PCR - Fall 2025

Fall Issue

P C R

2025

Contents.

A Full-Cell Investigation of Na3V2(PO4)3 and Hard Carbon Electrodes for Sodium-Ion Batteries 5 A Traditional Approach to Symbolic Piano Continuation 10 GENETICS: A Deep Dive 14 Gene Editing Technology Conquers Deadly Disease 15 Development of a Hematology Workflow Video at NYU Tisch Hospital 16 Disseminated Histoplasmosis Mimicking Colonic Carcinoma in a Patient with Crohn’s Disease 16 Organizing and Mentoring for the Inaugural EleutherAI Summer of Open AI Research 19 Psychological and Musical Innovation in the Beatles’ Work 22

2

Editor’s Note.

Welcome to the 2025 Fall edition of the Pingry Community Research (PCR) Jour-

nal. We are excited to showcase Pingry’s top scientific talent, both in terms of research skills and knowledge of scientific concepts and discoveries.

The PCR journal provides students the opportunity to publish novel research. Through a written medium, students demonstrate their in-depth understanding of complex, collegiate-level scientific topics, and their applications in research at Pingry.

The fall edition of PCR highlights work in three categories. Reporter articles are written by students on a recent scientific advancement of their choosing. Summer research articles summarize the findings of, and/or provide a reflection on, novel research conducted by students over the preceding summer. Research papers present original scientific investigations conducted by students in any form.

This year, we are trying something new by interleaving all three types throughout the journal to create a more dynamic reading experience. Please let us know your thoughts!

Through the PCR journal, we hope to spark intellectual curiosity and promote scientific inquiry amongst the next generation of Pingry researchers.

Dive into the wonders of Pingry Research through this edition of PCR: Pingry’s foremost journal of scientific research.

Christian Zhou-Zheng (VI), Editor-in-Chief Katherine Jung (VI), Editor-in-Chief

3

Editorial Staff.

Editors-in-Chief: Christian Zhou-Zheng (VI) Katherine Jung (VI)

Copy Editors: Alan Huang (IV) Amelia Liu (V) Anavi Sinha (III) Cecilia Caligiuri (VI) Edward Huang (VI) Haeley Cole-Boksner (III) James Draper (VI) Ryan Hao (V) Samaya Shah (V) Sarah Yu (V) Shanti Swadia (V) Suvid Bordia (IV)

Head Layout Editor: Julia Ronnen (VI) Head Copy Editor: Tingting Luo (VI) Faculty Advisor: Mr. Maxwell Layout Editors: Aiden Suh (V) Amelia Liu (V) Anna Ojo (V) Jasmine Zhou (V) Ryan Hao (V) Sarah Yu (V) Shanti Swadia (V)

Art Funds Cartoon Credits: Aanav Shah (III) Adam Mazo (III) Anika Gupta (III) Brendan Lin (III) Caroline Ouyang (III) Celia Lowenstein (III) Levi Pearl (III) Noah Maloney (III) Quinn Nolan (III) Santiago Galvan (III)

4

A Full-Cell Investigation of Na3V2(PO4)3 and Hard Carbon Electrodes for Sodium-Ion Batteries

by Alan Huang (IV), Junlin Wu, Mia Ge Abstract Sodium-ion batteries (SIBs) have emerged as promising alternatives to lithium-ion batteries due to the abundance and low cost of sodium. Na3V2(PO4)3 (NVP) serves as a proficient cathode material, offering a reasonably high capacity and cycling stability [1] [2]. Additionally, two of the three Na+ per formula unit are extracted, resulting in a high potential (~3.4 V vs. Na+/Na) with an oxidation of V3+/V4+ [3]. Re-insertion of these two Na+ on discharge yields a reversible specific capacity of 117.6 mAh/g [3]. Further intercalation is possible but occurs only at a lower potential (~1.4 V vs. Na+/Na), where vanadium is reduced from V³+/V²+, contributing ~60 mAh/g [3]. Hard carbon inversely provides viable anode performance despite graphite’s incompatibility with sodium storage [1]. This investigation demonstrates that spherical hard carbon (HC) anodes exhibit ~80 mAh/g higher capacity than irregular variants. Both suffer from poor initial coulombic efficiency (ICE) due to irreversible sodium trapping for defects within the materials. A presodiation methodology is proposed to address hard carbon’s low ICE problem. The NVP/HC system shows fair rate capability, but cycling challenges persist due to limited sodium inventory and residual carbon fragments, causing ongoing side reactions.

V for lithium) [1] [2]. Effective electrode materials must accommodate sodium’s larger size while maintaining structural integrity and high coulombic efficiency. This investigation focuses on the systematic development of Na3V2(PO4)3 (NVP) cathode and hard carbon (HC) anode materials, examining structural characterization, electrochemical performance, and full-cell integration while addressing key limitations through presodiation.

Cathode Characterization and Performance NVP represents the optimal polyanion compound for SIB cathodes [4]. XRD analysis confirms the rhombohedral NASICON phase (space group R-3c), characterized by sharp, well-defined peaks that indicate good crystallinity (fig. 1) [5]. Energy-dispersive X-ray spectroscopy mapping confirms uniform elemental distribution throughout the particles (fig. 2). The actual atomic composition closely matches theoretical values: Na (15.5% actual vs. 15% theoretical), P (13.9% vs. 10%), O (61.3% vs.

60%), and V (9.3% vs. 15%) (table 1). The slight deviation in vanadium content is likely due to analytical challenges. SEM imaging reveals uniform particles ranging from 0.5-1.5 μm with a mean size of 1.1 μm; the white particle color in the image indicates a rather low electronic conductivity, suggesting flaws in the production process (fig.

Introduction Sodium-ion batteries address limitations of lithium-ion systems through sodium’s unique advantages. Sodium is the sixth most abundant element in Earth’s crust while lithium is scarce, allowing for long-term resource security and cost stability [1]. The larger sodium ionic radius (1.02 Å vs. 0.76 Å for lithium) enables the incorporation of more 3d transition metals and space for complex materials design [1] [2]. Sodium exhibits weak solvation energy in polar solvents, contributing to higher ionic conductivity in Na+-based electrolytes.[1] However, challenges remain, including lower energy density due to sodium’s higher atomic mass (23.00 vs. 6.94 for lithium) and somewhat lower standard electrode potential (-2.71 V vs. -3.04

Figure 1. X-Ray diffraction (XRD) pattern of Na3V2(PO4)3 (NVP). The diffraction peaks correspond to the rhombohedral NASICON phase, showing sharp and well-defined reflections that confirm good crystallinity.

5

Table 1. Comparison of theoretical and experimental atomic composition of NVP. Elemental analysis yields Na (15.5% vs.

15%), V (9.3% vs. 15%), P (13.9% vs. 10%), and O (61.3% vs.

60%), with minor deviation in vanadium content likely due to measurement limitations.

Figure 2. Energy-dispersive X-ray spectroscopy (EDX) elemental mapping of NVP particles. Sodium, vanadium, phosphorous, and oxygen are uniformly distributed across the particles, indicating homogenous composition.

Anode Characterization and Performance Hard carbon serves as the primary anode material since graphite cannot accommodate sodium intercalation due to size constraints. Two morphologies were investigated: irregular (angular, plate-like particles) and spherical variants (fig. 5). Spherical hard carbon provided ~80 mAh/g higher capacity through improved material utilization and particle packing efficiency (fig. 6). XRD analysis reveals characteristic amorphous features with broad humps at 24-26° and 42-44°, plus sharp impurity peaks suggesting residual inorganics (fig. 7). These impurities most likely consist of inorganic residues or graphitized domains that survive carbonization (e.g. quartz SiO2 (202/020) or anatase TiO2 (200)). Both materials demonstrate good rate capability and cycling stability after initial cycles.

3). Electrochemical testing demonstrates high ICE (92.793.6%) and excellent cycling stability over 80 cycles. The characteristic flat plateau around 3.4 V confirms the twophase transformation, while the rate capability remains good from 0.1C to 2C, depending on the electrode loading (fig. 4).

High-loading NVP electrodes exhibit significantly inferior performance compared to low-loading configurations, indicated by reduced capacity and poor rate capability. This degradation results from compromised ion/electron transport in thick electrodes, leading to an incomplete utilization of the active material and delamination from the current collectors. These findings necessitate a careful balance between energy density requirements and electrochemical performance.

Fig. 6 also indicates that both hard carbon variants suffer from poor ICE: spherical (73.9%) and irregular (72.3%).

Figure 3. Scanning electron microscopy (SEM) image of NVP particles. The micrographs show uniform morphology with particle sizes ranging from 0.5 to 1.5 μm, with the average particle diameter of approximately 1.1 μm.

Figure 4. 1st cycle GCD (a) and rate capability (b) graphs. (a) initial charge—discharge curves demonstrating high initial Coulombic efficiency (92.7—93.6%). (b) Cycling performance showing excellent capacity retention over 80 cycles

6

Figure 5. Morphology of hard carbon anodes. SEM images of irregular hard carbon (angular, plate-like particles) and spherical hard carbon (smooth, rounded particles) highlight the distinct morphologies investigated.

Figure 7. X-ray diffraction (XRD) patterns of hard carbon anodes. Both morphologies display broad amorphous humps at 24–26° and 42–44°, characteristic of turbostratic carbon, along with sharp impurity peaks indicating residual inorganic phases.

This limitation arises from hard carbon’s dual storage mechanism, which combines reversible sodium storage in graphitic domains with irreversible trapping in pores, defects, and microporous structures.[6] Additional ICE loss occurs through the formation of solid electrolyte interphase (SEI) and side reactions with electrolyte components. Hard carbon maintains reversible capacity after the second cycle, with coulombic efficiency improving to 98.0%.

irreversible sites before full-cell assembly. This approach proved highly effective, directly addressing poor ICE by pre-cycling hard carbon anodes in a half-cell, occupying irreversible sites with excess sodium from metal electrodes, and preserving sodium inventory in final full-cell configurations.

Presodiation Implementation & Full Cell Design and Performance The full-cell design required careful capacity matching To address ICE limitations, an electrochemical presodibetween NVP cathode and hard carbon anode, targetation strategy was employed involving pre-cycling hard ing N/P ratio ~1.0 with ~0.5 mAh/cm² areal capacity carbon anodes in half-cells against sodium metal to fill using 11 mm diameter electrodes (e.g., 4.5 mg/cm² NVP paired with 2 mg/cm² HC). Without presodiation, NVP/

Figure 6. Galvanostatic charge–discharge (GCD) profiles of hard carbon anodes. Spherical hard carbon exhibits ~80 mAh g-¹ higher reversible capacity than irregular hard carbon, attributed to improved material utilization and particle packing efficiency.

Figure 8. First-cycle performance of NVP/HC full-cells with and without presodiation. Initial coulombic efficiency improves from 53.3% without presodiation to 95.4% after presodiation (~42% increase), with first-cycle capacity boosted by ~50 mAh g-¹.

7

Conclusion and Outlook Comprehensive optimization encompasses multiple approaches, including foreign-ion doping/substitution for NVP with single and multiple-ion strategies to improve electrical conductivity, carbon modification approaches to optimize crystal structure, and nanostructure design extending contact area while enhancing Na+ diffusion kinetics. Immediate solutions include implementing high-quality hard carbon with reduced surface area and fewer residual fragments, developing sodium-rich NVP variants, and doping transitional elements (Mn, Cr, Mg) into NVP structures.

Figure 9. Rate capability and cycling stability of presodiated NVP/HC full-cells. The cells deliver stable capacity from 0.1C to 2C, but show significant capacity degradation over extended cycling with an average coulombic efficiency of 99.3%.

HC full-cells achieved only 53.3% ICE due to irreversible sodium consumption. Presodiation optimization dramatically improved performance, increasing ICE to 95.4% (~42% improvement) while increasing first-cycle capacity by ~50 mAh/g (fig. 8). The optimized NVP/HC system demonstrates fair rate capability from 0.1C to 2C, but exhibits poor cycling stability with substantial capacity degradation and low average coulombic efficiency (99.3%) required for prolonged cycling (fig. 9). This degradation results from limited Na+ inventory in full-cells compared to half-cells and residual spherical hard carbon fragments likely stemming from production errors, which have a high surface area promoting continuous SEI formation. EIS confirms progressive interface degradation, with the 70th cycle showing significantly higher resistance than the initial cycles (fig. 10). Previous SEM analysis in fig. 5 reveals nanosized carbon fragments continuously consuming sodium inventory through ongoing electrolyte decomposition.

This investigation demonstrates significant promise for NVP/HC sodium-ion battery systems while identifying critical optimization pathways. Presodiation strategies successfully address ICE limitations, improving full-cell performance from 53.3% to 95.4% efficiency. However, cycling stability challenges persist due to interface degradation and ongoing side reactions. The foundation established provides clear advancement pathways emphasizing materials’ optimization, interface engineering, and full-cell integration strategies essential for commercially viable energy storage solutions. Continued development of advanced materials and cell designs will be crucial for achieving the performance standards required for widespread SIB deployment.

References [1] Yabuuchi, Naoaki, et al. “Research Development on Sodium-Ion Batteries.” Chemical Reviews, vol. 114, no.

23, 2014, pp. 11636–11682. ACS Publications, https:// doi.org/10.1021/cr500192f.

[2] Shoaib, Muhammad, and Venkataraman Thangadurai. “Exploring the Anionic Redox Chemistry in Cathode Materials for High-Energy-Density Sodium-Ion Batteries.” ACS Omega, vol. 7, no. 39, 2022, pp. 34710–34717. https://doi.org/10.1021/acsomega.2c03883.

[3] Sayahpour, Behzad, et al. “Perspective: Design of Cathode Materials for Sustainable Sodium-Ion Batteries.”

MRS Energy & Sustainability, vol. 9, 2022, pp. 183–197. https://doi.org/10.1557/s43581-022-00029-9.

[4] Jiang, Xiaoyu, et al. “An All-Phosphate and Zero-Strain Sodium-Ion Battery Based on Na3V2(PO4)3 Cathode, NaTi2(PO4)3 Anode, and Trimethyl Phosphate Electrolyte with Intrinsic Safety and Long Lifespan.” ACS Applied Materials & Interfaces, vol. 9, no. 50, 2017, pp. 43733–

Figure 10, Electrochemical impedance spectroscopy (EIS) of NVP/HC full-cells. Nyquist plots demonstrate progressive interfacial degradation, with the 70th cycle showing markedly higher resistance compared to early cycles.

8

43738. https://doi.org/10.1021/acsami.7b14946.

[5] Wang, Jingyang, et al. “Design Principles for NASICON Super-Ionic Conductors.” Nature Communications, vol. 14, 2023. https://doi.org/10.1038/s41467-02340669-0.

[6] Ghimbeu, Camélia Matei, et al. “Insights on the Na+ Ion Storage Mechanism in Hard Carbon: Discrimination between the Porosity, Surface Functional Groups and Defects.” Nano Energy, vol. 44, 2018, pp. 327–335. https://doi.org/10.1016/j.nanoen.2017.12.013.

9

A Traditional Approach to Symbolic Piano Continuation

By Christian Zhou-Zheng, John Backsund, Dun Li Chan, Alex Coventry, Avid Eslami, Jyotin Goel, Xingwen Han, Danysh Soomro, Galen Wei Abstract We present a traditional approach to symbolic piano music continuation for the MIREX 2025 Symbolic Music Generation challenge. While computational music generation has recently focused on developing large foundation models with sophisticated architectural modifications, we argue that simpler approaches remain more effective for constrained, single-instrument tasks. We thus return to a simple, unaugmented next-token-prediction objective on tokenized raw MIDI, aiming to outperform large foundation models by using better data and better fundamentals. We release model weights and code at https://github.com/christianazinn/mirex2025.

1. Introduction The generation of continuations of piano music has a long history in computational music generation, largely due to the propensity of piano music readily available in symbolic formats. Recent developments in sequence modeling have allowed continuation to be viewed as an autoregressive task, to be modeled with a suitable tokenization scheme and a powerful sequence model like the ubiquitous Transformer [1]. A nonexhaustive list of prior work in this vein includes the Music Transformer [2], Museformer [3], FIGARO [4], and MuseCoco [5].

Most research in symbolic music modeling is focused on generalizing these techniques to—and improving performance on—long-sequence, multitrack, multi-instrument, and/or text- or attribute-controllable generative tasks. Typically, specialized techniques must be developed for these foundation models to handle these harder tasks, such as fine- and coarse-grained attention for long sequences [3], and text feature extraction techniques [4] and attribute augmentation [5] for controllability.

However, when restricted to a simple, short-form, and single-instrument task, such as that posed by the MIREX 2025 Symbolic Music Generation challenge [6], these sophisticated techniques may not be necessary, or may even be excessively complicated for the task. We conjecture

that a return to the language-modeling approach taken by the Music Transformer [2], consisting of straightforward next-token prediction on tokenized raw musical data without any augmentations, will outperform foundation models when trained specifically for this task.

2. System Description As indicated previously, an accurate intuition for our approach may be obtained by taking the pretraining methodology of a standard large language model and replacing domain-specific components with their symbolic music counterparts (e.g. tokenized text with tokenized MIDI). We structure the remainder of this extended abstract as a technical report on our methodology, in the interest of reproducibility for future work in this vein.

2.1 Objective We adopt the objective of the MIREX 2025 Symbolic Music Generation challenge, paraphrased as follows: Given a 4-measure piano prompt, with an optional pickup measure, generate a musically coherent 12-measure continuation. Assume all music is in 4/4 time and quantized to a sixteenth-note resolution.

The input and output are JSON objects, containing “prompt” and “generation” keys, which in turn contain lists of objects of the form { “start”: 16, “pitch”: 72, “duration”: 6 }, where “start” ranges from 0-79 for the prompt and 80271 for the generation, and “pitch” ranges from 0-127, corresponding to MIDI pitch numbers. We model the music in MIDI instead of JSON for ease of use with existing computational music tooling, and convert between the formats using scripts provided in the competition baseline.

10

2.2. Data selection and preprocessing We used the recently released Aria-MIDI dataset [7] for training, under the premise that better data is always one of the most effective ways to improve performance.

We used the pruned split, discarded examples with an audio score of less than 0.9, and held out 10% of the remainder for validation. The dataset comprises Type 0 MIDI files automatically transcribed from solo piano performances, which are rarely quantized, so we quantized all files to a sixteenth-note resolution ahead of training.

learning rate scheduler from 1e-4 to 1e-5 and optimized with respect to the standard cross-entropy loss.

Each choice was informed by official recommendations and the first author’s previous experience [13]. Training took 46 hours to complete.

We used the MidiTok library [8] to implement a simplified REMI tokenization [9], resulting in a vocabulary size of 228. In particular, we disabled note velocity encodings (use_velocities=False) and the splitting of tokens at bars or beats (encode_ids_splits=”no”). These settings were chosen to efficiently model the simplified musical representation specified by the competition, as described in Section 2.1.

2.5. Inference We used the rwkv.cpp library [16] to perform inference. The sampling parameters were tuned as follows: temperature=1.0, top-p=0.95, repetition penalty=1.0, top-k=40. We published an end-to-end system as a Docker image, which can be run with the following bash script with command bash script.sh in.json out/ n_sample:

Each training sample was generated by loading a MIDI file from the dataset with the symusic library [10], encoding it with our REMI tokenizer, and selecting a random range of 16 consecutive bars that contained at least 100 tokens, which was passed to the model. Training samples typically ranged from 200 to 1200 tokens in length.

2.3. Model Architecture We adopted a decoder-only RWKV-7 backbone [11] as our sequence model, as it provides better data efficiency and easier training in a resource-constrained environment over the quadratic Transformer [1]. We adopted the “deep and narrow” strategy suggested by Tay et al.

[12] and Zhou-Zheng and Pasquier [13], resulting in a model with 12 layers, a hidden dimension of 384, and a feedforward dimension of 1536, for about 20 million total parameters. This small size allowed remarkably fast training and inference in a resource-constrained consumer environment.

2.4. Training We trained for 50 epochs on a single RTX 4090 using the RWKV-LM library [14], using the Adam optimizer [15], weight decay of 0.1, a batch size of 32, and a sequence length of 1024, as we found that the training sequences rarely exceeded this length. We used a cosine

For testing, we took 8 samples on each of 7 test prompts at intervals of 4 epochs and qualitatively analyzed them.

We found the model from epoch 32 to perform best, as later checkpoints showed a slight decline in quality and earlier checkpoints seemed undertrained.

#!/bin/bash IN_ABS=$(realpath “$1”) OUT_ABS=$(realpath “$2”) mkdir -p $OUT_ABS USER=”christianzhouzheng” IMAGE=”$USER/rwkv-mirex:latest” docker pull “$IMAGE” docker run --rm \ -v “$IN_ABS:/app/input.json:ro” \ -v “$OUT_ABS:/app/output” \ “$IMAGE” “/app/input.json” \ “/app/output” “$3” 3. Results For the competition, a double-blind subjective listening test was performed, in which each of four anonymized models generated 8 samples for each of 8 prompts; from these 8 samples, each team cherry-picked one “best” sample for use in the listening test. The models compared were our 20M RWKV, another entry, and two baselines: the 780M Large Anticipatory Transformer [17] and 1.2B xlarge MuseCoco [5] foundation models.

Each survey participant was asked to rate each sample on a scale from 1 to 5 in each of the four following metrics: coherency, how well the continuation aligns with the style of the prompt; structure, how effectively it main-

11

tains a natural form and phrasing; creativity, the degree of novelty without losing musical sense; and musicality, the overall musical quality. Twenty participants with diverse musical backgrounds responded to the survey, of which 14 completed all eight pages with an average completion time of 32 minutes. The results are reported in Table 1.

5. Acknowledgements We thank the MIREX 2025 organizers and the Symbolic Music Generation task captains for providing us with this opportunity. We also thank the members of the Metacreation Lab for their feedback and support.

This extended abstract was produced by a project under the inaugural EleutherAI Summer of Open AI Research (SOAR), led by project captain Christian Zhou-Zheng. We thank the anonymous Discord user genetyx8 for coordinating SOAR, and Stella Biderman and EleutherAI for generously providing compute for other experiments.

Table 1. Model performance comparison. Each entry is reported as mean ± SEM^s, where SEM stands for the standard error of mean. Superscript s is a letter: within each column, different letters indicate significant differences (p < 0.05) according to a Wilcoxon signed-rank test.

As visible in Table 1, our 20M RWKV-7 model clearly beat the PixelGen submission and the MuseCoco model, and performed on par with the strong Anticipatory Transformer baseline, which has 39 times more parameters (780M). The results reinforce our longstanding intuitions that simple, small, task-specific models can still perform on par or better than much larger foundation models at the specific task in question, and that the sophisticated techniques of foundation models may not be necessary for strong single-task performance.

4. Conclusion We presented a simple, traditional approach to piano music continuation, inspired by recent advances in language modeling. By leveraging state-of-the-art resources and techniques, such as the RWKV-7 architecture [11] and Aria-MIDI dataset [7], we demonstrated that large, generalized foundation models can be matched or outperformed by a smaller, specialized model with a strong foundation. We conclude that there remains a niche for purpose-trained \emph{small models} in an era of progression defined by scaling parameter counts, especially in the field of music.

6. References [1] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” in Proceedings of the 31st International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2017, p. 6000–6010.

[2] C.-Z. A. Huang, A. Vaswani, J. Uszkoreit, N. M. Shazeer, I. Simon, C. Hawthorne, A. M. Dai, M. D. Hoffman, M. Dinculescu, and D. Eck, “Music transformer: Generating music with long-term structure,” in International Conference on Learning Representations, 2018.

[3] B. Yu, P. Lui, R. Wang, W. Hu, X. Tan, W. Ye, S. Zhang, T. Qin, and T.-Y. Liu, “Museformer: transformer with fineand coarse-grained attention for music generation,” in Proceedings of the 36th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2022.

[4] D. von Rütte, L. Biggio, Y. Kilcher, and T. Hofmann, “FIGARO: Controllable music generation using learned and expert features,” in The Eleventh International Conference on Learning Representations, 2023.

[5] P. Lu, X. Xu, C. W. Kang, B. Yu, C. Xing, X. Tan, and J. Bian, “MuseCoco: Generating symbolic music from text,” arXiv preprint arXiv:2306.00110, 2023.

[6] G. Xia, J. Jiang, A. Maezawa, Z. Wang, Y. Zhang, R. Yuan, and J. S. Downie, “MIREX,” https://www.music-ir. org/mirex/wiki/MIREX_HOME, Future MIREX Team, 2025, part of ISMIR Conference.

[7] L. Bradshaw and S. Colton, “Aria-MIDI: A dataset of piano MIDI files for symbolic music modeling,” in In-

12

ternational Conference on Learning Representations, 2025.

[8] N. Fradet, J.-P. Briot, F. Chhel, A. El Fallah Seghrouchni, and N. Gutowski, “MidiTok: A python package for MIDI file tokenization,” in Extended Abstracts for the Late-Breaking Demo Session of the 22nd International Society for Music Information Retrieval Conference, 2021.

[9] Y.-S. Huang and Y.-H. Yang, “Pop music transformer: Beat-based modeling and generation of expressive pop piano compositions,” in Proceedings of the 28th ACM International Conference on Multimedia, New York, NY, USA, 2020, p. 1180–1188.

[10] Y. Liao and Z. Luo, “symusic: A swift and unified toolkit for symbolic music processing,” in Extended Abstracts for the Late-Breaking Demo Session of the 25th International Society for Music Information Retrieval Conference, 2024.

[11] B. Peng, R. Zhang, D. Goldstein, E. Alcaide, X. Du, H. Hou, J. Lin, J. Liu, J. Lu, W. Merrill, G. Song, K. Tan, S. Utpala, N. Wilce, J. S. Wind, T. Wu, D. Wuttke, and C. Zhou-Zheng, “RWKV-7 “Goose” with expressive dynamic state evolution,” arXiv preprint arXiv:2503.14456, 2025.

[12] Y. Tay, M. Dehghani, J. Rao, W. Fedus, S. Abnar, H. W. Chung, S. Narang, D. Yogatama, A. Vaswani, and D. Metzler, “Scale efficiently: Insights from pretraining and finetuning transformers,” in International Conference on Learning Representations, 2022.

[13] C. Zhou-Zheng and P. Pasquier, “Personalizable long-context symbolic music infilling with MIDI-RWKV,” arXiv preprint arXiv:2506.13001, 2025.

[14] B. Peng, “RWKV-LM,” https://github.com/BlinkDL/RWKV-LM, 2025.

[15] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014.

[16] RWKV, “rwkv.cpp,” https://github.com/RWKV/ rwkv.cpp, 2025.

[17] J. Thickstun, D. Hall, C. Donahue, and P. Liang, “Anticipatory music transformer,” Transactions on Machine Learning Research, 2024.

© C. Zhou-Zheng, J. Backsund, D. Li Chan, A. Coventry, A. Eslami, J. Goel, X. Han, D. Soomro, and G. Wei. Licensed under a Creative Commons Attribution 4.0 International License (CC BY 4.0).

Attribution: C. Zhou-Zheng, J. Backsund, D. Li Chan, A. Coventry, A. Eslami, J. Goel, X. Han, D. Soomro, and G. Wei, “A Traditional Approach to Symbolic Piano Continuation”, in Extended Abstracts for the Late-Breaking Demo Session of the 26th Int. Society for Music Information Retrieval Conf., Daejeon, South Korea, 2025.

This paper is available online at arXiv ID 2509.12267 and was published at ISMIR 2025 in September.

13

GENETICS: A Deep Dive

by Saveer Grewal (IV) This summer, I took a Harvard pre-college course that introduced me to a new perspective on how our bodies function through genetics. The class started with the basics of DNA and the human genome, later moving to advanced topics such as cancer genetics, population studies, and new technology like CRISPR. Initially, I thought that the study of genetics just involved DNA, but I learned that it’s really about how genetic information shapes everything from simple traits to serious diseases within our bodies.

The course also incorporated real-life examples of its topics. One student wondered if they had an early risk of breast cancer after a relative was diagnosed, so the instructor explored their odds using Punnett squares. Another example described someone who only developed lactose intolerance later in life, showing that the way you live your life can affect your genetics, in a process called epigenetics.

Later in the course, we studied traits, like height or risk of diabetes, that are not controlled by a single gene but by polygenic inheritance, a pattern where a single trait is influenced by the combined effects of multiple genes. I learned about genome-wide association studies (GWAS), which look at thousands of people to distinguish links between DNA and disease. Enlarging our scale, we also studied population genetics, showing how groups of people have different risks for certain conditions based on their DNA history.

The first topic was about the central dogma, which explains that DNA is transcribed into RNA, which is then translated into an amino acid chain, or a protein.

I also learned how RNA is edited before becoming mRNA. Finally, I learned how small changes in our genetic sequence, known as mutations, can either be harmful or harmless to the resulting protein. Mutations might change just one amino acid, cut a protein short, or not affect the protein at all. Even tiny changes in DNA can have harmful consequences for the body.

The course ended with an outlook on the medical industry. We compared traditional approaches of treatment to precision medicine, which uses a person’s DNA to guide treatment. We also learned how genes affect drug responses and were introduced to CRISPR, a tool that can edit DNA. Learning about CRISPR was really fun and new because it showed me how science is moving toward solving problems that used to seem impossible.

Next, I learned how genes are inherited. I now understand the difference between a genotype, an organism’s complete set of genes, and a haplotype, a set of closely linked genetic markers, and how recombination during meiosis creates variability in inheritance.

I examined studies that enable scientists to identify regions of DNA associated with specific diseases.

This summer course allowed me to experience firsthand how genes work and how they affect our lives.

I now understand how DNA turns into proteins, how genetic variation can cause disease, and how scientists study populations to learn more about health. I also got a glimpse into the future with genome editing and precision medicine. My experience made me more curious about biology and opened my eyes to the power of genetics in both biology and medicine.

The course also covered chromosomes, describing how mistakes in cell division can cause extra or missing chromosomes, such as trisomy 21 (three copies of chromosome 21), or Down syndrome. We also studied structural changes in DNA, which can be found using tests like karyotypes, FISH, or microarrays. The class then proceeded into cancer genetics, explaining tumor suppressor genes, oncogenes, and how mutations cause cancer. An interesting thing I learned was how scientists separate “driver” mutations that cause cancer from “passenger” ones that don’t directly cause it.

14

Gene Editing Technology Conquers Deadly Disease

by Joe Cridge (IV)

Carbamoyl phosphate synthetase 1 (CPS1) deficiency is a genetic disorder in which a necessary enzyme that removes nitrogen from the body is either significantly reduced or absent. Without this enzyme, nitrogen accumulates in the form of ammonia in the blood, causing vomiting, appetite loss, progressive lethargy, coma, and brain damage.

crisprtx.com/gene-editing.

“In Landmark Study, CHOP-Penn Team Treats Newborn With Base-Editing Therapy.” Research Institute at CHOP, www.research.chop.edu/cornerstone-blog/ in-landmark-study-chop-penn-team-treats-newbornwith-base-editing-therapy.

“What is Carbamoyl Phosphate Synthetase 1 Deficiency?” News-Medical, 3 May 2023, www.news-medical. net/health/What-is-Carbamoyl-Phosphate-Synthe-

CPS1 deficiency has two forms: a severe neonatal form, which shows symptoms shortly after birth, and a less severe form that appears later in life and can cause psychiatric illness and developmental delay. Since its discovery in 1962, many attempts have been made to find a cure for this disease; however, none have been successful.

Dietary restrictions, medication to remove ammonia, and hemodialysis during severe flare-ups have been the main treatments. In the most severe cases, liver transplant is considered; however, liver transplants carry major health risks, such as internal bleeding, blood clots, and organ rejection. They may also be unavailable for neonatal patients, who often have the most severe disease.

Figure 1. CRISPR gene-editing visualization.

Source: National Institutes of Health on Flickr.

In February 2025, studies conducted at the University of Pennsylvania demonstrated the potential of CRISPR, a personalized gene-editing technology. CRISPR works by removing the abnormal section of DNA in a gene and replacing it with the correct DNA sequence, resulting in normal nitrogen metabolism. This research represents a major step forward in the field of genetic medicine. Now that scientists have shown that this technology works for CPS1 deficiency, CRISPR can be further studied and potentially used in the future to treat other ultra-rare genetic conditions.

Works Cited “Carbamoyl Phosphate Synthetase 1 Deficiency.” rarediseases.org, rarediseases.org/rare-diseases/carbamoyl-phosphate-synthetase-i-deficiency/.

“Gene Editing.” CRISPR Therapeutics, 10 Nov. 2023,

Figure 2. Protein interaction network for CPS1 deficiency. Source: STRING database.

15

Development of a Hematology Workflow Video at NYU Tisch Hospital

by Julia Ronnen (VI) This past summer, I worked with NYU Tisch Hospital to develop a video that visually represented the entire journey of a patient specimen with​‍​‌‍​‍‌​‍​‌‍​‍‌ the goal to enhance the transparency and the understanding of the hospital hematology lab processes. The main aim of the project was to connect the laboratory staff with the clinical ones so that nurses, doctors, medical students, and new employees can get an insight into the diagnostic pipeline of the hospital.

I collaborated with lab technologists, pathologists, and the hospital’s education office to plan out the whole journey of the specimen. We followed the blood draw, saw the transport through the pneumatic tube system, and recorded the testing stages: from accessioning through the instrument analyzer to manual differentials under the microscope. We also focused heavily on quality-control procedures and looked over the

criteria that were used to notify senior technologists or pathologists about the abnormal findings.

We took videos in the hematology lab where the staff were performing with the advanced instruments in high throughput analyzers, and preparing the slides for the microscope. We also talked with different staff members to get them featured in explaining their roles and how communication between the clinical floors and the lab facilitates the quick and accurate diagnostics.

The video we produced is an easy-to-follow, through, and sequential depiction of the hematology unit’s daily routine, to be used during the orientation of staff, laboratory tours, and educational workshops. Some of the clinicians have already responded that this tool makes the testing process less bewildering and that the communication among different departments is getting better.

Disseminated Histoplasmosis Mimicking Colonic Carcinoma in a Patient with Crohn’s Disease

By Julia Ronnen (VI) , Afreen Karimkhan, Sui Zee, Rong Xia Department of Pathology, New York University, Langone Health, New York, NY, USA Background: Patients with Crohn’s disease are at increased risk for both colorectal carcinoma and opportunistic infections due to chronic inflammation and long-term immunosuppressive therapy. Fungal infections such as histoplasmosis may mimic malignancy or inflammatory exacerbation, leading to diagnostic delay.

Case Presentation: We report a woman in her fifties with a 10-year history of Crohn’s disease treated with a tumor necrosis factor

(TNF) inhibitor and an interleukin-23 antagonist who presented with abdominal pain, diarrhea, and colonic masses. Colonoscopic biopsies showed dense inflammation with necrotizing granulomata. Grocott’s methenamine silver stain and immunohistochemistry revealed intracytoplasmic budding yeasts consistent with Histoplasma capsulatum, confirmed by polymerase chain reaction. The urinary Histoplasma galactomannan antigen test was negative, while serum 1,3 beta-D-glucan was positive. The patient was successfully treated with amphotericin B followed by itraconazole, with complete clinical recovery.

16

Conclusion: Gastrointestinal histoplasmosis can closely resemble Crohn’s flare or malignancy in immunosuppressed patients. Careful histopathologic evaluation and integration of clinical and laboratory findings are essential for timely diagnosis and effective antifungal therapy.

Introduction Crohn’s disease affects an estimated 6–8 million people worldwide and is characterized by chronic, relapsing inflammation of the gastrointestinal tract. (1) Patients with long-standing disease are at increased risk of developing dysplasia and colorectal carcinoma, necessitating regular endoscopic surveillance and vigilant clinicopathologic follow-up. (2) While biologic and immunosuppressive therapies— particularly tumor necrosis factor (TNF) inhibitors and interleukin antagonists—have revolutionized disease management, (3, 4) they also heighten susceptibility to opportunistic infections, including Mycobacterium tuberculosis, cytomegalovirus (CMV), and fungal pathogens. (5) When patients with Crohn’s disease present with active colitis or obstructive lesions, it is critical to differentiate between an inflammatory flare, infection, or neoplastic transformation. Opportunistic infections such as histoplasmosis, tuberculosis, and CMV can produce ulcerative or mass-forming lesions that closely mimic malignancy both endoscopically and radiologically. (5)

treated with intravenous corticosteroids, resulting in marked improvement. Colonoscopy at that time demonstrated severe inflammation at the hepatic flexure.

Given her suboptimal response to adalimumab, therapy was transitioned to risankizumab-rzaa, an interleukin-23 (IL-23) antagonist. After three months of treatment, she again developed worsening abdominal pain and diarrhea.

On examination, she was febrile and mildly hypotensive.

Laboratory studies revealed an elevated C-reactive protein (50.4 mg/L) and anemia (hemoglobin 10.7 g/dL).

Magnetic resonance enterography showed small-bowel strictures and circumferential wall thickening of the ascending colon, which had slightly progressed compared with prior imaging and raised concern for neoplasm.

Colonoscopy revealed circumferential, polypoid, and obstructive masses in the hepatic flexure and cecum (Figure 1A). Based on the clinical presentation and imaging findings, the differential diagnosis included colorectal adenocarcinoma, intestinal tuberculosis, fungal infection, and exacerbation of Crohn’s disease.

Histologic examination of the colonoscopic biopsies demonstrated dense lymphoplasmacytic inflammation with histiocytic aggregates and necrotizing granulomata (Figure 1B). Grocott’s methenamine silver (GMS) staining highlighted numerous 2–4 µm intracytoplasmic budding yeasts consistent with Histoplasma species (Figure 1C). Immunohistochemistry for Histoplasma was positive (Figure 1D).

Here, we report a case of disseminated histoplasmosis in a woman with a 10-year history of Crohn’s disease who was receiving combination biologic therapy with a tumor necrosis factor (TNF) inhibitor and an interleukin-23 antagonist, and who developed colonic masses initially suggestive of carcinoma.

Case Report A woman in her fifties with a 10-year history of Crohn’s disease presented with several weeks of persistent diarrhea, fever, and abdominal pain. Her treatment regimen included adalimumab, a tumor necrosis factor (TNF) inhibitor, administered every two weeks for nine years and later increased to weekly dosing due to partial response.

Although she initially experienced symptom relief, she subsequently developed recurrent severe abdominal pain, prompting an emergency department visit. She was

17

Figure 1.

Patients treated with TNF inhibitors are particularly susceptible to opportunistic infections because tumor necrosis factor is crucial for macrophage activation and granuloma maintenance—key components of host defense against intracellular pathogens such as Histoplasma capsulatum and Mycobacterium tuberculosis. (8) The addition of IL-23 antagonists may further suppress cell-mediated immunity and increase the risk of disseminated fungal infection. (9, 10)

Figure 2.

Despite corticosteroid and antibiotic therapy, the patient developed worsening hypotension, necessitating emergent right hemicolectomy with en bloc liver wedge resection. Surgical pathology revealed extensive fungal infection, chronic inflammation, and necrotizing granulomata (Figure 2 A-C) involving both the colon and liver.

Computed tomography of the chest demonstrated multiple pulmonary nodules consistent with disseminated fungal infection. Polymerase chain reaction (PCR) testing on the surgical specimen confirmed Histoplasma capsulatum.

Interestingly, the urinary Histoplasma galactomannan antigen test remained negative, while the serum 1,3 beta-D-glucan was positive. The constellation of histopathologic findings and confirmatory PCR results supported the diagnosis of disseminated histoplasmosis.

The patient was treated with intravenous amphotericin B followed by oral itraconazole for maintenance therapy.

She showed marked clinical and biochemical improvement and remained clinically stable without recurrence during 12 months of follow-up (Figure 2D).

Discussion This case highlights the diagnostic complexity of gastrointestinal histoplasmosis in patients with Crohn’s disease receiving biologic immunosuppressive therapy. Colonic histoplasmosis can closely mimic carcinoma or inflammatory pseudotumor both clinically and radiologically, often leading to delayed diagnosis and unnecessary surgical intervention. The overlapping presentations of Crohn’s flare, infectious colitis, and treatment-related complications further complicate interpretation. (6, 7)

The differential diagnosis for mass-forming or granulomatous colonic lesions in an immunosuppressed patient is broad. Neoplastic causes include Crohn’s disease–associated colorectal carcinoma and inflammatory pseudotumor. Infectious etiologies encompass intestinal tuberculosis, which can produce circumferential wall thickening and necrotizing granulomata; cytomegalovirus (CMV) colitis, which can mimic an inflammatory flare with ulcerative or pseudotumoral lesions; and systemic mycoses such as coccidioidomycosis, cryptococcosis, and blastomycosis, each distinguishable by its unique fungal morphology. Noninfectious entities such as sarcoidosis, Yersinia enterocolitica infection, and intestinal lymphoma can also present with granulomatous inflammation and must be considered.

Accurate diagnosis depends heavily on histopathologic evaluation. Fungal organisms are often sparse, and extensive necrosis or mixed inflammation can obscure them on hematoxylin and eosin (H&amp;E) sections.

Special stains such as Grocott’s methenamine silver (GMS) and periodic acid–Schiff (PAS) are essential for detecting yeast forms, while immunohistochemistry and fungal polymerase chain reaction (PCR) provide confirmatory evidence, particularly when cultures are negative or unavailable.

An important diagnostic pitfall is the limited sensitivity of antigen testing. The Histoplasma urinary antigen assay is less sensitive in localized gastrointestinal disease than in disseminated infection. (11,12) Conversely, serum 1,3 beta-D-glucan, though nonspecific, can support the diagnosis of invasive fungal infection when interpreted in clinical context. (12) A comprehensive diagnostic approach that integrates clinical findings, laboratory data, imaging, and histopathology is therefore critical.

This case also underscores the importance of obtaining a detailed history—including prior exposures, environmental risks, or unexplained systemic symptoms—even

18

when travel to endemic regions is not documented, as such details can reveal key clues to opportunistic infection.

From a management standpoint, prompt recognition and antifungal therapy are lifesaving. For moderate to severe disease, induction with intravenous amphotericin B followed by oral itraconazole for 6–12 months remains standard. Lifelong prophylaxis may be considered in patients requiring ongoing immunosuppression. (13) The decision to resume biologic therapy should balance the need for Crohn’s disease control against the risk of fungal relapse. Multidisciplinary coordination among gastroenterology, infectious disease, and pathology teams is essential for optimal care.

Long-term monitoring should include clinical assessment, laboratory evaluation, and imaging to detect recurrence. This patient’s favorable response to antifungal therapy and withdrawal of biologics underscores the importance of early diagnosis and comprehensive follow-up.

In summary, disseminated histoplasmosis should be included in the differential diagnosis of new or obstructive colonic lesions in Crohn’s disease patients receiving biologic therapy. Maintaining a high index of suspicion for infection and recognizing histologic clues can prevent misdiagnosis as carcinoma or refractory inflammation and lead to timely, targeted treatment.

Organizing and Mentoring for the Inaugural EleutherAI Summer of Open AI Research

by Christian Zhou-Zheng (VI)

I. Introduction This summer, I had the privilege of mentoring a group of students in AI research under a program my friend was running at EleutherAI. For some background, EleutherAI is an open science lab, meaning that anyone—even you and me—can come in and contribute to original research. Major, impactful research projects have spun out of people with good ideas simply coming to the open forum and finding collaborators!

Examples include The Pile, the first implementation of “why don’t we just train AI on the entire Internet;” the LM Evaluation Harness, used to provide the performance benchmarks in the press releases of OpenAI and friends; the YaRN context length extension technique; and the RWKV linear attention architecture, which I worked on. I am a huge proponent of open science and encourage others to get involved as well, and if you’re interested in hands-on learning and making an impact in machine learning research, come join us on Discord at the link on www.eleuther.ai.

For some background, I had been helping organize a research program at Eleuther where we would connect interested students with mentors who had project

ideas, providing experience for the students and manpower for the mentors. This turned into the Summer of Open AI Research (SOAR; not to be confused with OpenAI), a month-long program during August.

At the same time, I had been notified by a colleague about a competition being held by ISMIR 2025, the Conference of the International Society for Music Information Retrieval, to train models for computational music tasks; given the matching timelines, I thought it would be a great idea to run a SOAR project to participate in this competition, in conjunction with running the program overall. Therefore I’ll talk some about both the administration and the mentoring I did.

II. Administering the Program After setting up the website (www.eleuther.ai/soar), we advertised in a few large research communities, including Eleuther itself as well as Cohere Labs and Apart Research, to gain exposure. The program ultimately received over 500 applications for 50-100 total spots across 12 projects, wildly exceeding our expectations.

Each applicant was asked to provide some personal background, technical background, and answer a few

19

questions about themselves. They then ranked their top 5 projects, and each project mentor subsequently graded each applicant to their project on a scale from 1 to 10. The hospital resident matching algorithm was used to match applicants to projects, and mentors could take as many applicants as they wished. The acceptance rate ended up around 30%, as the mentors accepted more applicants than needed to account for expected dropouts over the course of the program.

Some statistics are provided in Figure 1. The vast majority of applicants were in their 20s, and they were predominantly male. Most applicants held either a bachelor’s or masters’ degree, or were in the process of studying for one. 90.3% claimed to have some prior research experience, but only 41.9% had published in a peer-reviewed venue prior to attending the program, and only 16.5% had participated in a similar event previously (events like this are not very common). Importantly, 85.9% were of the opinion that Canada geese are based, while the other 14.1% called them cringe.

The application process took up most of the administrative time, as each project was meant to be self-sufficient. The only remaining tasks were to distribute and manage sponsored GPU clusters among the projects that required them, and to organize weekly talks for the mentees to share their progress with the program as a whole, which took negligible time.

III. Mentoring For the Program I accepted 11 mentees, of whom 3 dropped out within the first week; there were no dropouts after that. Our preprint is elsewhere in this edition of PCR, entitled “A Traditional Approach to Symbolic Piano Continuation”, so following is a discussion of the process that gave rise to that preprint and some of my thoughts on it.

My mentees were mostly familiar with text modeling but unfamiliar with music, so I provided them some resources to read up on ahead of the program’s start. I was pleased to see that they had all read and internalized the core concepts by our first meeting, which was good foreshadowing for the agency they would display throughout the program; others I mentored previously had not taken such initiative without some fairly significant prodding.

20

Figure 1. Distribution of gender, age bracket, and education level among applicants to SOAR 2025.

I ended up splitting the group into two for the first week.

One half did a literature review on select papers, while the other half began coding a framework that we could use to design, run, and evaluate experiments once we received access to the 8x NVIDIA A100 pod assigned by the program. I provided four papers and a seed codebase, then let my mentees work from there. The two “teams” engaged in more cross-team communication than I had anticipated, and by halfway through the week, the lines had blurred enough that I called off the separation.

Their effective collaboration resulted in initial code for each of their several ideas well ahead of when I had expected, and they took it upon themselves to separate the ideas into a hierarchy so that they could attempt solutions of various levels of complexity by the competition’s end: a naive sequence modeling approach with Transformers, a naive sequence modeling approach with RWKV, and a more involved retrieval-augmented approach. They were clearly well-versed in the scientific process - or at least those who were not were quickly brought up to speed by those who were - and their gusto in setting up a training and evaluation pipeline ahead of time paid off once we were able to begin experimenting, as they made significant progress.

Over those two weeks, I mostly stepped back from the technical aspects of the project, relegating myself to answering the occasional question about methodology and priors and hosting weekly update meetings.

This was on account of a few factors: the team was effectively self-sufficient at this point, organizing among themselves, which I was delighted to see; and in any case, I had some real-life affairs to attend to. I will thus fast-forward to the submission process, wherein we collectively rated the outputs of each of the three models we had trained to select one for final submission, which ended up being the RWKV model (preprint available in this edition of PCR). To spoil the outcome, we won the competition!

in expecting to need to do a considerable amount of hand-holding; I attribute the lack of need for this to the remarkable amount of effort each of the team members put into the project, but also to the depth of their existing knowledge, as a large part of the team’s success was their ability to transfer skills in the scientific process - particularly in machine learning, or at least computer science - from prior experience to this project. It is not clear to me how the team would have performed without such background, but I expect they would not have done nearly as well, nor learned as much from it as they did.

As both mentor and administrator, this made me wonder: should we optimize for impact by selecting the most experienced applicants, or for educational value by accepting those with less background who might benefit more from the mentorship? The answer was not clear during the application process, but I think the sweet spot for something like SOAR lies in the middle, where participants already had strong fundamentals but limited domain-specific knowledge. This way, mentors don’t need to bother with the very basics, and can instead focus on helping their mentees do meaningful work in a more focused area. This need not necessarily have less educational value than the alternative, either.

From an organizational perspective, the program succeeded beyond our expectations. The original impetus for creating SOAR was the lack of accessible research opportunities in machine learning (outside “AI safety”, which is a mostly performative field, and has a whole host of other problems), and I think the overwhelming application response speaks to the demand for such opportunities. I think that if we advertised more widely, we would have received orders of magnitude more responses, and there needs to be more supply to address this demand.

Looking forward, I hope to see SOAR and open science programs like it continue and expand. For students seeking research experience, I cannot recommend programs like this highly enough!

IV. Reflections So what is to be learned from this experience? I found my involvement in the mentoring process to be remarkably hands-off: I had to touch very little code or papers after the initial setup process, and I served mostly an advisory role to my mentees. I had gone

21

Psychological Impact of the Beatles’ Music

by Shanti Swadia (V) Introduction Very few musical groups in history have inspired as much academic, cultural, and scientific study as the Beatles. Beginning in Liverpool in the early 1960s, John Lennon, Paul McCartney, George Harrison, and Ringo Starr achieved commercial success and soon reshaped the foundations of popular music. Their genius did not stem from catchy melodies or cultural timing, but from their intuitive grasp of musical principles that music theorists could only later explain. From unconventional chord progressions to experimental time signatures, the Beatles consistently expanded the genre of rock and pop, creating many new harmonic and rhythmic possibilities.

At the same time, their music deeply resonated with listeners on a psychological level. The Beatles managed to balance familiarity and novelty in their songs, rewarding audiences with moments of surprise while keeping them grounded in predictable patterns, which explains why their songs became so popular and influential. Studying the Beatles through a scientific lens reveals how their work exemplifies the principles of cognitive psychology, auditory perception, and music theory. Likewise, the group’s capacity to innovate within a popular genre that already existed demonstrates their artistic ability and instinctive talent for advanced musical science. The Beatles’ contributions will be explored in three categories: their novel approaches to songwriting, their theoretical innovations in harmony and rhythm, and the psychological effects that made their music so compelling to the audience.

The Genius of Songwriting One reason the Beatles stand out in music history is the originality of their songwriting. Lennon and McCartney’s partnership produced a plethora of songs that consistently experimented with form and content. Unlike many rock and roll groups of the early 1960s, whose songs often relied on repeated blues progressions or formulaic love lyrics, the Beatles’ approaches to structure and harmony had a great variety. A striking example is the way the band used unexpected chord progressions.

For instance, songs like “If I Fell” (1964) and “Something” (1969) introduced harmonic shifts that moved away from

the standard I-IV-V pattern common in early rock music.

Instead of treating chords as fixed patterns to repeat, they took advantage of these opportunities to shift emotional tone, leading listeners into surprising directions while keeping the melody consistent. Harrison’s songwriting also contributed significantly to this complexity. His exposure to Indian classical music after studying with Ravi Shankar introduced drone notes—sustained or continuously repeated tones that provide a constant harmonic foundation beneath the melody—and modal scales, which are scales that differ from the traditional major and minor patterns of Western music (3). These features gave songs like “Within You Without You” a meditative mood distinct from Western harmony. These elements were central aspects of Indian music, which emphasizes sustained tones and gradual melodic development rather than quick harmonic changes. The Beatles’ lyrics also became more sophisticated as the band matured over the years. Early hits like “I Want to Hold Your Hand” (1964) relied on simple romantic themes, but by the midto-late 1960s, the Beatles were integrating more surreal imagery and introspection, as demonstrated in “Yesterday” (1965) and “Lucy in the Sky with Diamonds” (1967).

Many scholars suggest that this shift mirrored larger cultural movements in the 1960s, but it also revealed the band’s ability to connect with listeners at different levels of emotional depth (2). The Beatles’ songwriting reflected an intuitive understanding of variation and contrast.

Many of their songs avoid repeating verses and choruses in the same way, instead introducing subtle rhythmic or melodic changes. This tech technique keeps songs engaging while preventing the listener from getting restless. It is one reason why even their simplest tracks often sound richer than one would expect. In short, the Beatles’ songwriting was not just successful because it was catchy, but because the band consistently subverted the listener’s expectations. They created songs that felt familiar enough to be accessible but inventive enough to stay interesting. Though these moments were initially regarded as spontaneous creativity, later analysis by music theorists revealed that the Beatles were unconsciously pushing the boundaries of musical structure in ways that could be explained through theory and science.

22

Innovations in Music Theory The Beatles’ impact can also be understood through their contributions to the technical language of music. While they were not formally trained in music theory, they repeatedly made choices that scholars now recognize as highly advanced. These innovations included the use of unconventional chord progressions, irregular time signatures, and new approaches to harmony. Many popular songs in the early 1960s followed the predictable I-IV-V blues progression, but the Beatles frequently altered this melodic formula. In their 1967 hit “Penny Lane”, the chorus cycles through a progression that moves unexpectedly from B major to A major before returning to the tonic, creating a sense of life and surprise. Similarly, in “Something” (1969), Harrison uses chromatic bass motion—a technique in which the bass line moves by half steps rather than by larger intervals within the scale, producing a connected sense of movement that adds tension without disrupting the overall harmony (5). The Beatles were very expressive with their chords and did not treat them as fixed rules, which allowed them to produce harmonies that stood apart from mainstream rock (6). They also explored rhythmic structures that were rare in popular music at the time. “All You Need Is Love” (1967) introduced bars in 7/4, defying the regular 4/4 pattern that dominated rock music, while “Happiness Is a Warm Gun” (1968) advanced this rhythmic experimentation even further by shifting between 4/4, 3/4, and 9/8. These experiments proved that complex rhythms could still gain mass popularity if balanced with strong melodic writing. Beyond chords and rhythm, the Beatles expanded the harmonic and modal devices of pop. For instance, Lennon and McCartney experimented with modal interchange, using chords from parallel minor keys, as in “Norwegian Wood” (1965). These choices enriched their sound and created textures typically associated with jazz or classical music rather than rock. What is most interesting about these innovations is how naturally they were incorporated into their songs. They introduced advanced harmonic and rhythmic ideas in ways that listeners could absorb even if they did not have formal music theory training (2).

Psychological Impact of the Beatles’ Music While the Beatles’ music can be analyzed in terms of chords, rhythms, and harmonies, its significance is also tied to how it affects listeners psychologically. Their popularity cannot only be explained by cultural timing, as it also reflects how their innovations engaged the human brain. Research in cognitive psychology and music per-

ception suggests that people enjoymusic that balances predictability with surprise (1). In a 2019 study published in Nature, Bianco found that melodies containing small, unpredictable variations generated greater physiological arousal, as measured by pupil dilation. This finding, shown in Figure 2, supports the idea that musical surprise activates the brain’s reward system and increases listener engagement.

As shown in Figure 2, unpredictable melodies in highly liked music produced the greatest pupil dilation, indicating that moments of surprise in sound evoke emotional responses. The Beatles’ frequent use of harmonic twists and rhythmic irregularities worked similarly, keeping listeners physiologically and emotionally engaged through the tension between their expectations and the novelty.

Psychologists describe this as a reward-based response: when a song slightly fulfills or slightly violates our expectations, it triggers a release of dopamine, enhancing enjoyment and emotional engagement. The Beatles mastered this intuitively. For example, they often used deceptive cadences, progressions that resolve to an unexpected chord instead of the tonic. In “Not a Second Time” (1963), they close phrases with harmonic turns that subvert listeners’ expectations, producing a moment of surprise that feels fresh and satisfying. The Beatles also understood the psychological pull of contrast. Songs like “A Day in the Life” (1967) move between dreamlike verses and jarring orchestral climaxes, creating strong emotional swings. Correspondingly, the lyrics amplified this effect. The band moved from simple love songs to more introspective and surreal songs, offering multiple layers of meaning. This broadened their audience appeal because listeners could interpret songs on both personal and cultural levels. Their lyrical growth paralleled social changes, which allowed audiences to connect with the music both intellectually and emotionally (2). Finally, the Beatles benefited from what psychologists call the “mere exposure effect.” The mere exposure effect details that the more people hear a song, the more they tend to like it. The Beatles avoided the problem of repetitious fatigue by varying song structures and introducing unexpected elements, thus making repeated listening more rewarding rather than monotonous. Their songs became global phenomena because they engaged both the cultural moment and the psychology of listening in ways that no band had ever done before (7).

23

Conclusion The success of the Beatles was largely the result of consistent creative risks. Their songwriting pushed beyond simple formulas, their music theory innovations redefined what rock could sound like, and their ability to affect listeners psychologically made their songs unforgettable. The band followed their instincts and science: they wrote from intuition, but psychologists have scientifically explained why those choices were so powerful.

The Beatles’ legacy does not exist only in their extensive

catalog of hit songs, but also in the way they transformed the relationship between complexity and accessibility in music. By studying the Beatles through a scientific lens, it is clear why they remain a central subject of analysis.

Their songs were experiments in harmony, rhythm, and human perception. Half a century later, many still consider them to be the greatest band of all time, and they continue to demonstrate that popular music can be as intellectually rich as it is emotionally powerful.

Figure 1. Comparison between a standard I-IV-V progression in early rock music and the Beatles’ altered progression in “Penny Lane” (1967), which introduces unexpected harmonic shifts. The left table shows the standard I-IV-V progression typical of early rock and roll, which cycles predictably through three primary chords (for example, A-D-E or C-F-G). The right table shows the chord sequence from “Penny Lane,” which diverges sharply from the simplicity of chords in the left table. Instead of remaining in one key and following a single tonal direction, the Beatles shift between B major and B minor, using chords such as G#m7, C#m7, and Bm6/G# to create chromatic motion and harmony.

Figure 2. Pupil dilation responses to predictable (P) and unpredictable (U) melodies across different levels of liking (High, Medium, Low). The x-axis represents 13 time bins during melody listening (5.6 seconds total), while the y-axis shows changes in pupil size relative to baseline. Each panel compares predictable and unpredictable melodies within a liking condition, and the scatter plot on the right shows the correlation between mean pupil dilation and reward sensitivity. Data is adapted from Bianco (2019), which examined physiological responses to musical predictability in non-musicians using eye-tracking measures of pupil dilation.

24

Works Cited 1. Bianco, R. “Music predictability and liking enhance pupil dilation and promote motor learning in non-musicians.” Nature.com, www.nature.com/articles/s41598019-53510-w/figures/3. Accessed 29 Sept. 2025.

2. The Bluze. “What Makes The Beatles So Special?” Thebluze.com, 13 Oct. 2017, thebluze.com/2017/10/13/ what-makes-the-beatles-so-special/.

3. Bose, Ajoy. “The ‘karmic connection’ between The Beatles’ George Harrison and Ravi Shankar.” Qz.com, 20 July 2022, qz.com/india/1210560/the-karmic-connection-between-the-beatles-george-harrison-and-ravishankar. Accessed 29 Sept. 2025.

4. “Clever Chord Progressions - Penny Lane (The Beatles).” Youtube.com, uploaded by Walk That Bass, Google, 10 Jan. 2020, www.youtube.com/watch?v=qPXc0urrpuA. Accessed 29 Sept. 2025.

5. Colbeck, Cameron. “The Genius of George Harrison As Told By Abbey Road’s Cameron Colbeck.” Abbeyroad.

com, 25 Feb. 2021, www.abbeyroad.com/news/the-genius-of-george-harrison-as-told-by-abbey-roads-cameron-colbeck-2737. Accessed 29 Sept. 2025.

6. Hartzog, Brian. “The Beatles Songwriting.” Brianhartzog.com, www.brianhartzog.com/beatles/beatles-songwriting.htm. Accessed 29 Sept. 2025.

7. Liverpool Museums. “What Made The Beatles Global Stars?” Liverpoolmuseums.org.uk, National Museums Liverpool, www.liverpoolmuseums.org.uk/stories/ what-made-beatles-global-stars. Accessed 29 Sept.

2025.

8. Richard, Maurice. “What Does It Mean When A Song Uses The 12 Bar Blues?” Halifaxguitarlessons.com, halifaxguitarlessons.com/what-does-it-mean-when-asong-uses-the-12-bar-blues/.

9. “10 Ways The Beatles Changed The World of Music Forever.” Soundoflife.com, 25 Mar. 2022, www.soundoflife.com/blogs/mixtape/6-ways-the-beatles-changedthe-world-of-music-forever. Accessed 29 Sept. 2025.

25

Art Fundamentals Cartoon Credits.

We are pleased to feature illustrations from Form III students of the Art Fundamentals course in this fall edition of PCR! Listed here are credits for each piece.

Aanav Shah (III) p. 2

Alexa Davies (III) p. 2 Quinn Nolan (III) p. 9

Celia Lowenstein (III) p. 9

Santiago Galvan (III) p. 9 Levi Pearl (III) p. 13 26

Art Fundamentals Cartoon Credits.

Noah Maloney (III)

Brendan Lin (III) p. 13

Caroline Ouyang (III) p. 25 Anika Gupta (III) p. 25 Anika Gupta (III) p. 25 27

The Pingry School 131 Martinsville Road Basking Ridge, NJ 07920 908-647-5555