ActBench says the opposite, and both can be right.

Shared entity: LLMs / Same source domain / Shared topic / Earlier coverage / Downstream implication

Human and LLM everyday reasoning are both pattern matching, not world models.

Both cover LLMs; reported by the same outlet (arxiv.org); overlapping topics (agent, assume, between).

Shared entity: LLMs / Same source domain / Shared topic / Earlier coverage

MerchantBench: the best LLM reached 27.3% of human net assets over a simulated year.

Both cover LLMs; reported by the same outlet (arxiv.org); overlapping topics (agent, best).

MOSAIC chains safe CLI commands into a 96.59% exploit rate

Both cover LLMs; reported by the same outlet (arxiv.org); overlapping topics (agent, attack).

Top AI model scores 72% on social intelligence test

Both cover LLMs; reported by the same outlet (arxiv.org); overlapping topics (best, harness).

DeCNIP strips LLM backdoors by pruning 0.1% of neurons

Both cover LLMs; reported by the same outlet (arxiv.org); overlapping topics (attack, between).

Researchers prune coding-agent context without a second model

Both cover LLMs; reported by the same outlet (arxiv.org); overlapping topics (agent, doesn).

Prompt Injection as Role Confusion: 60% Attack Success.

Both cover LLMs; reported by the same outlet (arxiv.org); overlapping topics (agent, attack).

Sleeper Cell Backdoors in Fine-Tuned Models: Poisoned LoRAs Pass All Benchmarks

Both cover LLMs; reported by the same outlet (arxiv.org); overlapping topics (agent, attack).