Jaewon Jang

Amarda Shehu, Adonyas Ababu, Asma Akbary et al.

Claims about whether large language model (LLM) chatbots "reason" are typically debated using curated benchmarks and laboratory-style evaluation protocols. This paper offers a complementary perspective: a student-led field experiment embedded as a midterm project in UNIV 182 (AI4All) at George Mason University, a Mason Core course designed for undergraduates across disciplines with no expected prior STEM exposure. Student teams designed their own reasoning tasks, ran them on widely used consumer chatbots representative of current capabilities, and evaluated both (i) answer correctness and (ii) the validity of the chatbot's stated reasoning (for example, cases where an answer is correct but the explanation is not, or vice versa). Across eight teams that reported standardized scores, students contributed 80 original reasoning prompts spanning six categories: pattern completion, transformation rules, spatial/visual reasoning, quantitative reasoning, relational/logic reasoning, and analogical reasoning. These prompts yielded 320 model responses plus follow-up explanations. Aggregating team-level results, OpenAI GPT-5 and Claude 4.5 achieved the highest mean answer accuracy (86.2% and 83.8%), followed by Grok 4 (82.5%) and Perplexity (73.1%); explanation validity showed a similar ordering (81.2%, 80.0%, 77.5%, 66.2%). Qualitatively, teams converged on a consistent error signature: strong performance on short, structured math and pattern items but reduced reliability on spatial/visual reasoning and multi-step transformations, with frequent "sound right but reason wrong" explanations. The assignment's primary contribution is pedagogical: it operationalizes AI literacy as experimental practice (prompt design, measurement, rater disagreement, and interpretability/grounding) while producing a reusable, student-generated corpus of reasoning probes grounded in authentic end-user interaction.

Taebong Kim, Youngsik Hong, Minsik Kim et al.

We present Darwin Family, a framework for training-free evolutionary merging of large language models via gradient-free weight-space recombination. We ask whether frontier-level reasoning performance can be improved without additional training, by reorganizing latent capabilities already encoded in existing checkpoints. Darwin introduces three key ideas: (i) a 14-dimensional adaptive merge genome enabling fine-grained component- and block-level recombination; (ii) MRI-Trust Fusion, which adaptively balances diagnostic layer-importance signals with evolutionary search through a learnable trust parameter; and (iii) an Architecture Mapper that enables cross-architecture breeding between heterogeneous model families. Empirically, the flagship Darwin-27B-Opus achieves 86.9% on GPQA Diamond, ranking #6 among 1,252 evaluated models, and outperforming its fully trained foundation model without any gradient-based training. Across scales from 4B to 35B parameters, Darwin models consistently improve over their parents, support recursive multi-generation evolution, and enable a training-free evolutionary merge that combines Transformer- and Mamba-based components. Together, the Darwin Family demonstrates that diagnostic-guided evolutionary merging is a practical and reproducible alternative to costly post-training pipelines for reasoning-centric language models.

Hongtao Lin, Haoyu Chen, Jaewon Jang et al.

User-to-item retrieval has been an active research area in recommendation system, and two tower models are widely adopted due to model simplicity and serving efficiency. In this work, we focus on a variant called \textit{conditional retrieval}, where we expect retrieved items to be relevant to a condition (e.g. topic). We propose a method that uses the same training data as standard two tower models but incorporates item-side information as conditions in query. This allows us to bootstrap new conditional retrieval use cases and encourages feature interactions between user and condition. Experiments show that our method can retrieve highly relevant items and outperforms standard two tower models with filters on engagement metrics. The proposed model is deployed to power a topic-based notification feed at Pinterest and led to +0.26\% weekly active users.

Jaewon Jang, Bonjun Choi

Federated learning ensures the privacy of clients by conducting distributed training on individual client devices and sharing only the model weights with a central server. However, in real-world scenarios, the heterogeneity of data among clients necessitates appropriate personalization methods. In this paper, we aim to address this heterogeneity using a form of parameter decoupling known as representation learning. Representation learning divides deep learning models into 'base' and 'head' components. The base component, capturing common features across all clients, is shared with the server, while the head component, capturing unique features specific to individual clients, remains local. We propose a new representation learning-based approach that suggests decoupling the entire deep learning model into more densely divided parts with the application of suitable scheduling methods, which can benefit not only data heterogeneity but also class heterogeneity. In this paper, we compare and analyze two layer scheduling approaches, namely forward (\textit{Vanilla}) and backward (\textit{Anti}), in the context of data and class heterogeneity among clients. Our experimental results show that the proposed algorithm, when compared to existing personalized federated learning algorithms, achieves increased accuracy, especially under challenging conditions, while reducing computation costs.

Premkumar Vincent, Gwenaelle Cunha Sergio, Jaewon Jang et al.

Thin-film solar cells are predominately designed similar to a stacked structure. Optimizing the layer thicknesses in this stack structure is crucial to extract the best efficiency of the solar cell. The commonplace method used in optimization simulations, such as for optimizing the optical spacer layers' thicknesses, is the parameter sweep. Our simulation study shows that the implementation of a meta-heuristic method like the genetic algorithm results in a significantly faster and accurate search method when compared to the brute-force parameter sweep method in both single and multi-layer optimization. While other sweep methods can also outperform the brute-force method, they do not consistently exhibit $100\%$ accuracy in the optimized results like our genetic algorithm. We have used a well-studied P3HT-based structure to test our algorithm. Our best-case scenario was observed to use $60.84\%$ fewer simulations than the brute-force method.