A Second Instruction Can Change How a Language Model Follows the First

Sent alone, the instruction Keep the first two items made GPT-OSS-20B return the first two items of every list tested. When an instruction to reverse the list or to remove duplicates followed it in the same request, the model often left those two items in place and edited the rest. On [9,2,7,7,2,6,6], the instruction alone returned [9,2], a second call that reversed this result returned [2,9], and the joined request returned [9,2,6,6,2,7,7]. The pattern appeared in a planned screen that compared four short cues, selected on single-operation examples, with English instructions on six pairs of list operations. Each test list gives five different arrays for the two orders, each operation alone and the unchanged list. The screen gave no consistent ranking. Selected cues and English passed both orders on 5/72 and 11/72 lists with Qwen3.8-27B (reasoning disabled) and on 61/72 and 51/72 with GPT-OSS. One operation pair decided the sign for each receiver. A fresh test then fixed the competing arrays in advance. On 24 new lists, all 20 errors under keep equaled a predicted prefix-preserving array, and an explicit discard wording gained 20 lists and lost none. A second panel of 36 new lists, balanced by length, repeated the result. Keep returned the intended array on 8/36 requests that ask for two items first, in each of two dispatches. Keep only returned it on 36/36, and so did the original wording when the second instruction named the resulting list. Prefix-preserving arrays appeared on 5/12 lists of five or six items and on 22/24 longer lists. The evidence covers arrays returned by one hosted model configuration. For evaluation, it shows that an instruction can pass every singleton check and still yield a predictable alternative when another instruction follows. A short cue appeared to fix this until one added English word did the same.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-19
DOI
https://doi.org/10.5281/zenodo.24366770
Primary Topic
Natural Language Processing Techniques
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

A Second Instruction Can Change How a Language Model Follows the First

Aryan Rajesh Ganjoo
Zenodo (CERN European Organization for Nuclear Research)
Natural Language Processing Techniques
preprint

A Second Instruction Can Change How a Language Model Follows the First

Aryan Rajesh Ganjoo
preprint en

Abstract

Sent alone, the instruction Keep the first two items made GPT-OSS-20B return the first two items of every list tested. When an instruction to reverse the list or to remove duplicates followed it in the same request, the model often left those two items in place and edited the rest. On [9,2,7,7,2,6,6], the instruction alone returned [9,2], a second call that reversed this result returned [2,9], and the joined request returned [9,2,6,6,2,7,7]. The pattern appeared in a planned screen that compared four short cues, selected on single-operation examples, with English instructions on six pairs of list operations. Each test list gives five different arrays for the two orders, each operation alone and the unchanged list. The screen gave no consistent ranking. Selected cues and English passed both orders on 5/72 and 11/72 lists with Qwen3.8-27B (reasoning disabled) and on 61/72 and 51/72 with GPT-OSS. One operation pair decided the sign for each receiver. A fresh test then fixed the competing arrays in advance. On 24 new lists, all 20 errors under keep equaled a predicted prefix-preserving array, and an explicit discard wording gained 20 lists and lost none. A second panel of 36 new lists, balanced by length, repeated the result. Keep returned the intended array on 8/36 requests that ask for two items first, in each of two dispatches. Keep only returned it on 36/36, and so did the original wording when the second instruction named the resulting list. Prefix-preserving arrays appeared on 5/12 lists of five or six items and on 22/24 longer lists. The evidence covers arrays returned by one hosted model configuration. For evaluation, it shows that an instruction can pass every singleton check and still yield a predictable alternative when another instruction follows. A short cue appeared to fix this until one added English word did the same.

Zenodo (CERN European Organization for Nuclear Research)
Natural Language Processing Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon