Sohee Yang

Joel Jang, Dongkeun Yoon, Sohee Yang et al. · deepmind, uw

Pretrained Language Models (LMs) memorize a vast amount of knowledge during initial pretraining, including information that may violate the privacy of personal lives and identities. Previous work addressing privacy issues for language models has mostly focused on data preprocessing and differential privacy methods, both requiring re-training the underlying LM. We propose knowledge unlearning as an alternative method to reduce privacy risks for LMs post hoc. We show that simply performing gradient ascent on target token sequences is effective at forgetting them with little to no degradation of general language modeling performances for larger LMs; it sometimes even substantially improves the underlying LM with just a few iterations. We also find that sequential unlearning is better than trying to unlearn all the data at once and that unlearning is highly dependent on which kind of data (domain) is forgotten. By showing comparisons with a previous data preprocessing method and a decoding method known to mitigate privacy risks for LMs, we show that unlearning can give a stronger empirical privacy guarantee in scenarios where the data vulnerable to extraction attacks are known a priori while being much more efficient and robust. We release the code and dataset needed to replicate our results at https://github.com/joeljang/knowledge-unlearning.

Joel Jang, Seonghyeon Ye, Changho Lee et al. · deepmind, uw

Language Models (LMs) become outdated as the world changes; they often fail to perform tasks requiring recent factual information which was absent or different during training, a phenomenon called temporal misalignment. This is especially a challenging problem because the research community still lacks a coherent dataset for assessing the adaptability of LMs to frequently-updated knowledge corpus such as Wikipedia. To this end, we introduce TemporalWiki, a lifelong benchmark for ever-evolving LMs that utilizes the difference between consecutive snapshots of English Wikipedia and English Wikidata for training and evaluation, respectively. The benchmark hence allows researchers to periodically track an LM's ability to retain previous knowledge and acquire updated/new knowledge at each point in time. We also find that training an LM on the diff data through continual learning methods achieves similar or better perplexity than on the entire snapshot in our benchmark with 12 times less computational cost, which verifies that factual knowledge in LMs can be safely updated with minimal training data via continual learning. The dataset and the code are available at https://github.com/joeljang/temporalwiki.

Seonghyeon Ye, Hyeonbin Hwang, Sohee Yang et al. · deepmind

In this paper, we present our finding that prepending a Task-Agnostic Prefix Prompt (TAPP) to the input improves the instruction-following ability of various Large Language Models (LLMs) during inference. TAPP is different from canonical prompts for LLMs in that it is a fixed prompt prepended to the beginning of every input regardless of the target task for zero-shot generalization. We observe that both base LLMs (i.e. not fine-tuned to follow instructions) and instruction-tuned models benefit from TAPP, resulting in 34.58% and 12.26% improvement on average, respectively. This implies that the instruction-following ability of LLMs can be improved during inference time with a fixed prompt constructed with simple heuristics. We hypothesize that TAPP assists language models to better estimate the output distribution by focusing more on the instruction of the target task during inference. In other words, such ability does not seem to be sufficiently activated in not only base LLMs but also many instruction-fine-tuned LLMs. All experiments are reproducible from https://github.com/seonghyeonye/TAPP.

Hyunji Lee, Jaeyoung Kim, Hoyeon Chang et al. · deepmind

The generative retrieval model depends solely on the information encoded in its model parameters without external memory, its information capacity is limited and fixed. To overcome the limitation, we propose Nonparametric Decoding (Np Decoding) which can be applied to existing generative retrieval models. Np Decoding uses nonparametric contextualized vocab embeddings (external memory) rather than vanilla vocab embeddings as decoder vocab embeddings. By leveraging the contextualized vocab embeddings, the generative retrieval model is able to utilize both the parametric and nonparametric space. Evaluation over 9 datasets (8 single-hop and 1 multi-hop) in the document retrieval task shows that applying Np Decoding to generative retrieval models significantly improves the performance. We also show that Np Decoding is data- and parameter-efficient, and shows high performance in the zero-shot setting.

Sohee Yang, Elena Gribovskaya, Nora Kassner et al. · deepmind

We study whether Large Language Models (LLMs) latently perform multi-hop reasoning with complex prompts such as "The mother of the singer of 'Superstition' is". We look for evidence of a latent reasoning pathway where an LLM (1) latently identifies "the singer of 'Superstition'" as Stevie Wonder, the bridge entity, and (2) uses its knowledge of Stevie Wonder's mother to complete the prompt. We analyze these two hops individually and consider their co-occurrence as indicative of latent multi-hop reasoning. For the first hop, we test if changing the prompt to indirectly mention the bridge entity instead of any other entity increases the LLM's internal recall of the bridge entity. For the second hop, we test if increasing this recall causes the LLM to better utilize what it knows about the bridge entity. We find strong evidence of latent multi-hop reasoning for the prompts of certain relation types, with the reasoning pathway used in more than 80% of the prompts. However, the utilization is highly contextual, varying across different types of prompts. Also, on average, the evidence for the second hop and the full multi-hop traversal is rather moderate and only substantial for the first hop. Moreover, we find a clear scaling trend with increasing model size for the first hop of reasoning but not for the second hop. Our experimental findings suggest potential challenges and opportunities for future development and applications of LLMs.

Joel Jang, Seonghyeon Ye, Sohee Yang et al.

Large Language Models (LMs) are known to encode world knowledge in their parameters as they pretrain on a vast amount of web corpus, which is often utilized for performing knowledge-dependent downstream tasks such as question answering, fact-checking, and open dialogue. In real-world scenarios, the world knowledge stored in the LMs can quickly become outdated as the world changes, but it is non-trivial to avoid catastrophic forgetting and reliably acquire new knowledge while preserving invariant knowledge. To push the community towards better maintenance of ever-changing LMs, we formulate a new continual learning (CL) problem called Continual Knowledge Learning (CKL). We construct a new benchmark and metric to quantify the retention of time-invariant world knowledge, the update of outdated knowledge, and the acquisition of new knowledge. We adopt applicable recent methods from literature to create several strong baselines. Through extensive experiments, we find that CKL exhibits unique challenges that are not addressed in previous CL setups, where parameter expansion is necessary to reliably retain and learn knowledge simultaneously. By highlighting the critical causes of knowledge forgetting, we show that CKL is a challenging and important problem that helps us better understand and train ever-changing LMs. The benchmark datasets, evaluation script, and baseline code to reproduce our results are available at https://github.com/joeljang/continual-knowledge-learning.

Jung-Woo Ha, Kihyun Nam, Jingu Kang et al.

Automatic speech recognition (ASR) via call is essential for various applications, including AI for contact center (AICC) services. Despite the advancement of ASR, however, most publicly available call-based speech corpora such as Switchboard are old-fashioned. Also, most existing call corpora are in English and mainly focus on open domain dialog or general scenarios such as audiobooks. Here we introduce a new large-scale Korean call-based speech corpus under a goal-oriented dialog scenario from more than 11,000 people, i.e., ClovaCall corpus. ClovaCall includes approximately 60,000 pairs of a short sentence and its corresponding spoken utterance in a restaurant reservation domain. We validate the effectiveness of our dataset with intensive experiments using two standard ASR models. Furthermore, we release our ClovaCall dataset and baseline source codes to be available via https://github.com/ClovaAI/ClovaCall.

Dongkeun Yoon, Seungone Kim, Sohee Yang et al. · cmu, deepmind

Despite their strengths, large language models (LLMs) often fail to communicate their confidence accurately, making it difficult to assess when they might be wrong and limiting their reliability. In this work, we demonstrate that reasoning models that engage in extended chain-of-thought (CoT) reasoning exhibit superior performance not only in problem-solving but also in accurately expressing their confidence. Specifically, we benchmark six reasoning models across six datasets and find that they achieve strictly better confidence calibration than their non-reasoning counterparts in 33 out of the 36 settings. Our detailed analysis reveals that these gains in calibration stem from the slow thinking behaviors of reasoning models (e.g., exploring alternative approaches and backtracking) which enable them to adjust their confidence dynamically throughout their CoT, making it progressively more accurate. In particular, we find that reasoning models become increasingly better calibrated as their CoT unfolds, a trend not observed in non-reasoning models. Moreover, removing slow thinking behaviors from the CoT leads to a significant drop in calibration. Lastly, we show that non-reasoning models also demonstrate enhanced calibration when simply guided to slow think via in-context learning, fully isolating slow thinking as the source of the calibration gains.

Sohee Yang, Nora Kassner, Elena Gribovskaya et al. · deepmind

We evaluate how well Large Language Models (LLMs) latently recall and compose facts to answer multi-hop queries like "In the year Scarlett Johansson was born, the Summer Olympics were hosted in the country of". One major challenge in such evaluation is that LLMs may have developed shortcuts by encountering the head entity "Scarlett Johansson" and the answer entity "United States" in the same training sequences or merely guess the answer based on frequency-based priors. To prevent shortcuts, we exclude test queries where the head and answer entities might have co-appeared during training. Through careful selection of relations and facts and systematic removal of cases where models might guess answers or exploit partial matches, we construct an evaluation dataset SOCRATES (ShOrtCut-fRee lATent rEaSoning). We observe that LLMs demonstrate promising latent multi-hop reasoning abilities without exploiting shortcuts, but only for certain types of queries. For queries requiring latent recall of countries as the intermediate answer, the best models achieve 80% latent composability, but this drops to just 5% for the recall of years. Comparisons with Chain-of-Thought highlight a significant gap between the ability of models to reason latently versus explicitly. Analysis reveals that latent representations of the intermediate answer are constructed more often in queries with higher latent composability, and shows the emergence of latent multi-hop reasoning during pretraining.

Sohee Yang, Sang-Woo Lee, Nora Kassner et al. · deepmind

Recent reasoning models show the ability to reflect, backtrack, and self-validate their reasoning, which is crucial in spotting mistakes and arriving at accurate solutions. A natural question that arises is how effectively models can perform such self-reevaluation. We tackle this question by investigating how well reasoning models identify and recover from four types of unhelpful thoughts: uninformative rambling thoughts, thoughts irrelevant to the question, thoughts misdirecting the question as a slightly different question, and thoughts that lead to incorrect answers. We show that models are effective at identifying most unhelpful thoughts but struggle to recover from the same thoughts when these are injected into their thinking process, causing significant performance drops. Models tend to naively continue the line of reasoning of the injected irrelevant thoughts, which showcases that their self-reevaluation abilities are far from a general "meta-cognitive" awareness. Moreover, we observe non/inverse-scaling trends, where larger models struggle more than smaller ones to recover from short irrelevant thoughts, even when instructed to reevaluate their reasoning. We demonstrate the implications of these findings with a jailbreak experiment using irrelevant thought injection, showing that the smallest models are the least distracted by harmful-response-triggering thoughts. Overall, our findings call for improvement in self-reevaluation of reasoning models to develop better reasoning and safer systems.

Joshua Ong Jun Leang, Zheng Zhao, Aryo Pradipta Gema et al.

Best-of-n sampling improves the accuracy of large language models (LLMs) and large reasoning models (LRMs) by generating multiple candidate solutions and selecting the one with the highest reward. The key challenge for reasoning tasks is designing a scoring function that can identify correct reasoning chains without access to ground-truth answers. We propose Probabilistic Confidence Selection And Ranking (PiCSAR): a simple, training-free method that scores each candidate generation using the joint log-likelihood of the reasoning and final answer. The joint log-likelihood of the reasoning and final answer naturally decomposes into reasoning confidence and answer confidence. PiCSAR achieves substantial gains across diverse benchmarks (+10.18 on MATH500, +9.81 on AIME2025), outperforming baselines with at least 2x fewer samples in 16 out of 20 comparisons. Our analysis reveals that correct reasoning chains exhibit significantly higher reasoning and answer confidence, justifying the effectiveness of PiCSAR.

Sang-Woo Lee, Sohee Yang, Donghyun Kwak et al.

Many recent papers have studied the development of superforecaster-level event forecasting LLMs. While methodological problems with early studies cast doubt on the use of LLMs for event forecasting, recent studies with improved evaluation methods have shown that state-of-the-art LLMs are gradually reaching superforecaster-level performance, and reinforcement learning has also been reported to improve future forecasting. Additionally, the unprecedented success of recent reasoning models and Deep Research-style models suggests that technology capable of greatly improving forecasting performance has been developed. Therefore, based on these positive recent trends, we argue that the time is ripe for research on large-scale training of superforecaster-level event forecasting LLMs. We discuss two key research directions: training methods and data acquisition. For training, we first introduce three difficulties of LLM-based event forecasting training: noisiness-sparsity, knowledge cut-off, and simple reward structure problems. Then, we present related ideas to mitigate these problems: hypothetical event Bayesian networks, utilizing poorly-recalled and counterfactual events, and auxiliary reward signals. For data, we propose aggressive use of market, public, and crawling datasets to enable large-scale training and evaluation. Finally, we explain how these technical advances could enable AI to provide predictive intelligence to society in broader areas. This position paper presents promising specific paths and considerations for getting closer to superforecaster-level AI technology, aiming to call for researchers' interest in these directions.

Hyeonbin Hwang, Byeongguk Jeon, Seungone Kim et al. · cmu, deepmind

Autoregressive language models (LMs) generate one token at a time, yet human reasoning operates over higher-level abstractions - sentences, propositions, and concepts. This contrast raises a central question- Can LMs likewise learn to reason over structured semantic units rather than raw token sequences? In this work, we investigate whether pretrained LMs can be lifted into such abstract reasoning spaces by building on their learned representations. We present a framework that adapts a pretrained token-level LM to operate in sentence space by autoregressively predicting continuous embeddings of next sentences. We explore two embedding paradigms inspired by classical representation learning: 1) semantic embeddings, learned via autoencoding to preserve surface meaning; and 2) contextual embeddings, trained via next-sentence prediction to encode anticipatory structure. We evaluate both under two inference regimes: Discretized, which decodes each predicted embedding into text before re-encoding; and Continuous, which reasons entirely in embedding space for improved efficiency. Across four domains - mathematics, logic, commonsense, and planning - contextual embeddings under continuous inference show competitive performance with Chain-of-Thought (CoT) while reducing inference-time FLOPs on average by half. We also present early signs of scalability and modular adaptation. Finally, to visualize latent trajectories, we introduce SentenceLens, a diagnostic tool that decodes intermediate model states into interpretable sentences. Together, our results indicate that pretrained LMs can effectively transition to abstract, structured reasoning within latent embedding spaces.

Eden Biran, Daniela Gottesman, Sohee Yang et al.

Large language models (LLMs) can solve complex multi-step problems, but little is known about how these computations are implemented internally. Motivated by this, we study how LLMs answer multi-hop queries such as "The spouse of the performer of Imagine is". These queries require two information extraction steps: a latent one for resolving the first hop ("the performer of Imagine") into the bridge entity (John Lennon), and another for resolving the second hop ("the spouse of John Lennon") into the target entity (Yoko Ono). Understanding how the latent step is computed internally is key to understanding the overall computation. By carefully analyzing the internal computations of transformer-based LLMs, we discover that the bridge entity is resolved in the early layers of the model. Then, only after this resolution, the two-hop query is solved in the later layers. Because the second hop commences in later layers, there could be cases where these layers no longer encode the necessary knowledge for correctly predicting the answer. Motivated by this, we propose a novel "back-patching" analysis method whereby a hidden representation from a later layer is patched back to an earlier layer. We find that in up to 66% of previously incorrect cases there exists a back-patch that results in the correct generation of the answer, showing that the later layers indeed sometimes lack the needed functionality. Overall, our methods and findings open further opportunities for understanding and improving latent reasoning in transformer-based LLMs.

Hoyeon Chang, Jinho Park, Seonghyeon Ye et al.

Despite the recent observation that large language models (LLMs) can store substantial factual knowledge, there is a limited understanding of the mechanisms of how they acquire factual knowledge through pretraining. This work addresses this gap by studying how LLMs acquire factual knowledge during pretraining. The findings reveal several important insights into the dynamics of factual knowledge acquisition during pretraining. First, counterintuitively, we observe that pretraining on more data shows no significant improvement in the model's capability to acquire and maintain factual knowledge. Next, there is a power-law relationship between training steps and forgetting of memorization and generalization of factual knowledge, and LLMs trained with duplicated training data exhibit faster forgetting. Third, training LLMs with larger batch sizes can enhance the models' robustness to forgetting. Overall, our observations suggest that factual knowledge acquisition in LLM pretraining occurs by progressively increasing the probability of factual knowledge presented in the pretraining data at each step. However, this increase is diluted by subsequent forgetting. Based on this interpretation, we demonstrate that we can provide plausible explanations for recently observed behaviors of LLMs, such as the poor performance of LLMs on long-tail knowledge and the benefits of deduplicating the pretraining corpus.

Chaeeun Kim, Soyoung Yoon, Hyunji Lee et al.

Benchmarking the performance of information retrieval (IR) is mostly conducted with a fixed set of documents (static corpora). However, in realistic scenarios, this is rarely the case and the documents to be retrieved are constantly updated and added. In this paper, we focus on Generative Retrievals (GR), which apply autoregressive language models to IR problems, and explore their adaptability and robustness in dynamic scenarios. We also conduct an extensive evaluation of computational and memory efficiency, crucial factors for real-world deployment of IR systems handling vast and ever-changing document collections. Our results on the StreamingQA benchmark demonstrate that GR is more adaptable to evolving knowledge (4-11%), robust in learning knowledge with temporal information, and efficient in terms of inference FLOPs (x2), indexing time (x6), and storage footprint (x4) compared to Dual Encoders (DE), which are commonly used in retrieval systems. Our paper highlights the potential of GR for future use in practical IR systems within dynamic environments.

Sohee Yang, Jonghyeon Kim, Joel Jang et al.

Previous works in prompt engineering for large language models have introduced different gradient-free probability-based prompt selection methods that aim to choose the optimal prompt among the candidates for a given task but have failed to provide a comprehensive and fair comparison between each other. In this paper, we propose a unified framework to interpret and evaluate the existing probability-based prompt selection methods by performing extensive experiments on 13 common and diverse NLP tasks. We find that each of the existing methods can be interpreted as some variant of the method that maximizes mutual information between the input and the predicted output (MI). Utilizing this finding, we develop several other combinatorial variants of MI and increase the effectiveness of the oracle prompt selection method from 87.79% to 94.98%, measured as the ratio of the performance of the selected prompt to that of the optimal oracle prompt. Furthermore, considering that all the methods rely on the output probability distribution of the model that might be biased, we propose a novel calibration method called Calibration by Marginalization (CBM) that is orthogonal to the existing methods and helps increase the prompt selection effectiveness of the best method to 96.85%, achieving 99.44% of the oracle prompt F1 without calibration.

Sohee Yang, Minjoon Seo

In open-domain question answering (QA), retrieve-and-read mechanism has the inherent benefit of interpretability and the easiness of adding, removing, or editing knowledge compared to the parametric approaches of closed-book QA models. However, it is also known to suffer from its large storage footprint due to its document corpus and index. Here, we discuss several orthogonal strategies to drastically reduce the footprint of a retrieve-and-read open-domain QA system by up to 160x. Our results indicate that retrieve-and-read can be a viable option even in a highly constrained serving environment such as edge devices, as we show that it can achieve better accuracy than a purely parametric model with comparable docker-level system size.

Sewon Min, Jordan Boyd-Graber, Chris Alberti et al.

We review the EfficientQA competition from NeurIPS 2020. The competition focused on open-domain question answering (QA), where systems take natural language questions as input and return natural language answers. The aim of the competition was to build systems that can predict correct answers while also satisfying strict on-disk memory budgets. These memory budgets were designed to encourage contestants to explore the trade-off between storing retrieval corpora or the parameters of learned models. In this report, we describe the motivation and organization of the competition, review the best submissions, and analyze system predictions to inform a discussion of evaluation for open-domain QA.

Sohee Yang, Minjoon Seo

The state of the art in open-domain question answering (QA) relies on an efficient retriever that drastically reduces the search space for the expensive reader. A rather overlooked question in the community is the relationship between the retriever and the reader, and in particular, if the whole purpose of the retriever is just a fast approximation for the reader. Our empirical evidence indicates that the answer is no, and that the reader and the retriever are complementary to each other even in terms of accuracy only. We make a careful conjecture that the architectural constraint of the retriever, which has been originally intended for enabling approximate search, seems to also make the model more robust in large-scale search. We then propose to distill the reader into the retriever so that the retriever absorbs the strength of the reader while keeping its own benefit. Experimental results show that our method can enhance the document recall rate as well as the end-to-end QA accuracy of off-the-shelf retrievers in open-domain QA tasks.

Wonseok Hwang, Jinyeong Yim, Seunghyun Park et al.

Information Extraction (IE) for semi-structured document images is often approached as a sequence tagging problem by classifying each recognized input token into one of the IOB (Inside, Outside, and Beginning) categories. However, such problem setup has two inherent limitations that (1) it cannot easily handle complex spatial relationships and (2) it is not suitable for highly structured information, which are nevertheless frequently observed in real-world document images. To tackle these issues, we first formulate the IE task as spatial dependency parsing problem that focuses on the relationship among text tokens in the documents. Under this setup, we then propose SPADE (SPAtial DEpendency parser) that models highly complex spatial relationships and an arbitrary number of information layers in the documents in an end-to-end manner. We evaluate it on various kinds of documents such as receipts, name cards, forms, and invoices, and show that it achieves a similar or better performance compared to strong baselines including BERT-based IOB taggger.

Sungdong Kim, Sohee Yang, Gyuwan Kim et al.

Recent works in dialogue state tracking (DST) focus on an open vocabulary-based setting to resolve scalability and generalization issues of the predefined ontology-based approaches. However, they are inefficient in that they predict the dialogue state at every turn from scratch. Here, we consider dialogue state as an explicit fixed-sized memory and propose a selectively overwriting mechanism for more efficient DST. This mechanism consists of two steps: (1) predicting state operation on each of the memory slots, and (2) overwriting the memory with new values, of which only a few are generated according to the predicted state operations. Our method decomposes DST into two sub-tasks and guides the decoder to focus only on one of the tasks, thus reducing the burden of the decoder. This enhances the effectiveness of training and DST performance. Our SOM-DST (Selectively Overwriting Memory for Dialogue State Tracking) model achieves state-of-the-art joint goal accuracy with 51.72% in MultiWOZ 2.0 and 53.01% in MultiWOZ 2.1 in an open vocabulary-based DST setting. In addition, we analyze the accuracy gaps between the current and the ground truth-given situations and suggest that it is a promising direction to improve state operation prediction to boost the DST performance.

Sang-Woo Lee, Tong Gao, Sohee Yang et al.

Answerer in Questioner's Mind (AQM) is an information-theoretic framework that has been recently proposed for task-oriented dialog systems. AQM benefits from asking a question that would maximize the information gain when it is asked. However, due to its intrinsic nature of explicitly calculating the information gain, AQM has a limitation when the solution space is very large. To address this, we propose AQM+ that can deal with a large-scale problem and ask a question that is more coherent to the current context of the dialog. We evaluate our method on GuessWhich, a challenging task-oriented visual dialog problem, where the number of candidate classes is near 10K. Our experimental results and ablation studies show that AQM+ outperforms the state-of-the-art models by a remarkable margin with a reasonable approximation. In particular, the proposed AQM+ reduces more than 60% of error as the dialog proceeds, while the comparative algorithms diminish the error by less than 6%. Based on our results, we argue that AQM+ is a general task-oriented dialog algorithm that can be applied for non-yes-or-no responses.