From 06009e45a2396c8325959a15b60b3cbecf343975 Mon Sep 17 00:00:00 2001 From: Tobias Gerstenberg Date: Fri, 20 Sep 2024 10:54:25 -0700 Subject: [PATCH] added gandhi affect paper --- .../publication/amemiya2024disagreement.md | 5 +- content/publication/du2024robotic.md | 2 +- content/publication/franken2024sami.md | 2 +- content/publication/gandhi2024affective.md | 33 ++ .../gerstenberg2024counterfactual.md | 2 +- docs/404.html | 8 +- docs/bibtex/cic_papers.bib | 12 +- docs/index.html | 6 +- docs/index.xml | 20 +- docs/member/tobias_gerstenberg/index.html | 51 ++ .../amemiya2024disagreement/index.html | 8 + docs/publication/du2024robotic/index.html | 4 +- docs/publication/franken2024sami/index.html | 4 +- .../gandhi2024affective/index.html | 488 ++++++++++++++++++ .../gerstenberg2024counterfactual/index.html | 4 + docs/publication/index.html | 80 ++- docs/publication/index.xml | 9 + docs/publication_types/1/index.html | 29 +- docs/publication_types/1/index.xml | 29 +- docs/publication_types/2/index.html | 38 +- docs/publication_types/2/index.xml | 20 +- docs/publication_types/2/page/2/index.html | 38 +- docs/publication_types/2/page/3/index.html | 38 +- docs/publication_types/2/page/4/index.html | 20 +- docs/publication_types/index.html | 8 +- docs/publication_types/index.xml | 22 +- docs/sitemap.xml | 34 +- static/bibtex/cic_papers.bib | 12 +- 28 files changed, 874 insertions(+), 152 deletions(-) create mode 100644 content/publication/gandhi2024affective.md create mode 100644 docs/publication/gandhi2024affective/index.html diff --git a/content/publication/amemiya2024disagreement.md b/content/publication/amemiya2024disagreement.md index fb212db..d3a8347 100644 --- a/content/publication/amemiya2024disagreement.md +++ b/content/publication/amemiya2024disagreement.md @@ -24,7 +24,10 @@ url_slides = "" url_video = "" url_poster = "posters/amemiya2023disagreement-poster.pdf" url_source = "" -url_custom = [{name = "Github", url = "https://github.com/cicl-stanford/children_disagree"}] +url_custom = [{name = "Link", url = "https://www.sciencedirect.com/science/article/pii/S0010027724001227"}, +{name = "Github", url = "https://github.com/cicl-stanford/children_disagree"}, +{name = "Press: Psypost", url = "https://www.psypost.org/childrens-ability-to-detect-ambiguity-in-disagreements-sharpens-between-ages-7-and-11/"} +] math = true highlight = true [header] diff --git a/content/publication/du2024robotic.md b/content/publication/du2024robotic.md index e8b7b4b..e33613b 100644 --- a/content/publication/du2024robotic.md +++ b/content/publication/du2024robotic.md @@ -9,7 +9,7 @@ title = "To Err is Robotic: Rapid Value-Based Trial-and-Error during Deployment" date = "2024-06-26" authors = ["M. Du","A. Khazatsky","T. Gerstenberg","C. Finn"] -publication_types = ["2"] +publication_types = ["1"] publication_short = "_arXiv_" publication = "Du, M., Khazatsky, A., Gerstenberg, T., Finn, C. (2024). To Err is Robotic: Rapid Value-Based Trial-and-Error during Deployment. _arXiv_." abstract = "When faced with a novel scenario, it can be hard to succeed on the first attempt. In these challenging situations, it is important to know how to retry quickly and meaningfully. Retrying behavior can emerge naturally in robots trained on diverse data, but such robot policies will typically only exhibit undirected retrying behavior and may not terminate a suboptimal approach before an unrecoverable mistake. We can improve these robot policies by instilling an explicit ability to try, evaluate, and retry a diverse range of strategies. We introduce Bellman-Guided Retrials, an algorithm that works on top of a base robot policy by monitoring the robot's progress, detecting when a change of plan is needed, and adapting the executed strategy until the robot succeeds. We start with a base policy trained on expert demonstrations of a variety of scenarios. Then, using the same expert demonstrations, we train a value function to estimate task completion. During test time, we use the value function to compare our expected rate of progress to our achieved rate of progress. If our current strategy fails to make progress at a reasonable rate, we recover the robot and sample a new strategy from the base policy while skewing it away from behaviors that have recently failed. We evaluate our method on simulated and real-world environments that contain a diverse suite of scenarios. We find that Bellman-Guided Retrials increases the average absolute success rates of base policies by more than 20% in simulation and 50% in real-world experiments, demonstrating a promising framework for instilling existing trained policies with explicit trial and error capabilities. For evaluation videos and other documentation, go to https://sites.google.com/view/to-err-robotic/home" diff --git a/content/publication/franken2024sami.md b/content/publication/franken2024sami.md index 9821be6..b02dbd7 100644 --- a/content/publication/franken2024sami.md +++ b/content/publication/franken2024sami.md @@ -9,7 +9,7 @@ title = "Self-supervised alignment with mutual information: Learning to follow principles without preference labels" date = "2024-04-22" authors = ["J. Fränken","E. Zelikman","R. Rafailov","K. Gandhi","T. Gerstenberg","N. D. Goodman"] -publication_types = ["2"] +publication_types = ["1"] publication_short = "_arXiv_" publication = "Fränken, J., Zelikman, E., Rafailov, R., Gandhi, K., Gerstenberg, T., Goodman, N. D. (2024). Self-supervised alignment with mutual information: Learning to follow principles without preference labels. _arXiv_." abstract = "When prompting a language model (LM), users frequently 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 into a model can be resource-intensive and technically challenging, generally requiring human preference labels or examples. We introduce SAMI, a method for teaching a pretrained LM to follow behavioral principles that does not require any preference labels or demonstrations. SAMI is an iterative algorithm that finetunes a pretrained LM to increase the conditional mutual information between constitutions and self-generated responses given queries from a datasest. 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 'principle writer' model; to avoid dependence on stronger models, we further evaluate aligning a strong pretrained model (mixtral-8x7b) using constitutions written by a weak instruction-finetuned model (mistral-7b-instruct). The SAMI-trained mixtral-8x7b outperforms both the initial model and the instruction-finetuned model, achieving a 65% win rate on summarization. Our results indicate that a pretrained LM can learn to follow constitutions without using preference labels, demonstrations, or human oversight." diff --git a/content/publication/gandhi2024affective.md b/content/publication/gandhi2024affective.md new file mode 100644 index 0000000..6fa1d3b --- /dev/null +++ b/content/publication/gandhi2024affective.md @@ -0,0 +1,33 @@ ++++ +# 0 -> 'Forthcoming', +# 1 -> 'Preprint', +# 2 -> 'Journal', +# 3 -> 'Conference Proceedings', +# 4 -> 'Book chapter', +# 5 -> 'Thesis' + +title = "Human-like Affective Cognition in Foundation Models" +date = "2024-09-20" +authors = ["K. Gandhi","Z. Lynch","J. Fränken","K. Patterson","S. Wambu","T. Gerstenberg","D. C. Ong","N. D. Goodman"] +publication_types = ["1"] +publication_short = "_arXiv_" +publication = "Gandhi K., Lynch Z., Fränken J., Patterson K., Wambu S., Gerstenberg T., Ong D. C., Goodman N. D. (2024). Human-like Affective Cognition in Foundation Models. _arXiv_." +abstract = "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." +image_preview = "" +selected = false +projects = [] +#url_pdf = "papers/gandhi2024affective.pdf" +url_preprint = "https://arxiv.org/abs/2409.11733" +url_code = "" +url_dataset = "" +url_slides = "" +url_video = "" +url_poster = "" +url_source = "" +#url_custom = [{name = "Github", url = ""}] +math = true +highlight = true +[header] +# image = "publications/gandhi2024affective.png" +caption = "" ++++ \ No newline at end of file diff --git a/content/publication/gerstenberg2024counterfactual.md b/content/publication/gerstenberg2024counterfactual.md index c917437..027df86 100644 --- a/content/publication/gerstenberg2024counterfactual.md +++ b/content/publication/gerstenberg2024counterfactual.md @@ -24,7 +24,7 @@ url_slides = "" url_video = "" url_poster = "" url_source = "" -url_custom = [{name = "Press: HAI News", url = "https://hai.stanford.edu/news/humans-use-counterfactuals-reason-about-causality-can-ai"}] +url_custom = [{name = "Link", url = "https://www.sciencedirect.com/science/article/pii/S1364661324001074"}, {name = "Press: HAI News", url = "https://hai.stanford.edu/news/humans-use-counterfactuals-reason-about-causality-can-ai"}] math = true highlight = true [header] diff --git a/docs/404.html b/docs/404.html index a0933d6..45a943e 100644 --- a/docs/404.html +++ b/docs/404.html @@ -237,6 +237,10 @@

Page not found

Publications

+ + @@ -253,10 +257,6 @@

Publications

  • Whodunnit? Inferring what happened from multimodal evidence
    • - - diff --git a/docs/bibtex/cic_papers.bib b/docs/bibtex/cic_papers.bib index 058e7e7..1c24c36 100644 --- a/docs/bibtex/cic_papers.bib +++ b/docs/bibtex/cic_papers.bib @@ -1,13 +1,23 @@ %% This BibTeX bibliography file was created using BibDesk. %% https://bibdesk.sourceforge.io/ -%% Created for Tobias Gerstenberg at 2024-06-26 12:25:09 -0700 +%% Created for Tobias Gerstenberg at 2024-09-20 10:50:37 -0700 %% Saved with string encoding Unicode (UTF-8) +@article{gandhi2024affective, + abstract = {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.}, + author = {Kanishk Gandhi and Zoe Lynch and Jan-Philipp Fr{\"a}nken and Kayla Patterson and Sharon Wambu and Tobias Gerstenberg and Desmond C. Ong and Noah D. Goodman}, + date-added = {2024-09-20 10:50:33 -0700}, + date-modified = {2024-09-20 10:50:37 -0700}, + journal = {arXiv}, + note = {https://arxiv.org/abs/2409.11733}, + title = {Human-like Affective Cognition in Foundation Models}, + year = {2024}} + @article{du2024robotic, abstract = {When faced with a novel scenario, it can be hard to succeed on the first attempt. In these challenging situations, it is important to know how to retry quickly and meaningfully. Retrying behavior can emerge naturally in robots trained on diverse data, but such robot policies will typically only exhibit undirected retrying behavior and may not terminate a suboptimal approach before an unrecoverable mistake. We can improve these robot policies by instilling an explicit ability to try, evaluate, and retry a diverse range of strategies. We introduce Bellman-Guided Retrials, an algorithm that works on top of a base robot policy by monitoring the robot's progress, detecting when a change of plan is needed, and adapting the executed strategy until the robot succeeds. We start with a base policy trained on expert demonstrations of a variety of scenarios. Then, using the same expert demonstrations, we train a value function to estimate task completion. During test time, we use the value function to compare our expected rate of progress to our achieved rate of progress. If our current strategy fails to make progress at a reasonable rate, we recover the robot and sample a new strategy from the base policy while skewing it away from behaviors that have recently failed. We evaluate our method on simulated and real-world environments that contain a diverse suite of scenarios. We find that Bellman-Guided Retrials increases the average absolute success rates of base policies by more than 20% in simulation and 50% in real-world experiments, demonstrating a promising framework for instilling existing trained policies with explicit trial and error capabilities. For evaluation videos and other documentation, go to https://sites.google.com/view/to-err-robotic/home}, author = {Du, Maximilian and Khazatsky, Alexander and Gerstenberg, Tobias and Finn, Chelsea}, diff --git a/docs/index.html b/docs/index.html index 6cba24d..ed450d8 100644 --- a/docs/index.html +++ b/docs/index.html @@ -110,7 +110,7 @@ - + @@ -1660,6 +1660,10 @@

    Selected
    publications

    + + Link + + Press: HAI News diff --git a/docs/index.xml b/docs/index.xml index d239a44..4a88fc9 100644 --- a/docs/index.xml +++ b/docs/index.xml @@ -6,9 +6,18 @@ Hugo -- gohugo.io en-us © 2024 Tobias Gerstenberg - Wed, 26 Jun 2024 00:00:00 +0000 + Fri, 20 Sep 2024 00:00:00 +0000 + + Human-like Affective Cognition in Foundation Models + https://cicl.stanford.edu/publication/gandhi2024affective/ + Fri, 20 Sep 2024 00:00:00 +0000 + + https://cicl.stanford.edu/publication/gandhi2024affective/ + + + To Err is Robotic: Rapid Value-Based Trial-and-Error during Deployment https://cicl.stanford.edu/publication/du2024robotic/ @@ -135,14 +144,5 @@ - - Anticipating the risks and benefits of counterfactual world simulation models - https://cicl.stanford.edu/publication/kirfel2023anticipating/ - Mon, 30 Oct 2023 00:00:00 +0000 - - https://cicl.stanford.edu/publication/kirfel2023anticipating/ - - - diff --git a/docs/member/tobias_gerstenberg/index.html b/docs/member/tobias_gerstenberg/index.html index 55d780f..71da238 100644 --- a/docs/member/tobias_gerstenberg/index.html +++ b/docs/member/tobias_gerstenberg/index.html @@ -356,6 +356,45 @@

    Publications

    + + + (2024). + + Human-like Affective Cognition in Foundation Models. + arXiv. + + + + +

    + + + + + + Preprint + + + + + + + + + + + + + + + + + +

    + +
    +
    @@ -437,6 +476,10 @@

    Publications

    + + Link + + Press: HAI News @@ -488,10 +531,18 @@

    Publications

    + + Link + + Github + + Press: Psypost + +

    diff --git a/docs/publication/amemiya2024disagreement/index.html b/docs/publication/amemiya2024disagreement/index.html index 6139b36..0873448 100644 --- a/docs/publication/amemiya2024disagreement/index.html +++ b/docs/publication/amemiya2024disagreement/index.html @@ -341,10 +341,18 @@

    Abstract

    + + Link + + Github + + Press: Psypost + +
    diff --git a/docs/publication/du2024robotic/index.html b/docs/publication/du2024robotic/index.html index 654268d..66691d4 100644 --- a/docs/publication/du2024robotic/index.html +++ b/docs/publication/du2024robotic/index.html @@ -266,8 +266,8 @@

    Abstract

    Type
    - - Journal + + Preprint
    diff --git a/docs/publication/franken2024sami/index.html b/docs/publication/franken2024sami/index.html index 57060aa..9dce50f 100644 --- a/docs/publication/franken2024sami/index.html +++ b/docs/publication/franken2024sami/index.html @@ -266,8 +266,8 @@

    Abstract

    Type
    - - Journal + + Preprint
    diff --git a/docs/publication/gandhi2024affective/index.html b/docs/publication/gandhi2024affective/index.html new file mode 100644 index 0000000..e14f2f5 --- /dev/null +++ b/docs/publication/gandhi2024affective/index.html @@ -0,0 +1,488 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + Human-like Affective Cognition in Foundation Models | Causality in Cognition Lab + + + + + + +
    + + + + +
    +
    +

    Human-like Affective Cognition in Foundation Models

    + + + + + +
    + + + + +

    Abstract

    +

    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.

    + + + + + + +
    +
    +
    +
    +
    Type
    +
    + + + Preprint + + +
    +
    +
    +
    +
    +
    + + +
    +
    +
    +
    +
    Publication
    +
    Gandhi K., Lynch Z., Fränken J., Patterson K., Wambu S., Gerstenberg T., Ong D. C., Goodman N. D. (2024). Human-like Affective Cognition in Foundation Models. arXiv.
    +
    +
    +
    +
    +
    + +
    +
    +
    +
    +
    Date
    + +
    +
    +
    +
    +
    + + +
    +
    +
    +
    +
    Links
    +
    + + + + + + Preprint + + + + + + + + + + + + + + + + + +
    +
    +
    +
    +
    +
    + + +
    + +
    + +

    << Back to list of publications

     +
    + +
    + + + + + +
    + + +
    + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/docs/publication/gerstenberg2024counterfactual/index.html b/docs/publication/gerstenberg2024counterfactual/index.html index 8b80b57..edc0690 100644 --- a/docs/publication/gerstenberg2024counterfactual/index.html +++ b/docs/publication/gerstenberg2024counterfactual/index.html @@ -337,6 +337,10 @@

    Abstract

    + + Link + + Press: HAI News diff --git a/docs/publication/index.html b/docs/publication/index.html index 1a635de..a71beba 100644 --- a/docs/publication/index.html +++ b/docs/publication/index.html @@ -1584,6 +1584,19 @@

    Publications

    + + + + + + + + + + + + + @@ -1720,7 +1733,58 @@

    Publications

    -
    +
    + +
    + + + (2024). + + Human-like Affective Cognition in Foundation Models. + arXiv. + + + + +

    + + + + + + Preprint + + + + + + + + + + + + + + + + + +

    + +
    + + +
    + + + + + + + +