E01¶


Research	R0054 — Prompt Claims v2
Run	2026-03-31
Claim	C002
Source	SRC03
Evidence	SRC03-E01
Type	Reported

LLMs have systematic weaknesses in processing negation, explaining why positive-only instructions are insufficient.

URL: https://gadlet.com/posts/negative-prompting/

Extract¶

Key findings from the search results:

Research from KAIST found that larger AI models actually perform worse on negated prompts
GPT-3, GPT-Neo, and other models consistently struggle with negation across multiple benchmarks
LLMs "may struggle to distinguish between facts and their negations, misunderstand the semantic impact of negative particles, and fail to generalize negation handling robustly, even with instruction tuning"
Claude documentation suggests "prefer general instructions over prescriptive steps" — but this applies to simple tasks, not complex multi-step analytical processes
The implication: positive instructions are the primary vehicle, but negative constraints are needed to catch failure modes that positive instructions cannot prevent

Relevance to Hypotheses¶

Hypothesis	Relationship	Strength
H1	Supports	Explains the mechanism: LLMs have documented weaknesses that require both instruction types
H2	Supports	The Claude documentation note provides weak support for the idea that positive-only might sometimes work
H3	Contradicts	Directly contradicts — LLMs have systematic negation weaknesses that cannot be overcome by positive instructions alone

Context¶

The KAIST finding about larger models performing worse on negation is counterintuitive and particularly relevant — it suggests that as models become more capable, they do not automatically become better at following negative constraints, which reinforces the need for careful prompt design combining both instruction types.