Finding the right north-star metrics is highly critical for advancing mathematical reasoning capabilities of foundation models, especially given that existing evaluations are either too easy or only focusing on getting correct short answers. To address these issues, we present IMO-Bench, a suite of advanced reasoning benchmarks that specifically targets the level of the International Mathematical Olympiad (IMO), the most prestigious venue for young mathematicians. IMOAnswerBench first tests models on 400 diverse Olympiad problems with verifiable short answers. IMO-ProofBench is the next-level evaluation for proof-writing capabilities, which includes both basic and advanced IMO problems as well as detailed grading guidelines to facilitate automatic grading. These benchmarks played a crucial role in our historic achievement of the gold-level performance at IMO 2025 with Gemini Deep Think (Luong and Lockhart, 2025). Our model achieved 80.0% on IMO-AnswerBench and 65.7% on the advanced IMO-ProofBench, surpassing the best non-Gemini models by large margins of 6.9% and 42.4% respectively. We also showed that autograders built with Gemini reasoning correlate well with human evaluations and construct IMO-GradingBench, with 1000 human gradings on proofs, to enable further progress in automatic evaluation of long-form answers. We hope that IMO-Bench will help the community towards advancing robust mathematical reasoning and release it at https://github.com/google-deepmind/superhuman/imobench.
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
Knowledge Graph Question Answering (KGQA) aims to answer natural language questions based on knowledge graphs. Recent approaches apply the Retrieval-Augmented Generation (RAG) paradigm to incorporate Large Language Models (LLMs) to this task, where a retriever selects a question-related subgraph and an LLM-based generator is then adopted to predict answers based on the retrieved subgraph. However, the subgraph selection process is non-differentiable, preventing end-to-end training of the retriever and the generator in these approaches, which leads to sub-optimal performance. To overcome this limitation, this paper proposes a Differentiable RAG (D-RAG) approach that jointly optimizes the retriever and the generator for KGQA. Via reformulating the optimization objective as an expectation over a subgraph distribution with respect to answer generation likelihood, D-RAG makes the joint optimization feasible. Specifically, it implements this joint optimization through a differentiable subgraph sampling and prompting module that integrates Gumbel-Softmax reparameterization for sampling and a neural prompt construction process that fuses semantic and structural information. Experimental results on WebQSP and CWQ demonstrate that D-RAG outperforms state-of-the-art approaches.
Detecting hateful content is a challenging and important problem. Automated tools, like machine‐learning models, can help, but they require continuous training to adapt to the ever-changing landscape of social media. In this work, we evaluate eight open-source LLMs’ capability to detect antisemitic content, specifically leveraging in-context definition as a policy guideline. We explore various prompting techniques and design a new CoT-like prompt, Guided-CoT. Guided‐CoT handles the in-context policy well, increasing performance across all evaluated models, regardless of decoding configuration, model sizes, or reasoning capability. Notably, Llama 3.1 70B outperforms fine-tuned GPT-3.5. Additionally, we examine LLM errors and introduce metrics to quantify semantic divergence in model-generated rationales, revealing notable differences and paradoxical behaviors among LLMs. Our experiments highlight the differences observed across LLMs’ utility, explainability, and reliability.
Methods for lexical semantic-change detection quantify changes in the meaning of words over time. Prior methods have excelled on established benchmarks consisting of pre-selected target words, chosen ahead of time due to the prohibitive cost of manually annotating all words. However, performance measured on small curated wordsets cannot reveal how well these methods perform at discovering semantic changes among the full corpus vocabulary, which is the actual end goal for many applications. In this paper, we implement a top-k setup to evaluate semantic-change discovery despite lacking complete annotations. (At the same time, we also extend the annotations in the commonly used LiverpoolFC and SemEval-EN benchmarks by 85% and 90%, respectively). We deploy our evaluation setup on a battery of semantic-change detection methods under multiple variations. We find that when presented with a natural distribution of instances, all the methods struggle at ranking known large changes higher than other words in the vocabulary. Furthermore, we manually verify that the majority of words with high detected-change scores in LiverpoolFC do not actually experience meaning changes. In fact, for most of the methods, less than a half of the highest-ranked changes were determined to have changed in meaning. Given the large performance discrepancies between existing benchmark results and discovery “in the wild”, we recommend that researchers direct more attention to semantic-change discovery and include it in their suite of evaluations. Our annotations and code for running evaluations are available at https://github.com/khonzoda/semantic-change-discovery-emnlp2025.
Layer-wise Minimal Pair Probing Reveals Contextual Grammatical-Conceptual Hierarchy in Speech Representations
PDF ↗Transformer-based speech language models (SLMs) have significantly improved neural speech recognition and understanding. While existing research has examined how well SLMs encode shallow acoustic and phonetic features, the extent to which SLMs encode nuanced syntactic and conceptual features remains unclear. By drawing parallels with linguistic competence assessments for large language models, this study is the first to systematically evaluate the presence of contextual syntactic and semantic features across SLMs for self-supervised learning (S3M), automatic speech recognition (ASR), speech compression (codec), and as the encoder for auditory large language models (AudioLLMs). Through minimal pair designs and diagnostic feature analysis across 71 tasks spanning diverse linguistic levels, our layer-wise and time-resolved analysis uncovers that 1) all speech encode grammatical features more robustly than conceptual ones. 2) Despite never seeing text, S3M match or surpass ASR encoders on every linguistic level, demonstrating that rich grammatical and even conceptual knowledge can arise purely from audio. 3) S3M representations peak mid-network and then crash in the final layers, whereas ASR and AudioLLM encoders maintain or improve, reflecting how pre-training objectives reshape late-layer content. 4) Temporal probing further shows that S3Ms encode grammatical cues 500 ms before a word begins, whereas AudioLLMs distribute evidence more evenly—indicating that objectives shape not only where but also when linguistic information is most salient. Together, these findings establish the first large-scale map of contextual syntax and semantics in speech models and highlight both the promise and the limits of current SLM training paradigms.
The memorization of training data in large language models (LLMs) poses significant privacy and copyright concerns. Existing data extraction methods, particularly heuristic-based divergence attacks, often exhibit limited success and offer limited insight into the fundamental drivers of memorization leakage. This paper introduces Confusion-Inducing Attacks (CIA), a principled framework for extracting memorized data by systematically maximizing model uncertainty. We empirically demonstrate that the emission of memorized text during divergence is preceded by a sustained spike in token-level prediction entropy. CIA leverages this insight by optimizing input snippets to deliberately induce this consecutive high-entropy state. For aligned LLMs, we further propose Mismatched Supervised Fine-tuning (SFT) to simultaneously weaken their alignment and induce targeted confusion, thereby increasing susceptibility to our attacks. Experiments on various unaligned and aligned LLMs demonstrate that our proposed attacks outperform existing baselines in extracting verbatim and near-verbatim training data without requiring prior knowledge of the training data. Our findings highlight persistent memorization risks across various LLMs and offer a more systematic method for assessing these vulnerabilities.
Not Your Typical Government Tipline: LLM-Assisted Routing of Environmental Protection Agency Citizen Tips
PDF ↗Regulatory agencies often operate with limited resources and rely on tips from the public to identify potential violations. However, processing these tips at scale presents significant operational challenges, as agencies must correctly identify and route relevant tips to the appropriate enforcement divisions. Through a case study, we demonstrate how advances in large language models can be utilized to support overburdened agencies with limited capacities. In partnership with the U.S. Environmental Protection Agency, we leverage previously unstudied citizen tips data from their “Report a Violation” system to develop an LLM-assisted pipeline for tip routing. Our approach filters out 80.5% of irrelevant tips and increases overall routing accuracy from 31.8% to 82.4% compared to the current routing system. At a time of increased focus on government efficiencies, our approach provides a constructive path forward by using technology to empower civil servants.
LLMs are increasingly applied in the legal domain in tasks such as summarizing legal texts and providing basic legal advice. Yet, their capacity to draft full judicial analyses in U.S. court opinions is still largely uncharted, such as generating entire judicial reasoning sections in U.S. court decisions, remain under-explored. Given the continued adoption of LLMs and the significance of law to society at large, measurement of LLM’s legal reasoning capabilities is a pressing task. We propose CourtReasoner, a novel expert-annotated judicial reasoning benchmark for evaluating LLM agents’ capabilities in complex legal reasoning. Sourcing U.S. court opinions, we construct benchmarks that measure the LLMs ability to construct goal-oriented legal reasoning. CourtReasoner measured the agent’s ability to argue both ways in a legal dispute, rather than simple Q/A. Our results show that more than 60% of frontier model outputs contain invalid arguments and more than 53% of frontier model produced irrelevant citations when conducting complex legal reasoning. We also introduce a meta-evaluation benchmark to provide insights into the capabilities of LLMs as evaluators of legal reasoning. We will release our data, code and full annotation guidelines publicly for future research.
SimulatorArena: Are User Simulators Reliable Proxies for Multi-Turn Evaluation of AI Assistants?
PDF ↗Large language models (LLMs) are increasingly used in interactive applications, and human evaluation remains the gold standard for assessing their performance in multi-turn conversations. Since human studies are costly, time-consuming, and hard to reproduce, recent work explores using LLMs to simulate users for automatic assistant evaluation. However, there is no benchmark or systematic study to evaluate whether these simulated users are reliable stand-ins for real users. To address this, we introduce SimulatorArena, a benchmark of 909 annotated human–LLM conversations on two interactive tasks—math tutoring and document creation. SimulatorArena evaluates simulators based on how closely their messages match human behavior and how well their assistant ratings align with human judgments. Experiments on various simulator methods show that simulators conditioned on user profiles, capturing traits like background and message styles, align closely with human judgments. They reach Spearman’s \rho of 0.7 on both tasks, providing a practical, scalable alternative to human evaluation. Using the best simulator for each task, we benchmark 18 assistants, including the latest LLMs such as GPT-5, Claude 4.1 Opus, and Gemini 2.5 Pro.
Large language models (LLMs) excel at many NLP tasks, yet their multi-step logical reasoning remains unreliable. Existing solutions such as Chain-of-Thought prompting generate intermediate steps but provide no internal check of their logical coherence. In this paper, we use the “QK-score”, a lightweight metric based on query–key alignments within transformer attention heads, to evaluate the logical reasoning capabilities of LLMs. Our method automatically identifies attention heads that play a key role in distinguishing valid from invalid logical inferences, enabling efficient inference-time evaluation via a single forward pass. It reveals latent reasoning structure in LLMs and provides a scalable mechanistic alternative to ablation-based analysis. Across three benchmarks: ProntoQA-OOD, PARARULE-Plus, and MultiLogicEval, with models ranging from 1.5B to 70B parameters, the selected heads predict logical validity up to 14% better than the model probabilities, and remain robust under distractors and increasing reasoning depth of d\leq 6.
The exposure of large language models (LLMs) to copyrighted material during pre-training raises concerns about unintentional copyright infringement post deployment. This has driven the development of “copyright takedown” methods—post-training approaches aimed at preventing models from generating content substantially similar to copyrighted ones. While current mitigation approaches are somewhat effective for average-case risks, we demonstrate that they overlook worst-case copyright risks exhibited by the existence of long, verbatim quotes from copyrighted sources. We propose BloomScrub, a remarkably simple yet highly effective inference-time approach that provides certified copyright takedown. Our method repeatedly interleaves quote detection with rewriting techniques to transform potentially infringing segments. By leveraging efficient data sketches (Bloom filters), our approach enables scalable copyright screening—even for large-scale real-world corpora. When quotes beyond a length threshold cannot be removed, the system can abstain from responding, offering certified risk reduction. Experimental results show that BloomScrub reduces infringement risk, preserves utility, and accommodates different levels of enforcement stringency with adaptive abstention. Our results suggest that lightweight, inference-time methods can be surprisingly effective for copyright prevention.
Whitespace is a critical component of poetic form, reflecting both adherence to standardized forms and rebellion against those forms. Each poem’s whitespace distribution reflects the artistic choices of the poet and is an integral semantic and spatial feature of the poem. Yet, despite the popularity of poetry as both a long-standing art form and as a generation task for large language models (LLMs), whitespace has not received sufficient attention from the NLP community. Using a corpus of 19k English-language published poems from Poetry Foundation, we investigate how 4k poets have used whitespace in their works. We release a subset of 2.8k public-domain poems with preserved formatting to facilitate further research in this area. We compare whitespace usage in the published poems to (1) 51k LLM-generated poems, and (2) 12k unpublished poems posted in an online community. We also explore whitespace usage across time periods, poetic forms, and data sources. Additionally, we find that different text processing methods can result in significantly different representations of whitespace in poetry data, motivating us to use these poems and whitespace patterns to discuss implications for the processing strategies used to assemble pretraining datasets for LLMs.
Hawaiian orthography employs two distinct spelling systems, both of which are used by communities of speakers today. These two spelling systems are distinguished by the presence of the ‘okina letter and kahakō diacritic, which represent glottal stops and long vowels, respectively. We develop several models ranging in complexity to convert between these two orthographies. Our results demonstrate that simple statistical n-gram models surprisingly outperform neural seq2seq models and LLMs, highlighting the potential for traditional machine learning approaches in a low-resource setting.
A Simple Yet Effective Method for Non-Refusing Context Relevant Fine-grained Safety Steering in LLMs
PDF ↗Fine-tuning large language models (LLMs) to meet evolving safety policies is costly and impractical. Mechanistic interpretability enables inference-time control through latent activation steering, but its potential for precise, customizable safety adjustments remains underexplored. We propose SafeSteer, a simple and effective method to guide LLM outputs by (i) leveraging category-specific steering vectors for fine-grained control, (ii) applying a gradient-free, unsupervised approach that enhances safety while preserving text quality and topic relevance without forcing explicit refusals, and (iii) eliminating the need for contrastive safe data. Across multiple LLMs, datasets, and risk categories, SafeSteer provides precise control, avoids blanket refusals, and directs models to generate safe, relevant content, aligning with recent findings that simple activation-steering techniques often outperform more complex alternatives.
The detection of audio deepfakes (ADD) has become increasingly important due to the rapid evolution of generative speech models. However, progress in this field remains uneven across languages, particularly for low-resource languages like Portuguese, which lack high-quality datasets. In this paper, we introduce BRSpeech-DF, the first publicly available ADD dataset for Portuguese, encompassing both Brazilian and European variants. The dataset contains over 458,000 utterances, including a smaller portion of real speech from 62 speakers and a large collection of synthetic samples generated using multiple zero-shot text-to-speech (TTS) models, each conditioned on the original speaker’s voice. By providing this resource, our objective is to support the development of robust, multilingual detection systems, thereby advancing equity in speech forensics and security research. BRSpeech-DF addresses a significant gap in annotated data for underrepresented languages, facilitating more inclusive and generalizable advancements in synthetic speech detection.
Large language models show promising performance on reasoning tasks, yet evaluation methods for low-resource languages remain limited, particularly for complex STEM problem-solving. We introduce Trojsten Benchmark, a Slovak-language dataset of 1,108 high-school competition problems with reference solutions across mathematics, physics, and programming, and a rubric-based LLM grading framework. Using GPT-4 to generate rubrics and grade solutions, we observe 1.05 average absolute deviation from human graders (5-point scale), while benchmarking GPT-3.5-Turbo, GPT-4, GPT-4o, and open-weight models (Llama 3, Phi-3). We quantify multistep reasoning performance by difficulty, show consistent underperformance on harder items, and demonstrate language sensitivity: accuracy drops on English translations of Slovak statements, evidencing challenges beyond translation. Trojsten Benchmark complements English-centric math datasets (e.g., MATH, GSM8K) by targeting open-response, rubric-gradable reasoning under low-resource linguistic framing. We release code and data to enable reproducible evaluation and human-aligned auto-grading for STEM in under-served languages.
We introduce a novel Question Answering (QA) architecture that enhances answer selection by retrieving targeted supporting evidence. Unlike traditional methods, which retrieve documents or passages relevant only to a query q, our approach retrieves content relevant to the combined pair (q, a), explicitly emphasizing the supporting relation between the query and a candidate answer a. By prioritizing this relational context, our model effectively identifies paragraphs that directly substantiate the correctness of a with respect to q, leading to more accurate answer verification than standard retrieval systems. Our neural retrieval method also scales efficiently to collections containing hundreds of millions of paragraphs. Moreover, this approach can be used by large language models (LLMs) to retrieve explanatory paragraphs that ground their reasoning, enabling them to tackle more complex QA tasks with greater reliability and interpretability.
Large language models have made tremendous progress in recent years, but low-resource languages, like Tibetan, remain significantly underrepresented in their evaluation. Despite Tibetan being spoken by over seven million people, it has largely been neglected in the development and assessment of LLMs. To address this gap, we present a Tibetan Language Understanding Evaluation Benchmark, TLUE, which is also the first large-scale benchmark for measuring the proficiency of large language models in the Tibetan language. TLUE comprises two major components: a comprehensive multi-task understanding benchmark spanning 5 domains and 67 subdomains, and a safety benchmark encompassing 7 subdomains. Finally, we evaluate a diverse set of state-of-the-art LLMs. Experimental results demonstrate that most large language models perform below the random baseline, especially highlighting the considerable challenges they face in Tibetan language processing. TLUE provides a crucial foundation for advancing future research in Tibetan language understanding and highlights the importance of promoting greater inclusivity in the development of large language models.
Spectral Scaling Laws in Language Models: emphHow Effectively Do Feed-Forward Networks Use Their Latent Space?
PDF ↗As Large Language Models (LLMs) scale, the question is not just how large they become, but how much of their capacity is effectively utilized. Existing scaling laws relate model size to loss, yet overlook how components exploit their latent space. In this work, we focus on Feed-Forward Networks (FFNs) and recast width selection as a spectral utilization optimization problem. Using a lightweight diagnostic suite: Hard Rank (participation ratio), Soft Rank (Shannon Rank), Spectral Concentration, and the composite Spectral Utilization Index (SUI), we quantify how many latent directions are meaningfully activated across LLaMA, GPT-2, and nGPT families. Our key finding is an Asymmetric Spectral Scaling Law: soft rank follows an almost perfect power law with FFN width, while hard rank grows only sublinearly, with high variance. This asymmetry suggests that widening FFNs mostly adds low-energy tail directions, while dominant-mode subspaces saturate early. Moreover, at larger widths, variance further collapses into a narrow subspace, leaving much of the latent space under-utilized. These results recast FFN width selection as a principled trade-off between tail capacity and dominant-mode capacity, offering concrete guidance for inference-efficient LLM design.
Standard Byte-Pair Encoding (BPE) tokenization compresses text by pairing a learned token vocabulary with a detailed merge list. Recent work has shown that this merge list exposes a potential attack surface for extracting information about language model’s training data. In this paper, we explore the downstream impact of BPE inference algorithms that do not rely on this merge list at all, and hence differ from the encoding process during the BPE training. To address this question, we investigate two broad classes of BPE inference schemes that differ from BPE application during training: a) targetted deviation from merge-lists including random merge orders, and various corruptions of merge list involving deletion/truncation, and b) non-targetted BPE inference algorithms that do not depend on the merge list but focus on compressing the text either greedily or exactly. Extensive experiments across diverse language modeling tasks like accuracy-based QA bench- marks, machine translation, and open-ended generation reveal that while the targetted deviation from the merge lists exhibit significant degradation in language model performance, the non-targetted merge-list free inference algorithms result in minimal impact on downstream performance that is often much smaller than expected. These findings pave way for simpler and potentially more privacy-preserving tokenization schemes that do not catastrophically compromise model performance.