state-of-the-art LMMs exposes a critical “binding hallucination,” revealing that current high performance is often built on fragile visual shortcuts rather than true understanding. KnowDR-REC thus serves as a pivotal diagnostic instrument, steering future research toward the genuine integration of perception and reasoning.
Retrieval-Augmented Generation (RAG) has been widely adopted to enhance large language models (LLMs) by incorporating external knowledge. However, the two main existing paradigms struggle with multi-hop reasoning: aggregate-first approaches suffer from high construction costs and limited adaptability to dynamic knowledge, while dynamic-first approaches rely heavily on LLM reasoning and are prone to error propagation across reasoning steps. To address these limitations, we propose SR-RAG, a symbolic reasoning framework for multi-hop question answering. SR-RAG integrates the advantages of both paradigms by dynamically generating sub-questions, performing information retrieval and symbolic encoding based on an on-the-fly graph, and using a symbolic verifier to formally validate intermediate reasoning steps to ensure the correctness of intermediate answers and the completeness of the reasoning chain . We evaluate SR-RAG on multiple multi-hop benchmarks and a medical dataset. Experimental results demonstrate that it significantly improves both accuracy and robustness.
As language models accelerate scientific research by automating hypothesis generation and implementation, a new bottleneck emerges: evaluating and filtering hundreds of AI-generated ideas without exhaustive experimentation. We ask whether LMs can learn to forecast the empirical success of research ideas before any experiments are run. We study comparative empirical forecasting: given a benchmark-specific research goal and two candidate ideas, predict which will achieve better benchmark performance. We construct a dataset of 11,488 idea pairs grounded in objective outcomes from PapersWithCode. While off-the-shelf 8B-parameter models struggle (30% acc.), SFT dramatically boosts performance to 77.1%, outperforming GPT-5 (61.1%). By framing evaluation as a reasoning task via Reinforcement Learning with Verifiable Rewards (RLVR), we train models to discover latent reasoning paths, achieving 71.35% acc. with interpretable justifications. Through additional ablations and out-of-distribution tests, we show robustness to surface-level heuristics and transfer to both a cross-domain time-split test set and an independently constructed test set. Our results demonstrate that compute-efficient small language models can serve as effective, objective verifiers, offering a scalable path for autonomous scientific discovery.
AnchorAlign: A Novel Anchor Alignment-enhanced Generative Method for Joint Named Entity Recognition and Relation Extraction
PDF ↗Named Entity Recognition (NER) and Relation Extraction (RE) are two fundamental and interdependent tasks in information extraction (IE), aiming to identify entities and relations from unstructured text. Recently, generative methods have become mainstream instead of discriminative methods for IE, especially joint multi-task IE, due to their promising performance and flexibility. For joint NER and RE, existing methods suffer from misalignment between entities and relations, as well as misalignment among relations. To address these issues, we propose AnchorAlign, a novel generative method enhanced by anchor alignment. Specifically, we first introduce an anchor entity selection mechanism to identify key entities in the text as anchor points, which serve as semantic pivots to bridge the two tasks. Then, we design a dual-level anchor alignment module: at the semantic level, we construct a cross-task semantic alignment space to align the semantic representations of anchor entities and their associated relations; at the generation level, we introduce an anchor-guided generation constraint to guide the model to generate entities and relations with strict alignment based on the anchor points. Extensive experiments on five benchmark datasets show that AnchorAlign outperforms state-of-the-art baselines, demonstrating its effectiveness. Our work provides a new perspective for optimizing the joint modeling of NER and RE, and has potential to be extended to more complex multi-task IE such as NER and Event Extraction (EE).
A Learnable Skill Combination Strategy for Multi-task Learning in Natural Language Understanding
PDF ↗In the realm of domain-specific natural language understanding (NLU) tasks, acquiring high-quality labeled data is often arduous, thereby posing significant challenges for effective model training. Multi-task learning (MTL) addresses these limitations by jointly optimizing multiple tasks within a unified framework. In this paper, we introduce a novel sparse NLU multi-task learning framework that decomposes the language model into modular skill components and employs a dynamic, learnable skill-combination mechanism to adaptively handle diverse tasks. Extensive experiments on benchmark NLU datasets demonstrate that our proposed method surpasses conventional multi-task learning approaches in performance.
Large language model (LLM) routing assigns each query to the most suitable model from an ensemble. We introduce LLMRouterBench, a large-scale benchmark and unified framework for LLM routing. It comprises over 400K instances from 21 datasets and 33 models. Moreover, it provides comprehensive metrics for both performance-oriented and performance-cost trade-off routing, and integrates 10 representative routing baselines. Using LLMRouterBench, we systematically re-evaluate the field. While confirming strong model complementarity—the central premise of LLM routing—we find that many routing methods exhibit similar performance under unified evaluation, and several recent approaches, including commercial routers, fail to reliably outperform a simple baseline. Meanwhile, a substantial gap remains to the Oracle, driven primarily by persistent model-recall failures. We further show that backbone embedding models have limited impact, that larger ensembles exhibit diminishing returns compared to careful model curation, and that the benchmark also enables latency-aware analysis. All code and data are available at https://github.com/ynulihao/LLMRouterBench.
The Development Knowledge Question Answering (Dev Knowledge QA) task aims to provide accurate natural language answers to knowledge-seeking questions during software development. To investigate the importance of Dev Knowledge QA in AI-assisted software development and the extent to which it has been explored, we conduct a preliminary analysis of real user–LLM dialogues from WildChat. Our findings indicate that Dev Knowledge QA plays a significant role in real-world software development scenarios, and these raw dialogues cannot be directly used to construct a Dev Knowledge QA benchmark. Existing Dev Knowledge QA benchmarks are limited in development knowledge scope and often not built from real user queries. To bridge this gap, we design a three-phase pipeline that transforms real-world dialogue into simple development knowledge-seeking QA pairs. Through this pipeline, we introduce SimpleDevQA, a multilingual Dev Knowledge QA benchmark inspired by real user dialogues. This dataset covers three languages (English, Chinese, and Russian), and focuses on questions with unique, short, and verifiable answers, making evaluation more accurate and simple. Extensive experiments with 18 mainstream LLMs show that closed-source models generally perform best on SimpleDevQA. We also find that RAG-based knowledge injection improves accuracy, and that Dev Knowledge QA performance correlates with both model confidence and code-generation capability. To facilitate the replication study, we have released our data and code at: https://github.com/DeepSoftwareAnalytics/SimpleDevQA.
Memory serves as a pivotal component in interactive response generation, supplying essential background information and referential knowledge for dialogues. Conventional interactive algorithms have predominantly treated memory as a merely contextual element, largely neglecting the nuanced cognitive processes involved in individualized memory encoding and retrieval. This conceptual gap has led to the prevailing schema where memory-enhanced dialogue datasets incorporate monolithic, undifferentiated memory content, failing to capture the personalized nature of persoa memory processing. Grounded in the self-reference effect from cognitive psychology, we introduce a Multi-Turn Dialogue Dataset with Personalized Contextual Memory (), establishing a comprehensive benchmark to facilitate advanced research on personalized memory processing algorithms.
Understanding and Mitigating Spurious Signal Amplification in Test-Time Reinforcement Learning for Math Reasoning
PDF ↗Test-time reinforcement learning (TTRL) always adapts models at inference time via pseudo-labeling, leaving it vulnerable to spurious optimization signals from label noise.Through an empirical study, we observe that responses with medium consistency form an ambiguity region and constitute the primary source of reward noise.Crucially, we find that such spurious signals can be even amplified through group-relative advantage estimation.Motivated by these findings, we propose a unified framework, Debiased and Denoised test-time Reinforcement Learning (DDRL), to mitigate spurious signals.Concretely, DDRL first applies a frequency-based sampling strategy to exclude ambiguous samples while maintaining a balanced set of positive and negative examples.It then adopts a debiased advantage estimation with fixed advantages, removing the bias introduced by group-relative policy optimization.Finally, DDRL incorporates a consensus-based off-policy refinement stage, which leverages the rejection-sampled dataset to enable efficient and stable model updates.Experiments on three large language models across multiple mathematical reasoning benchmarks demonstrate that DDRL consistently outperforms existing TTRL baselines.The code is available at https://github.com/yuyongcan/DDRL.
DANCE: Diversity-attended Dynamic Caching with Asymmetric Quantization for Test-time Adaptation of Vision-Language Models
PDF ↗This paper studies the problem of test-time adaptation for vision-language models (VLMs). Recent approaches typically measure the prediction entropy to store a confident cache for logit refinement. However, these confident samples tend to approach prototypes with limited coverage of data distribution, which could result in biased predictions as the distribution evolves. Towards this end, we propose a novel approach named Diversity-attended Dynamic Caching with Asymmetric Quantization (DANCE) for test-time adaptation of VLMs. The core of our DANCE is to maintain a dynamic cache to store diversity-aware test samples, which support efficient logit adjustment via asymmetric quantization. In particular, we first generate multiple augmented views of each sample and aggregate their outputs from pre-trained VLMs via a consistency-aware mechanism. More importantly, we construct a dynamic cache, which stores the most reliable and diverse samples to cover evolving test distributions. To measure the diversity efficiently, we quantize cached samples and compute the asymmetric similarity across query samples and memory samples, which guide the cache updating via replacing samples with the lowest scores iteratively. Finally, we leverage the asymmetric similarity between the quantized prototype representations from the dynamic cache to update logits under distribution shifts. Extensive experiments on various benchmark datasets validate the superiority of the proposed DANCE in different settings.
Toxic language includes content that is offensive, abusive, or that promotes harm. Progress in preventing toxic output from large language models (LLMs) is hampered by inconsistent definitions of toxicity. We introduce TRuST, a large-scale dataset that unifies and expands prior resources through a carefully synthesized definition of toxicity, and corresponding annotation scheme. It consists of ∼300k annotations, with high-quality human annotation on ∼11k. To ensure high-quality, we designed a rigorous, multi-stage human annotation process, and evaluated the diversity of the annotators. Then we benchmarked state-of-the-art LLMs and pre-trained models on three tasks: toxicity detection, identification of the target group, and of toxic words. Our results indicate that fine-tuned PLMs outperform LLMs on the three tasks, and that current reasoning models do not reliably improve performance. TRuST constitutes one of the most comprehensive resources for evaluating and mitigating LLM toxicity, and other research in socially-aware and safer language technologies.
Sports have witnessed growing global enthusiasm in recent years, serving as a vital force for physical health, cultural exchange, social connection, and economic growth. The rapid advancement of large models, particularly (multimodal) large language models (M)LLMs, has demonstrated transformative potential to reshape sports understanding, analysis, and interaction across diverse domains. This paper presents a comprehensive survey of large models in sports, including (i) an overview of tasks and applications across different participant groups; (ii) a detailed analysis of sports-related datasets and benchmarks; and (iii) a critical discussion of current challenges and future directions. Our goal is to establish a foundation for advancing research and practical development of large-model-driven sports intelligence. An open-source GitHub repository is maintained at: https://github.com/Road2Redemption/Awesome_Large_Models_In_Sports1.
Automatic evaluation metrics are essential for building multilingual translation systems. The common practice of evaluating these systems is averaging metric scores across languages, yet this is suspicious since metrics may suffer from cross-lingual scoring bias, where translations of equal quality receive different scores across languages. This problem has not been systematically studied because no benchmark exists that provides parallel-quality instances across languages, and expert annotation is not realistic. In this work, we propose XQ-MEval, a semi-automatically built dataset covering nine translation directions, to benchmark translation metrics. Specifically, we inject MQM-defined errors into gold translations automatically, filter them by native speakers for reliability, and merge errors to generate pseudo translations with controllable quality. These pseudo translations are then paired with corresponding sources and references to form triplets used in assessing the qualities of translation metrics. Using XQ-MEval, our experiments on nine representative metrics reveal the inconsistency between averaging and human judgment and provide the first empirical evidence of cross-lingual scoring bias. Finally, we propose a normalization strategy derived from XQ-MEval that aligns score distributions across languages, improving the fairness and reliability of multilingual metric evaluation.
Suicide memes are memes used to express suicide-related thoughts or comment on suicide-related issues. Suicide memes are increasingly common on social media, yet remain poorly understood and potentially harmful. There is an urgent need to better understand their characteristics and to develop appropriate content moderation strategies that limits users’ exposure to potentially harmful content. Currently, the absence of annotated datasets of suicide memes remains a key barrier to developing and evaluating automated moderation approaches. In this paper, we introduce FigSIM, the first dataset designed for fine-grained analysis of suicide memes. The dataset consists of 1049 memes, each annotated for (1) fine-grained suicide severity levels, (2) figurative phenomena (e.g. metaphors), and (3) suicide-related content (e.g. suicide method depiction). We benchmark 16 unimodal and multimodal models across three tasks: figurative language, suicide severity, and suicide-related content detection. Overall, FigSIM demonstrates that suicide memes pose unique challenges for both modeling and content moderation. Analysis revealed biases, such as underprediction of higher suicide severity levels, especially for figurative memes.
We present **ViLegalLM**, comprising **ViLegalBERT** and **ViLegalQwen**, the first suite of Vietnamese pretrained language models for legal text understanding and generation. It includes one encoder-only model (ViLegalBERT, 135M parameters) and two decoder-only models (ViLegalQwen2.5-1.5B-Base and ViLegalQwen3-1.7B-Base), all continually pretrained on a newly curated 16GB Vietnamese legal corpus, significantly larger than previous work. To mitigate data scarcity, we construct three synthetic datasets using LLM-based generation and hard negative mining for True/False QA, Multiple Choice QA, and Natural Language Inference. We establish state-of-the-art results among open-source models on four main Vietnamese legal downstream tasks spanning ten benchmarks, demonstrating that continual pretraining from base models consistently outperforms instruction-tuned adaptation. Source codes, corpus, datasets, and model checkpoints are publicly available at https://github.com/ntphuc149/ViLegalLM.
Large language models are increasingly used in inventive problem-solving, but effective support requires more than open-ended idea generation. Inventive problem-solving requires improving one aspect of a technical system without unintentionally worsening another. TRIZ (Theory of Inventive Problem Solving) provides a unique and structured framework for this setting by representing engineering trade-offs as contradictions and linking them to standardized inventive principles. However, prior TRIZ–LLM evaluations are typically small-scale, case studies in focused areas of technology, and rarely grounded in patent text, which makes it difficult to assess structured reasoning at scale. We introduce TRIZBench, a dataset and benchmark for TRIZ reasoning grounded in open technical sources and U.S. patents. TRIZBench evaluates the core TRIZ workflow through three tasks: contradiction prediction, inventive principle prediction, and grounded TRIZ reasoning. Experiments with multiple LLM baselines show that detecting contradictions is easier than recovering correct trade-off pairs, while principle prediction benefits from explicitly exploiting TRIZ structure. Our findings further underscore the importance of grounding. We show that semantic retrieval enables evidence-based justifications and helps explain why LLMs fail. Dataset and code are available at https://github.com/ellenzhuwang/trizbench.