arXiv subjects

Arjun Verma

Publications and source records attributed to Arjun Verma.

4 recordsLinked to original sources

arXiv2026

RETROSPECT: RETROsynthesis via Sequential Prediction, and Chemically Transformed-ranking

Single-step retrosynthesis needs both accurate first-ranked suggestions and candidate lists that are rich enough for downstream selection. We study this as a proposal-selection decomposition. Our system, RETROSPECT, combines a single Transformer proposal model, which we call the ChemAlign Transformer, with a LambdaMART reranker over structural, reaction-template, upstream-score, and optional DFT-derived descriptors. The generator is trained with hybrid root-aligned and random SMILES augmentation, Pre-LayerNorm, tied embeddings, exponential moving average weights, and a differentiable atom-balance auxiliary loss. On the full USPTO-50K test set of 5,007 reactions, the generator reaches 55.00% top-1 and 86.18% top-10 exact-match accuracy with 99.86% top-1 validity. On the merged candidate-pool benchmark used for reranking, which contains 5,007 test products and about 111 candidates per product, a LambdaMART model trained on the structural feature set reaches 59.4% top-1 with 0.7171 mean reciprocal rank. Feature ablations show that upstream proposal score and template-frequency statistics provide most of the reranking signal, while DFT and reaction-center DFT features provide smaller and less consistent gains. These results support a modular view of retrosynthesis: stronger single-model proposal and learned candidate selection are complementary, and the proposal model can serve as a drop-in component for ensemble systems such as RetroChimera (Maziarz et al., 2024)

cs.LG↗

arXiv2025

Computational Economics in Large Language Models: Exploring Model Behavior and Incentive Design under Resource Constraints

Large language models (LLMs) are limited by substantial computational cost. We introduce a "computational economics" framework that treats an LLM as an internal economy of resource-constrained agents (attention heads and neuron blocks) that must allocate scarce computation to maximize task utility. First, we show empirically that when computation is scarce, standard LLMs reallocate attention toward high-value tokens while preserving accuracy. Building on this observation, we propose an incentive-driven training paradigm that augments the task loss with a differentiable computation cost term, encouraging sparse and efficient activations. On GLUE (MNLI, STS-B, CoLA) and WikiText-103, the method yields a family of models that trace a Pareto frontier and consistently dominate post-hoc pruning; for a similar accuracy we obtain roughly a forty percent reduction in FLOPS and lower latency, together with more interpretable attention patterns. These results indicate that economic principles offer a principled route to designing efficient, adaptive, and more transparent LLMs under strict resource constraints.

cs.CL↗

arXiv2021

Modeling Prejudice and Its Effect on Societal Prosperity

Existing studies on prejudice, which is important in multi-group dynamics in societies, focus on the social-psychological knowledge behind the processes involving prejudice and its propagation. We instead create a multi-agent framework that simulates the propagation of prejudice and measures its tangible impact on the prosperity of individuals as well as of larger social structures, including groups and factions within. Groups in society help us define prejudice, and factions represent smaller tight-knit circles of individuals with similar opinions. We model social interactions using the Continuous Prisoner's Dilemma (CPD) and a type of agent called a prejudiced agent, whose cooperation is affected by a prejudice attribute, updated over time based both on the agent's own experiences and those of others in its faction. Our simulations show that modeling prejudice as an exclusively out-group phenomenon generates implicit in-group promotion, which eventually leads to higher relative prosperity of the prejudiced population. This skew in prosperity is shown to be correlated to factors such as size difference between groups and the number of prejudiced agents in a group. Although prejudiced agents achieve higher prosperity within prejudiced societies, their presence degrades the overall prosperity levels of their societies. Our proposed system model can serve as a basis for promoting a deeper understanding of origins, propagation, and ramifications of prejudice through rigorous simulative studies grounded in apt theoretical backgrounds. This can help conduct impactful research on prominent social issues such as racism, religious discrimination, and unfair immigrant treatment. This model can also serve as a foundation to study other socio-psychological phenomena in tandem with prejudice such as the distribution of wealth, social status, and ethnocentrism in a society.

cs.MA↗

arXiv2017

Nonstandard Convergence to Jamming in Random Sequential Adsorption: The Case of Patterned One-Dimensional Substrates

We study approach to the large-time jammed state of the deposited particles in the model of random sequential adsorption. The convergence laws are usually derived from the argument of Pomeau which includes the assumption of the dominance, at large enough times, of small landing regions into each of which only a single particle can be deposited without overlapping earlier deposited particles and which, after a certain time are no longer created by depositions in larger gaps. The second assumption has been that the size distribution of gaps open for particle-center landing in this large-time small-gaps regime is finite in the limit of zero gap size. We report numerical Monte Carlo studies of a recently introduced model of random sequential adsorption on patterned one-dimensional substrates that suggest that the second assumption must be generalized. We argue that a region exists in the parameter space of the studied model in which the gap-size distribution in the Pomeau large-time regime actually linearly vanishes at zero gap sizes. In another region, the distribution develops a threshold property, i.e., there are no small gaps below a certain gap size. We discuss the implications of these findings for new asymptotic power-law and exponential-modified-by-a-power-law convergences to jamming in irreversible one-dimensional deposition.

cond-mat.stat-mech↗