In-Context Examples Suppress Scientific Knowledge Recall in LLMs

Chaemin Jang, Woojin Park, Hyeok Yun, Dongman Lee, Jihee Kim

KAIST · Shanghai Jiao Tong University

Models recall the right formula on their own. Show them a few examples of it, and they stop deriving. They start pattern-fitting instead.

Read the paper ↗

At a glance

60Tasks

5Domains

6,000Trials

4Models

Main result

The strategy shift

Few-shot examples pull every domain toward pattern-fitting, whether or not it helps.

Formula recall by domain, zero-shot versus ten-shot — Same-size shift, opposite outcomes: Economics −15.7 accuracy pp, Geoscience +13.3 accuracy pp. Shown at 10-shot; the direction holds from k=1 through k=20.

Validation

A causal check

Is this real, or just noisy chain-of-thought labels? A paired log-probability test on three open models says real.

Paired log-probability shift across three models — Probability mass moves toward the example answer and away from the knowledge answer, on the same input.

Example

How a task looks

Example task from the economics domain, shown zero-shot and few-shot. What is the consumer surplus at the Cournot-Nash equilibrium?

Zero-shot: correct

Step 1: Identify Cournot model
  “Two firms, linear demand P = a − bQ”
Step 2: Derive equilibrium
  q* = (a−c)/3b = 20, Q* = 40
Step 3: Compute CS from the formula
  CS = ½(a−P*)Q* = 2(a−c)²/9b
     = 1600 ✓

Derived from the Cournot equilibrium

Few-shot: wrong

Step 1: Fit pattern from examples
  “CS scales roughly with market size”
Step 2: Apply fitted rule
  CS ≈ 1522 ... or 1688
  (varies by trial) ✕

Fitted heuristic, inconsistent across trials

Interpretation

Why recall is so easy to suppress

Swap “price” for a meaningless code word. Keep the same math. Zero-shot accuracy: 55–88% → 3–14%. The model isn't deriving the answer. It's recognizing keywords. That's why a few examples are enough to knock it off course.

Takeaways

What we find

Uniform direction. Every domain loses 6–20pp of formula-based reasoning, even where accuracy barely moves.
Strong priors lose most. Economics (88.3% prior) drops most; Geoscience (55.3% prior) looks best: a weak prior is just easy to beat, not improved on.
Recall is a keyword trigger. Swap the vocabulary, keep the math. Accuracy collapses to single digits.

Reference

Citation

@inproceedings{jang2026incontext,
  title     = {In-Context Examples Suppress Scientific Knowledge Recall in LLMs},
  author    = {Jang, Chaemin and Park, Woojin and Yun, Hyeok and Lee, Dongman and Kim, Jihee},
  booktitle = {Conference on Language Modeling (COLM)},
  year      = {2026}
}