While LLMs are widely used for generic tasks like question answering and search, they struggle to adapt to specialized knowledge, such as industrial workflows in healthcare, legal, and agricultural sectors, as well as knowledge-driven tasks such as news journalism, investigative research, and consulting without expensive fine-tuning or sub-optimal retrieval methods. Existing retrieval-augmented models, such as RAG, offer improvements but fail to account for structured domain knowledge, leading to suboptimal context generation. Ontologies, which conceptually organize domain knowledge by defining entities and their interrelationships, offer a structured representation to address this gap. This paper presents OG-RAG, an Ontology-Grounded Retrieval Augmented Generation method designed to enhance LLM-generated responses by anchoring retrieval processes in domain-specific ontologies. OG-RAG constructs a hypergraph representation of domain documents, where each hyperedge encapsulates clusters of factual knowledge grounded using domain-specific ontology and retrieves a minimal set of hyperedges for a given query using an optimization algorithm. Our evaluations demonstrate that OG-RAG increases the recall of accurate facts by 55% and improves response correctness by 40% across four different LLMs. Additionally, OG-RAG enables 30% faster attribution of responses to context and boosts fact-based reasoning accuracy by 27% compared to baseline methods. We release the code at [https://github.com/microsoft/ograg2](https://github.com/microsoft/ograg2).
论文检索
输入标题、作者或关键词，从 7,591 篇学术成果中精准定位
SinhalaMMLU: A Comprehensive Benchmark for Evaluating Multitask Language Understanding in Sinhala
PDF ↗Large Language Models (LLMs) demonstrate impressive general knowledge and reasoning abilities, yet their evaluation has predominantly focused on global or anglocentric subjects, often neglecting low-resource languages and culturally specific content. While recent multilingual benchmarks attempt to bridge this gap, many rely on automatic translation, which can introduce errors and misrepresent the original cultural context. To address this, we introduce SinhalaMMLU, the first multiple-choice question answering benchmark designed specifically for Sinhala, a low-resource language. The dataset includes over 7,000 questions spanning secondary to collegiate education levels, aligned with the Sri Lankan national curriculum, and covers six domains and 30 subjects, encompassing both general academic topics and culturally grounded knowledge. We evaluate 26 LLMs on SinhalaMMLU and observe that, while Claude 3.5 sonnet and GPT-4o achieve the highest average accuracies at 67% and 62% respectively, overall model performance remains limited. In particular, models struggle in culturally rich domains such as the Humanities, revealing substantial room for improvement in adapting LLMs to low-resource and culturally specific contexts.
Fill-in-the-Middle (FIM), or infilling, has become integral to code language models, enabling generation of missing code given both left and right contexts. However, the current FIM training paradigm which performs next-token prediction (NTP) over reordered sequence often leads to models struggling to generate content that aligns well with the surrounding context. We hypothesize that NTP alone is insufficient for models to learn effective planning conditioned on the distant right context, a critical factor for successful code infilling. To overcome this, we propose Horizon-Length Prediction (HLP), a novel training objective that teaches models to predict the number of remaining middle tokens at each step. HLP advances FIM with lookahead planning, enabling models to inherently learn infilling boundaries for arbitrary left and right contexts without relying on dataset-specific post-processing. Our evaluation across different model families and sizes shows that HLP significantly improves FIM performance by up to 24% relatively on diverse benchmarks, across file-level and repository-level. Furthermore, the enhanced planning capability gained through HLP boosts model performance on code reasoning. Importantly, HLP incurs negligible training overhead and no additional inference cost, ensuring its practicality for real-world scenarios.
The growing deployment of deep learning models in real-world applications necessitates not only high predictive accuracy, but also mechanism to identify unreliable predictions, especially in high-stakes scenarios where decision risk must be minimized. Existing methods estimate uncertainty by leveraging predictive confidence (e.g., Softmax Response), structural characteristics of representation space (e.g., Mahalanobis distance), or stochastic variation in model outputs (e.g., Bayesian inference techniques such as Monte Carlo Dropout). In this work, we propose a novel uncertainty estimation (UE) framework based on sparse dictionary learning by identifying dictionary atoms associated with misclassified samples. We leverage pointwise mutual information (PMI) to quantify the association between sparse features and predictive failure. Our method – Sparsity-based Uncertainty Estimation (SUE) – is computationally efficient, offers interpretability via atom-level analysis of the dictionary, has no assumption about the class distribution (unlike Mahalanobis distance). We evaluated SUE on several NLU benchmarks (GLUE and ANLI tasks) and sentiment analysis benchmarks (Twitter, ParaDetox, and Jigsaw). In general, SUE outperforms or matches the performance of other methods. SUE performs particularly well when there is considerable uncertainty in the model, i.e., when the model lacks high precision.
Non-English dialogue datasets are scarce, and models are often trained or evaluated on translations of English-language dialogues, an approach which can introduce artifacts that reduce their naturalness and cultural appropriateness. This work proposes Dialogue Act Script (DAS), a structured framework for encoding, localizing, and generating multilingual dialogues from abstract intent representations. Rather than translating dialogue utterances directly, DAS enables the generation of new dialogues in the target language that are culturally and contextually appropriate. By using structured dialogue act representations, DAS supports flexible localization across languages, mitigating translationese and enabling more fluent, naturalistic conversations. Human evaluations across Italian, German, and Chinese show that DAS-generated dialogues consistently outperform those produced by both machine and human translators on measures of cultural relevance, coherence, and situational appropriateness.
Language Models (LMs) are typically tuned with human preferences to produce helpful responses, but the impact of preference tuning on the ability to handle culturally diverse queries remains understudied. In this paper, we systematically analyze how native human cultural preferences can be incorporated into the preference learning process to train more culturally aware LMs. We introduce CARE, a multilingual resource containing 3,490 culturally specific questions and 31.7k responses with human judgments. We demonstrate how a modest amount of high-quality native preferences improves cultural awareness across various LMs, outperforming larger generic preference data. Our analyses reveal that models with stronger initial cultural performance benefit more from alignment, leading to gaps among models developed in different regions with varying access to culturally relevant data. CARE is publicly available at https://github.com/Guochry/CARE.
The affective attitude of liking a recommended item reflects just one category in a wide spectrum of affective phenomena that also includes emotions such as entranced or intrigued, moods such as cheerful or buoyant, as well as more fine-grained affective states, such as “pleasantly surprised by the conclusion”. In this paper, we introduce a novel recommendation task that can leverage a virtually unbounded range of affective states sought explicitly by the user in order to identify items that, upon consumption, are likely to induce those affective states. Correspondingly, we create a large dataset of user preferences containing expressions of fine-grained affective states that are mined from book reviews, and propose ACRec, a Transformer-based architecture that leverages such affective expressions as input. We then use the resulting dataset of affective states preferences, together with the linked users and their histories of book readings, ratings, and reviews, to train and evaluate multiple recommendation models on the task of matching recommended items with affective preferences. Experimental comparisons with a range of state-of-the-art baselines demonstrate ACRec’s superior ability to leverage explicit affective preferences.
Conversational agents powered by large language models (LLMs) are rapidly becoming integral to our daily interactions, generating unprecedented amounts of conversational data. Such datasets offer a powerful lens into societal interests, trending topics, and collective concerns. Yet existing approaches typically treat these interactions as independent, missing critical insights that could emerge from aggregating and reasoning across large-scale conversation logs. In this paper, we introduce Aggregative Question Answering, a novel task requiring models to reason explicitly over thousands of user-chatbot interactions to answer aggregational queries, such as identifying emerging concerns among specific demographics. To enable research in this direction, we construct a benchmark, WildChat-AQA, comprising 6,027 aggregative questions derived from 182,330 real-world chatbot conversations. Experiments show that existing methods either struggle to reason effectively or incur prohibitive computational costs, underscoring the need for new approaches capable of extracting collective insights from large-scale conversational data.
Large language models (LLMs) have improved significantly in their reasoning through extensive training on massive datasets. However, relying solely on additional data for improvement is becoming increasingly impractical, highlighting the need for models to autonomously enhance their reasoning without external supervision. In this paper, we propose \textbf{Debate, Train, Evolve (DTE)}, a novel ground truth-free training framework that uses multi-agent debate traces to evolve a single language model. We also introduce a new prompting strategy \textbf{Reflect-Critique-Refine}, to improve debate quality by explicitly instructing agents to critique and refine their reasoning. Extensive evaluations on \textbf{seven} reasoning benchmarks with \textbf{six} open-weight models show that our DTE framework achieve substantial improvements, with an average accuracy gain of \textbf{8.92\%} on the challenging GSM-PLUS dataset. Furthermore, we observe strong cross-domain generalization, with an average accuracy gain of \textbf{5.8\%} on all other benchmarks, suggesting that our method captures general reasoning capabilities. Our framework code and trained models are publicly available at https://github.com/ctrl-gaurav/Debate-Train-Evolve.
From Shortcuts to Balance: Attribution Analysis of Speech-Text Feature Utilization in Distinguishing Original from Machine-Translated Texts
PDF ↗Neural text-based models for detecting machine-translated texts can rely on named entities (NEs) as classification shortcuts. While masking NEs encourages learning genuine translationese signals, it degrades the classification performance. Incorporating speech features compensates for this loss, but their interaction with NE reliance requires careful investigation. Through systematic attribution analysis across modalities, we find that bimodal integration leads to more balanced feature utilization, reducing the reliance on NEs in text while moderating overemphasis attribution patterns in speech features.
Large language models (LLMs) hold promise for therapeutic interventions, yet most existing datasets rely solely on text, overlooking non-verbal emotional cues essential to real-world therapy. To address this, we introduce a multimodal dataset of 1,441 publicly sourced therapy session videos containing both dialogue and non-verbal signals such as facial expressions and vocal tone. Inspired by Hochschild’s concept of emotional labor, we propose a computational formulation of emotional dissonance—the mismatch between facial and vocal emotion—and use it to guide emotionally aware prompting. Our experiments show that integrating multimodal cues, especially dissonance, improves the quality of generated interventions. We also find that LLM-based evaluators misalign with expert assessments in this domain, highlighting the need for human-centered evaluation. Data and code will be released to support future research.
Text anonymization is essential for responsibly developing and deploying AI in high-stakes domains such as healthcare, social services, and law. In this work, we propose a novel methodology for privacy-preserving synthetic text generation that leverages the principles of de-identification and the Hiding In Plain Sight (HIPS) theory. Our approach introduces entity-aware control codes to guide controllable generation using either in-context learning (ICL) or prefix tuning. The ICL variant ensures privacy levels consistent with the underlying de-identification system, while the prefix tuning variant incorporates a custom masking strategy and loss function to support scalable, high-quality generation. Experiments on legal and clinical datasets demonstrate that our method achieves a strong balance between privacy protection and utility, offering a practical and effective solution for synthetic text generation in sensitive domains.
SemCSE: Semantic Contrastive Sentence Embeddings Using LLM-Generated Summaries For Scientific Abstracts
PDF ↗We introduce SemCSE, an unsupervised method for learning semantic embeddings of scientific texts. Building on recent advances in contrastive learning for text embeddings, our approach leverages LLM-generated summaries of scientific abstracts to train a model that positions semantically related summaries closer together in the embedding space. This resulting objective ensures that the model captures the true semantic content of a text, in contrast to traditional citation-based approaches that do not necessarily reflect semantic similarity. To validate this, we propose a novel benchmark designed to assess a model’s ability to understand and encode the semantic content of scientific texts, demonstrating that our method enforces a stronger semantic separation within the embedding space. Additionally, we evaluate SemCSE on the comprehensive SciRepEval benchmark for scientific text embeddings, where it achieves state-of-the-art performance among models of its size, thus highlighting the benefits of a semantically focused training approach.
LLM query-passage relevance assessment is typically studied using a one-by-one pointwise (PW) strategy where each LLM call judges one passage at a time. However, this strategy requires as many LLM calls as there are passages while also preventing information sharing between passages. We thus hypothesize that batched PW methods, which evaluate multiple passages per LLM call, can improve not only efficiency but also judgment quality — by enabling content from multiple passages to be seen jointly. Moreover, batched PW methods may be better suited to harness the test-time scaling benefits of self-consistency — the ensembling technique of repeating (potentially perturbed) LLM tasks in parallel and aggregating results — since batching can naturally enable prompt diversification through varied batch permutations and compositions to create more robust ensembles. We evaluate several batched PW methods against one-by-one PW and listwise ranking baselines on LLM relevance assessment and ranking tasks, using three passage retrieval datasets and GPT-4o, Claude Sonnet 3, and Amazon Nova Pro. We show that batching can greatly amplify self-consistency benefits, making batched PW methods achieve the best performance while often reducing latency by an order of magnitude or more compared to one-by-one PW methods. For instance, on legal search, batched PW ranking with GPT-4o improves from 43.8% to 51.3% NDCG@10 when using 1 vs. 15 self-consistency calls, compared to one-by-one PW ranking improving from 44.9% to 46.8% and being 15.3x slower.
Large language model (LLM) reasoning can be improved by scaling test-time compute with aggregation, i.e., generating multiple samples and aggregating over them. While improving performance, this strategy often reaches a saturation point beyond which additional compute provides no return. Refinement offers an alternative by using model-generated feedback to improve answer quality. However, refinement faces three key challenges: (1) Excessive refinement: Uniformly refining all instances can cause over-correction and reduce overall performance. (2) Inability to localize and address errors: LLMs struggle to identify and correct their own mistakes. (3) Insufficient refinement: Stopping refinement too soon could leave errors unaddressed. To tackle these issues, we propose MAgICoRe, a framework for Multi-Agent Iteration for Coarse-to-fine Refinement. MAgICoRe mitigates excessive refinement by categorizing problems as easy or hard, solving easy problems with coarse-grained aggregation, and solving the hard ones with fine-grained multi-agent refinement. To better localize errors, we incorporate external step-wise reward model scores, and to ensure sufficient refinement, we iteratively refine the solutions using a multi-agent setup. We evaluate MAgICoRe on Llama-3-8B and GPT- 3.5 and show its effectiveness across seven reasoning datasets. One iteration of MAgICoRe beats Self-Consistency by 3.4%, Best-of-k by 3.2%, and Self-Refine by 4.0% even when these baselines use k = 120, and MAgICoRe uses less than 50% of the compute.
Reasoning has long been viewed as an emergent property of large language models (LLMs). However, recent studies challenge this assumption, showing that small language models (SLMs) can also achieve competitive reasoning performance. This paper introduces \textbf{ThinkSLM}, the first extensive benchmark to systematically evaluate and study the reasoning abilities of SLMs trained from scratch or derived from LLMs through quantization, pruning, and distillation. We first establish a reliable evaluation criterion comparing available methods and LLM judges against our human evaluations. Then we present a study evaluating \textbf{72} diverse SLMs from \textbf{six} major model families across \textbf{17 reasoning benchmarks}. We repeat all our experiments \textbf{three} times to ensure a robust assessment. Our findings show that: \textbf{\textit{1)}} reasoning ability in SLMs is strongly influenced by training methods and data quality rather than solely model scale; \textbf{\textit{2)}} quantization preserves reasoning capability, while pruning significantly disrupts it;\textbf{\textit{ 3)}} larger models consistently exhibit higher robustness against adversarial perturbations and intermediate reasoning, but certain smaller models closely match or exceed the larger models’ performance. Our findings challenge the assumption that scaling is the only way to achieve strong reasoning. Instead, we foresee a future where SLMs with strong reasoning capabilities can be developed through structured training or post-training compression. Our \textbf{ThinkSLM} Leaderboard is publicly available at: https://ctrl-gaurav.github.io/thinkslm.github.io/.
Large language models (LLMs) have shown promise in robotic procedural planning, yet their human-centric reasoning often omits the low-level, grounded details needed for robotic execution. Vision-language models (VLMs) offer a path toward more perceptually grounded plans, but current methods either rely on expensive, large-scale models or are constrained to narrow simulation settings. We introduce SelfReVision, a lightweight and scalable self-improvement framework for vision-language procedural planning. SelfReVision enables small VLMs to iteratively critique, revise, and verify their own plans, without external supervision or teacher models, drawing inspiration from chain-of-thought prompting and self-instruct paradigms. Through this self-distillation loop, models generate higher-quality, execution-ready plans that can be used both at inference and for continued fine-tuning. Using models varying from 3B to 72B, our results show that SelfReVision not only boosts performance over weak base VLMs but also outperforms models 100X the size, yielding improved control in downstream embodied tasks.
Which components in transformer language models are responsible for discourse understanding? We hypothesize that sparse computational graphs, termed as discursive circuits, control how models process discourse relations. Unlike simpler tasks, discourse relations involve longer spans and complex reasoning. To make circuit discovery feasible, we introduce a task called Completion under Discourse Relation (CuDR), where a model completes a discourse given a specified relation. To support this task, we construct a corpus of minimal contrastive pairs tailored for activation patching in circuit discovery. Experiments show that sparse circuits (≈0.2% of a full GPT-2 model) recover discourse understanding in the English PDTB-based CuDR task. These circuits generalize well to unseen discourse frameworks such as RST and SDRT. Further analysis shows lower layers capture linguistic features such as lexical semantics and coreference, while upper layers encode discourse-level abstractions. Feature utility is consistent across frameworks (e.g., coreference supports Expansion-like relations).
With the rise of long-context language models (LMs) capable of processing tens of thousands of tokens in a single context window, do multi-stage retrieval-augmented generation (RAG) pipelines still offer measurable benefits over simpler, single-stage approaches? To assess this question, we conduct a controlled evaluation for QA tasks under systematically scaled token budgets, comparing two recent multi-stage pipelines, ReadAgent and RAPTOR, against three baselines, including DOS RAG (Document’s Original Structure RAG), a simple retrieve-then-read method that preserves original passage order. Despite its straightforward design, DOS RAG consistently matches or outperforms more intricate methods on multiple long-context QA benchmarks. We trace this strength to a combination of maintaining source fidelity and document structure, prioritizing recall within effective context windows, and favoring simplicity over added pipeline complexity. We recommend establishing DOS RAG as a simple yet strong baseline for future RAG evaluations, paired with state-of-the-art embedding and language models, and benchmarked under matched token budgets, to ensure that added pipeline complexity is justified by clear performance gains as models continue to improve.
While Vision Language Models (VLMs) learn conceptual representations, in the form of generalized knowledge, during training, they are typically used to analyze individual instances. When evaluation instances are atypical, this paradigm results in tension between two priors in the model. The first is a pragmatic prior that the textual and visual input are both relevant, arising from VLM finetuning on congruent inputs; the second is a semantic prior that the conceptual representation is generally true for instances of the category. In order to understand how VLMs trade off these priors, we introduce a new evaluation dataset, VISaGE, consisting of both typical and exceptional images. In carefully balanced experiments, we show that conceptual understanding degrades when the assumption of congruency underlying the pragmatic prior is violated with incongruent images. This effect is stronger than the effect of the semantic prior when querying about individual instances