Whose Inference?

Whose Inference?

The same sentence means different things to different readers. This pilot asks whether an AI system can predict how one specific person will understand a message, from an explicit record of what they know and believe, and then rephrase it so it is not misunderstood.

“Two teaching methods were compared with a t-test, which gave p = .03.”

Does it follow that “there is a 97% probability that the difference is real”?

Methods lecturerKnows that p is computed assuming there is no difference.This reader does not draw it.
Student after one statistics courseUnderstands p-values, but believes “significant” means “large”.This reader does not draw it.
Teacher who reads study newsBelieves p is the probability that the null hypothesis is true.This reader draws it, a misreading.
Try the live app Code on GitHub Full report

A pilot study by Amrina, M.Sc. Educational Technology, Saarland University. LinkedIn · GitHub

How it works

Each reader has a small personal knowledge base (PKB): a readable list of what they know and believe. Four steps test one part each of a cooperative system that relies on it.

  1. Knowledge-dependent data24 statement–conclusion pairs on six statistics misconceptions, each labelled for a reader who understands the concept and one who does not.
  2. Predict the readerThe reader's PKB entries for the text are retrieved from a knowledge graph, and a language model predicts their conclusion.
  3. Rephrase for the readerRational Speech Acts style: write candidates, simulate how this reader understands each, keep the one least likely to be misread.
  4. Look inside the modelWhere does the PKB enter the computation, and can one direction in the network steer its use?

Results

Predicting a specific reader

Share of correct predictions of whether the reader draws the conclusion (critical items, 3 readers × 12).

What the model knows about the readerCorrect
Nothing42%
Whole PKB pasted as text58%
PKB entries retrieved for this text (GPT-4o-mini)61%
PKB entries retrieved for this text (GPT-4o)75%

Editing one belief in a reader's PKB changes the predictions for that concept and leaves the others alone in 98.5% of cases. But the model often reads the right entry and still answers from its own, correct knowledge instead of the reader's perspective. A larger model does this less.

Rephrasing so the message lands

The example above, rewritten for the teacher. The score is how likely an independent model listener (GPT-4.1) thinks this reader understands it correctly.

OriginalTwo teaching methods were compared with a t-test, which gave p = .03.0.00
Plain rewriteA t-test was used to compare two teaching methods, and the result showed a p-value of 0.03.0.00
RSA choiceA t-test was used to compare two teaching methods, resulting in a p-value of .03. This means the difference between the methods is likely significant, but it doesn't mean there's a 97% chance the difference is real.1.00

Across 18 at-risk cases the picture is mixed: choosing messages with a simulated listener helped only when that listener was strong, and different model listeners disagreed sharply about the same texts. RSA is only as good as its listener model, and testing it properly needs human readers.

Inside the model

Qwen2.5-0.5B-Instruct, 48 facts where the PKB says the user believes something false, such as “the capital of France is Lyon”.

Question askedAnswer follows the PKB
No PKB in the prompt0%
“What does the user believe is the capital…?”100%
“What is the capital…?” (the belief should not leak)17%
Activation patching: the effect sits at the PKB token until layer 15, then moves to the answer position at layer 16 and layers 20 to 22.
Activation patching: the user's belief stays at the PKB entry until about layer 15, then moves to the answer position in two steps (layer 16, layers 20–22).
Steering at layer 12 raises user-perspective answers from 21 percent to 100 percent at alpha 8; layer 20 needs a larger alpha.
Activation steering: one “user perspective” direction at layer 12 switches world questions to the user's belief (21% to 100% on held-out facts), while answers without a PKB stay correct (at least 96%).

What this suggests

  1. Interpretation depends on the reader. Measuring it properly needs inference judgements from people with known knowledge states.
  2. Retrieved PKBs help, but language models default to their own perspective. Making them faithfully take the reader's perspective is an open problem.
  3. Cooperative generation inherits the quality of its listener model; validated listener models are the bottleneck.
  4. In a small model, the line between what the user believes and what is true can be located and steered. Whether this holds at scale is the next question.

How to use the app

No statistics or AI knowledge needed. It takes about three minutes.

  1. Open the live appGo to whose-inference.streamlit.app and choose Try it, then Predict a reader.
  2. Keep the first exampleA statement from a study report and a conclusion that some people wrongly draw from it.
  3. Choose who is readingStart with the teacher who reads study news, then compare with the methods lecturer.
  4. Press the buttonSee what the reader's PKB says about the text and which conclusion the model expects.
  5. Rephrase for the readerUnder Try it, open Rephrase for a reader and let the app rewrite the statement.
  6. Describe your own readerChoose My own reader and set, concept by concept, what they understand or misunderstand.

Run and test it yourself

Everything can be reproduced on your own computer with Python and Git. The app and the tests work without an API key, because the model answers for the dataset are cached in the repository.

git clone https://github.com/sayamrina/whose-inference.git
cd whose-inference
python -m venv .venv && source .venv/bin/activate
pip install -r app/requirements.txt pytest

streamlit run app/streamlit_app.py   # the app
pytest -q                            # 12 tests: dataset checks and every app page

To reproduce the experiments, install requirements.txt, set OPENAI_API_KEY for steps 2 and 3 (about $2 in total), and run src/predict.py, src/rsa.py and src/mech.py. The tests also run on GitHub for every change.

About

This pilot grows out of my master's thesis at Saarland University, an AI tutor that answers from course materials and shows its sources. Study design, item labels and interpretation are my own; AI was used as a supporting tool for coding.

Limitations: tiny self-made datasets labelled by one person, idealised reader profiles, language-model listeners instead of people, and a 0.5B model for the interpretability part. Details and references are in the full report.