Insert New
Slide
​
Roboto
A
Slide Break Defaults…
​
Cell Actions
Window Options
Start Presentation
▼
100%
Slide
1
of
25
​
​

AI Agents for
Higher-Order Thinking
(and Literature Review)*

John McNally
Wolfram Research
ACS Fall Meeting 2025
Slide
2
of
25
​
​
Agenda

A paradigm for developing skills

The AI toolkit

AI for Literature Review*
Slide
3
of
25
​
​

FAQ About AI Agents


  • How many AI agents are you talking about?
    • 
  • In this talk: 1. The assertion is to want one agent per lesson activity.
    • 
  • What do you mean by AI agent?
    • 
  • In my parlance: Agent = Language Model + Prompting + Tools
    • 
  • Do you mean a Wolfram agent or OpenAI agent?
    • 
  • In this talk: both. You can build agents as an amalgam of resources.
    • Slide
    4
    of
    25
    ​
    ​

    A Paradigm for AI Agents in Skills-Based Pedagogy

    Slide
    5
    of
    25
    ​
    ​

    The Backstory - Graded Autonomy

    Slide
    6
    of
    25
    ​
    ​

    Differentiated Instruction

    “Differentiated instruction involves teaching in a way that meets the different needs and interests of students using varied course content, activities, and assessments.”
    - https://ctl.stanford.edu/differentiated-instruction

    How the Paradigm Supports Differentiated Instruction

    Code One ⇒ Procedural and Factual Knowledge
    Prompt One ⇒ Strategic Planning and Execution
    Build One ⇒ Systems Thinking and Multi-disciplinary Creativity
    Slide
    7
    of
    25
    ​
    ​

    Code One - Get Some Lab Data

    
  • A lesson “from the wild”
    • 
  • Non-linear regression and modeling of Thermogravimetric Analysis data
    • 
  • Thanks to Ghada Raba from NC State University for sharing the data
    • In[]:=
    TGAdata=
    Import[
    ]
    Out[]=
    Tabular
    Row count: 1452
    Column count: 3
    
    Slide
    8
    of
    25
    ​
    ​

    Code One - Explore a Model

    In[]:=
    params={w0,w1,w2,w3,a1,a2,a3,b1,b2,b3};​​model=w0+{w1,w2,w3}.LogisticSigmoid[{a1,a2,a3}-{b1,b2,b3}t];
    Out[]=
    ​
    w0
    w1
    w2
    w3
    a1
    a2
    a3
    b1
    b2
    b3
    Observed
    Model
    Slide
    9
    of
    25
    ​
    ​

    Code One - Fit and Visualize a Model

    In[]:=
    nlm=
    NonlinearModelFit[
    ]
    ;​​TGAfitted=
    TransformColumns[
    ]
    ;​​Column
    ListPlot[
    ]
    ,
    ListPlot[
    ]
    ,
    
    Out[]=
    10
    20
    30
    40
    5
    10
    15
    Observed
    Model Prediction
    10
    20
    30
    40
    -0.02
    0.02
    0.04
    0.06
    Percent Error
    Slide
    10
    of
    25
    ​
    ​

    Prompt One - Planning

    ​
    I want to do the following using the provided tool:
    1. Import the lab data found at https://www.wolframcloud.com/obj/jmcnally0/ACSDemo/preformattedTGA.csv.
    2. Fit the decomposition reaction data to a 3-step logistic sigmoid model.
    3. Visualize the model predictions and observed data.
    4. Visualize the percent error of the fitted model.
    5. Identify regions where the percent error is greater than 2 %.
    Slide
    11
    of
    25
    ​
    ​

    Prompt One - Exploration

    Make me a manipulative where I can adjust all those parameters in the empirical model
    Slide
    12
    of
    25
    ​
    ​

    Prompt One - Extension

    Wonderful. Now I want to visualize the percent error and identify regions where it is greater than 1%.
    Slide
    13
    of
    25
    ​
    ​

    Prompt One - Secret Sauce

    The example works reliably when the agent is given specific tools.
    Performance is hit-or-miss if the language model has to invent tools.
    https://chatgpt.com/share/68a53adb-9014-800d-9c01-d08f72ae1267
    Slide
    14
    of
    25
    ​
    ​

    Build One - Community Design Suggestions

    from Sigot and Tassoti in J. Chem. Educ. 2025, 102, 2151−2159
    Slide
    15
    of
    25
    ​
    ​

    Build One - Community Design Checklist

    How did we do today?
    Slide
    16
    of
    25
    ​
    ​

    Build One - Extension Activity for Differentiation

    Gory Details:
    Slide
    17
    of
    25
    ​
    ​
    The AI Toolkit
    
    Computational Tools
    
    Semantic Search
    
    Language Models
    Slide
    18
    of
    25
    ​
    ​

    The AI Toolkit - Language Models

    Stochasticity is an inherent part of language model text generation!
    Slide
    19
    of
    25
    ​
    ​

    The AI Toolkit - Semantic Search

    Slide
    20
    of
    25
    ​
    ​

    The AI Toolkit - Computational Tools

    Slide
    21
    of
    25
    ​
    ​

    Recap

    
  • Ask if you have the right agent for your lesson, not for your course
    • 
  • AI agents today have a trade-off between generality and reliability
    • Slide
    22
    of
    25
    ​
    ​

    Agents for Literature Review

    * Time permitting
    Slide
    23
    of
    25
    ​
    ​

    Shortest Paths Between Literature Results

    Slide
    24
    of
    25
    ​
    ​

    Serve the Results as a Notebook

    Slide
    25
    of
    25
    ​
    ​
    Thanks!
    Some code in the slides requires permissions
    Please reach out if you want to a runnable copy