I think it depends on how you define energy efficiency and the workload you are testing with - ARM is designed for mobile and embedded, but often does not have completely comparable features.  The newer quad-core mobile Atom is more comparable to ARM SoCs than other Intel micro-architecture solutions.  So, to really compare well, I would look to compare mobile Atom to ARM SoC (e.g. TI OMAP or NVIDIA Tegra4) and make sure you select SoCs of similar capability, then compare in terms of throughput and hours of operation on a battery - not just idle time.  For example, with a workload to decode video, drive graphics, and process network packets - or something that is a representative workload for your actual application.  I'm not sure a blanket statement that ARM SoCs are more power efficient than mobile Atom is fair (not sure it isn't either) - the key is to test with real workload - it's a bit like comparing a economy car to a van in some cases - if you include hauling or towing 2 tons of payload, the economy car does not look good (may not work at all) and the previously inefficient van starts to sound pretty good.  Of course you can compare specs and TDP, but I think workload and real application testing with battery lifetime is what really matters.  For that matter, off-load to an FPGA or ASIC co-processor is often the most efficient (e.g. for video decode and computer vision transforms).

Very briefly, ARM processors are RISC based. Which means that their architecture incorporates less clocks per Instruction. Whereas the CISC processors such as Intel, use a higher clock rate to speed thigs up.

  Greater clocks per instruction = Greater power consumption