Tobias Gerstenberg

Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao et al. · allen-ai, amazon-science

Language models demonstrate both quantitative improvement and new qualitative capabilities with increasing scale. Despite their potentially transformative impact, these new capabilities are as yet poorly characterized. In order to inform future research, prepare for disruptive new model capabilities, and ameliorate socially harmful effects, it is vital that we understand the present and near-future capabilities and limitations of language models. To address this challenge, we introduce the Beyond the Imitation Game benchmark (BIG-bench). BIG-bench currently consists of 204 tasks, contributed by 450 authors across 132 institutions. Task topics are diverse, drawing problems from linguistics, childhood development, math, common-sense reasoning, biology, physics, social bias, software development, and beyond. BIG-bench focuses on tasks that are believed to be beyond the capabilities of current language models. We evaluate the behavior of OpenAI's GPT models, Google-internal dense transformer architectures, and Switch-style sparse transformers on BIG-bench, across model sizes spanning millions to hundreds of billions of parameters. In addition, a team of human expert raters performed all tasks in order to provide a strong baseline. Findings include: model performance and calibration both improve with scale, but are poor in absolute terms (and when compared with rater performance); performance is remarkably similar across model classes, though with benefits from sparsity; tasks that improve gradually and predictably commonly involve a large knowledge or memorization component, whereas tasks that exhibit "breakthrough" behavior at a critical scale often involve multiple steps or components, or brittle metrics; social bias typically increases with scale in settings with ambiguous context, but this can be improved with prompting.

Helena Vasconcelos, Matthew Jörke, Madeleine Grunde-McLaughlin et al. · uw

Prior work has identified a resilient phenomenon that threatens the performance of human-AI decision-making teams: overreliance, when people agree with an AI, even when it is incorrect. Surprisingly, overreliance does not reduce when the AI produces explanations for its predictions, compared to only providing predictions. Some have argued that overreliance results from cognitive biases or uncalibrated trust, attributing overreliance to an inevitability of human cognition. By contrast, our paper argues that people strategically choose whether or not to engage with an AI explanation, demonstrating empirically that there are scenarios where AI explanations reduce overreliance. To achieve this, we formalize this strategic choice in a cost-benefit framework, where the costs and benefits of engaging with the task are weighed against the costs and benefits of relying on the AI. We manipulate the costs and benefits in a maze task, where participants collaborate with a simulated AI to find the exit of a maze. Through 5 studies (N = 731), we find that costs such as task difficulty (Study 1), explanation difficulty (Study 2, 3), and benefits such as monetary compensation (Study 4) affect overreliance. Finally, Study 5 adapts the Cognitive Effort Discounting paradigm to quantify the utility of different explanations, providing further support for our framework. Our results suggest that some of the null effects found in literature could be due in part to the explanation not sufficiently reducing the costs of verifying the AI's prediction.

Allen Nie, Yuhui Zhang, Atharva Amdekar et al. · stanford

Human commonsense understanding of the physical and social world is organized around intuitive theories. These theories support making causal and moral judgments. When something bad happens, we naturally ask: who did what, and why? A rich literature in cognitive science has studied people's causal and moral intuitions. This work has revealed a number of factors that systematically influence people's judgments, such as the violation of norms and whether the harm is avoidable or inevitable. We collected a dataset of stories from 24 cognitive science papers and developed a system to annotate each story with the factors they investigated. Using this dataset, we test whether large language models (LLMs) make causal and moral judgments about text-based scenarios that align with those of human participants. On the aggregate level, alignment has improved with more recent LLMs. However, using statistical analyses, we find that LLMs weigh the different factors quite differently from human participants. These results show how curated, challenge datasets combined with insights from cognitive science can help us go beyond comparisons based merely on aggregate metrics: we uncover LLMs implicit tendencies and show to what extent these align with human intuitions.

Kanishk Gandhi, Jan-Philipp Fränken, Tobias Gerstenberg et al.

As Large Language Models (LLMs) become increasingly integrated into our everyday lives, understanding their ability to comprehend human mental states becomes critical for ensuring effective interactions. However, despite the recent attempts to assess the Theory-of-Mind (ToM) reasoning capabilities of LLMs, the degree to which these models can align with human ToM remains a nuanced topic of exploration. This is primarily due to two distinct challenges: (1) the presence of inconsistent results from previous evaluations, and (2) concerns surrounding the validity of existing evaluation methodologies. To address these challenges, we present a novel framework for procedurally generating evaluations with LLMs by populating causal templates. Using our framework, we create a new social reasoning benchmark (BigToM) for LLMs which consists of 25 controls and 5,000 model-written evaluations. We find that human participants rate the quality of our benchmark higher than previous crowd-sourced evaluations and comparable to expert-written evaluations. Using BigToM, we evaluate the social reasoning capabilities of a variety of LLMs and compare model performances with human performance. Our results suggest that GPT4 has ToM capabilities that mirror human inference patterns, though less reliable, while other LLMs struggle.

Jan-Philipp Fränken, Sam Kwok, Peixuan Ye et al. · stanford

We explore the idea of aligning an AI assistant by inverting a model of users' (unknown) preferences from observed interactions. To validate our proposal, we run proof-of-concept simulations in the economic ultimatum game, formalizing user preferences as policies that guide the actions of simulated players. We find that the AI assistant accurately aligns its behavior to match standard policies from the economic literature (e.g., selfish, altruistic). However, the assistant's learned policies lack robustness and exhibit limited generalization in an out-of-distribution setting when confronted with a currency (e.g., grams of medicine) that was not included in the assistant's training distribution. Additionally, we find that when there is inconsistency in the relationship between language use and an unknown policy (e.g., an altruistic policy combined with rude language), the assistant's learning of the policy is slowed. Overall, our preliminary results suggest that developing simulation frameworks in which AI assistants need to infer preferences from diverse users can provide a valuable approach for studying practical alignment questions.

Britt Besch, Tobias Gerstenberg

Explanations are inherently contrastive: E happened rather than E' because of C rather than C'. However, these contrasts, or "foils", are rarely mentioned explicitly but have to be inferred in context. Here, we investigate how people select the intended foil E' of a why-question. Participants read vignettes and judged, for each foil, their prior expectation (what will happen next), closeness (what is most similar to what happened), and hindsight expectation (what could have happened instead), as well as which foil they thought the question asker had in mind when they asked the why-question. We found that foil selections were best predicted by hindsight expectation judgments. This suggests that people infer the foil by considering what a question asker finds surprising after the outcome occurred. Since correct foil selection is relevant not only in human-human interaction but also increasingly in dialogues with large language models, we investigated their performance on the same task. The coupling between LLMs' explicit expectation judgments and their foil selections is inconsistent.

Kanishk Gandhi, Zoe Lynch, Jan-Philipp Fränken et al.

Understanding emotions is fundamental to human interaction and experience. Humans easily infer emotions from situations or facial expressions, situations from emotions, and do a variety of other affective cognition. How adept is modern AI at these inferences? We introduce an evaluation framework for testing affective cognition in foundation models. Starting from psychological theory, we generate 1,280 diverse scenarios exploring relationships between appraisals, emotions, expressions, and outcomes. We evaluate the abilities of foundation models (GPT-4, Claude-3, Gemini-1.5-Pro) and humans (N = 567) across carefully selected conditions. Our results show foundation models tend to agree with human intuitions, matching or exceeding interparticipant agreement. In some conditions, models are ``superhuman'' -- they better predict modal human judgements than the average human. All models benefit from chain-of-thought reasoning. This suggests foundation models have acquired a human-like understanding of emotions and their influence on beliefs and behavior.

Behnam Rahdari, Sameer Shaikh, Jonathan H Chen et al.

LLMs are popular among clinicians for decision-support because of simple text-based interaction. However, their impact on clinicians' performance is ambiguous. Not knowing how clinicians use this new technology and how they compare it to traditional clinical decision-support systems (CDSS) restricts designing novel mechanisms that overcome existing tool limitations and enhance performance and experience. This qualitative study examines how clinicians (n=12) perceive different interaction modalities (text-based conversation with LLMs, interactive and static UI, and voice) for decision-support. In open-ended use of LLM-based tools, our participants took a tool-centric approach using them for information retrieval and confirmation with simple prompts instead of use as active deliberation partners that can handle complex questions. Critical engagement emerged with changes to the interaction setup. Engagement also differed with individual cognitive styles. Lastly, benefits and drawbacks of interaction with text, voice and traditional UIs for clinical decision-support show the lack of a one-size-fits-all interaction modality.

Neha Balamurugan, Sarah Wu, Adam Chun et al.

Humans excel at visual social inference, the ability to infer hidden elements of a scene from subtle behavioral cues such as other people's gaze, pose, and orientation. This ability drives everyday social reasoning in humans and is critical for developing more human-like AI agents. We introduce Spot The Ball, a challenging benchmark for evaluating visual social inference in vision-language models (VLMs) using sports as a test domain. The task is to localize a removed sports ball from soccer, basketball, and volleyball images. We present a curated evaluation set with human baselines and a scalable pipeline for generating additional test items. We evaluate four state-of-the-art VLMs (Gemini, GPT, LLaMA, Qwen) using three prompting strategies, finding that humans are consistently two to three times more accurate (20-34%) than models ($\leq$ 17%) across all sports. Our analyses show that models rely on superficial spatial heuristics--such as guessing near the image center or nearby players--while humans leverage social cues like gaze direction and body pose. These findings reveal a persistent human-model gap in visual social reasoning and underscore the need for architectures that explicitly encode structured behavioral cues to achieve robust, human-like inference.

Chinmaya Andukuri, Jan-Philipp Fränken, Tobias Gerstenberg et al.

When prompting language models to complete a task, users often leave important aspects unsaid. While asking questions could resolve this ambiguity (GATE; Li et al., 2023), models often struggle to ask good questions. We explore a language model's ability to self-improve (STaR; Zelikman et al., 2022) by rewarding the model for generating useful questions-a simple method we dub STaR-GATE. We generate a synthetic dataset of 25,500 unique persona-task prompts to simulate conversations between a pretrained language model-the Questioner-and a Roleplayer whose preferences are unknown to the Questioner. By asking questions, the Questioner elicits preferences from the Roleplayer. The Questioner is iteratively finetuned on questions that increase the probability of high-quality responses to the task, which are generated by an Oracle with access to the Roleplayer's latent preferences. After two iterations of self-improvement, the Questioner asks better questions, allowing it to generate responses that are preferred over responses from the initial model on 72% of tasks. Our results indicate that teaching a language model to ask better questions leads to better personalized responses.

Hanqi Zhou, Britt Besch, Charley M. Wu et al.

Explaining how to get from A to B can be challenging. It requires mentally simulating what the listener will do based on what they are told. To capture this process, we propose a computational model that converts utterances into action plans: a large language model translates an explanation into program-like guidance (a policy prior and value map), and a planning agent executes it under partial observability. We score explanations by the efficiency and reliability of the resulting paths, penalizing replanning. Across four preregistered experiments, we collect a corpus of 1,200 explanations over 24 maps, elicit helpfulness judgments, measure baseline navigation, and test behavior with explanations of differing quality. Higher-scored explanations are judged more helpful and improve navigation: participants with explanations outperform those without, and high-scoring explanations help more than low-scoring ones. Together, these results show procedural explanation as utility-guided communication shaped by how language can be grounded into action under uncertainty.

Jan-Philipp Fränken, Eric Zelikman, Rafael Rafailov et al.

When prompting a language model (LM), users often expect the model to adhere to a set of behavioral principles across diverse tasks, such as producing insightful content while avoiding harmful or biased language. Instilling such principles (i.e., a constitution) into a model is resource-intensive, technically challenging, and generally requires human preference labels or examples. We introduce SAMI, an iterative algorithm that finetunes a pretrained language model (without requiring preference labels or demonstrations) to increase the conditional mutual information between constitutions and self-generated responses given queries from a dataset. On single-turn dialogue and summarization, a SAMI-trained mistral-7b outperforms the initial pretrained model, with win rates between 66% and 77%. Strikingly, it also surpasses an instruction-finetuned baseline (mistral-7b-instruct) with win rates between 55% and 57% on single-turn dialogue. SAMI requires a model that writes the principles. To avoid dependence on strong models for writing principles, we align a strong pretrained model (mixtral-8x7b) using constitutions written by a weak instruction-finetuned model (mistral-7b-instruct), achieving a 65% win rate on summarization. Finally, we investigate whether SAMI generalizes to diverse summarization principles (e.g., "summaries should be scientific") and scales to stronger models (llama3-70b), finding that it achieves win rates of up to 68% for learned and 67% for held-out principles compared to the base model. Our results show that a pretrained LM can learn to follow constitutions without using preference labels, demonstrations, or human oversight.

Samuel G. B. Johnson, Amir-Hossein Karimi, Yoshua Bengio et al.

Although AI has become increasingly smart, its wisdom has not kept pace. In this article, we examine what is known about human wisdom and sketch a vision of its AI counterpart. We analyze human wisdom as a set of strategies for solving intractable problems-those outside the scope of analytic techniques-including both object-level strategies like heuristics [for managing problems] and metacognitive strategies like intellectual humility, perspective-taking, or context-adaptability [for managing object-level strategies]. We argue that AI systems particularly struggle with metacognition; improved metacognition would lead to AI more robust to novel environments, explainable to users, cooperative with others, and safer in risking fewer misaligned goals with human users. We discuss how wise AI might be benchmarked, trained, and implemented.

Lionel Wong, Katherine M. Collins, Lance Ying et al.

When faced with novel situations, people are able to marshal relevant considerations from a wide range of background knowledge and put these to use in inferences and predictions. What permits us to draw in globally relevant information and reason over it coherently? Here, we explore the hypothesis that people use a combination of distributed and symbolic representations to construct bespoke mental models tailored to novel situations. We propose a computational implementation of this idea -- a ``Model Synthesis Architecture'' (MSA) -- using language models to implement global relevance-based retrieval and model synthesis and probabilistic programs to implement bespoke, coherent world models. We evaluate our MSA as a model of human judgments on a novel reasoning dataset. The dataset -- built around a `Model Olympics` domain of sports vignettes -- tests models' capacity for human-like, open-ended reasoning by requiring (i) judgments about novel causal structures described in language; (ii) drawing on large bodies of background knowledge; and (iii) doing both in light of observations that introduce arbitrary novel variables. Our MSA approach captures human judgments better than language model-only baselines, under both direct and chain-of-thought generations from the LM that supports model synthesis. These results suggest that MSAs can be implemented in a way that mirrors people's ability to deliver locally coherent reasoning over globally relevant variables, offering a path to understanding and replicating human reasoning in open-ended domains.

Carlota Parés-Morlans, Michelle Yi, Claire Chen et al.

Tasks that involve complex interactions between objects with unknown dynamics make planning before execution difficult. These tasks require agents to iteratively improve their actions after actively exploring causes and effects in the environment. For these type of tasks, we propose Causal-PIK, a method that leverages Bayesian optimization to reason about causal interactions via a Physics-Informed Kernel to help guide efficient search for the best next action. Experimental results on Virtual Tools and PHYRE physical reasoning benchmarks show that Causal-PIK outperforms state-of-the-art results, requiring fewer actions to reach the goal. We also compare Causal-PIK to human studies, including results from a new user study we conducted on the PHYRE benchmark. We find that Causal-PIK remains competitive on tasks that are very challenging, even for human problem-solvers.

Jan-Philipp Fränken, Kanishk Gandhi, Tori Qiu et al.

As AI systems like language models are increasingly integrated into decision-making processes affecting people's lives, it's critical to ensure that these systems have sound moral reasoning. To test whether they do, we need to develop systematic evaluations. We provide a framework that uses a language model to translate causal graphs that capture key aspects of moral dilemmas into prompt templates. With this framework, we procedurally generated a large and diverse set of moral dilemmas -- the OffTheRails benchmark -- consisting of 50 scenarios and 400 unique test items. We collected moral permissibility and intention judgments from human participants for a subset of our items and compared these judgments to those from two language models (GPT-4 and Claude-2) across eight conditions. We find that moral dilemmas in which the harm is a necessary means (as compared to a side effect) resulted in lower permissibility and higher intention ratings for both participants and language models. The same pattern was observed for evitable versus inevitable harmful outcomes. However, there was no clear effect of whether the harm resulted from an agent's action versus from having omitted to act. We discuss limitations of our prompt generation pipeline and opportunities for improving scenarios to increase the strength of experimental effects.

Logan Cross, Erik Brockbank, Tobias Gerstenberg et al.

How do we predict others from patterns in their behavior and what are the computational constraints that limit this ability? We investigate these questions by modeling human behavior over repeated games of rock, paper, scissors from Brockbank & Vul (2024). Against algorithmic opponents that varied in strategic sophistication, people readily exploit simple transition patterns (e.g., consistently playing rock after paper) but struggle to detect more complex sequential dependencies. To understand the cognitive mechanisms underlying these abilities and their limitations, we deploy Hypothetical Minds (HM), a large language model-based agent that generates and tests hypotheses about opponent strategies, as a cognitive model of this behavior (Cross et al., 2024). We show that when applied to the same experimental conditions, HM closely mirrors human performance patterns, succeeding and failing in similar ways. To better understand the source of HM's failures and whether people might face similar cognitive bottlenecks in this context, we performed a series of ablations and augmentations targeting different components of the system. When provided with natural language descriptions of the opponents' strategies, HM successfully exploited 6/7 bot opponents with win rates >80% suggesting that accurate hypothesis generation is the primary cognitive bottleneck in this task. Further, by systematically manipulating the model's hypotheses through pedagogically-inspired interventions, we find that the model substantially updates its causal understanding of opponent behavior, revealing how model-based analyses can produce testable hypotheses about human cognition.

Kailas Vodrahalli, Tobias Gerstenberg, James Zou

In many practical applications of AI, an AI model is used as a decision aid for human users. The AI provides advice that a human (sometimes) incorporates into their decision-making process. The AI advice is often presented with some measure of "confidence" that the human can use to calibrate how much they depend on or trust the advice. In this paper, we present an initial exploration that suggests showing AI models as more confident than they actually are, even when the original AI is well-calibrated, can improve human-AI performance (measured as the accuracy and confidence of the human's final prediction after seeing the AI advice). We first train a model to predict human incorporation of AI advice using data from thousands of human-AI interactions. This enables us to explicitly estimate how to transform the AI's prediction confidence, making the AI uncalibrated, in order to improve the final human prediction. We empirically validate our results across four different tasks--dealing with images, text and tabular data--involving hundreds of human participants. We further support our findings with simulation analysis. Our findings suggest the importance of jointly optimizing the human-AI system as opposed to the standard paradigm of optimizing the AI model alone.

Kailas Vodrahalli, Roxana Daneshjou, Tobias Gerstenberg et al.

In decision support applications of AI, the AI algorithm's output is framed as a suggestion to a human user. The user may ignore this advice or take it into consideration to modify their decision. With the increasing prevalence of such human-AI interactions, it is important to understand how users react to AI advice. In this paper, we recruited over 1100 crowdworkers to characterize how humans use AI suggestions relative to equivalent suggestions from a group of peer humans across several experimental settings. We find that participants' beliefs about how human versus AI performance on a given task affects whether they heed the advice. When participants do heed the advice, they use it similarly for human and AI suggestions. Based on these results, we propose a two-stage, "activation-integration" model for human behavior and use it to characterize the factors that affect human-AI interactions.

Ilker Yildirim, Basil Saeed, Grace Bennett-Pierre et al.

The ability to estimate task difficulty is critical for many real-world decisions such as setting appropriate goals for ourselves or appreciating others' accomplishments. Here we give a computational account of how humans judge the difficulty of a range of physical construction tasks (e.g., moving 10 loose blocks from their initial configuration to their target configuration, such as a vertical tower) by quantifying two key factors that influence construction difficulty: physical effort and physical risk. Physical effort captures the minimal work needed to transport all objects to their final positions, and is computed using a hybrid task-and-motion planner. Physical risk corresponds to stability of the structure, and is computed using noisy physics simulations to capture the costs for precision (e.g., attention, coordination, fine motor movements) required for success. We show that the full effort-risk model captures human estimates of difficulty and construction time better than either component alone.

Ilker Yildirim, Tobias Gerstenberg, Basil Saeed et al.

In this paper, we present a new task that investigates how people interact with and make judgments about towers of blocks. In Experiment~1, participants in the lab solved a series of problems in which they had to re-configure three blocks from an initial to a final configuration. We recorded whether they used one hand or two hands to do so. In Experiment~2, we asked participants online to judge whether they think the person in the lab used one or two hands. The results revealed a close correspondence between participants' actions in the lab, and the mental simulations of participants online. To explain participants' actions and mental simulations, we develop a model that plans over a symbolic representation of the situation, executes the plan using a geometric solver, and checks the plan's feasibility by taking into account the physical constraints of the scene. Our model explains participants' actions and judgments to a high degree of quantitative accuracy.