nucleation

LIT-942: Do different prompting methods yield a common task representation in language models?

2 min read

Properties5
StatusActive — read 2026-10-14 (NOTE-803) in full; worth reading as the direct test of whether a task given by demonstrations and the same task given by an instruction share one internal representation. Under function vectors they do not: mostly different heads, vectors that add, and an asymmetric cost to swapping them. Read it knowing that it tests prompt forms, not paraphrase or stance, on simple lexical tasks in small models, and that the authors say they cannot prove the negative. Filed THEORY-264; bears on QUESTION-024 without answering it.
Version2
Tagsrepresentation-learning, pragmatics, anthology-candidate
Cited as “paper” byNOTE-803 — Davidson et al. — Do different prompting methods yield a common task representation?
Cited as “source” byTHEORY-264 — A task given by demonstrations and the same task given by an instruction are carried in a language model by mostly different attention heads, so function vectors find no common task representation across the two presentations

Graph View

Guy Davidson, Todd M. Gureckis, Brenden M. Lake and Adina Williams (2025), arXiv:2505.12075 (NeurIPS 2025).

Key takeaways

From the reading, NOTE-803.

  • Instructions yield function vectors too. Built from the 20 heads whose patching most helps, they raise zero-shot accuracy from below 20% to above 50% in the best models, and work far better in post-trained models.
  • Mostly different heads. Of the top 20, demonstrations and instructions share 7 in the Llama-3.2-3B models, 4 in the Llama-3.1-8B models, and 2 to 7 elsewhere. Instruction representations are more diffuse.
  • The two vectors add. Adding both beats either alone, without interference; part of the gain is the larger intervention.
  • Swapping costs, asymmetrically. Instruction heads read with demonstration activations lose less than the reverse, which the authors take as preliminary evidence that demonstration inference is more basic.
  • Post-training builds the instruction route. It appears at SFT and DPO in OLMo-2, and post-trained instruction vectors steer base models; demonstration vectors are unchanged by post-training.
  • The limit. Simple lexical tasks, small models, one extraction method, and in the authors’ words “we cannot conclusively prove the negative”.

Standing in the record

Filed 2026-10-12 for QUESTION-024 and read 2026-10-14 in NOTE-803. It is the record’s direct evidence that one task given in two prompt forms is not carried by one internal representation, held as THEORY-264 with LIT-946. It does not test what QUESTION-024 asks, a stance induced from examples surviving paraphrase. LIT-946 points the same way by another method, not the opposite way as the triage of 2026-10-14 had it. Its prompt-design advice belongs to the anthology; the anthology-candidate flag stands.

Backlinks