Beyond the Imitation Game: Quantifying and extrapolating the capabilities of language models — part 5/9

Cleaned-text code points [62890, 78787). Bibliography omitted here; source.md includes it.

A worrying finding is that model performance on social bias metrics often grows worse with increasing scale (Figure 12 ). One potential explanation for this may be that larger models do a better job of matching biases in their training set. This result emphasizes the importance of research, engineering, and policy efforts directed at fairness in machine-learning systems, especially at scale. In particular, we believe that techniques of the type employed in the LaMDA model ( Thoppilan et al., 2022 ) to improve model safety and reduce harmful biases will become a crucial component in future language models. Paired with this worrying finding, we also find that social bias can decrease with model scale in settings where unambiguous context makes it clear that the bias is not applicable or desirable (Figure 13 ). We experimented with prepending pro-social prefixes to tasks, and found that this intervention also acts to improve measures of social bias.

The models we evaluated performed better on English than non-English language tasks. Models performed particularly poorly on tasks involving low resource languages. In several cases, performance on low resource language tasks failed to improve with model scale, even when performance on corresponding English language tasks improved reliably with scale (Figures 14 and 15 ).

The performance of all model classes (excluding PaLM, a more modern model trained with a significantly improved training set) was remarkably similar at the same scale (Figures 1 , 1 ), whether trained by Google or OpenAI and with sparse or dense architectures. Despite this broad similarity between model classes, we observed some benefits from a sparse model architecture, with BIG-G sparse models achieving the performance of dense models with roughly twice their effective parameter count on BIG-bench as a whole, and producing multiple choice predictions that were as well calibrated as models about ten times larger (Figures 5 , 1 ).

Models demonstrate many surprising capabilities as they grow. One capability we explore in depth is an improved ability to propose legal chess moves with increasing scale, which suggests that models are learning the rules of chess from transcribed games in their training data (Figure 16 ). This provides a simple example of language models capturing structure that is only incidentally and implicitly encoded in their training data. Also surprising was the ability of language models to identify a movie from a string of emojis representing the plot.

### 6.2 Conclusion

It is our hope that BIG-bench will remain a benchmark that can measure progress well beyond the current state of the art. We therefore end by emphasizing that BIG-bench is a living benchmark. We continue to accept submission of new tasks for peer review on a rolling basis, and new evaluations to be included in the BIG-bench leaderboard. BIG-bench tasks can be specified using a lightweight human-readable JSON format, or defined by Python code. Contributors of accepted tasks will qualify for authorship on future BIG-bench papers and releases.

### Broader Impact Statement

Language models continue to exhibit remarkable, surprising, and cautionary capabilities as they are scaled. These capabilities are sure to have broad and difficult to predict impact on society. BIG-bench is an effort to build a benchmark useful for quantifying and extrapolating these capabilities. In constructing BIG-bench we made an effort to appeal to as broad a swath of the research community as possible -- bringing together authors from different fields and with different backgrounds. Despite this, BIG-bench certainly has holes in what is covered. Some of these holes were apparent when creating BIG-bench. For identified shortcomings we reached out to task authors with expertise to submit tasks, for others -- for instance probing domains beyond text -- we have explicitly limited BIG-bench’s scope. As with any benchmark, however, there are certainly gaps that were not identified. These gaps in coverage will prevent BIG-bench from fully characterizing how the field is progressing, and for those using BIG-bench to guide model development, the limitations of BIG-bench could lead to development focusing on particular areas and neglecting others. We hope to address this to some extent, by treating BIG-bench as a living benchmark 6 6 6 We will continue to follow our review criteria that was used to evaluate and accept our current set of tasks. We plan to snapshot specific version used in the analysis of this paper for easy comparison. – continuing to include tasks targeting broader capabilities and welcome discussion of the benchmark’s limitations (see Appendix F ).

## 7 Author contributions

### 7.1 Core contributors

Here we outline contributions by members of the BIG-bench core project team. The reader is encouraged to attribute credit generously! Everyone who contributed to more than one item spent between weeks and months of their time working on BIG-bench, and in some cases those that contributed the most spoke the most softly about their contributions.

The BIG-bench github code infrastructure and documentation was developed by Guy Gur-Ari, Ethan Dyer, Ambrose Slone, Jascha Sohl-Dickstein, Noah Fiedel, Jaehoon Lee, Daniel Freeman, Aitor Lewkowycz, Anders Andreassen, Gaurav Mishra, and Vedant Misra.

The Google-internal code infrastructure for BIG-bench was built by Ambrose Slone, Ethan Dyer, Gaurav Mishra, Jaehoon Lee, Daniel Freeman, Aitor Lewkowycz, and Anders Andreassen.

The OpenAI-internal code infrastructure for BIG-bench was built by Alex Ray, Vedant Misra, and Alethea Power.

Public interfaces for BIG-bench – especially the SeqIO Bridge and Hugging Face dataset implementations were developed by Gaurav Mishra, Noah Fiedel, Ethan Dyer, and Anders Andreassen.

BIG-G and BIG-G sparse models were trained specifically for use with BIG-bench by Ambrose Slone, William Fedus, Guy Gur-Ari, Barret Zoph, and Jaehoon Lee.

Debugging and improvement of tasks after merging was performed by Ethan Dyer, Ambrose Slone, Jascha Sohl-Dickstein, Guy Gur-Ari, Anders Andreassen, Rosanne Liu, Daniel Freeman, Jaehoon Lee, Vinay Ramasesh, and Vedant Misra.

Professional English language editing of task content was performed by Sarah A. Rous, with support from Anders Andreassen, Vinay Ramasesh, and Kristen Chiafullo.

BIG-bench Lite was developed by Ethan Dyer.

Significant effort was spent liaising with other teams , both inside and external to Google, by Noah Fiedel, Ethan Dyer, Guy Gur-Ari, Rosanne Liu, Gaurav Mishra, Vedant Misra, Jascha Sohl-Dickstein, and Ambrose Slone.

Significant effort was spent actively soliciting BIG-bench task contributions by Jascha Sohl-Dickstein, Ethan Dyer, Rosanne Liu, and Vinay Ramasesh.

Data analysis was performed by Ethan Dyer, Vinay Ramasesh, Ambrose Slone, Jaehoon Lee, Guy Gur-Ari, Noah Constant, Aitor Lewkowycz, Daniel Freeman, Jascha Sohl-Dickstein, Rosanne Liu, Noah Fiedel, Gaurav Mishra, and Vedant Misra.

Measurement of human performance on BIG-bench tasks was organized by Daniel Freeman and Danielle Perszyk.

The humans who performed BIG-bench tasks , and whose scores are reported in this paper, are Vinay Ramasesh, Daniel Freeman, Vedant Misra, and an expert evaluation team consisting of Andrew La, Angela Jiang, Stephen Prasad, Joseph Boudeman, Joseph Guerr, Joyce Chua, Melvin McElrath, Ryan Stovall, Melody Arnaud, Katerina Ignatyeva, Tyler Shultz, Michael Starritt, Sean Casey, Cindy Ramirez, Hayden Bogar, Anh Vuong, Cameron Dour, Michael Gu, and Theodore Rothschild, and managed by Courtney Ashcraft.

The initial project idea was developed in conversation between Guy Gur-Ari, Ethan Dyer, and Jascha Sohl-Dickstein. The initial design doc for the project was written by Jascha Sohl-Dickstein, and further developed by Guy Gur-Ari, Ethan Dyer, Daniel Freeman, Aitor Lewkowycz, and others.

A Google-internal “call for tasks” project was created by Noah Fiedel and Gaurav Mishra and then merged into BIG-bench. This “call for tasks” project is the source of the JSON task API in BIG-bench.

BIG-bench was managed and organized by Guy Gur-Ari, Jascha Sohl-Dickstein, Noah Fiedel, and Ethan Dyer.

The BIG-bench paper was written by Jascha Sohl-Dickstein, Guy Gur-Ari, Ethan Dyer, Daniel Freeman, Noah Fiedel, Vinay Ramasesh, Anders Andreassen, Jaehoon Lee, Noah Constant, Rosanne Liu, Aitor Lewkowycz, Vedant Misra, and Ambrose Slone.

BIG-bench was affiliated with the Workshop on Enormous Language Models at ICLR 2021 . The BIG-bench portions of the workshop were organized by Ethan Dyer, Jascha Sohl-Dickstein, Daniel Freeman, Vedant Misra, Aitor Lewkowycz, Ambrose Slone, and Jaehoon Lee.

Administrative support was provided by Kristen Chiafullo and Vedant Misra.

Meta-review of task submissions was performed by Vinay Ramasesh, Jascha Sohl-Dickstein, Guy Gur-Ari, Ethan Dyer, Anders Andreassen, Aitor Lewkowycz, Daniel Freeman, Jaehoon Lee, Vedant Misra, Rosanne Liu, Noah Fiedel, Gaurav Mishra, and Ambrose Slone. Vinay Ramsesh, Kristen Chiafullo, Guy Gur-Ari, and Jascha Sohl-Dickstein acted at various times as meta-meta reviewers, and coordinated the review process.

Review of BIG-bench tasks was performed by Abhishek Rao, Adam Brown, Aditya Gupta, Adrià Garriga-Alonso, Aitor Lewkowycz, Akshat Agarwal, Alethea Power, Alicia Parrish, Allen Nie, Ambrose Slone, Anders Andreassen, Andrea Madotto, Andrew Dai, Anna Golubeva, Avia Efrat, Behnam Hedayatnia, Behnam Neyshabur, Behrooz Ghorbani, Boris Hanin, Chandan Singh, Charles Rathkopf, Colin Raffel, Daniel Freeman, Daniel Khashabi, Dar Gilboa, David Dohan, Denis Emelin, Derek Tam, Ernie Chang, Ethan Dyer, Gaurav Mishra, Genta Indra Winata, Gerard de Melo, Gregor Betz, Guy Gur-Ari, Harsh Mehta, Hongming Zhang, Igor Babuschkin, Ilan Buzzetti, Isaac Noble, Jack Geissinger, Jaehoon Lee, James B. Simon, James Koppel, James Sully, Jascha Sohl-Dickstein, Jekaterina Novikova, Jesse Michael Han, Joshua Rule, Karthik Gopalakrishnan, Kaustubh D. Dhole, Kyle Aitken, Ludwig Schmidt, Luke Metz, Marcelo Menegali, Marco Maru, Michele Bevilacqua, Michihiro Yasunaga, Mor Geva, Mozhdeh Gheini, Nayeon Lee, Neha Wadia, Nicholas Roberts, Nick Doiron, Nitish Shirish Keskar, Noah Fiedel, Paul Liang, Qing Lyu, Richard Barnes, Rosanne Liu, Samuel Schoenholz, Shixiang Shane Gu, Shubham Toshniwal, Shyam Upadhyay, Sneha Makini, Stanislav Fort, Stella Biderman, Tal Linzen, Tal Schuster, Tao Yu, Thomas Phan, Timothy Nguyen, Vedant Misra, Vikas Raunak, Vinay Prabhu, Vinay Ramasesh, William Saunders, Xudong Shen, Zijian Wang, and Zijie Jay Wang.

The following highly active community members made for a more vibrant and extensive project: Evgenii Zheltonozhskii, Roman Sitelew, Daphne Ippolito, Karthik Gopalakrishnan, Stella Biderman, James B. Simon, Kaustubh D. Dhole, and Sebastian Gehrmann.

### 7.2 Task authors, and task-specific citations

abstract_narrative_understanding : Sayan Ghosh and Shashank Srivastava, UNC Chapel Hill (sayghosh, ssrivastava)@cs.unc.edu. Task cites: ( Ghosh & Srivastava, 2021 ; Holyoak, 2012 ; Nippold et al., 2001 ; Tan et al., 2015 ; Wang et al., 2020b ; Mostafazadeh et al., 2016 )

abstraction_and_reasoning_corpus : Evgenii Zheltonozhskii, Technion – Israel Institute of Technology (evgeniizh@campus.technion.ac.il), and Francois Chollet, Google (fchollet@google.com). Task cites: ( Chollet, 2019 ; Brown et al., 2020 ; Chollet, 2020 )

anachronisms : Peter Chang, Mike Cain, Spencer Torene, Eleanor Hagerman, Zachary Seid, Berk Ekmekci, and Blake Howald, Thomson Reuters Special Services (peter.chang, michael.cain, s.torene, eleanor.hagerman, zach.seid, berk.ekmekci, blake.howald)@thomsonreuters.com. Task cites: ( Otterbacher et al., 2002 ; Popescu & Strapparava, 2015 ; Llorens et al., 2015 ; Meng et al., 2017 ; Geva et al., 2021 )

analogical_similarity : Denis Kleyko, UC Berkeley and Research Institutes of Sweden (denkle@berkeley.edu), and Cameron Diao, Rice University (cwd2@rice.edu). Task cites: ( Plate, 2003 ; Plate, 1994 ; Thagard et al., 1990 ; Gentner et al., 1993 )

analytic_entailment : Rachel Etta Rudolph, Auburn University (rer0037@auburn.edu); Alexander W. Kocurek, Cornell University (awk78@cornell.edu); and Ethan Jerzak, National University of Singapore (phiejj@nus.edu.sg). Task cites: ( Hume, 1739–1740 ; Kant, 1781/1787 ; Wittgenstein, 1953 ; Quine, 1951 ; Grice & Strawson, 1956 ; Bolukbasi et al., 2016 ; Kocurek et al., 2020 ; Rudolph & Kocurek, 2020 ; Kocurek & Jerzak, 2021 )

arithmetic : Guy Gur-Ari, Google (guyga@google.com). Task cites: ( Brown et al., 2020 ; Saxton et al., 2019 )

ascii_word_recognition : Nitish Shirish Keskar, Salesforce (nkeskar@salesforce.com). Task cites: ( Child et al., 2019 ; Chen et al., 2020 )

authorship_verification : Niklas Deckers and Martin Potthast, Leipzig University and ScaDS.AI (niklas.deckers, martin.potthast)@uni-leipzig.de; Sebastian Bischoff, Leipzig University, ScaDS.AI and University of Tübingen (sebastian@salzreute.de); Benno Stein, Bauhaus-Universität Weimar (benno.stein@uni-weimar.de); and Matthias Hagen, Friedrich-Schiller-Universität Jena (matthias.hagen@uni-jena.de). Task cites: ( Bischoff et al., 2020 ; Koppel & Schler, 2004 )

auto_categorization : Fanyue Xia, Carnegie Mellon University (fanyuex@andrew.cmu.edu)

auto_debugging : Mo Tiwari, Medina Orduna Baitemirova, and Chris Waites, Stanford University (motiwari, medinab, waites)@stanford.edu. Task cites: ( Zaremba & Sutskever, 2014 )

bbq_lite : Alicia Parrish, Angelica Chen, Phu Mon Htut, Nikita Nangia, Vishakh Padmakumar, Jason Phang, Jana Thompson, and Samuel R. Bowman, New York University (alicia.v.parrish, ac5968, pmh330, nikitanangia, vp1271, zp489, jt4407, bowman)@nyu.edu. Task cites: ( Crawford, 2017 ; Khashabi et al., 2020b ; Li et al., 2020a )

bbq_lite_json : Alicia Parrish, Angelica Chen, Phu Mon Htut, Nikita Nangia, Vishakh Padmakumar, Jason Phang, Jana Thompson, and Samuel R. Bowman, New York University (alicia.v.parrish, ac5968, pmh330, nikitanangia, vp1271, zp489, jt4407, bowman)@nyu.edu

bias_from_probabilities : Yasaman Bahri and Jascha Sohl-Dickstein, Google (yasamanb, jaschasd)@google.com. Task cites: ( Bender et al., 2021 ; Abid et al., 2021 )

boolean_expressions : Roman Novak, Google (romann@google.com). Task cites: ( Habernal et al., 2018 ; Yu et al., 2020 ; Dua et al., 2019 ; Liu et al., 2020a ; Sinha et al., 2019 ; Wang et al., 2019a ; Steinbach & Kohut, 2002 ; Saxton et al., 2019 ; Payani & Fekri, 2019 ; Trask et al., 2018 ; Selsam et al., 2018 ; Allamanis et al., 2016 ; Evans et al., 2018 ; Shi et al., 2020 )

bridging_anaphora_resolution_barqa : Yufang Hou, IBM Research Europe, Ireland (bnuxiaofang@gmail.com); Katja Markert, Heidelberg University (markert@cl.uni-heidelberg.de); and Michael Strube, Heidelberg Institute for Theoretical Studies (michael.strube@h-its.org). Task cites: ( Hou, 2020 ; Hou et al., 2013 ; Markert et al., 2012 ; Rajpurkar et al., 2016 )

causal_judgment : Allen Nie, Tatsu Hashimoto, and Tobias Gerstenberg, Stanford University (anie, thashim, gerstenberg)@stanford.edu. Task cites: ( Gordon, 2010 ; Bosselut et al., 2019 ; Halpern, 2016 ; Knobe, 2003 )

cause_and_effect : Guy Gur-Ari, Google (guyga@google.com), and Neta Gur-Ari Krakover (krakover@gmail.com). Task cites: ( Gordon, 2010 )

checkmate_in_one : Nitish Shirish Keskar, Salesforce (nkeskar@salesforce.com). Task cites: ( Alexander, 2020 ; Ammanabrolu et al., 2019 ; Dambekodi et al., 2020 ; Ammanabrolu et al., 2020 )

chess_state_tracking : Karen Livescu, Kevin Gimpel, and Shubham Toshniwal, TTI Chicago (klivescu, kgimpel, shtoshni)@ttic.edu; and Sam Wiseman, Duke University (swiseman@cs.duke.edu). Task cites: ( Weston et al., 2015 ; Côté et al., 2018 ; Toshniwal et al., 2021 ; Alexander, 2020 ; Chen, 2020 ; Noever et al., 2020 ; Swingle et al., 2021 )