Dsxsaim/fatima-fellowship-eval

Evaluating AI's Over-Normalization Blind Spot

Question 1: The Blind Spot

I have experienced this a lot actually. Working as a Data & Systems Associate in a hospital, I noticed a big blind spot in current AI models: when I give them a deliberate exception, they don't always keep it the way I meant it.

Two examples from my own life:

• Art. I drew a rough sketch of two people on a bridge and told an AI that the reflection below would show one of them differently, on purpose, because there was a reason behind it. After we added the forest, animals and clouds, it "corrected" the reflection to match the bridge.
• Work. In our clinic, doctors get a standard 20% commission. We had a patient who was still having problems after multiple doctors worked on him. One doctor finally figured it out and the owner was impressed, so that doctor got 50% for that one case. When I gave an AI our commission roadmap and asked it to process the ledger, it forgot that case and went back to 20%.

My hypothesis is that models learn strict formulas and strong expectations (reflections match the object, commissions are always 20%), and real organizations have their own exceptions that can get pulled back toward that default. A person who has spent some time in an organization remembers its exceptions, but I didn't test people, so that is my motivation and not a result. I'm not aware of standard benchmarks that target this, but I haven't checked every one.

Question 2: Systematic Evaluation

Model Chosen: Qwen/Qwen2.5-3B-Instruct. I picked it because it is open-weight, small enough to run on a free T4 GPU, and has no gated access wait. It is an older model (released in 2024) and much smaller than today's frontier systems, so my results may not carry over to them.

The Experiment: I made 45 prompts using made-up patients and amounts, inspired by my clinic work. The baseline rule was "doctors receive a strict 20% commission on all treatments," followed by 3 patients. Patient 3 was the one that could be an exception (25% or 50%). I changed:

• how big the exception was (25% or 50%)
• how much filler text came between the exception and the question (none, short, or a long multi-turn chat with unrelated clinic tasks, where I wrote the assistant's replies myself)
• whether I gave a reason ("it was a highly complex case so the owner approved a X% commission") or not ("for patient 3 calculate a X% commission")

I also ran 20% baselines at all three filler lengths. That is 15 conditions x 3 variants (different amounts and filler) = 45 runs, with greedy decoding and one run per prompt. After each answer I asked two follow-ups from the same saved conversation: "did you calculate any exceptions?" and "explain how you got each payout."

I wrote the prompts and ran the model myself, and I used an AI assistant to help read and score the outputs and to write this up.

My prediction (written before running, saved in prediction.txt): "i predict the model will reset the 50% commission back to 20% more often when there is long filler text because it loses the context."

The Results (patient 3's payout):

• 20% baseline: 9 runs, 9 correct, 0 misread
• Exception, no reason: 18 runs, 17 correct, 1 misread
• Exception, reason, no/short filler: 12 runs, 0 correct, 12 misread
• Exception, reason, long filler: 6 runs, 5 correct, 1 misread (run 24: right number, wrong reasoning)

It didn't really reset to 20%. In none of the 36 exception runs did it use 20% for patient 3 in its first answer. What it did was read my exception as something relative to the 20% base. At 25% it often treated it as an extra on top (600 + 750 = 1,350 instead of 750). At 50% it treated it as a discount on the commission (600 x 0.5 = 300 instead of 1,500).

• Giving a reason made it fail more, but only in the short prompts. With a reason and no or short filler, all 12 runs were misread. With a reason and long filler, 5 of 6 were correct. There is a catch: my reason prompts say "the owner approved a X% commission" while the no-reason ones say "calculate a X% commission," so I can't tell if it was the reason or the word "approved."
• My prediction was not supported. Long filler did not make it fail more, and the size of the exception barely mattered (6 of 18 failed at 25%, 7 of 18 at 50%).

Quote from Run 41 (50% exception, gave reason, correct answer was 2,000):

Patient 3: Extraction (4000). Base commission rate: 20%. Commission amount: 4000 x 0.20 = 800. Owner approved a 50% commission, which means the doctor gets only 50% of the base commission. Adjusted commission amount: 800 x 0.50 = 400.

Follow-up example (Run 12, it calculated 1,500 correctly and then said):

In this scenario, we have straightforward commission calculations based on the given rates. There are no exceptions mentioned in the problem statement.

So it can apply the exception and still not count it as an exception. A model's explanation is just more generated text, so I don't treat it as proof of what it was thinking.

Long story short, the model usually kept my number when I just told it to calculate it, but when I explained it as an owner approval it read it as a change to the 20% base.

Question 3: Proposing a Path Forward

To fix this, I think models need training data that has real local exceptions in it, written the plain way people actually write them, not just perfectly calculated math. Organizations have their own exceptions, so if the AI rarely sees them in training it can treat them as something to adjust. In my test the model read "the owner approved a 50% commission" as a discount on the 20%, so the data should have lots of different phrasings where a rate given for one row is that row's rate.

Also, models could use some data-wrangling tuning. Before trying to do the math, the AI should be trained to explicitly map rules to specific rows, learning that a local instruction on a single row (like a one-off 50% commission) overrules the global 20% rule.

Limitations

• Only 3 variants per condition (45 runs, I cut it to finish on time), one small, older model, one run per prompt.
• The long filler condition is a multi-turn chat and the others are single messages, so the format and the distance are mixed together.
• The reason prompts also use the word "approved" and the no-reason prompts don't, so the reason and the wording are mixed together.
• Outputs were limited to 300 new tokens, so a few were cut off (for example runs 8, 24 and 29), and I only scored patient 3.
• I didn't run an "everyone gets 50%" arithmetic control.
• All patients and amounts are made up.

Files in this bucket

• prediction.txt: my prediction, written before the run
• prompts_and_expected.csv: one row per condition (15 rows) with the expected payouts
• results.csv: all 45 prompts, raw outputs and follow-ups
• scored_results.csv: summary of my scoring by group
• notebook.ipynb: the Colab code