Gal Yona
Gal Yona is a Research Scientist at Google Research, Tel Aviv, where she is working on improving factuality in large language models, with an emphasis on robustness and uncertainty. Before joining Google, Gal completed her PhD in Computer Science at the Weizmann Institute of Science, developing definitions and algorithms for preventing discrimination in machine learning models. Gal received numerous award during her PhD, including the Blavatnik Prizes for Outstanding Israeli Doctoral Students in Computer Science (2022) and the Google PhD Fellowship in Machine Learning (2021).
Research Areas
Authored Publications
Sort By
Preview abstract
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 faithful confidence calibration of LLMs, benchmarking models' ability to use linguistic expressions of uncertainty that 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.
View details
Confidence Improves Self-Consistency in LLMs
Tom Sheffer
Eran Ofek
Ariel Goldstein
Zorik Gekhman
ACL 2025, Findings (2025)
Preview abstract
Self-consistency decoding enhances LLMs’ performance on reasoning tasks by sampling diverse reasoning paths and selecting the most frequent answer. However, it is computationally expensive, as sampling many of these (lengthy) paths is required to increase the chances that the correct answer emerges as the most frequent one. To address this, we introduce Confidence-Informed Self-Consistency (CISC). CISC performs a weighted majority vote based on confidence scores obtained directly from the model. By prioritizing high-confidence paths, it can identify the correct answer with a significantly smaller sample size. When tested on nine models and four datasets, CISC outperforms self-consistency in nearly all configurations, reducing the required number of reasoning paths by over 40% on average. In addition, we introduce the notion of within-question confidence evaluation, after showing that standard evaluation methods are poor predictors of success in distinguishing correct and incorrect answers to the same question. In fact, the most calibrated confidence method proved to be the least effective for CISC. Lastly, beyond these practical implications, our results and analyses show that LLMs can effectively judge the correctness of their own outputs, contributing to the ongoing debate on this topic.
View details
Preview abstract
Recent work suggested utilizing inference compute, showing that scaling of number of samples consistently improves the fractions of problems solved by any attempt, namely the coverage. In this work, we suggest that inference scaling gains should be compared with proper baselines, as some datasets become degenerate when allowing a large number of attempts. We focus on two domains - mathematical reasoning and factual knowledge, showing that for the MATH and Entity Questions datasets, informed answer enumeration obtains similar or even better results than repeated model sampling, with a much lower sample budget. While we believe that inference scaling is a promising approach for unlocking the potential of language models, we recommend carefully selecting models and datasets when applying this method. Otherwise, the results of inference scaling should be interpreted with caution.
View details
Preview abstract
Recent work suggested utilizing inference compute, showing that scaling of number of samples consistently improves the fractions of problems solved by any attempt, namely the coverage. In this work, we suggest that inference scaling gains should be compared with proper baselines, as some datasets become degenerate when allowing a large number of attempts. We focus on two domains - mathematical reasoning and factual knowledge, showing that for the MATH and Entity Questions datasets, informed answer enumeration obtains similar or even better results than repeated model sampling, with a much lower sample budget. While we believe that inference scaling is a promising approach for unlocking the potential of language models, we recommend carefully selecting models and datasets when applying this method. Otherwise, the results of inference scaling should be interpreted with caution.
View details
Preview abstract
Factual questions typically can be answered correctly at different levels of granularity. For example, both “August 4, 1961” and “1961” are correct answers to the question “When was Barack Obama born?”. Standard question answering (QA) evaluation protocols, however, do not explicitly take this into account and compare a predicted answer against answers of a single granularity level.
In this work, we propose GRANOLA QA, a novel evaluation setting where a predicted answer is evaluated in terms of accuracy and informativeness against a set of multi-granularity answers. We present a simple methodology for enriching existing datasets with multi-granularity answers, and create GRANOLA-EQ, a multi-granularity version of the EntityQuestions dataset.
We evaluate a range of decoding methods on GRANOLA-EQ, including a new algorithm, called Decoding with Response Aggregation (DRAG), that is geared towards aligning the response granularity with the model's uncertainty. Our experiments show that large language models with standard decoding tend to generate specific answers, which are often incorrect. In contrast, when evaluated on multi-granularity answers, DRAG yields a nearly 20 point increase in accuracy on average, which further increases for rare entities.Overall, this reveals that standard evaluation and decoding schemes may significantly underestimate the knowledge encapsulated in LMs.
View details
Preview abstract
We posit that large language models (LLMs) should be capable of expressing their intrinsic uncertainty in natural language. For example, if the LLM is equally likely to output two contradicting answers to the same question, then its generated response should reflect this uncertainty by hedging its answer (e.g., "I'm not sure, but I think..."). We formalize faithful response uncertainty based on the gap between the model's intrinsic confidence in the assertions it makes and the decisiveness by which they are conveyed. This example-level metric reliably indicates whether the model reflects its uncertainty, as it penalizes both excessive and insufficient hedging. We evaluate a variety of aligned LLMs at faithfully communicating uncertainty on several knowledge-intensive question answering tasks. Our results provide strong evidence that modern LLMs are poor at faithfully conveying their uncertainty, and that better alignment is necessary to improve their trustworthiness.
View details
Useful Confidence Measures: Beyond the Max Score
NeurIPS 2022 Workshop on Distribution Shifts (DistShift) (2022) (to appear)
Preview abstract
An important component in deploying machine learning (ML) in safety-critic applications is having a reliable measure of confidence in the ML's predictions. For a classifier $f$ producing a probability vector $f(x)$ over the candidate classes, the confidence is typically taken to be $\max_i f(x)_i$. This approach is potentially limited, as it disregards the rest of the probability vector. In this work, we derive several confidence measures that depend on information beyond the maximum score, such as margin-based and entropy-based measures, and empirically evaluate their usefulness. We focus on NLP tasks and Transformer-based models. We show that in the "out of the box" regime (where the scores of $f$ are used as is), using only the maximum score to inform the confidence measure is highly suboptimal. In the post-processing regime (where the scores of $f$ can be improved using additional held-out data), this remains true (though the differences are less pronounced), with entropy-based confidence emerging as a surprisingly useful measure.
View details
Active Learning with Label Comparisons
Shay Moran
Uncertainty in Artificial Intelligence (submitted) (2022)
Preview abstract
Supervised learning typically relies on manual annotation of the true labels. However, when there are many potential labels, it will be time consuming for a human annotator to search these for the best one. On the other hand, comparing two candidate labels is often much easier. In this paper, we focus on this type of pairwise supervision, and ask how it can be used effectively in learning, and in particular active learning. We obtain several surprising results in this context. In principle, finding the best label out of $k$ can be done with $k-1$ active queries. However, we show that there is a natural class where this approach is in fact sub-optimal, and that there is a more comparison-efficient active learning scheme. A key element in our analysis is the ``label neighborhood graph'' of the true distribution, which has an edge between two classes if they share a decision boundary. We also show that in the PAC setting, pairwise comparisons cannot provide improved sample complexity in the worst case. We complement our theoretical results with experiments, clearly demonstrating the effect of the neighborhood graph on sample complexity.
View details
