Papers

Large Language Models (LLMs) have shown impressive capabilities in complex reasoning tasks. However, current approaches employ uniform language density for both intermediate reasoning and final answers, leading to computational inefficiency. Our observation found that reasoning process serves a computational function for the model itself, while answering serves a communicative function for human understanding. This distinction enables the use of compressed, symbol-rich language for intermediate computations while maintaining human-readable final explanations. To address this inefficiency, we present Denser: \underline{D}ual-d\underline{ens}ity inf\underline{er}ence, a novel framework that optimizes information density separately for reasoning and answering phases. Our framework implements this through three components: a query processing module that analyzes input problems, a high-density compressed reasoning mechanism for efficient intermediate computations, and an answer generation component that translates compressed reasoning into human-readable solutions. Experimental evaluation across multiple reasoning question answering benchmarks demonstrates that Denser reduces token consumption by up to 62% compared to standard Chain-of-Thought methods while preserving or improving accuracy. These efficiency gains are particularly significant for complex multi-step reasoning problems where traditional methods generate extensive explanations.

Recent evidence shows that the versification of prompts constitutes a highly effective adversarial mechanism against aligned LLMs. The study 'Adversarial poetry as a universal single-turn jailbreak mechanism in large language models' demonstrates that instructions routinely refused in prose become executable when rewritten as verse, producing up to 18 x more safety failures in benchmarks derived from MLCommons AILuminate. Manually written poems reach approximately 62% ASR, and automated versions 43%, with some models surpassing 90% success in single-turn interactions. The effect is structural: systems trained with RLHF, constitutional AI, and hybrid pipelines exhibit consistent degradation under minimal semiotic formal variation. Versification displaces the prompt into sparsely supervised latent regions, revealing guardrails that are excessively dependent on surface patterns. This dissociation between apparent robustness and real vulnerability exposes deep limitations in current alignment regimes. The absence of evaluations in Portuguese, a language with high morphosyntactic complexity, a rich metric-prosodic tradition, and over 250 million speakers, constitutes a critical gap. Experimental protocols must parameterise scansion, metre, and prosodic variation to test vulnerabilities specific to Lusophone patterns, which are currently ignored.

This study applies Brookes' Measure of Categorical Dispersion ({\Delta}) to analyze the thematic structure of contemporary Arabic Applied Linguistics research. Using a comprehensive, real-world dataset of 1,564 publications from 2019 to 2025, classified into eight core sub-disciplines, we calculate a dispersion index of {\Delta} = 0.194. This remarkably low value indicates extreme thematic dispersion, revealing that the field is characterized by pronounced heterogeneity rather than concentration. The analysis identifies Computational Linguistics as a dominant but non-hegemonic force, coexisting with robust research in Sociolinguistics, Language Teaching, and other subfields. This study clarifies the correct application of Brookes' original formula, demonstrates its utility for field characterization, and provides a replicable bibliometric methodology for assessing disciplinary structure across domains.

Extracting structured information from zeolite synthesis experimental procedures is critical for materials discovery, yet existing methods have not systematically evaluated Large Language Models (LLMs) for this domain-specific task. This work addresses a fundamental question: what is the efficacy of different prompting strategies when applying LLMs to scientific information extraction? We focus on four key subtasks: event type classification (identifying synthesis steps), trigger text identification (locating event mentions), argument role extraction (recognizing parameter types), and argument text extraction (extracting parameter values). We evaluate four prompting strategies - zero-shot, few-shot, event-specific, and reflection-based - across six state-of-the-art LLMs (Gemma-3-12b-it, GPT-5-mini, O4-mini, Claude-Haiku-3.5, DeepSeek reasoning and non-reasoning) using the ZSEE dataset of 1,530 annotated sentences. Results demonstrate strong performance on event type classification (80-90% F1) but modest performance on fine-grained extraction tasks, particularly argument role and argument text extraction (50-65% F1). GPT-5-mini exhibits extreme prompt sensitivity with 11-79% F1 variation. Notably, advanced prompting strategies provide minimal improvements over zero-shot approaches, revealing fundamental architectural limitations. Error analysis identifies systematic hallucination, over-generalization, and inability to capture synthesis-specific nuances. Our findings demonstrate that while LLMs achieve high-level understanding, precise extraction of experimental parameters requires domain-adapted models, providing quantitative benchmarks for scientific information extraction.

The deployment of Large Language Models (LLMs) in interactive systems necessitates a deep alignment with the nuanced and dynamic preferences of individual users. Current alignment techniques predominantly address universal human values or static, single-turn preferences, thereby failing to address the critical needs of long-term personalization and the initial user cold-start problem. To bridge this gap, we propose PersonalAgent, a novel user-centric lifelong agent designed to continuously infer and adapt to user preferences. PersonalAgent constructs and dynamically refines a unified user profile by decomposing dialogues into single-turn interactions, framing preference inference as a sequential decision-making task. Experiments show that PersonalAgent achieves superior performance over strong prompt-based and policy optimization baselines, not only in idealized but also in noisy conversational contexts, while preserving cross-session preference consistency. Furthermore, human evaluation confirms that PersonalAgent excels at capturing user preferences naturally and coherently. Our findings underscore the importance of lifelong personalization for developing more inclusive and adaptive conversational agents. Our code is available here.

The rapid development of Generative AI is bringing innovative changes to education and assessment. As the prevalence of students utilizing AI for assignments increases, concerns regarding academic integrity and the validity of assessments are growing. This study utilizes the Earth Science I section of the 2025 Korean College Scholastic Ability Test (CSAT) to deeply analyze the multimodal scientific reasoning capabilities and cognitive limitations of state-of-the-art Large Language Models (LLMs), including GPT-4o, Gemini 2.5 Flash, and Gemini 2.5 Pro. Three experimental conditions (full-page input, individual item input, and optimized multimodal input) were designed to evaluate model performance across different data structures. Quantitative results indicated that unstructured inputs led to significant performance degradation due to segmentation and Optical Character Recognition (OCR) failures. Even under optimized conditions, models exhibited fundamental reasoning flaws. Qualitative analysis revealed that "Perception Errors" were dominant, highlighting a "Perception-Cognition Gap" where models failed to interpret symbolic meanings in schematic diagrams despite recognizing visual data. Furthermore, models demonstrated a "Calculation-Conceptualization Discrepancy," successfully performing calculations while failing to apply the underlying scientific concepts, and "Process Hallucination," where models skipped visual verification in favor of plausible but unfounded background knowledge. Addressing the challenge of unauthorized AI use in coursework, this study provides actionable cues for designing "AI-resistant questions" that target these specific cognitive vulnerabilities. By exploiting AI's weaknesses, such as the gap between perception and cognition, educators can distinguish genuine student competency from AI-generated responses, thereby ensuring assessment fairness.

Reinforcement Learning with Verifiable Rewards (RLVR) significantly enhances the reasoning capability of Large Language Models (LLMs). Current RLVR approaches typically conduct training across all generated tokens, but neglect to explore which tokens (e.g., prefix tokens) actually contribute to reasoning. This uniform training strategy spends substantial effort on optimizing low-return tokens, which in turn impedes the potential improvement from high-return tokens and reduces overall training effectiveness. To address this issue, we propose a novel RLVR approach called Progressive Prefix-token Policy Optimization (PPPO), which highlights the significance of the prefix segment of generated outputs. Specifically, inspired by the well-established human thinking theory of Path Dependence, where early-stage thoughts substantially constrain subsequent thinking trajectory, we identify an analogous phenomenon in LLM reasoning termed Beginning Lock-in Effect (BLE). PPPO leverages this finding by focusing its optimization objective on the prefix reasoning process of LLMs. This targeted optimization strategy can positively influence subsequent reasoning processes, and ultimately improve final results. To improve the learning effectiveness of LLMs on how to start reasoning with high quality, PPPO introduces two training strategies: (a) Progressive Prefix Retention, which shapes a progressive learning process by increasing the proportion of retained prefix tokens during training; (b) Continuation Accumulated Reward, which mitigates reward bias by sampling multiple continuations for one prefix token sequence, and accumulating their scores as the reward signal. Extensive experimental results on various reasoning tasks demonstrate that our proposed PPPO outperforms representative RLVR methods, with the accuracy improvements of 18.02% on only 26.17% training tokens.

We introduce SynGP500, a clinician-curated collection of 500 synthetic Australian general practice medical notes. The dataset integrates curriculum-based clinical breadth (RACGP 2022 Curriculum), epidemiologically-calibrated prevalence (BEACH study), and diverse consultation contexts. This approach systematically includes both common presentations and less-common curriculum-specified conditions that GPs must recognize but appear infrequently in single practice populations, potentially supporting more generalizable model training than datasets constrained by naturally occurring case distributions. SynGP500 is messy by design, reflecting the authentic complexity of healthcare delivery: telegraphic documentation, typos, patient non-adherence, socioeconomic barriers, and clinician-patient disagreements, unlike sanitized synthetic datasets that obscure clinical realities. Multi-faceted validation demonstrates dataset quality through epidemiological alignment with real Australian GP consultation patterns (BEACH study), stylometric analysis confirming high linguistic variation, semantic diversity analysis demonstrating broad coverage, and exploratory downstream evaluation using self-supervised medical concept extraction, showing F1 improvements. SynGP500 addresses a critical national gap, providing researchers and educators with a resource for developing and evaluating clinical NLP methods for Australian general practice while inherently protecting patient privacy.

Moralizations - arguments that invoke moral values to justify demands or positions - are a yet underexplored form of persuasive communication. We present the Moralization Corpus, a novel multi-genre dataset designed to analyze how moral values are strategically used in argumentative discourse. Moralizations are pragmatically complex and often implicit, posing significant challenges for both human annotators and NLP systems. We develop a frame-based annotation scheme that captures the constitutive elements of moralizations - moral values, demands, and discourse protagonists - and apply it to a diverse set of German texts, including political debates, news articles, and online discussions. The corpus enables fine-grained analysis of moralizing language across communicative formats and domains. We further evaluate several large language models (LLMs) under varied prompting conditions for the task of moralization detection and moralization component extraction and compare it to human annotations in order to investigate the challenges of automatic and manual analysis of moralizations. Results show that detailed prompt instructions has a greater effect than few-shot or explanation-based prompting, and that moralization remains a highly subjective and context-sensitive task. We release all data, annotation guidelines, and code to foster future interdisciplinary research on moral discourse and moral reasoning in NLP.

LLMs trained on web-scale data raise concerns about privacy and the right to be forgotten. To address these issues, Machine Unlearning provides techniques to remove specific information from trained models without retraining from scratch. However, existing benchmarks for evaluating unlearning in LLMs face two major limitations: they focus only on English and support only entity-level forgetting (removing all information about a person). We introduce FAME (Fictional Actors for Multilingual Erasure), a synthetic benchmark for evaluating Machine Unlearning across five languages: English, French, German, Italian, and Spanish. FAME contains 1,000 fictional actor biographies and 20,000 question-answer pairs. Each biography includes information on 20 topics organized into structured categories (biography, career, achievements, personal information). This design enables both entity-level unlearning (i.e., forgetting entire identities) and instance-level unlearning (i.e., forgetting specific facts while retaining others). We provide two dataset splits to support these two different unlearning scenarios and enable systematic comparison of unlearning techniques across languages. Since FAME uses entirely fictional data, it ensures that the information was never encountered during model pretraining, allowing for a controlled evaluation of unlearning methods.

This paper presents the systems submitted by the Yes-MT team for the Low-Resource Indic Language Translation Shared Task at WMT 2024 (Pakray et al., 2024), focusing on translating between English and the Assamese, Mizo, Khasi, and Manipuri languages. The experiments explored various approaches, including fine-tuning pre-trained models like mT5 (Xue et al., 2020) and IndicBart (Dabre et al., 2021) in both multilingual and monolingual settings, LoRA (Hu et al., 2021) fine-tuning IndicTrans2 (Gala et al., 2023), zero-shot and few-shot prompting (Brown, 2020) with large language models (LLMs) like Llama 3 (Dubey et al., 2024) and Mixtral 8x7b (Jiang et al., 2024), LoRA supervised fine-tuning of Llama 3 (Mecklenburg et al., 2024), and training Transformer models (Vaswani, 2017) from scratch. The results were evaluated on the WMT23 Low-Resource Indic Language Translation Shared Task test data using SacreBLEU (Post, 2018) and CHRF (Popovic, 2015), highlighting the challenges of low-resource translation and the potential of LLMs for these tasks, particularly with fine-tuning.

Large language models (LLMs) often generate hallucinations in knowledge-intensive QA due to parametric knowledge limitations. While existing methods like KG-CoT improve reliability by integrating knowledge graph (KG) paths, they suffer from rigid hop-count selection (solely question-driven) and underutilization of reasoning paths (lack of guidance). To address this, we propose RFKG-CoT: First, it replaces the rigid hop-count selector with a relation-driven adaptive hop-count selector that dynamically adjusts reasoning steps by activating KG relations (e.g., 1-hop for direct "brother" relations, 2-hop for indirect "father-son" chains), formalized via a relation mask. Second, it introduces a few-shot in-context learning path guidance mechanism with CoT (think) that constructs examples in a "question-paths-answer" format to enhance LLMs' ability to understand reasoning paths. Experiments on four KGQA benchmarks show RFKG-CoT improves accuracy by up to 14.7 pp (Llama2-7B on WebQSP) over KG-CoT. Ablations confirm the hop-count selector and the path prompt are complementary, jointly transforming KG evidence into more faithful answers.

This paper explores the growing epistemic parallel between NLG evaluation and grading of students in a Finnish University. We argue that both domains are experiencing a Great Misalignment Problem. As students increasingly use tools like ChatGPT to produce sophisticated outputs, traditional assessment methods that focus on final products rather than learning processes have lost their validity. To address this, we introduce the Pedagogical Multi-Factor Assessment (P-MFA) model, a process-based, multi-evidence framework inspired by the logic of multi-factor authentication.

Large language models (LLMs) are evolving into agentic systems that reason, plan, and operate external tools. The Model Context Protocol (MCP) is a key enabler of this transition, offering a standardized interface for connecting LLMs with heterogeneous tools and services. Yet MCP's openness and multi-server workflows introduce new safety risks that existing benchmarks fail to capture, as they focus on isolated attacks or lack real-world coverage. We present MCP-SafetyBench, a comprehensive benchmark built on real MCP servers that supports realistic multi-turn evaluation across five domains: browser automation, financial analysis, location navigation, repository management, and web search. It incorporates a unified taxonomy of 20 MCP attack types spanning server, host, and user sides, and includes tasks requiring multi-step reasoning and cross-server coordination under uncertainty. Using MCP-SafetyBench, we systematically evaluate leading open- and closed-source LLMs, revealing that all models remain vulnerable to MCP attacks, with a notable safety-utility trade-off. Our results highlight the urgent need for stronger defenses and establish MCP-SafetyBench as a foundation for diagnosing and mitigating safety risks in real-world MCP deployments.

This paper presents a large-scale corpus of Rakuten Travel Reviews. Our collection contains 7.29 million customer reviews for 16 years, ranging from 2009 to 2024. Each record in the dataset contains the review text, its response from an accommodation, an anonymized reviewer ID, review date, accommodation ID, plan ID, plan title, room type, room name, purpose, accompanying group, and user ratings from six aspect categories, as well as an overall score. We present statistical information about our corpus and provide insights into factors driving data drift between 2019 and 2024 using statistical approaches.

Test-time reinforcement learning mitigates the reliance on annotated data by using majority voting results as pseudo-labels, emerging as a complementary direction to reinforcement learning with verifiable rewards (RLVR) for improving reasoning ability. However, this voting strategy often induces confirmation bias and suffers from sparse rewards, limiting the overall performance. In this work, we propose subgroup-specific step-wise confidence-weighted pseudo-label estimation (SCOPE), a framework integrating model confidence and dynamic subgroup partitioning to address these issues. Specifically, SCOPE integrates the proposed step-wise confidence into pseudo label estimation, prioritizing high-quality reasoning paths over simple frequency count. Furthermore, it dynamically partitions the candidate outputs pool into independent subgroups by balancing reasoning quality against exploration diversity. By deriving local consensus via repeat sampling for each sub group, SCOPE provides diverse supervision targets to encourage broader exploration. We conduct experiments across various models and benchmarks, experimental results show that SCOPE consistently outperforms recent baselines. Notably, SCOPE achieving relative improvements of 13.1% on challenging AIME 2025 and 8.1% on AMC. The code is released at https://github.com/szu-tera/SCOPE.

A goal of interpretability is to recover disentangled representations of latent concepts (features) from the activations of neural networks. The quality of features is typically evaluated in isolation, and under implicit independence assumptions that may not hold in practice. Thus, it is unclear to what extent common featurization methods such as sparse autoencoders (SAEs) and probes disentangle one concept from another. We propose a multi-concept evaluation setting using concepts including sentiment, domain, voice, and tense. We evaluate how well featurizers produce disentangled representations of each concept, observing that features are typically sensitive to only one concept, but also that concepts are distributed across many features. Then, we steer these features, measuring whether each concept is independently manipulable, and whether features interact. Even in idealized settings, steering a feature often affects many concepts, despite a near absence of interaction effects. These results suggest that correlational metrics are insufficient to establish steering selectivity, and that demonstrating that two features operate in separate spaces is insufficient to claim that they will be selective for one concept. These results underscore the importance of multi-concept evaluations in interpretability research.

Although large language models (LLMs) are increasingly used in security-critical workflows, practitioners lack quantitative guidance on which safeguards are worth deploying. This paper introduces a decision-oriented framework and reproducible methodology that together quantify residual risk, convert adversarial probe outcomes into financial risk estimates and return-on-control (RoC) metrics, and enable monetary comparison of layered defenses for LLM-based systems. A retrieval-augmented generation (RAG) service is instantiated using the DeepSeek-R1 model over a corpus containing synthetic personally identifiable information (PII), and subjected to automated attacks with Garak across five vulnerability classes: PII leakage, latent context injection, prompt injection, adversarial attack generation, and divergence. For each (vulnerability, control) pair, attack success probabilities are estimated via Laplace's Rule of Succession and combined with loss triangle distributions, calibrated from public breach-cost data, in 10,000-run Monte Carlo simulations to produce loss exceedance curves and expected losses. Three widely used mitigations, attribute-based access control (ABAC); named entity recognition (NER) redaction using Microsoft Presidio; and NeMo Guardrails, are then compared to a baseline RAG configuration. The baseline system exhibits very high attack success rates (>= 0.98 for PII, latent injection, and prompt injection), yielding a total simulated expected loss of $313k per attack scenario. ABAC collapses success probabilities for PII and prompt-related attacks to near zero and reduces the total expected loss by ~94%, achieving an RoC of 9.83. NER redaction likewise eliminates PII leakage and attains an RoC of 5.97, while NeMo Guardrails provides only marginal benefit (RoC of 0.05).

Retrieval-Augmented Generation (RAG) systems remain susceptible to hallucinations despite grounding in retrieved evidence. While current detection methods leverage embedding similarity and natural language inference (NLI), their reliability in safety-critical settings remains unproven. We apply conformal prediction to RAG hallucination detection, transforming heuristic scores into decision sets with finite-sample coverage guarantees (1-alpha). Using calibration sets of n=600, we demonstrate a fundamental dichotomy: on synthetic hallucinations (Natural Questions), embedding methods achieve 95% coverage with 0% False Positive Rate (FPR). However, on real hallucinations from RLHF-aligned models (HaluEval), the same methods fail catastrophically, yielding 100% FPR at target coverage. We analyze this failure through the lens of distributional tails, showing that while NLI models achieve acceptable AUC (0.81), the "hardest" hallucinations are semantically indistinguishable from faithful responses, forcing conformal thresholds to reject nearly all valid outputs. Crucially, GPT-4 as a judge achieves 7% FPR (95% CI:[3.4%, 13.7%]) on the same data, proving the task is solvable via reasoning but opaque to surface-level semantics–a phenomenon we term the "Semantic Illusion."

The transition of Large Language Models (LLMs) from stochastic chat interfaces to reliable software components necessitates a fundamental re-engineering of interaction paradigms. Current methodologies, predominantly heuristic-based "prompt engineering," fail to provide the deterministic guarantees required for mission-critical applications. We introduce the Meta-Prompting Protocol, a rigorous theoretical framework that formalizes the orchestration of LLMs as a programmable, self-optimizing system. Central to this protocol is the Adversarial Trinity, a tripartite topology comprising a Generator (P), an Auditor (A), and an Optimizer (O). By treating natural language instructions as differentiable variables within a semantic computation graph and utilizing textual critiques as gradients, this architecture mitigates hallucination and prevents model collapse. We demonstrate the theoretical viability of this approach using declarative programming paradigms (DSPy) and automatic textual differentiation (TextGrad), establishing a foundation for "Observable Software Engineering" in the era of probabilistic computing.

Disclaimer: Samples in this paper may be harmful and cause discomfort. Multimodal large language models (MLLMs) enable multimodal understanding but inherit toxic signals from weakly curated pretraining corpora, leading to explicitly toxic outputs, especially under adversarial triggers that late, opaque training-free detoxification methods struggle to handle. We propose SGM, a white-box neuron-level multimodal intervention that acts like safety glasses for toxic neurons: it recalibrates a set of toxicity-associated neurons via expertise-weighted soft suppression, neutralizing harmful cross-modal activations without any parameter updates. We establish MM-TOXIC-QA, a multimodal toxicity data framework, and compare SGM with existing detoxification techniques. Experiments on open-source MLLMs show that SGM mitigates explicit toxicity in standard and adversarial conditions, cutting average harmful rates from 45.0% to 4.5% while preserving fluency and multimodal reasoning. SGM is extensible, and its combined defenses, denoted as SGM*, integrate with existing detoxification methods for stronger safety performance, providing an interpretable, low-cost solution for toxicity-controlled multimodal generation.

3D Scene Graphs (3DSGs) constitute a powerful representation of the physical world, distinguished by their abilities to explicitly model the complex spatial, semantic, and functional relationships between entities, rendering a foundational understanding that enables agents to interact intelligently with their environment and execute versatile behaviors. Embodied navigation, as a crucial component of such capabilities, leverages the compact and expressive nature of 3DSGs to enable long-horizon reasoning and planning in complex, large-scale environments. However, prior works rely on a static-world assumption, defining traversable space solely based on static spatial layouts and thereby treating interactable obstacles as non-traversable. This fundamental limitation severely undermines their effectiveness in real-world scenarios, leading to limited reachability, low efficiency, and inferior extensibility. To address these issues, we propose HERO, a novel framework for constructing Hierarchical Traversable 3DSGs, that redefines traversability by modeling operable obstacles as pathways, capturing their physical interactivity, functional semantics, and the scene's relational hierarchy. The results show that, relative to its baseline, HERO reduces PL by 35.1% in partially obstructed environments and increases SR by 79.4% in fully obstructed ones, demonstrating substantially higher efficiency and reachability.

Dialogue Topic Segmentation (DTS) is crucial for understanding task-oriented public-channel communications, such as maritime VHF dialogues, which feature informal speech and implicit transitions. To address the limitations of traditional methods, we propose DASH-DTS, a novel LLM-based framework. Its core contributions are: (1) topic shift detection via dialogue handshake recognition; (2) contextual enhancement through similarity-guided example selection; and (3) the generation of selective positive and negative samples to improve model discrimination and robustness. Additionally, we release VHF-Dial, the first public dataset of real-world maritime VHF communications, to advance research in this domain. DASH-DTS provides interpretable reasoning and confidence scores for each segment. Experimental results demonstrate that our framework achieves several sota segmentation trusted accuracy on both VHF-Dial and standard benchmarks, establishing a strong foundation for stable monitoring and decision support in operational dialogues.

Process Reward Models (PRMs) have become essential for improving Large Language Models (LLMs) via test-time scaling, yet their effectiveness in coding remains limited due to the lack of meaningful step decompositions in code and the noise of Monte-Carlo-generated partial labels. We propose DreamPRM-Code, a coding-focused PRM that treats functions as reasoning steps using a Chain-of-Function prompting strategy to induce modular code generation, enabling PRM training and application analogous to mathematical reasoning tasks. To address label noise, DreamPRM-Code introduces a meta-learning-based correction mechanism that leverages clean final-solution unit-test labels and performs bi-level optimization to refine intermediate labels. Applying on test-time scaling, DreamPRM-Code achieved state-of-the-art performance on LiveCodeBench with 80.9 pass@1 rate, surpassing OpenAI o4-mini.

Multimodal scientific reasoning remains a significant challenge for large language models (LLMs), particularly in chemistry, where problem-solving relies on symbolic diagrams, molecular structures, and structured visual data. Here, we systematically evaluate 40 proprietary and open-source multimodal LLMs, including GPT-5, o3, Gemini-2.5-Pro, and Qwen2.5-VL, on a curated benchmark of Olympiad-style chemistry questions drawn from over two decades of U.S. National Chemistry Olympiad (USNCO) exams. These questions require integrated visual and textual reasoning across diverse modalities. We find that many models struggle with modality fusion, where in some cases, removing the image even improves accuracy, indicating misalignment in vision-language integration. Chain-of-Thought prompting consistently enhances both accuracy and visual grounding, as demonstrated through ablation studies and occlusion-based interpretability. Our results reveal critical limitations in the scientific reasoning abilities of current MLLMs, providing actionable strategies for developing more robust and interpretable multimodal systems in chemistry. This work provides a timely benchmark for measuring progress in domain-specific multimodal AI and underscores the need for further advances at the intersection of artificial intelligence and scientific reasoning.

When not using reasoning, repeating the input prompt improves performance for popular models (Gemini, GPT, Claude, and Deepseek) without increasing the number of generated tokens or latency.

Advances in large language models have accelerated progress in text-to-SQL, methods for converting natural language queries into valid SQL queries. A key bottleneck for developing generalizable text-to-SQL models is the lack of large-scale datasets with sufficient schema and query complexity, domain coverage, and task diversity. We introduce SQaLe: a large-scale semi-synthetic text-to-SQL dataset built on 135,875 relational database schemas expanded from a collection of real-world schemas, SchemaPile. We establish a principled generation pipeline which combines schema sampling, question synthesis, and SQL construction, and produce 517,676 high-quality (question, schema, query) triples. The SQaLe dataset captures realistic schema size variability, diverse query patterns, and natural language ambiguity while maintaining execution validity. We provide an analysis of its contents and characteristics, and find that SQaLe introduces the most realistic large-scale text-to-SQL dataset to date in comparison with existing benchmarks and datasets. We discuss how SQaLe enables our vision for data scaling and model generalization in text-to-SQL research. The dataset is accessible at: https://huggingface.co/datasets/trl-lab/SQaLe-text-to-SQL-dataset.

Natural language information needs over symbolic music scores rarely reduce to a single step lookup. Many queries require compositional Music Information Retrieval (MIR) that extracts multiple pieces of evidence from structured notation and aggregates them to answer the question. This setting remains challenging for Large Language Models due to the mismatch between natural language intents and symbolic representations, as well as the difficulty of reliably handling long structured contexts. Existing benchmarks only partially capture these retrieval demands, often emphasizing isolated theoretical knowledge or simplified settings. We introduce CSyMR-Bench, a benchmark for compositional MIR in symbolic music reasoning grounded in authentic user scenarios. It contains 126 multiple choice questions curated from community discussions and professional examinations, where each item requires chaining multiple atomic analyses over a score to derive implicit musical evidence. To support diagnosis, we provide a taxonomy with six query intent categories and six analytical dimension tags. We further propose a tool-augmented retrieval and reasoning framework that integrates a ReAct-style controller with deterministic symbolic analysis operators built with music21. Experiments across prompting baselines and agent variants show that tool-grounded compositional retrieval consistently outperforms Large Language Model-only approaches, yielding 5-7% absolute accuracy gains, with the largest improvements on analysis-heavy categories.

Safety mechanisms in LLMs remain vulnerable to attacks that reframe harmful requests through culturally coded structures. We introduce Adversarial Tales, a jailbreak technique that embeds harmful content within cyberpunk narratives and prompts models to perform functional analysis inspired by Vladimir Propp's morphology of folktales. By casting the task as structural decomposition, the attack induces models to reconstruct harmful procedures as legitimate narrative interpretation. Across 26 frontier models from nine providers, we observe an average attack success rate of 71.3%, with no model family proving reliably robust. Together with our prior work on Adversarial Poetry, these findings suggest that structurally-grounded jailbreaks constitute a broad vulnerability class rather than isolated techniques. The space of culturally coded frames that can mediate harmful intent is vast, likely inexhaustible by pattern-matching defenses alone. Understanding why these attacks succeed is therefore essential: we outline a mechanistic interpretability research agenda to investigate how narrative cues reshape model representations and whether models can learn to recognize harmful intent independently of surface form.

Large Language Models exhibit sycophancy: prioritizing agreeableness over correctness. Current remedies evaluate reasoning outcomes: RLHF rewards correct answers, self-correction critiques outputs. All require ground truth, which is often unavailable at inference time and vulnerable to the same biases. We explore evaluating the reasoning process instead. Regulated Causal Anchoring (RCA) verifies whether outputs follow from their reasoning traces, without requiring ground truth. Sycophancy manifests as trace-output inconsistency: models derive one answer but output another to please users. RCA detects this inconsistency, achieving 0.0% sycophancy while accepting 88% of valid hints. We identify two failures invisible to outcome evaluation: Inverse Scaling (frontier models sycophant more because rationalization requires capability) and the Final Output Gap (correct reasoning precedes sycophantic output). Traditional self-correction reduces these failures to 7-9% but cannot eliminate them because the model critiques itself with the same biases. RCA's process evaluation operates at inference time, requires no ground truth, and uses an independent judge that breaks the self-reinforcing bias loop: three properties that outcome evaluation lacks.