When evaluating Large Language Models (LLMs) in question answering domains, it is common to ask the model to choose among a fixed set of choices (so-called multiple-choice question-answering, or MCQA). Although downstream tasks of interest typically do not provide systems with explicit options among which to choose, this approach is nevertheless widely used because it makes automatic grading straightforward and has tended to produce challenging benchmarks that correlate sufficiently well with downstream performance. This paper investigates the extent to which this trend continues to hold for state-of-the-art reasoning models, describing a systematic evaluation of 15 different question-answering benchmarks (e.g., MMLU, GSM8K, MA TH, STEER-ME) and 27 different LLMs (including small models such as Qwen-2.5 7B Instruct, mid-sized models such as Llama-3.3 70B Instruct, and large state-of-the-art models such as OpenAI's o3). For each model-benchmark pair, we considered 5 ways of presenting the model with questions, including variations on whether multiple choices were offered to the model at all; whether "none of the above" sometimes replaced the right answer; and whether the model was permitted to perform chain-of-thought reasoning before and/or after the choices were presented. MCQA remained a good proxy for the downstream performance of models as long as they were allowed to perform chain-of-thought reasoning only before being presented with the options among which they had to select. On the other hand, large models that were able to perform reasoning after being given a set of options tended to significantly outperform their free-text performance due to exploiting the information in the options. We identify and quantify the signals models are using when answering MCQA questions, and offer practical guidelines when analyzing results from MCQA that better reflect LLMs' genuine reasoning capabilities.

A critical component in the trustworthiness of LLMs is reliable uncertainty communication, yet LLMs often use assertive language when conveying false claims, leading to over-reliance and eroded trust. We present the first systematic study of $\textit{faithful confidence calibration}$ of LLMs, benchmarking models' ability to use linguistic expressions of uncertainty that $\textit{faithfully reflect}$ their intrinsic uncertainty, across a comprehensive array of models, datasets, and prompting strategies. Our results demonstrate that LLMs largely fail at this task, and that existing interventions are insufficient: standard prompt approaches provide only marginal gains, and existing, factuality-based calibration techniques can even harm faithful calibration. To address this critical gap, we introduce MetaFaith, a novel prompt-based calibration approach inspired by human metacognition. We show that MetaFaith robustly improves faithful calibration across diverse models and task domains, enabling up to 61% improvement in faithfulness and achieving an 83% win rate over original generations as judged by humans.

The internet serves as a common source of training data for generative AI (genAI) models but is increasingly populated with AI-generated content. This duality raises the possibility that future genAI models may be trained on other models' generated outputs. Prior work has studied consequences of models training on their own generated outputs, but limited work has considered what happens if models ingest content produced by other models. Given society's increasing dependence on genAI tools, understanding such data-mediated model interactions is critical. This work provides empirical evidence for how data-mediated interactions might unfold in practice, develops a theoretical model for this interactive training process, and experimentally validates the theory. We find that data-mediated interactions can benefit models by exposing them to novel concepts perhaps missed in original training data, but also can homogenize their performance on shared tasks.

This enables learning on new tasks using far less data than is required to learn them in isolation.

The Information Bottleneck (IB) objective uses information theory to formulate a task-performance versus robustness trade-off. It has been successfully applied in the standard discriminative classification setting. We pose the question whether the IB can also be used to train generative likelihood models such as normalizing flows. Since normalizing flows use invertible network architectures (INNs), they are information-preserving by construction. This seems contradictory to the idea of a bottleneck.

We study the relationship between the notions of differentially private learning and online learning in games. Several recent works have shown that differentially private learning implies online learning, but an open problem of Neel, Roth, and Wu [27] asks whether this implication is efficient. Specifically, does an efficient differentially private learner imply an efficient online learner? In this paper we resolve this open question in the context of pure differential privacy. We derive an efficient black-box reduction from differentially private learning to online learning from expert advice.