arXiv Machine Learning

MemGuard-Alpha: Limits of Membership Inference for Detecting and Filtering Memorization-Contaminated Signals in LLM-Based Financial Forecasting

MemGuard-Alpha evaluates whether membership inference attacks (MIA) can detect memorization in large language models (LLMs) used for financial alpha signals. The study combines five MIA methods with a temporal proximity feature and a cross-model disagreement metric, then audits them across seven LLMs, 50 S&P 100 stocks, and 299,600 prompt-model pairs. Findings show that temporal proximity alone perfectly predicts in-sample status, MIA discriminative power largely stems from model scale differences, and filtering based on contamination scores does not improve risk-adjusted performance once transaction costs are considered.

arXiv Machine Learning
Aug 19

Temporal Leakage in Financial News NLP: A Multi-Architecture Audit with a Regime-Specific M&A Signal

The paper audits the impact of temporal leakage on financial-news direction prediction across 49,799 articles and 16 feature-model combinations, including TF‑IDF, MiniLM, FinBERT, and fine‑tuned RoBERTa‑large / DeBERTa‑v3‑large, as well as zero/few‑shot and LoRA probes of Llama‑3 and Qwen2.5. Random train‑test splits inflate MCC scores by 1.1× to 6.5×, with larger models and richer features showing greater gains, while end‑to‑end FinBERT fine‑tuning actually increases the gap. Only the mergers and acquisitions (M&A) category shows a positive locked‑test signal under near‑temporal chronological evaluation, with the signal localized to 2024‑2025 European‑tilted M&A semantics and not transferring to a 2009‑2020 U.S. corpus.

By Chenhao Xue, Raslen Guesmi, Siwei Feng, Yucheng Gong, Jacob Xavier Sundram, Jordan Pang, Lan Wang, Julian Kaljuvee
arXiv Computation and Language
Oct 12

AMLworld-Compact: Importance-Weighted Downsampling for Cost-Effective LLM Evaluation and Error Diagnosis

AMLworld-Compact introduces an importance‑weighted downsampling protocol for evaluating large language models on AMLworld’s test splits, retaining all illicit transactions while sampling benign ones from supervised‑score strata. The HT‑Coreset method reduces evaluation targets by 271× on HI‑Small and 611× on LI‑Small, providing unbiased confusion‑count estimates with exact recall and ratio estimates for precision and F1. Experiments with boosted‑tree strata and Graph Feature Preprocessor features show lower F1 RMSE compared to uniform sampling, and audit tools reveal that lexical checks do not reliably indicate correctness of model outputs.

By Natapong Nitarach, Phume Ngampornsukswadi, Pittawat Taveekitworachai, Surapon Nonesung, Warit Sirichotedumrong, Kunat Pipatanakul
arXiv Computation and Language
6d ago

IFRS Compliance Verification: An Open Multi-LLM Benchmark and Content-Conditioned Ensemble Classifier

The paper introduces a new benchmark of 19,998 IFRS regulation‑disclosure pairs and a multi‑stage framework that combines domain‑adapted sentence embeddings with a learned ensemble classifier to verify corporate financial disclosures against IFRS. The ensemble, trained on GPT‑4.1, Claude Sonnet 4, and Gemini 2.5 Pro, achieves 81.7% weighted F1 on the disagreement set—outperforming the best single LLM by 10.4 points and majority voting by 14.6 points—and the retrieval pipeline reaches 45.3% Recall@15 on unseen regulations. All data, models, and code will be publicly released.

By Aghoghomena Akasukpe, Li Yang, Xinyao Zhou, Chen Shen, Miguel Vargas Martin
arXiv Computation and Language
Sep 4

Contamination Inflates Scores but Rarely Reorders Large Language Model Leaderboards

The paper investigates how benchmark contamination—leakage of test items into training data—affects large language model (LLM) leaderboards. By comparing original test items with semantically equivalent paraphrases, the authors measure contamination as a violation of anchor-item invariance and find that it inflates absolute scores but rarely changes model rankings. Across 47 public models and 74 finetuned models on four benchmarks, the rank correlation between standard and paraphrase-controlled leaderboards is 0.997, with only a handful of cases showing differential contamination that could alter rankings.

By Xingyao Xiao (Stanford University), Yihong Cheng (City University of Macau)
arXiv Computation and Language
Oct 5

OLMo-Detect: A Multi-Stage, Confounder-Controlled Benchmark for Membership Inference on Large Language Models

OLMo-Detect is a new benchmark for membership inference attacks (MIAs) on large language models, covering pre‑training, mid‑training, and post‑training stages. It aligns member and non‑member samples on key axes and rigorously filters non‑members using infini‑gram, also offering a shifted variant to test distribution robustness. Evaluation of 15 unsupervised and 3 supervised MIAs on the OLMo 2 family shows limited overall performance (best AUC 0.68), peak detection at mid‑training driven by data type, and sensitivity to distribution shifts, with findings generalizing to OLMo 3 and other models.

By Tao Shi, Chaoyi Xiang, Qiongkai Xu, Jey Han Lau
arXiv Machine Learning
Jul 15

Extractable Memorization From First Principles

arXiv:2607. 12649v1 Announce Type: new Abstract: Recent work on extractable memorization in LLMs suffers from two contrasting validity problems.

By A. Feder Cooper, Marika Swanberg, Jamie Hayes, Lea Duesterwald, Christopher De Sa, Daniel E. Ho, Mark A. Lemley, Percy Liang
arXiv Machine Learning
Nov 5

Information-Matched Auditing of Conversational Emotion-Shift Forecasting: A Leakage-Safe Evaluation Protocol

The paper introduces an information‑matched, leakage‑safe auditing protocol for evaluating conversational emotion‑shift forecasting models. It addresses common distortions such as unequal feature richness, noisy current‑emotion labels, and differing hyperparameter budgets by using explicit commitment points, causality tests, balanced configuration searches, and seed‑dialogue‑aware inference. Applying the protocol to six causal model families across four corpora shows that a deployable train‑only current‑emotion classifier improves over a predicted‑label baseline, especially on DailyDialog, while matched‑gold conditions reveal baseline competitiveness on IEMOCAP and stronger performance on MELD.

By Bin Wen, Dai-Qiao Zhang, Tien-Ping Tan
arXiv Machine Learning
1d ago

How Much Is Left Matters: False Positives and False Negatives in Partial-Prompt Contamination Tests

The paper evaluates partial‑prompt completion tests—where a model is given the first part of a benchmark problem and checked for reproducing the rest—on seven open‑weight models using MATH‑500 and recent AIME 2025/2026 problems as true negatives. It finds that memorization flags are highly sensitive to the amount of text left: problems with fewer than eight remaining words are flagged far more often than those with 25 or more, and most ROUGE‑based flags arise from short leftovers. On true negatives, even a high ROUGE‑L threshold flags up to 6.7% of problems at a 60% prefix, and exact‑match flags 8.6% at an 80% prefix, while models distilled from Qwen2.5‑Math show almost no recall of contamination. The authors recommend length‑stratified reporting, limiting prefixes to 60%, using style‑matched post‑cutoff controls, and cross‑family calibration to reduce false positives.

By Abhishek Anand