hi, i'm adrian Adrian Erlikhman

Do LLM Fingerprints Survive Outside English?

Accepted · IEEE BigData 2026 High School Symposium · Phoenix, 17 Dec

A 21-feature classifier names the model 59–72% of the time in all seven languages, against 20% chance. Asked the same question, the models score 20–25%.

Attributing text to a specific model has been studied almost only in English. We collect 840 essays from five frontier models (GPT-5.5, Claude Opus 4.7, Gemini 3.5 Flash, Grok 4.3, DeepSeek V4 Pro) in seven languages from English to Hindi, and attribute them with 21 interpretable features computed from Universal Dependencies parses. A five-way classifier reaches 59–72% in every language against 20% chance, and we detect no decline as a language's resources fall. Character n-grams are more accurate within a language, but the interpretable features transfer between languages better (49% against 41%). Attribution holds at 57–70% on machine-translated text, while the five models, asked which of them wrote each text, score only 20–25%.

Curated to Win: Certifying Reward-Model and Judge Selection Against Evaluation-Set Curation

RewardBench's leader survives the removal of any 11 of its 2,985 items. Removing 12 particular items makes the runner-up the winner.

Automated research and self-improvement pipelines pick a reward model or LLM judge by its accuracy on a validation set, and that set is often curated by the pipeline or its designers. The paper's curation certificate says how much of the set could be removed before the winner changes. In benchmark weight it is at least the gap between the top two models; in items it is exact, from one sort per challenger. Linear programs show the bound is tight near the top of RewardBench (118 models) and RewardBench 2 (188). RewardBench's leader survives any 11 removed items, and RewardBench 2's any 42. On 200-item RewardBench validation sets, picks whose top two are more than 2 points apart match the full-benchmark winner 87% of the time, against 43% for the rest.

Whose Line Is It Anyway? Auditing Self- and Other-Recognition in LLM Judges

The usual test flags self-recognition in 15 of 37 cases. 8 survive a false-alarm control, 3 beat the pooled peers, and none beats its strongest single peer.

LLM judges are anonymized on the assumption that they cannot tell who wrote a text. The usual check is whether a judge names itself on its own writing more often than chance. We add two controls from the same experiment: how often the judge names itself on text it did not write, and how often the other judges name it on its text. Across nine judges, 1,270 texts and over 19,000 judgments, the usual check flags 15 of 37 cases. In 7 the judge names itself as often on others' text. Of the other 8, all frontier models, 3 beat the pooled peers and none beats its strongest single peer: Claude names itself on 33 of its 38 chat-app essays, and GPT names Claude on 32. The authorship signal is in the text (a style classifier attributes 86% of the essays); the judges' answers do not follow it.

Does Portfolio Optimization Beat Splitting Evenly?

Accepted with revisions · Oxford Journal of Student Scholarship · revision submitted

Splitting evenly beat the Markowitz portfolio in 74% of 250 calibrated simulations (Sharpe 0.70 against 0.51).

Modern portfolio theory says to hold assets in the proportions that maximize reward-to-risk — the Markowitz portfolio. We compare it to simply holding equal amounts of each asset (1/N) on a factor model calibrated to real market data. Across 250 simulations the 1/N portfolio earns a higher Sharpe (0.70 vs 0.51) and wins 74% of the time, a gap worth about 3 points of return a year at the same risk. Given the true covariance but estimated means, the optimizer still loses; given true means, it is nearly optimal. Estimating expected returns is the whole problem, and shrinkage only fixes it with about two decades of stationary data.

How Robust Are the Legatum Prosperity Index Rankings? A Monte Carlo and Sensitivity Analysis of Weighting, Normalization, and Aggregation

The ranking holds as a whole (median Spearman ≥ 0.99) but not mid-table, where the median 90% rank interval spans 36 places.

The Legatum Prosperity Index ranks 167 countries by an equal-weighted mean of twelve pillar scores. We reproduce every published score and rank exactly, then vary the weights, normalization and aggregation: 10,000 draws under each of three weight priors and a 131,072-run Sobol' design. The ranking holds as a whole (median Spearman ≥ 0.9899) but not in mid-table: the median 90% rank interval is 22 places overall and 36 for ranks 68–100. The paper's new result is rank leverage, c²(1 − ρ²): a weight can reorder countries only through the share of its pillar's variance that the composite doesn't share, times the pillar's squared relative spread. With arithmetic aggregation it tracks the weights' total effects under all four normalizations (Spearman 0.91–0.99), and equal weights of 1/12 give leverages that differ sixfold.

Advanced Mathematics Research

Ongoing study of induction, number theory, combinatorics, and game theory with a PhD mentor, applied to algorithmic problem-solving and computational mathematics. My working session notes are below.