Abstract

The rapid advancement of Large Language Models (LLMs) has en- abled them to generate complex, multi-step plans and itineraries. However, these generated plans often lack temporal and spatial consistency, particularly in scenarios involving physical travel con- straints. This research aims to study the temporal performance of di!erent LLMs and presents a validation framework that eval- uates and improves the temporal consistency of LLM-generated travel itineraries. The system employs multiple state-of-the-art LLMs to generate travel plans and validates them against real-world "ight duration constraints using the AeroDataBox API. This work contributes to the understanding of LLM capabilities in handling complex temporal reasoning tasks like itinerary generation and provides a framework to rectify any temporal inconsistencies like overlapping journeys or unrealistic transit times in the itineraries generated by LLMs before the itinerary is given to the user. Our experiments reveal that while current LLMs frequently produce temporally inconsistent itineraries, these can be systematically and reliably corrected using our framework, enabling their practical deployment in large-scale travel planning.

Abstract

Background Music Emotion Recognition (MER) systems primarily rely on audio features (Wang et al., 2021), with recent approaches incorporating lyrics to analyze sentiment and structure for improved accuracy (Agrawal et al., 2021). Musical content, particularly melody, has been found to have a stronger capacity to convey perceived emotional expression than lyrics, although lyrics tend to enhance negative emotions more easily than positive emotions (Ali & Peynircioğlu, 2006). This variability raises important questions: Which component carries the greatest emotional information? Do listeners rely on music, vocals, or lyrics when assessing emotional content? Aim Here, we systematically analyse the contribution of music, vocals, and lyrics to perception of emotion in music. Methods For this study, we used the DEAM (Soleymani et al., 2013) and PMEmo (Zhang et al., 2018) datasets, comprising a total of 2,700 songs. The DEAM dataset includes 1,802 audio items annotated for both dynamic and static emotion using continuous valence and arousal ratings. Participants utilised a two-dimensional interface based on the Self‑Assessment Manikins (SAM) to continuously rate the emotional content at a 2‑Hz sampling rate. The PMEmo dataset contains 794 popular music choruses sourced from international music charts (Billboard Hot 100, iTunes Top 100, and UK Top 40) and was annotated using a similar interface to capture continuous dynamic ratings for valence and arousal at the same sampling rate. Songs were source-separated using Ultimate Vocal Remover (Takahashi et al., 2017) to isolate music and vocals, while lyrics were transcribed using OpenAI’s Whisper model (Radford et al., 2022) and verified manually. Deep learning models trained on diverse song and speech datasets predicted VA values for musical and vocal components, while lyrics were analyzed using models trained on general and lyrically annotated texts (Çano, 2017). Spearman's correlation was calculated between predicted and human-annotated (HA) VA values. Additionally, quadrant-based (Q1: positive V, high A; Q2: negative V, high A; Q3: negative V, low A; Q4: positive V, low A) concurrency analyses evaluated prediction alignment with human-identified emotional quadrants. Results The musical component exhibited the highest correlation with HA ratings (valence: r = 0.70, arousal: r = 0.74), followed by vocals (valence: r = 0.54, arousal: r = 0.65; all p < .001), while lyrics contributed the least (valence: r = 0.11, arousal: r = 0.01). In terms of overall quadrant concurrency, the musical component showed the highest concurrency (60.41%) followed by vocals (48.84%) and then lyrics (31.45%). Specifically, the highest quadrant concurrency was observed in Q3 (musical: 90.53%, vocal: 93.53%, lyrics: 66.16%) followed by Q1 for musical (64.51%) and vocals (39.10%), and Q4 for lyrics (42.50%). Discussion Our findings emphasize the dominant role of the musical component in shaping perceived emotional expression in Western tonal music, aligning with prior work highlighting melody’s emotional salience (Ali & Peynircioğlu, 2006). Furthermore, for music signifying negative valence and low arousal (Q3), there is a higher degree of congruence between the components in conveying sadness and its related emotions. These results suggest that MER systems could prioritize musical and vocal components over lyrical content (Wang et al., 2021; Agrawal et al., 2021), as they appear to primarily drive the perception of emotional expression. References Agrawal, Y., Shanker, R. G. R., & Alluri, V. (2021). Transformer-based approach towards music emotion recognition from lyrics. In V. U. Rao & R. Kailash (Eds.), Advances in information retrieval (pp. 167–175). Springer. https://doi.org/10.1007/978-3-030-72240-1_12 Ali, S. O., & Peynircioğlu, Z. F. (2006). Songs and emotions: Are lyrics and melodies equal partners? Psychology of Music, 34(4), 511–534. https://doi.org/10.1177/0305735606067168 Çano, E. (2017). MoodyLyrics: A sentiment-annotated lyrics dataset. In Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2017) (pp. 118–124). Association for Computational Linguistics. https://doi.org/10.18653/v1/S17-1017 Radford, A., Kim, J. W., Xu, T., Brockman, G., McLeavey, C., & Sutskever, I. (2022). Robust speech recognition via large-scale weak supervision. arXiv. https://doi.org/10.48550/arXiv.2212.04356 Soleymani, M., Caro, M., Schmidt, E., Sha, C.-Y., & Yang, Y.-H. (2013). 1000 songs for emotional analysis of music. In P. Herrera, E. Gómez, & T. Kaneko (Eds.), Proceedings of the 2nd ACM International Workshop on Music Information Retrieval with User-Centered and Multimodal Strategies (pp. 1–6). Association for Computing Machinery. Takahashi, N., Huang, P.-S., & Lee, P. (2017). Multi-scale multi-band DenseNets for audio source separation. arXiv. https://doi.org/10.48550/arXiv.1706.09588 Wang, S., Xu, C., Ding, A. S., & Tang, Z. (2021). A Novel Emotion-Aware Hybrid Music Recommendation Method Using Deep Neural Network. Electronics, 10(15), 1769. https://doi.org/10.3390/electronics10151769 Zhang, K., Zhang, H., Li, S., Yang, C., & Sun, L. (2018). The PMEmo dataset for music emotion recognition. In Proceedings of the 2018 ACM International Conference on Multimedia Retrieval (pp. 135–142). Association for Computing Machinery. https://doi.org/10.1145/3206025.3206037

Abstract

The pyrimidine-pyrimidone (6-4) photoproduct (6-4PP) is a UV-induced DNA lesion implicated in skin disorders and cancers. The repair protein XPC/RAD4 detects this lesion and initiates nucleotide excision repair to safeguard genomic integrity. While the X-ray crystallographic structure of the 6–4PP containing DNA-RAD4/XPC complex reveals DNA distortion and extrusion of the lesion and partner bases, the precise mechanism by which RAD4/XPC initiates lesion repair remains unclear. To investigate this, we employed molecular dynamics simulations, umbrella sampling, and the nudged elastic band method to map the minimum energy path (MEP) from RAD4’s initial encounter with damaged DNA to its bound state. Our results reveal that the initial interrogation phase involves partial opening of the DNA, marked by the partial extrusion of the lesion while its partner bases largely remain intrahelical, accompanied by significant DNA unwinding near the damage site. This partially opened state represents the rate-limiting distortion, with 5′ base flipping as the bottleneck for downstream events leading to the bound state. Upon overcoming the bottleneck, the DNA adopts a final untwisted conformation, with the lesion flipping out first, followed by the complete sequential flipping of the 5′ base and then the 3′ partner bases, while full β-hairpin insertion ultimately stabilizes the final bound DNA complex. The energetics and structural intermediates along the MEP provide key insights into conformational changes that drive lesion recognition, extrusion, and stable binding, advancing our understanding of nucleotide excision repair.

Abstract

Determining spin state energy gaps (SSE) of 3d transition metal complexes (TMCs) is a major challenge in theoretical chemistry, as high-level quantum methods, though reliable, are computationally impractical for large-scale studies. This work explores a machine learning (ML)-based approach to predict DFT adiabatic SSE gaps using descriptors derived from a single high-spin DFT calculation. This approach is adopted to eliminate the differential treatment of electronic correlation between high-spin and low-spin structures. Our descriptors aim to incorporate the knowledge of crystal field theory into the ML model. They include atomic energy levels of bare metal ions, natural charges of ligating atoms, d-orbital molecular orbital eigenvalues derived from an high spin calculation, HOMO–LUMO gaps of free ligands, and simple identity-based features. We train ML models on 1434 SSE values spanning 934 complexes and demonstrate their transferability to more challenging complexes having bidentate π-bonding ligands despite being trained on simpler Werner-type monodentate complexes. We achieved a minimum MAE of 4.0 kcal mol−1 on the monodentate test set, and maintained a comparable MAE of 6.6 kcal mol−1 in the transferability assessment. This approach bypasses the need for multi-reference low-spin optimizations while retaining predictive accuracy, offering a cost-effective strategy for SSE estimation in transition metal chemistry. We hope the insights covered in this study will contribute to the development of additional electronic structure-based descriptors for SSE predictions.

Abstract

Recognizing textual attributes such as bold, italic, underline, and strikeout is essential for understanding text semantics, structure, and visual presentation. These attributes highlight key information, making them crucial for document analysis. Existing methods struggle with com putational efficiency or adaptability in noisy, multilingual settings. To address this, we introduce TexTAR, a multi-task, context-aware Trans former for Textual Attribute Recognition (TAR). Our novel data selec tion pipeline enhances context awareness, and our architecture employs a 2D RoPE (Rotary Positional Embedding)-style mechanism to incorpo rate input context for more accurate attribute predictions. We also in troduce MMTAD, a diverse, multilingual, multi-domain dataset annotated with text attributes across real-world documents such as legal records, notices, and textbooks. Extensive evaluations show TexTAR outperforms existing methods, demonstrating that contextual awareness contributes to state-of-the-art TAR performance

Abstract

This paper presents a dual-loop reference-less Clock and Data Recovery (CDR) architecture implemented in TSMC 28 nm CMOS technology for PCIe Gen3 applications. The proposed design features an extended BiDirectional Frequency Detector (BFD) that enables rapid frequency acquisition in both directions over a wide locking range. By utilizing DNFAST and DNSLOW signals, the BFD dynamically adjusts the correction rate based on frequency error, ensuring accurate frequency tracking and a seamless transition from frequency acquisition to phase tracking. This approach significantly enhances the locking performance and stability of the CDR system. The proposed CDR achieves a locking time of 4.1μs when transitioning from PCIe Gen1 (2. 5 G b p s) to PCIe Gen3 (8 G b p s) under a PRBS7 pattern with a power consumption of 3.9 mW at 0.9 V power supply.

Abstract

In this paper, we propose an efficient and robust framework for estimating leakage power while considering statistical variations in process parameters, supply voltage, and temperature (PVT) in FinFET-based digital circuits. Extensive testing reports an average error of less than 1 % in the ML models for standard digital cells, with an average MAPE (Mean Absolute Percentage Error) of 0.764%,0.770%, and 0.814% in the 7 nm,10 nm, and 16 nm FinFET technology nodes, respectively, which ensures a highly accurate leakage estimate for larger circuits. Additionally, we investigated the transistor-level approach, which distinctively estimates subthreshold, gate, and body leakage power using only transistor-level models and stacks, and investigated its trade-offs to gate-level models. Using ML techniques, we achieve a speedup of over 400 times compared to HSPICE for the characterization of digital blocks, using which complex cell designs can be analyzed across various input test combinations, significantly saving the time required for extensive simulations.

Abstract

Document layout analysis is essential for downstream tasks such as information retrieval, extraction, OCR, and digitization. However, existing large-scale datasets like PubLayNet and DocBank lack fine-grained region labels and multilingual diversity, making them insufficient for representing complex document layouts. Human-annotated datasets such as M6Doc and D4LA offer richer labels and greater domain diversity, but are too small to train robust models and lack adequate multilingual coverage. This gap is especially pronounced for Indic documents, which encompass diverse scripts yet remain underrepresented in current datasets, further limiting progress in this space. To address these shortcomings, we introduce IndicDLP, a large-scale foundational document layout dataset spanning 11 representative Indic languages alongside English and 12 common document domains. Additionally, we curate UED-mini, a dataset derived from DocLayNet and M6Doc, to enhance pretraining and provide a solid foundation for Indic layout models. Our experiments demonstrate that fine-tuning existing English models on IndicDLP significantly boosts performance, validating its effectiveness. Moreover, models trained on IndicDLP generalize well beyond Indic layouts, making it a valuable resource for document digitization. This work bridges gaps in scale, diversity, and annotation granularity, driving inclusive and efficient document understanding

Abstract

A 16-Gb/s continuous-time linear equalizer (CTLE) is designed in 28-nm CMOS technology, featuring an active inductor load and feedforward compensation to enhance high-frequency performance. The proposed design features a power-efficient architecture, achieving a remarkable eye opening of 45.75 ps and a vertical differential eye opening of 781.19 mV. The design achieves reliable operation for channel losses of 9 dB, consuming a total power of 3.3 mW from a 0.9 V supply. This CTLE provides an energy-efficient solution for high-speed SerDes applications, with a calculated figure-of-merit of 0.022 pJ/hit/dB.

Abstract

Segment matching is an important intermediate task in computer vision that establishes correspondences between semantically or geometrically coherent regions across images. Unlike keypoint matching, which focuses on localized features, segment matching captures structured regions, offering greater robustness to occlusions, lighting variations, and viewpoint changes. In this paper, we leverage the spatial understanding of 3D foundation models to tackle wide-baseline segment matching, a challenging setting involving extreme viewpoint shifts. We propose an architecture that uses the inductive bias of these 3D foundation models to match segments across image pairs with up to 180 degree view-point change rotation. Extensive experiments show that our approach outperforms state-of-the-art methods, including the SAM2 video propagator and local feature matching methods, by up to 30% on the AUPRC metric, on ScanNet++ and Replica datasets. We further demonstrate benefits of the proposed model on relevant downstream tasks, including 3D instance mapping and object-relative navigation. Project Page: https://segmast3r.github.io/