The Readout Shortcut: Positional Number Copying Dominates Arithmetic CoT Readout in Small Language Models
The paper says small models often copy the last number in a chain-of-thought instead of truly computing the answer.
Intelligence analysis by GPT-5.4 Mini
In three small instruction-tuned models, the paper finds a positional shortcut during arithmetic readout: the model often outputs the number nearest the answer slot, not the result of reasoning. The authors argue this can make chain-of-thought look faithful when it is mostly transport, not computation.
A small model is being tested on math problems. The surprising part is that it often does not solve the math at the end. It just copies the number it saw most recently, like a student who glances at the last line of notes and writes that down.
That means the fancy step-by-step explanation can look helpful while the final answer is really coming from a shortcut. The model may still know some math, but the shortcut can take over when a number is sitting in the wrong place.
It is like following a treasure map but grabbing the coin closest to the door instead of walking the route. The paper says this matters because people may think the model is reasoning well when it is sometimes only copying.
Analysis
What the paper tests
The paper looks at arithmetic chain-of-thought in three 1-3B instruction-tuned language models on GSM8K. Instead of asking whether the models can produce a correct final answer after reasoning, it isolates the answer-readout stage with prefix completion to see what happens right before the answer is emitted.
Main finding
The core result is a positional shortcut. The models often copy the number sitting closest to the answer delimiter, even when the intermediate reasoning is correct or when the trailing number is a distractor. The abstract says that the presence of the gold answer explains 54-92 percentage points of accuracy and accounts for 89-92% of each model's teacher-forcing ceiling. On wrong items, the final answer still matches the last CoT number 95-96% of the time.
What happens when the shortcut is removed
The paper reports that replacing the trailing number with an incorrect one drives accuracy close to zero, even if earlier steps are right. Removing that copyable number restores 5-32 percentage points above that floor, showing that the model can do some arithmetic but may be suppressed when a copy target is present. Qwen and Llama copy novel distractors at high rates, while Gemma appears to gate that behavior more selectively.
Broader implication
The authors argue that step-level faithfulness evaluations can confuse answer transport with actual computation. That is a warning for CoT oversight: a model may look like it is reasoning because it is faithfully repeating the most recent number, not because it has preserved and used the full chain of logic.
Key points
- The paper studies arithmetic chain-of-thought readout in three small instruction-tuned language models.
- A positional shortcut often dominates: the model copies the number nearest the answer slot.
- Correct intermediate reasoning can be overridden by a copyable trailing number.
- The behavior appears on GSM8K and also replicates on GSM-Symbolic.
- The authors warn that faithfulness metrics may confuse copying with real reasoning.
