Large language models (LLMs) trained purely on text ostensibly lack any direct perceptual experience, yet their internal representations are implicitly shaped by multimodal regularities encoded in language. We test the hypothesis that explicit sensory prompting can surface this latent structure, bringing a text-only LLM into closer representational alignment with specialist vision and audio encoders. When a sensory prompt tells the model to 'see' or 'hear', it cues the model to resolve its next-token predictions as if they were conditioned on latent visual or auditory evidence that is never actually supplied. Our findings reveal that lightweight prompt engineering can reliably activate modality-appropriate representations in purely text-trained LLMs.

Cinematography understanding refers to the ability to recognize not only the visual content of a scene but also the cinematic techniques that shape narrative meaning. This capability is attracting increasing attention, as it enhances mul-timodal understanding in real-world applications and underpins coherent content creation in film and media. However, our analysis reveals that ambiguous option design in ShotBench and ShotVL's shortcomings in reasoning consistency and instruction adherence undermine evaluation reliability, limiting fair comparison and hindering future progress. To overcome these issues, we systematically refine ShotBench through consistent option restructuring, conduct the first critical analysis of ShotVL's reasoning behavior, and introduce an extended evaluation protocol that jointly assesses task accuracy and core model competencies. These efforts lead to RefineShot, a refined and expanded benchmark that enables more reliable assessment and fosters future advances in cinematography understanding. We first analyze and refine the options in ShotBench to address their inconsistencies, then examine state-of-the-art models and reveal their reliability defects. Based on these findings, we propose a new evaluation protocol and demonstrate its effectiveness through comprehensive experiments.

LLMs have emerged as powerful tools for interpreting multimodal data. In medicine, they hold particular promise for synthesizing large volumes of clinical information into actionable insights and digital health applications. Yet, a major limitation remains their inability to handle time series. To overcome this gap, we present OpenTSLM, a family of Time Series Language Models (TSLMs) created by integrating time series as a native modality to pretrained LLMs, enabling reasoning over multiple time series of any length. We investigate two architectures for OpenTSLM. The first, OpenTSLM-SoftPrompt, models time series implicitly by concatenating learnable time series tokens with text tokens via soft prompting. Although parameter-efficient, we hypothesize that explicit time series modeling scales better and outperforms implicit approaches. We thus introduce OpenTSLM-Flamingo, which integrates time series with text via cross-attention. We benchmark both variants against baselines that treat time series as text tokens or plots, across a suite of text-time-series Chain-of-Thought (CoT) reasoning tasks. We introduce three datasets: HAR-CoT, Sleep-CoT, and ECG-QA-CoT. Across all, OpenTSLM models outperform baselines, reaching 69.9 F1 in sleep staging and 65.4 in HAR, compared to 9.05 and 52.2 for finetuned text-only models. Notably, even 1B-parameter OpenTSLM models surpass GPT-4o (15.47 and 2.95). OpenTSLM-Flamingo matches OpenTSLM-SoftPrompt in performance and outperforms on longer sequences, while maintaining stable memory requirements. By contrast, SoftPrompt grows exponentially in memory with sequence length, requiring around 110 GB compared to 40 GB VRAM when training on ECG-QA with LLaMA-3B. Expert reviews by clinicians find strong reasoning capabilities exhibited by OpenTSLMs on ECG-QA. To facilitate further research, we provide all code, datasets, and models open-source.

Objective: To develop an explainable multimodal large language model (MM-LLM) that (1) screens optic nerve head (ONH) OCT circle scans for quality and (2) generates structured clinical reports that include glaucoma diagnosis and sector-wise retinal nerve fiber layer (RNFL) thinning assessments. Design: Retrospective cohort study of 1,310 subjects contributing 43,849 Spectralis ONH OCT circle scans (1,331 glaucomatous and 867 healthy eyes) from the DIGS and ADAGES cohorts. Methods: A MM-LLM (Llama 3.2 Vision-Instruct model) was fine-tuned to generate clinical descriptions of OCT imaging data. Training data included paired OCT images and automatically generated, structured clinical reports that described global and sectoral RNFL thinning. Poor-quality scans were labeled as unusable and paired with a fixed refusal statement. The model was evaluated on a held-out test set for three tasks: quality assessment, glaucoma detection, and RNFL thinning classification across seven anatomical sectors. Evaluation metrics included accuracy, sensitivity, specificity, precision, and F1-score. Model description quality was also evaluated using standard text evaluation metrics. Results: The model achieved 0.90 accuracy and 0.98 specificity for quality triage. For glaucoma detection, accuracy was 0.86 (sensitivity 0.91, specificity 0.73, F1-score 0.91). RNFL thinning prediction accuracy ranged from 0.83 to 0.94, with highest performance in global and temporal sectors. Text generation scores showed strong alignment with reference reports (BLEU: 0.82; ROUGE-1: 0.94; ROUGE-2: 0.87; ROUGE-L: 0.92; BERTScore-F1: 0.99). Conclusions: The fine-tuned MM-LLM generated accurate clinical descriptions based on OCT imaging. The model achieved high accuracy in identifying image quality issues and detecting glaucoma. The model also provided sectoral descriptions of RNFL thinning to help support clinical OCT evaluation.

Abstract--The performance of Large Language Models (LLMs) for translating Natural Language (NL) queries into SQL varies significantly across databases (DBs). NL queries are often expressed using a domain specific vocabulary, and mapping these to the correct SQL requires an understanding of the embedded domain expressions, their relationship to the DB schema structure. Existing benchmarks rely on unrealistic, ad-hoc query specific textual hints for expressing domain knowledge. In this paper, we propose a systematic framework for associating structured domain statements at the database level. We present retrieval of relevant structured domain statements given a user query using sub-string level match. We evaluate on eleven realistic DB schemas covering diverse domains across five open-source and proprietary LLMs and demonstrate that (1) DB level structured domain statements are more practical and accurate than existing ad-hoc query specific textual domain statements, and (2) Our sub-string match based retrieval of relevant domain statements provides significantly higher accuracy than other retrieval approaches. The impressive natural language understanding and code generation capabilities of modern LLMs has led to significantly improved performance on NL-SQL semantic parsing [1], [2]. However, their accuracy varies widely with the database queried [3]. DBs in WikiSQL [4] or Spider [5] contain semantically meaningful table/column names and cell values making it easier for LLMs to accurately link domain expressions in the NL query with the DB schema/cell elements.

Large Language Models (LLMs) have shown remarkable performance on general Question Answering (QA), yet they often struggle in domain-specific scenarios where accurate and up-to-date information is required. Retrieval-Augmented Generation (RAG) addresses this limitation by enriching LLMs with external knowledge, but existing systems primarily rely on unstructured documents, while largely overlooking relational databases, which provide precise, timely, and efficiently queryable factual information, serving as indispensable infrastructure in domains such as finance, healthcare, and scientific research. Motivated by this gap, we conduct a systematic analysis that reveals three central observations: (i) databases and documents offer complementary strengths across queries, (ii) naively combining both sources introduces noise and cost without consistent accuracy gains, and (iii) selecting the most suitable source for each query is crucial to balance effectiveness and efficiency. We further observe that query types show consistent regularities in their alignment with retrieval paths, suggesting that routing decisions can be effectively guided by systematic rules that capture these patterns. Building on these insights, we propose a rule-driven routing framework. A routing agent scores candidate augmentation paths based on explicit rules and selects the most suitable one; a rule-making expert agent refines the rules over time using QA feedback to maintain adaptability; and a path-level meta-cache reuses past routing decisions for semantically similar queries to reduce latency and cost. Experiments on three QA benchmarks demonstrate that our framework consistently outperforms static strategies and learned routing baselines, achieving higher accuracy while maintaining moderate computational cost.

Leaderboards for large reasoning models (LRMs) have turned evaluation into a competition, incentivizing developers to optimize directly on benchmark suites. A shortcut to achieving higher rankings is to incorporate evaluation benchmarks into the training data, thereby yielding inflated performance, known as benchmark contamination. Despite that numerous contamination detection approaches have been proposed, surprisingly, our studies find that evading contamination detections for LRMs is alarmingly easy. We focus on the two scenarios where contamination may occur in practice: (I) when the base model evolves into LRM via supervised fine-tuning (SFT) and reinforcement learning (RL), we find that contamination during SFT can be originally identified by contamination detection methods. Y et, even a brief Group Relative Policy Optimization (GRPO) training can markedly conceal contamination signals that most detection methods rely on. Further empirical experiments and theoretical analysis indicate that Proximal Policy Optimization (PPO) style importance sampling and clipping objectives are the root cause of this detection concealment, indicating that a broad class of RL methods may inherently exhibit similar concealment capability; (II) when SFT contamination with CoT is applied to advanced LRMs as the final stage, most contamination detection methods perform near random guesses. Without exposure to non-members, contaminated LRMs would still have more confidence when responding to those unseen samples that share similar distributions to the training set, and thus, evade existing memorization-based detection methods. Together, our findings reveal the unique vulnerability of LRMs evaluations: Model developers could easily contaminate LRMs to achieve inflated leaderboards performance while leaving minimal traces of contamination, thereby strongly undermining the fairness of evaluation and threatening the integrity of public leaderboards. This underscores the urgent need for advanced contamination detection methods and trustworthy evaluation protocols tailored to LRMs. Our code is available at https://github.com/ASTRAL-Group/ Competition among model developers has intensified as Large Language Models (LLMs) have demonstrated remarkable capabilities in various real-world tasks (Achiam et al., 2023; Wang et al., 2024). The leaderboards for performance are becoming a competitive arena for all state-of-the-art (SOT A) LLMs. However, inadvertently, benchmark samples may appear during LLMs' pre-training due to vast amounts of web-scraped training data. In addition, in the pursuit of publicity, some model developers may even deliberately incorporate benchmark data into their training sets (Sun et al., 2025), resulting in inflated benchmark performance and leaderboard rankings. We refer to this as the benchmark contamination problem in LLMs (Xu et al., 2024; Balloccu et al., 2024).

Large Language Models (LLMs) have demonstrated exceptional capabilities, yet selecting the most reliable response from multiple LLMs remains a challenge, particularly in resource-constrained settings. Existing approaches often depend on costly external verifiers, human evaluators, or self-consistency techniques that require multiple samples from a single model. While multi-LLM systems produce more diverse responses than single models and thus have greater potential, they often underperform compared to single LLM self-consistency. We propose a principled, novel and computationally efficient method to select the best response from multiple different LLMs using a calibrated log-likelihood score, implicitly leveraging the inherent knowledge and confidence of these models. Our method demonstrates improvements of approx. 4%, 3%, and 5% across both debate (multi-round LLM discussions) and non-debate (Best-of-N with multiple LLMs) settings on GSM8K, MMLU (6 subsets), and ARC datasets respectively.

Completely Automated Public Turing tests to tell Computers and Humans Apart (CAPTCHAs) are a foundational component of web security, yet traditional implementations suffer from a trade-off between usability and resilience against AI-powered bots. This paper introduces a novel hybrid CAPTCHA system that synergizes the cognitive challenges posed by Large Language Models (LLMs) with the behavioral biometric analysis of keystroke dynamics. Our approach generates dynamic, unpredictable questions that are trivial for humans but non-trivial for automated agents, while simultaneously analyzing the user's typing rhythm to distinguish human patterns from robotic input. We present the system's architecture, formalize the feature extraction methodology for keystroke analysis, and report on an experimental evaluation. The results indicate that our dual-layered approach achieves a high degree of accuracy in bot detection, successfully thwarting both paste-based and script-based simulation attacks, while maintaining a high usability score among human participants. This work demonstrates the potential of combining cognitive and behavioral tests to create a new generation of more secure and user-friendly CAPTCHAs.

Large Language Model (LLM) agents use memory to learn from past interactions, enabling autonomous planning and decision-making in complex environments. However, this reliance on memory introduces a critical security risk: an adversary can inject seemingly harmless records into an agent's memory to manipulate its future behavior. This vulnerability is characterized by two core aspects: First, the malicious effect of injected records is only activated within a specific context, making them hard to detect when individual memory entries are audited in isolation. Second, once triggered, the manipulation can initiate a self-reinforcing error cycle: the corrupted outcome is stored as precedent, which not only amplifies the initial error but also progressively lowers the threshold for similar attacks in the future. To address these challenges, we introduce A-MemGuard (Agent-Memory Guard), the first proactive defense framework for LLM agent memory. The core idea of our work is the insight that memory itself must become both self-checking and self-correcting. Without modifying the agent's core architecture, A-MemGuard combines two mechanisms: (1) consensus-based validation, which detects anomalies by comparing reasoning paths derived from multiple related memories and (2) a dual-memory structure, where detected failures are distilled into ``lessons'' stored separately and consulted before future actions, breaking error cycles and enabling adaptation. Comprehensive evaluations on multiple benchmarks show that A-MemGuard effectively cuts attack success rates by over 95% while incurring a minimal utility cost. This work shifts LLM memory security from static filtering to a proactive, experience-driven model where defenses strengthen over time. Our code is available in https://github.com/TangciuYueng/AMemGuard