Large Language Models: An Applied Econometric Framework

The unintended consequences of large language models as a labor-augmenting technology in science

As a labor-augmenting technology, large language models (LLMs) have the potential to accelerate scientific activity across the research pipeline. But even if LLMs perform on par with human experts at selected tasks, their use will bring unintended consequences as they alter the balance of frictions and inducements that steer the allocation of research effort across projects. Here we develop a simple mathematical model to illustrate. In fields where LLMs are useful primarily as tools for discovering promising projects, researchers will become more selective about what they publish; where they facilitate the process of publishing existing data, researchers will become less selective. By allowing scientists to work more quickly, LLMs raise the opportunity cost of researcher time, creating incentives to refine papers less thoroughly before moving on. Enticing as it is to imagine that, by saving us time on mundane tasks, LLMs will provide us with more time to think deeply and develop projects completely, our results temper such hopes.

The unintended consequences of large language models as a labor-augmenting technology in science

As a labor-augmenting technology, large language models (LLMs) have the potential to accelerate scientific activity across the research pipeline. But even if LLMs perform on par with human experts at selected tasks, their use will bring unintended consequences as they alter the balance of frictions and inducements that steer the allocation of research effort across projects. Here we develop a simple mathematical model to illustrate. In fields where LLMs are useful primarily as tools for discovering promising projects, researchers will become more selective about what they publish; where they facilitate the process of publishing existing data, researchers will become less selective. By allowing scientists to work more quickly, LLMs raise the opportunity cost of researcher time, creating incentives to refine papers less thoroughly before moving on. Enticing as it is to imagine that, by saving us time on mundane tasks, LLMs will provide us with more time to think deeply and develop projects completely, our results temper such hopes.

Unlearning Traces the Influential Training Data of Language Models

Identifying the training datasets that influence a language model's outputs is essential for minimizing the generation of harmful content and enhancing its performance. Ideally, we can measure the influence of each dataset by removing it from training; however, it is prohibitively expensive to retrain a model multiple times. This paper presents UnTrac: unlearning traces the influence of a training dataset on the model's performance. UnTrac is extremely simple; each training dataset is unlearned by gradient ascent, and we evaluate how much the model's predictions change after unlearning. Furthermore, we propose a more scalable approach, UnTrac-Inv, which unlearns a test dataset and evaluates the unlearned model on training datasets. UnTrac-Inv resembles UnTrac, while being efficient for massive training datasets. In the experiments, we examine if our methods can assess the influence of pretraining datasets on generating toxic, biased, and untruthful content. Our methods estimate their influence much more accurately than existing methods while requiring neither excessive memory space nor multiple checkpoints.

Communication Bias in Large Language Models: A Regulatory Perspective

Large language models (LLMs) are increasingly central to many applications, raising concerns about bias, fairness, and regulatory compliance. This paper reviews risks of biased outputs and their...

Scientific production in the era of Large Language Models

Large Language Models (LLMs) are rapidly reshaping scientific research. We analyze these changes in multiple, large-scale datasets with 2.1M preprints, 28K peer review reports, and 246M online accesses to scientific documents. We find: 1) scientists adopting LLMs to draft manuscripts demonstrate a large increase in paper production, ranging from 23.7-89.3% depending on scientific field and author background, 2) LLM use has reversed the relationship between writing complexity and paper quality, leading to an influx of manuscripts that are linguistically complex but substantively underwhelming, and 3) LLM adopters access and cite more diverse prior work, including books and younger, less-cited documents. These findings highlight a stunning shift in scientific production that will likely require a change in how journals, funding agencies, and tenure committees evaluate scientific works.

LLM in a Flash: Efficient Large Language Model Inference with Limited Memory

Large language models (LLMs) are central to modern natural language processing, delivering exceptional performance in various tasks…

Training Compute-Optimal Large Language Models

We investigate the optimal model size and number of tokens for training a transformer language model under a given compute budget. We find that current large language models are significantly undertrained, a consequence of the recent focus on scaling language models whilst keeping the amount of training data constant. By training over 400 language models ranging from 70 million to over 16 billion parameters on 5 to 500 billion tokens, we find that for compute-optimal training, the model size and the number of training tokens should be scaled equally: for every doubling of model size the number of training tokens should also be doubled. We test this hypothesis by training a predicted compute-optimal model, Chinchilla, that uses the same compute budget as Gopher but with 70B parameters and 4$\times$ more more data. Chinchilla uniformly and significantly outperforms Gopher (280B), GPT-3 (175B), Jurassic-1 (178B), and Megatron-Turing NLG (530B) on a large range of downstream evaluation tasks. This also means that Chinchilla uses substantially less compute for fine-tuning and inference, greatly facilitating downstream usage. As a highlight, Chinchilla reaches a state-of-the-art average accuracy of 67.5% on the MMLU benchmark, greater than a 7% improvement over Gopher.

Language models struggle with compartmentalization

In the training data used by large language models (LLMs), the same latent concept is often presented in multiple distinct ways: the same facts appear in English and Swahili; many functions can be expressed in both Python and Haskell; we can express propositions in both formal and natural language. We show that LLMs can exhibit compartmentalization, where they fail to identify and share statistical strength between distinct presentations of unified concepts. In the worst case, LLMs simply learn parallel internal representations of each presentation of the concept, saturating model capacity with redundancies and decreasing sample efficiency with the number of such presentations. We also demonstrate that synthetic parallel data can fail to improve this despite being easily learned itself. Under this framework, we find that, for small models, early multilingual learning is nearly entirely compartmentalized. Finally, all interventions that we study exhibit a phase transition in which their effectiveness depends on the number of distinct presentations, suggesting that the language modeling objective may only inconsistently unify representations.

The diffusion of large language models in published academic articles

Large language models (LLMs) are rapidly changing academic research, raising questions of who is adopting these tools and under what conditions. This article analyzes full texts of 7.3 million journal articles published from 2020–2025 by four major publishers (Elsevier, Frontiers, MDPI, and PLoS) to track the prevalence of LLM-associated language and identify social and institutional correlates of adoption. A corpus of 228 focal words exhibiting sharp post-2022 frequency increases consistent with LLM output was developed; articles were scored on their rate of focal word usage. By 2025, an estimated 57% of published articles exhibited evidence of LLM influence, up from 12% in 2023. Among articles exhibiting LLM-influenced text, there is substantial heterogeneity, ranging from subtle linguistic influence to articles mostly or entirely LLM-generated. Difference-in-differences models reveal that LLM-associated language varies markedly across regions, institutional ranks, publishers, disciplines, and journal tiers. Economic development and proximity to English as a primary language are key predictors of regional variation. Lower-ranked institutions exhibit higher rates than elite universities, young for-profit publishers show elevated rates vis-à-vis competitors, and academic fields differ widely in adoption. LLM adoption in academic writing is pervasive but socially stratified. As models grow more powerful and their use becomes further entrenched in academic research, understanding social dynamics of adoption will be essential for governing the evolving relationship between AI and academic knowledge production.

Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting

Large Language Models (LLMs) can achieve strong performance on many tasks by producing step-by-step reasoning before giving a final output, often referred to as chain-of-thought reasoning (CoT). It is tempting to interpret these CoT explanations as the LLM's process for solving a task. This level of transparency into LLMs' predictions would yield significant safety benefits. However, we find that CoT explanations can systematically misrepresent the true reason for a model's prediction. We demonstrate that CoT explanations can be heavily influenced by adding biasing features to model inputs--e.g., by reordering the multiple-choice options in a few-shot prompt to make the answer always "(A)"--which models systematically fail to mention in their explanations. When we bias models toward incorrect answers, they frequently generate CoT explanations rationalizing those answers. This causes accuracy to drop by as much as 36% on a suite of 13 tasks from BIG-Bench Hard, when testing with GPT-3.5 from OpenAI and Claude 1.0 from Anthropic. On a social-bias task, model explanations justify giving answers in line with stereotypes without mentioning the influence of these social biases. Our findings indicate that CoT explanations can be plausible yet misleading, which risks increasing our trust in LLMs without guaranteeing their safety. Building more transparent and explainable systems will require either improving CoT faithfulness through targeted efforts or abandoning CoT in favor of alternative methods.

Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting

Large Language Models (LLMs) can achieve strong performance on many tasks by producing step-by-step reasoning before giving a final output, often referred to as chain-of-thought reasoning (CoT). It is tempting to interpret these CoT explanations as the LLM's process for solving a task. This level of transparency into LLMs' predictions would yield significant safety benefits. However, we find that CoT explanations can systematically misrepresent the true reason for a model's prediction. We demonstrate that CoT explanations can be heavily influenced by adding biasing features to model inputs--e.g., by reordering the multiple-choice options in a few-shot prompt to make the answer always "(A)"--which models systematically fail to mention in their explanations. When we bias models toward incorrect answers, they frequently generate CoT explanations rationalizing those answers. This causes accuracy to drop by as much as 36% on a suite of 13 tasks from BIG-Bench Hard, when testing with GPT-3.5 from OpenAI and Claude 1.0 from Anthropic. On a social-bias task, model explanations justify giving answers in line with stereotypes without mentioning the influence of these social biases. Our findings indicate that CoT explanations can be plausible yet misleading, which risks increasing our trust in LLMs without guaranteeing their safety. Building more transparent and explainable systems will require either improving CoT faithfulness through targeted efforts or abandoning CoT in favor of alternative methods.

Sparse Autoencoders for Hypothesis Generation

We describe HypotheSAEs, a general method to hypothesize interpretable relationships between text data (e.g., headlines) and a target variable (e.g., clicks). HypotheSAEs has three steps: (1) train a sparse autoencoder on text embeddings to produce interpretable features describing the data distribution, (2) select features that predict the target variable, and (3) generate a natural language interpretation of each feature (e.g., "mentions being surprised or shocked") using an LLM. Each interpretation serves as a hypothesis about what predicts the target variable. Compared to baselines, our method better identifies reference hypotheses on synthetic datasets (at least +0.06 in F1) and produces more predictive hypotheses on real datasets (~twice as many significant findings), despite requiring 1-2 orders of magnitude less compute than recent LLM-based methods. HypotheSAEs also produces novel discoveries on two well-studied tasks: explaining partisan differences in Congressional speeches and identifying drivers of engagement with online headlines.

Large language models outperform humans at estimating society's everyday norms—but hybrids are even better

As AI assistants and social robots enter human environments, their ability to navigate context-dependent social norms is essential to avoid harm and ensure successful collaboration. We evaluate six large language models (LLMs) on their ability to estimate American social norms across 555 everyday scenarios (measured in prior work) and compare these to estimates from 320 humans. LLMs achieve remarkably high accuracy, clearly outperforming the average human. However, the errors LLMs make are systematic; they are similar across runs of the same LLM and even across different LLMs. As a consequence of this homogeneity, aggregating estimates of LLMs produces little improvement. Individual humans make much worse estimates, often defaulting to extreme right-or-wrong judgments even when asked to estimate population averages, but their errors are idiosyncratic and, consequently, aggregating their estimates yields dramatic improvement through wisdom-of-crowds effects. As humans make different errors than LLMs, hybrid ensembles combining both substantially outperform either alone.

Topics, Authors, and Institutions in Large Language Model Research: Trends from 17K arXiv Papers

Large language models (LLMs) are dramatically influencing AI research, spurring discussions on what has changed so far and how to shape the field's future. To clarify such questions, we analyze a new dataset of 16,979 LLM-related arXiv papers, focusing on recent trends in 2023 vs. 2018-2022. First, we study disciplinary shifts: LLM research increasingly considers societal impacts, evidenced by 20x growth in LLM submissions to the Computers and Society sub-arXiv. An influx of new authors -- half of all first authors in 2023 -- are entering from non-NLP fields of CS, driving disciplinary expansion. Second, we study industry and academic publishing trends. Surprisingly, industry accounts for a smaller publication share in 2023, largely due to reduced output from Google and other Big Tech companies; universities in Asia are publishing more. Third, we study institutional collaboration: while industry-academic collaborations are common, they tend to focus on the same topics that industry focuses on rather than bridging differences. The most prolific institutions are all US- or China-based, but there is very little cross-country collaboration. We discuss implications around (1) how to support the influx of new authors, (2) how industry trends may affect academics, and (3) possible effects of (the lack of) collaboration.

Topics, Authors, and Institutions in Large Language Model Research: Trends from 17K arXiv Papers

Large language models (LLMs) are dramatically influencing AI research, spurring discussions on what has changed so far and how to shape the field's future. To clarify such questions, we analyze a new dataset of 16,979 LLM-related arXiv papers, focusing on recent trends in 2023 vs. 2018-2022. First, we study disciplinary shifts: LLM research increasingly considers societal impacts, evidenced by 20x growth in LLM submissions to the Computers and Society sub-arXiv. An influx of new authors -- half of all first authors in 2023 -- are entering from non-NLP fields of CS, driving disciplinary expansion. Second, we study industry and academic publishing trends. Surprisingly, industry accounts for a smaller publication share in 2023, largely due to reduced output from Google and other Big Tech companies; universities in Asia are publishing more. Third, we study institutional collaboration: while industry-academic collaborations are common, they tend to focus on the same topics that industry focuses on rather than bridging differences. The most prolific institutions are all US- or China-based, but there is very little cross-country collaboration. We discuss implications around (1) how to support the influx of new authors, (2) how industry trends may affect academics, and (3) possible effects of (the lack of) collaboration.

State media control influences large language models

Millions of people around the world query large language models (LLMs) for information. Although several studies have compellingly documented the persuasive potential of these models1–10, there is limited evidence of who or what influences the models themselves, leading to a flurry of concerns about which companies and governments build and regulate the models. Here we show through six studies that government control of the media across the world already influences the output of LLMs via their training data. We use a cross-national audit to show that LLMs exhibit a stronger pro-government valence in the languages of countries with lower media freedom than in those with higher media freedom. This result is correlational, so to triangulate the specific mechanism of how state media control can influence LLMs, we develop a multi-part case study on China’s media. We demonstrate that media scripted and curated by the Chinese state appears in LLM training datasets. To evaluate the plausible effect of this inclusion, we use an open-weight model to show that additional pretraining on Chinese state-coordinated media generates more positive answers to prompts about Chinese political institutions and leaders. We link this phenomenon to commercial models through two audit studies demonstrating that prompting models in Chinese generates more positive responses about China’s institutions and leaders than do the same queries in English. The combination of influence and persuasive potential across languages suggests the troubling conclusion that states and powerful institutions have increased strategic incentives to leverage media control in the hopes of shaping LLM output.