"Check My Work?": Measuring Sycophancy in a Simulated Educational Context

This study examines how user-provided suggestions affect Large Language Models (LLMs) in a simulated educational context, where sycophancy poses significant risks. Testing five different LLMs from the OpenAI GPT-4o and GPT-4.1 model classes across five experimental conditions, we show that response quality varies dramatically based on query framing. In cases where the student mentions an incorrect answer, the LLM correctness can degrade by as much as 15 percentage points, while mentioning the correct answer boosts accuracy by the same margin. Our results also show that this bias is stronger in smaller models, with an effect of up to 30% for the GPT-4.1-nano model, versus 8% for the GPT-4o model. Our analysis of how often LLMs "flip" their answer, and an investigation into token level probabilities, confirm that the models are generally changing their answers to answer choices mentioned by students in line with the sycophancy hypothesis. This sycophantic behavior has important implications for educational equity, as LLMs may accelerate learning for knowledgeable students while the same tools may reinforce misunderstanding for less knowledgeable students. Our results highlight the need to better understand the mechanism, and ways to mitigate, such bias in the educational context.

Identifying Legal Holdings with LLMs: A Systematic Study of Performance, Scale, and Memorization

As large language models (LLMs) continue to advance in capabilities, it is essential to assess how they perform on established benchmarks. In this study, we present a suite of experiments to assess the performance of modern LLMs (ranging from 3B to 90B+ parameters) on CaseHOLD, a legal benchmark dataset for identifying case holdings. Our experiments demonstrate ``scaling effects'' - performance on this task improves with model size, with more capable models like GPT4o and AmazonNovaPro achieving macro F1 scores of 0.744 and 0.720 respectively. These scores are competitive with the best published results on this dataset, and do not require any technically sophisticated model training, fine-tuning or few-shot prompting. To ensure that these strong results are not due to memorization of judicial opinions contained in the training data, we develop and utilize a novel citation anonymization test that preserves semantic meaning while ensuring case names and citations are fictitious. Models maintain strong performance under these conditions (macro F1 of 0.728), suggesting the performance is not due to rote memorization. These findings demonstrate both the promise and current limitations of LLMs for legal tasks with important implications for the development and measurement of automated legal analytics and legal benchmarks.

LLMForecaster: Improving Seasonal Event Forecasts with Unstructured Textual Data

Modern time-series forecasting models often fail to make full use of rich unstructured information about the time series themselves. This lack of proper conditioning can lead to obvious model failures; for example, models may be unaware of the details of a particular product, and hence fail to anticipate seasonal surges in customer demand in the lead up to major exogenous events like holidays for clearly relevant products. To address this shortcoming, this paper introduces a novel forecast post-processor -- which we call LLMForecaster -- that fine-tunes large language models (LLMs) to incorporate unstructured semantic and contextual information and historical data to improve the forecasts from an existing demand forecasting pipeline. In an industry-scale retail application, we demonstrate that our technique yields statistically significantly forecast improvements across several sets of products subject to holiday-driven demand surges.