Arctic ice is melting faster than computer models of climate calculate, according to a group of US researchers.
Since 1979, the Arctic has been losing summer ice at about 9% per decade, but models on average produce a melting rate less than half that figure.
The scientists suggest forecasts from the Intergovernmental Panel on Climate Change (IPCC) may be too cautious.
The latest observations indicate that Arctic summers could be ice-free by the middle of the century.
"Somewhere in the second half of the century, it would happen," said Ted Scambos of the National Snow and Ice Data Center (NSIDC) at the University of Colorado.
"Some computer models show periods of great sensitivity where the Arctic ice system collapses suddenly, and that trend may occur a bit earlier; that's the best guess, but exactly when it's hard to say," he told the BBC News website.
Dr Scambos co-authored the latest study, published in the journal Geophysical Research Letters, with other scientists from NSIDC and from the National Center for Atmospheric Research (NCAR), also in Boulder, Colorado.
They also calculate that about half, if not more, of the warming observed since 1979 originates in humanity's emissions of greenhouse gases.
There are measurements dating back about a century on the extent of Arctic ice, but satellite observations from 1979 onwards are generally thought to provide the most accurate dataset.
The new research involved analysing two periods, 1953-2006 and 1979-2006.
Records show a shrinkage over the longer period of 7.8% per decade. When only the more recent period is analysed, the rate rises to 9.1% per decade.
For comparison, the researchers looked at a collection of 18 computer models used by the IPCC and other institutions for making projections of future climates.
Models are always verified against real-world data from the recent past to see how well their output mimics reality.
The collection scrutinised here calculated an average decline of only 2.5% per decade for 1953-2006, and 4.3% per decade since 1979 - both well short of the real-world observations.
"There are lessons here for the climate modelling community," acknowledged NCAR's Marika Holland.
"The rate of ice loss, and the location of ice loss - these are things that the models need to improve, and there are physical processes such as the release of methane from melting permafrost that the models don't include."
This is the third time in the last few months that studies have suggested the IPCC's latest major global climate analysis, the Fourth Assessment Report, is too conservative.
In December, a German team published research suggesting that sea levels could rise by 50-140cm over the coming century. The IPCC, in February, gave a range of 28-43cm.
Then, also in February, came an analysis showing that temperature and sea level rises had been rising at or above the top end of IPCC projections since the panel's previous major assessment in 2001.
This is the opposite view from that put forward by many "climate sceptics", who view the whole field of computer modelling as deeply flawed, and the IPCC as an alarmist organisation.
Because of the way it works, the IPCC is bound to be conservative, as it assesses in considerable depth research already in the public domain. This process takes time, and means the panel's conclusions will always lag behind the latest publications.
Nevertheless, Marika Holland believes there is agreement on the major questions regarding Arctic ice; it is receding, and greenhouse gases of human origin are largely responsible.
"The fact that all models show ice loss over the observed period and all project large ice losses into the future is a very strong message," she said.