As X, Do Y: How Persona and Task Combine in Instruction-Tuned LLMs
1 distinct publishers across 2 articles (some outlets filed more than once).
Ownership mix: Other: 2
1 publishers · 2 articles · switch to 1-minute for disagreement and framing.
What happened
Role prompts of the form As X, do Y admit a clean linear decomposition at one specific site in the residual stream: the prompt-to-answer transition -- the last prompt token together with the first two generated tokens…
Why the coverage differs
AI-assisted comparison · labeled · generated Sep 9, 2026, 6:33 AM · not a verdict
What happened: As X, Do Y: How Persona and Task Combine in Instruction-Tuned LLMs
Where coverage diverges: Center: 2 (arXiv cs.AI, arXiv cs.AI).
Comparison summary
AI-assisted · Cerebras / Llama · Sep 9, 2026, 6:33 AM · inspect sources below rather than trusting this alone
What happened: As X, Do Y: How Persona and Task Combine in Instruction-Tuned LLMs
Where coverage diverges: Center: 2 (arXiv cs.AI, arXiv cs.AI).
What's missing: AI bias-comparison is temporarily offline. Configure Cerebras in admin to enable rich comparison summaries.
Report timeline
Oldest → newest among clustered members. Gaps may mean delayed pickup, not silence.
-
May 24, 2026, 9:00 PMarXiv cs.AI · CenterAs X, Do Y: How Persona and Task Combine in Instruction-Tuned LLMs
-
May 24, 2026, 9:00 PMarXiv cs.AI · CenterPoisonForge: Task-Level Targeted Poisoning Benchmark for Instruction-Tuned LLMs
Coverage by perspective
Perspective labels are external consensus ratings (AllSides / Ad Fontes / MBFC-style), not WeSearch truth scores. Center is not automatically more accurate.
Center · 2
Headline framing
Vocabulary fingerprints · not a political endorsement
AI framing analysis temporarily offline. Configure Cerebras in admin to enable framing comparison.
Bias/ownership: published methodology on source profiles · AI text always labeled · no reader paywall · no engagement ranking of news · transparency · contribute Ws · home