This paper investigates instrument classification using solo sheet music images rather than audio. It converts images into sequences of musical words via bootleg score representation and treats the task as text classification, training AWD‑LSTM, GPT‑2, and RoBERTa models on IMSLP data for eight instruments. Pretraining on unlabeled data and fine‑tuning improves RoBERTa’s accuracy from 34.5% to 42.9%, and two proposed data‑augmentation methods raise accuracy by an additional 15%.
By Kevin Ji, Daniel Yang, TJ Tsai
Safety-Flag is a unified benchmark that consolidates seven popular safety datasets into a single balanced flag/do‑not‑flag protocol, providing item‑level decisions and confidence scores for multiple large language models and dedicated guards. The benchmark evaluates moderator reliability across three dimensions—error direction, probability calibration, and confidence‑based error ranking—revealing that aggregate accuracy masks significant differences, such as one model flagging 85% of benign content while another misses 54% of harmful content. The study shows that general‑purpose models are overconfident, but temperature tuning can substantially improve calibration, and confidence‑based abstention can reduce selective risk, though performance varies with how well confidence ranks errors.
By Yibo Hu
The paper investigates whether reward hacking in large language models leaves identifiable patterns in their internal representations. By analyzing frontier open‑source LLMs such as Kimi K3, GLM 5.2, and Qwen 3.8 Max, the authors find that simple difference‑of‑means (DoM) vectors consistently capture a range of hacking behaviors across standard benchmarks like DeepSWE and SWE‑bench. These vectors are both generalizable and interpretable, enabling inexpensive online detection that can even predict future hacks before they occur.
By Leon Bergen, Usha Bhalla, Andrew Lee, Barak Widawsky, Linas Nasvytis, Connor Watts, Siddharth Boppana, Sidharth Baskaran, Dron Hazra, Michael Byun, Atticus Geiger, Owen Lewis, Matthew Kowal, Vasudev Shyam, Thomas Fel, Thomas McGrath, Ekdeep Singh Lubana, Jack Merullo
The paper introduces ADAPT, a lightweight machine‑learning force field that replaces graph neural networks with a direct coordinates‑in‑space Transformer encoder to model all pairwise atomic interactions. Applied to silicon point defects, ADAPT reduces force prediction error by about 22% and energy prediction error by roughly 40% compared to a state‑of‑the‑art GNN model, while also cutting computational cost. This approach addresses common GNN issues such as oversmoothing, oversquashing, and poor long‑range interaction representation, which are especially problematic for point defect modeling.
By Evan Dramko, Yihuang Xiong, Yizhi Zhu, Geoffroy Hautier, Thomas Reps, Christopher Jermaine, Anastasios Kyrillidis
The paper examines how large language models (LLMs) alter the expression of Dark Triad traits—Machiavellianism, narcissism, and psychopathy—when prompted to fake good or fake bad. Across seven state‑of‑the‑art models and two real‑world contexts (employment selection and forensic evaluation), most models lowered trait scores under fake‑good conditions and raised them under fake‑bad conditions, with varying consistency across traits and models. The study also finds that explicit fake‑bad instructions produce stronger distortions than contextual framing alone, underscoring the influence of motivational and situational context on personality‑related outputs.
By Victoria Popa, Guglielmo Cola, Caterina Senette, Maurizio Tesconi
The paper introduces STaR-CS, a method that generates multi‑party dialogues by modeling facilitator style and extracting structured skeletons, thereby easing data scarcity for cognitive stimulation therapy (CST) in low‑resource languages. Building on this corpus, the Reflective Cognitive Alignment (RCA) framework treats stimulation interactions as a sequential decision process, combining Protocol‑Constrained Chain‑of‑Cognition (PC‑CoC) for structured reasoning with Inference‑Time Value Alignment (IVA) to select responses that balance safety and engagement. Experiments with six large language models and two judges demonstrate that RCA improves protocol adherence, safety, and group facilitation compared to standard prompting.
By Jiyue Jiang, Ziyi Li, He Hu, Sheng Wang, Yuhan Chen, Yanyu Chen, Jingqi Zhou, Pengan Chen, Fei Ma, Irwin King, Yu Li, Chuan Wu
SAGE is a governed multi‑stage LLM pipeline that transforms enterprise guideline documents—containing narrative text, tables, and images—into structured artifacts. It uses a shared versioned rule store, schema‑validated contracts, and provenance tracking to validate, score, and reconcile extracted rules, automatically approving high‑confidence outputs while flagging uncertain items for human review. In a test on 120 documents, SAGE reduced processing time from days to 20–100 minutes and achieved a 96% success rate with only 3.2% hallucination.
By Mohammadreza Sediqin, Shivali Dalmia, Sumukha Thoppanahalli, Srinivasa Karthikeya Reddy Kovvuri, Abhishek Mukherji
The paper investigates whether large language models decide to gather safety-relevant evidence before acting. Using the SAFE benchmark, the authors evaluate models such as GPT‑5.5, o3, Claude Opus, and Claude Sonnet, finding distinct evidence‑acquisition strategies that vary with retrieval cost, severity, and presentation. Across models, expected‑value reasoning dominates Stage 1 rationales, and evidence framing can alter decisions near the inspection threshold while probability is often cited despite limited influence.
By Omer Tafveez
The paper investigates how Vision‑Language Models (VLMs) often report high confidence even after self‑correcting or arriving at wrong answers, a phenomenon the authors attribute to the verbalized confidence being largely independent of the model’s reasoning trajectory. By analyzing content variation, token masking, and hesitation markers, the authors demonstrate that confidence does not adequately reflect the actual reasoning process and that calibration training can sometimes worsen this disconnect. To address this blind spot, they introduce the Trajectory‑Grounding Score (TGS) in two forms—TGS‑self and TGS‑pair—and propose TGS‑Bench, a suite of 10 benchmarks that reveal divergences between conventional calibration metrics and trajectory‑grounded confidence.
By Jisoo Yang, Jaeho Han, Trung X. Pham, Junyeong Kim
The paper introduces EffiRAG, a graph-based retrieval‑augmented generation system that reduces the cost of building and querying a graph by using it only to locate relevant passages and generating answers from the original text. On the UltraDomain benchmark, EffiRAG outperforms LightRAG‑hybrid in 93 of 120 questions while cutting total system cost by 57 % (from USD 0.952 to USD 0.408). The study shows that graph‑based RAG can be both more accurate and cheaper, especially as the corpus grows, and recommends evaluating such systems on both answer quality and cost.
By Yuzhong Zhang, Haoyang Ma, Chao Peng, Lionel Briand, Boxi Yu, Jialun Cao
GuardEn is an executable safeguard framework that transforms complex safety policies into atomic propositions via Safety‑Rule Compilation and models their composition as executable code. At inference, Scene‑Grounded Execution uses visual scene graphs to instantiate these propositions, allowing rule‑grounded and interpretable safety reasoning. Experiments on SafetyVisionBench show GuardEn improves F1 by an average of 9.8 points over the strongest baseline.
By Jisoo Kim (Sungkyunkwan University), TaeYoon Kwack (Sungkyunkwan University), Jinwoo Jang (Sungkyunkwan University), Honguk Woo (Sungkyunkwan University)
The paper introduces Market Signal Injection (MSI), an attack that alters how market data is formatted or described—without changing its numerical values—to influence large language model (LLM) pricing agents. Experiments on nine open‑weight and three proprietary models in simulated duopoly and triopoly markets show that sentiment‑based formatting changes cause significant shifts in firm behavior, profits, and consumer surplus. The study also demonstrates that model susceptibility varies across families, that larger models are not always more robust, and that techniques such as input canonicalization and decision boundary anchoring can partially mitigate these attacks.
By Dohun Lee, Hyunwoo Park
The study reports that ChatGPT 5.2 successfully passed a Finnish-language Turing Test conducted in Finland, contrary to expectations that uneven representation of Finnish in training data would hinder performance. The authors attribute failures mainly to participants using colloquial Finnish cues to identify human authorship. They propose reinterpreting the Turing Test as a comparative method to assess an AI’s credible membership in a specific social world, highlighting its utility for probing the human‑machine boundary across domains.
By Otto Segersven, Pentti Henttonen
The paper investigates whether in-context learning (ICL) in large language model agents reflects genuine recursive reasoning or simply statistical extrapolation. By testing LLM agents in a public goods game with manipulated historical feedback, the authors compare decision quality to a rational expectations equilibrium benchmark. They find that disrupting historical patterns eliminates the benefits of longer context, especially in highly interdependent settings, indicating that ICL behavior aligns more with statistical extrapolation than strategic reasoning.
By Yu Liu, Wenwen Li, Yifan Dou, Guangnan Ye
The paper introduces a text-based, multi-agent version of the board game Clue to test multi-step deductive reasoning in large language models (LLMs). Six LLM-based agents (GPT‑4o‑mini and Gemini‑2.5‑Flash) play turn‑based games, and a tool‑augmented approach uses a structured possibility matrix to convert implicit game state into explicit remaining possibilities, thereby offloading memory and deductive constraints from the agents. The study compares this tool‑augmented method against a baseline to assess its impact on reasoning quality and task success in a strategic reasoning environment.
By Rebecca Ansell, Autumn Toney-Wails
The paper investigates whether knowledge editing truly erases original facts from language models. Using a linear trace probe, the authors find that after editing a fact in GPT‑2‑XL, the original object remains highly decodable from hidden states across three different editing methods, even when the model behaves correctly on edited prompts. This suggests that editing suppresses rather than removes the original association in representational space.
By Priyansh Srivastava, Romit Chatterjee
The paper introduces a five-task diagnostic experiment that separates perceptual and reasoning failures in multimodal large language models on physics and geometry benchmarks. It finds that misinterpreting diagrams hurts performance even on text-only solvable problems, and that accuracy improves when models receive human-authored captions. The study shows that correcting captions can recover many errors, revealing distinct reasoning bottlenecks that differ by domain, while a heavily pretrained model still underperforms and often truncates reasoning traces.
By Raj Jaiswal, Sree Krishna Uppalapati, Dhruvkumar Patel, Ria Khatoniar, Tanuja Ganu, Rajiv Ratn Shah
WARD is a runtime‑adaptive Vision Transformer framework designed for edge AI that combines channel‑wise subnetwork partitioning, reliability‑aware continual learning, and dynamic operating‑mode scheduling. It operates two physically isolated subnetworks across four modes—Full‑Precision, Low‑Power, High‑Reliability, and Adaptive—to balance computational cost and fault tolerance while maintaining uninterrupted inference. Implemented on a lightweight FPGA accelerator with minimal area overhead, WARD achieves a network‑level failure rate of 1.79% under high Bit Error Rates and supports rapid mode transitions within a few clock cycles.
By Mahdi Taheri, Pramit Kumar Bhaduri, Mohammad Masoumi, Ali Mahani
The paper introduces DISCERN, a two-tier protocol for certifying that updates to production models do not increase risk. It first uses unlabeled data to detect benign updates based on disagreement rates, then selectively labels only disagreements through an anytime-valid confidence sequence. The method achieves finite-sample validity with label-complexity bounds of order ρ²/ε², demonstrating significant label savings and strong empirical performance across 14,000+ audit streams.
By Vishnu Bindu Balachandran
The paper investigates whether domain-specific fine‑tuning benefits open‑ended scientific reasoning in astronomy. Using a curated 300‑question QA benchmark from 2017–2026 Olympiad‑style materials, the authors compare open‑weight, API‑served general‑purpose, multimodal, and astronomy‑specialized language models. Results show that strong general‑purpose models set the highest correctness baseline, but variations in metric agreement, judge sensitivity, benchmark composition, and modality suggest that domain specialization is task‑ and deployment‑dependent and that domain‑specific evaluation is crucial for scientific workflows.
By Vanessa Lama, Sanjay Das, Emily Herron, Yuan-Sen Ting, Tijmen de Haan, Junqi Yin, Tirthankar Ghosal, Feiyi Wang