We investigate the mechanistic sources of uncertainty in large language models (LLMs), an area with important implications for language model reliability and trustworthiness. To do so, we conduct a series of experiments designed to identify whether the factuality of generated responses and a model’s uncertainty originate in separate or shared circuits in the model architecture. We approach this question by adapting the well-established mechanistic interpretability techniques of causal tracing and zero-ablation to study the effect of different circuits on LLM generations. Our experiments on eight different models and five datasets, representing tasks predominantly requiring factual recall, provide strong evidence that a model’s uncertainty is produced in the same parts of the network that are responsible for the factuality of generated responses.

The automatic evaluation of instruction following typically involves using large language models (LLMs) to assess response quality. However, there is a lack of comprehensive evaluation of these LLM-based evaluators across two dimensions: the base LLMs and the evaluation protocols. Therefore, we present a thorough meta-evaluation of instruction following, including 25 base LLMs and 15 recently proposed evaluation protocols, on 4 human-annotated datasets, assessing the evaluation accuracy of the LLM-evaluators. Our evaluation allows us to identify the best-performing base LLMs and evaluation protocols with a high degree of robustness. Moreover, our evaluation reveals key findings: (1) Base LLM performance ranking remains largely consistent across evaluation protocols, with less capable LLMs showing greater improvement from protocol enhancements; (2) Robust evaluation of evaluation protocols requires many base LLMs with varying capability levels, as protocol effectiveness depends on the base LLM used; (3) Evaluation results on different datasets are not always consistent, so a rigorous evaluation requires multiple datasets with distinctive features. We release our meta-evaluation suite ReIFE, which provides the codebase and evaluation result collection for over 500 LLM-evaluators, laying groundwork for future research in instruction-following evaluation.

Direct Preference Optimization (DPO) has become a widely used training method for the instruction fine-tuning of large language models (LLMs). In this work, we explore an under-investigated aspect of DPO – its dependency on the reference model or policy. Such reference policies, typically instantiated as the model to be further fine-tuned, are important since they can impose an upper limit on DPO’s effectiveness. Therefore, we address three related research questions in this work. First, we explore the optimal strength of the KL divergence constraint in DPO, which penalizes deviations from the reference policy, and find that DPO is sensitive to this strength. Next, we examine the necessity of the KL-constraint from the reference policies in DPO by providing both theoretical and empirical comparisons between DPO and related learning objectives, demonstrating DPO’s superiority in this controlled setting. Additionally, we investigate whether DPO benefits from stronger reference policies, finding that a stronger reference policy can lead to improved performance, but only when it is similar to the model being fine-tuned. Our findings highlight the confounding role of reference policies in DPO and offer insights for best practices, while also identifying open research questions for future studies.

Evaluating and ranking the capabilities of different LLMs is crucial for understanding their performance and alignment with human preferences. Due to the high cost and time-consuming nature of human evaluations, an automatic LLM bencher (i.e., an automatic evaluation framework that aims to rank LLMs based on their alignment with human preferences) is indispensable. An automatic LLM bencher consists of four components: the input set (e.g., a user instruction), the evaluation model (e.g., an LLM), the evaluation type (e.g., pairwise comparison), and the aggregation method (e.g., the ELO rating system). However, previous work has not thoroughly explored how to select these components or how their different combinations influence the results. In this work, through controlled experiments, we provide a series of recommendations on how to choose each component to better automate the evaluation of LLMs. Furthermore, we discovered that when evaluating LLMs with similar performance, the performance of the automatic LLM bencher declines sharply, underscoring the limitations of current benchers and calling for future work. Lastly, we found that the evaluation models’ performance at the instance level (e.g., the accuracy of selecting the best output) does not always align with their effectiveness when used as a component of a bencher, highlighting the importance of dedicated system-level evaluation of benchers.

We introduce Physics, a comprehensive benchmark for university-level physics problem solving. It contains 1,297 expert-annotated problems covering six core areas: classical mechanics, quantum mechanics, thermodynamics and statistical mechanics, electromagnetism, atomic physics, and optics.Each problem requires advanced physics knowledge and mathematical reasoning.We develop a robust automated evaluation system for precise and reliable validation. Our evaluation of leading foundation models reveals substantial limitations. Even the most advanced model, o3-mini, achieves only 59.9% accuracy, highlighting significant challenges in solving high-level scientific problems.Through comprehensive error analysis, exploration of diverse prompting strategies, and Retrieval-Augmented Generation (RAG)-based knowledge augmentation, we identify key areas for improvement, laying the foundation for future advancements.

LLMs are increasingly applied in the legal domain in tasks such as summarizing legal texts and providing basic legal advice. Yet, their capacity to draft full judicial analyses in U.S. court opinions is still largely uncharted, such as generating entire judicial reasoning sections in U.S. court decisions, remain under-explored. Given the continued adoption of LLMs and the significance of law to society at large, measurement of LLM’s legal reasoning capabilities is a pressing task. We propose CourtReasoner, a novel expert-annotated judicial reasoning benchmark for evaluating LLM agents’ capabilities in complex legal reasoning. Sourcing U.S. court opinions, we construct benchmarks that measure the LLMs ability to construct goal-oriented legal reasoning. CourtReasoner measured the agent’s ability to argue both ways in a legal dispute, rather than simple Q/A. Our results show that more than 60% of frontier model outputs contain invalid arguments and more than 53% of frontier model produced irrelevant citations when conducting complex legal reasoning. We also introduce a meta-evaluation benchmark to provide insights into the capabilities of LLMs as evaluators of legal reasoning. We will release our data, code and full annotation guidelines publicly for future research.

We introduce AbGen, the first benchmark designed to evaluate the capabilities of LLMs in designing ablation studies for scientific research. AbGen consists of 2,000 expert-annotated examples derived from 677 NLP papers. In this benchmark, LLMs are tasked with generating detailed ablation study designs for a specified module or process based on the given research context. Our evaluation of leading LLMs, such as GPT-4o and Llama-3.1, highlights a significant performance gap between these models and human experts in terms of the importance, faithfulness, and soundness of the ablation study designs. Moreover, we demonstrate that current automated evaluation methods are not reliable for our task, as they show a significant discrepancy when compared to human assessment. To better investigate this, we develop AbGen-Eval, a meta-evaluation benchmark designed to assess the reliability of commonly used automated evaluation systems in measuring LLM performance on our task. We investigate various LLM-based evaluation methods on AbGen-Eval, providing insights for future research on developing more effective and reliable LLM-based evaluation systems for complex scientific tasks.

Existing evaluation benchmarks for foundation models in understanding scientific literature predominantly focus on single-document, text-only tasks. Such benchmarks often do not adequately represent the complexity of research workflows, which typically also involve interpreting non-textual data, such as figures and tables, and gathering information across multiple documents and related literature. To address this gap, we introduce M3SciQA, a multi-modal, multi-document scientific question answering benchmark designed for a more comprehensive evaluation of foundation models. M3Sci QA consists of 1452 expert-annotated questions spanning 70 natural language processing paper clusters, where each cluster represents a primary paper along with all its cited documents, mirroring the workflow of comprehending a single paper by requiring multi-modal and multi-document data. With M3SciQA, we conduct a comprehensive evaluation of 18 frontier foundation models. Our results indicate that current foundation models still significantly underperform compared to human experts in multi-modal information retrieval and in reasoning across multiple scientific documents. Additionally, we explore the implications of these findings for the future advancement of applying foundation models in multi-modal scientific literature analysis.

To enhance Large Language Models’ (LLMs) reliability, calibration is essential—the model’s confidence scores should align with the likelihood of its responses being correct. However, traditional calibration methods typically rely on a binary true/false assessment of response correctness, unsuitable for long-form generations where an answer can be partially correct. Addressing this gap, we introduce a unified calibration framework, in which both the correctness of the LLMs’ responses and their associated confidence levels are treated as distributions across a range of scores. We develop three metrics for assessing LLM calibration and propose confidence elicitation methods based on self-consistency and self-evaluation. Our experiments demonstrate that larger models don’t necessarily guarantee better calibration, that various calibration metrics complement each other, and that self-consistency methods excel in factoid datasets. We also find that calibration can be enhanced through techniques such as fine-tuning, scaling the temperature. Finally, we illustrate one application of long-form calibration through selective answering in long-form responses, optimizing correctness within a constrained API budget.

Large language models (LLMs) have achieved remarkable performance on a variety of natural language understanding tasks. However, existing benchmarks are inadequate in measuring the complex logical reasoning capabilities of a model. We present FOLIO, a human-annotated, logically complex and diverse dataset for reasoning in natural language (NL), equipped with first-order logic (FOL) annotations. FOLIO consists of 1,430 examples (unique conclusions), each paired with one of 487 sets of premises used to deductively reason for the validity of each conclusion. The logical correctness of the premises and conclusions is ensured by their FOL annotations, which are automatically verified by an FOL inference engine. In addition to the main NL reasoning task, NL-FOL pairs in FOLIO constitute a new NL-FOL translation dataset. Our experiments on FOLIO systematically evaluate the FOL reasoning ability of supervised fine-tuning on medium-sized language models. For both NL reasoning and NL-FOL translation, we benchmark multiple state-of-the-art language models. Our results show that a subset of FOLIO remains a challenge for one of the most capable Large Language Model (LLM) publicly available, GPT-4.

Abstractive summarization models are commonly trained using maximum likelihood estimation, which assumes a deterministic (one-point) target distribution in which an ideal model will assign all the probability mass to the reference summary. This assumption may lead to performance degradation during inference, where the model needs to compare several system-generated (candidate) summaries that have deviated from the reference summary. To address this problem, we propose a novel training paradigm which assumes a non-deterministic distribution so that different candidate summaries are assigned probability mass according to their quality. Our method achieves a new state-of-the-art result on the CNN/DailyMail (47.78 ROUGE-1) and XSum (49.07 ROUGE-1) datasets. Further analysis also shows that our model can estimate probabilities of candidate summaries that are more correlated with their level of quality.