![]()
When AI companies release new models, they often tout their scores on standardized tests called benchmarks. But traditional benchmark methodologies have recently faced questions about reliability from AI researchers as well as the companies that make the models.
kAm~?6 C62D@? 7@C E9@D6 5@F3ED :D E92E 6I:DE:?8 36?49>2C<D 42? D@>6E:>6D 4@?462= 3:8 5:776C6?46D 36EH66? >@56=D] u@C 6I2>A=6[ @? @?6 A@AF=2C 4@5:?8 36?49>2C<[ s66A$(t[ >@56=D 7C@> p?E9C@A:4[ ~A6?px[ 2?5 s66A$66< 2== D4@C6 H:E9:? 2 A@:?E @7 6249 @E96C] qFE H96? E96 D2>6 >@56=D H6C6 E6DE65 @? 2 ?6H 36?49>2C<[ E96 82A 36EH66? E96> 8C6H E@ ?62C=J bd A@:?ED[ D6A2C2E:?8 >@56=D E92E 925 =@@<65 6G6?=J >2E4965]k^Am
kAm%96 ?6H 36?49>2C<[ tIA6CErC@H5\$(t[ E6DE65 D6G6? =625:?8 px >@56=D @? ?6H=J HC:EE6? D@7EH2C6 6?8:?66C:?8 E2D<D 56D:8?65 E@ C67=64E C62=:DE:4[ >F=E:DE6A H@C<] k2 9C67lQ9EEADi^^HHH]E@AE2=]4@>^6IA6CE4C@H5^36?49>2C<D^6IA6CE4C@H5\DH6\36?49>2C<Qm%96 C6DF=ED D9@H65k^2m E92E p?E9C@A:4’D r=2F56 ~AFD d]d 925 E96 9:896DE 2G6C286 D4@C6[ ef]bT[ 7@==@H65 4=@D6=J 3J ~A6?px’D v!%\e pDEC2 2E eb]fT[ 2 82A H:E9:? E96 >2C8:? @7 6CC@C] r=2F56 u23=6 d]` D4@C65 ce]dT[ v6>:?: b]g u=2D9 D4@C65 c_T[ 2?5 "H6? b]g |2I D4@C65 bb]dT] s66A$66<’D >@56=[ H9:49 D4@C65 H:E9:? 2 A@:?E @7 r=2F56 ~AFD d]d 2?5 v!%\e pDEC2 @? s66A$(t[ 7:?:D965 D:IE9 2E ba]dT] $A246)$x’D vC@< c]e 42>6 :? D6G6?E9 2E ba]aT]k^Am
k7:8FC6mk:>8 DC4lQ9EEADi^^:>286D]DE24<6C]4@>^H^`ag_^:>286D^DE@C:6D^a_ae^``^`a^tIA6CErC@H50'a]A?8Q 2=ElQp 52E2 32C 492CE C6G62=:?8 tIA6CErC@H5\$(t C6DF=ED]Q DEJ=6lQH:5E9i `__TQ ^mk7:842AE:@?m%@AE2=k^7:842AE:@?mk^7:8FC6m
kAm%96 36?49>2C<:?8 E2D<D H6C6 49@D6? 7@C 2C62D H96C6 >@56=D E6?5 E@ DECF88=6[ D@ D4@C6D 2C6 =@H6C E92? E96J H@F=5 36 @? 2 3C@256C C2?86 @7 H@C<] %96 E2D<D H6C6 56D:8?65 E@ C6AC6D6?E C62= D@7EH2C6 6?8:?66C:?8 2DD:8?>6?ED 2?5 H6C6 ?@E 2CE:7:4:2==J 4@>A=:42E65 AC@3=6>D] %96 C6DF=E H2D 2 >2A @7 6249 >@56=’D =:>:ED[ H9:49 42? 96=A E62>D 564:56 H9:49 >@56=D 2C6 36DE 7@C H9:49 <:?5D @7 E2D<D]k^Am
kAmtIA6CErC@H5\$(t :D 2 4@==23@C2E:@? 36EH66? tIA6CErC@H5[ 2 A=2E7@C> 7@C EC2:?:?8 2?5 :>AC@G:?8 px >@56=D 2?5 DJDE6>D 56G6=@A65 3J px D6CG:46D 4@>A2?J k2 9C67lQ9EEADi^^HHH]E@AE2=]2:^Qm%@AE2= pxk^2m[Â 2?5 w&s[ 2 * r@>3:?2E@C\324<65 4@>A2?J E92E 3F:=5D C6:?7@C46>6?E =62C?:?8 6?G:C@?>6?ED]k^Am
kAmkDEC@?8m|6E9@5@=@8Ji w@H E96 %6DE (@C<Dk^DEC@?8m k^Am
kAmu@C tIA6CErC@H5\$(t[ D@7EH2C6 6?8:?66CD 7C@> %@AE2=’D E2=6?E ?6EH@C< 4C62E65 E96 36?49>2C< E2D<D 32D65 @? gf @A6?\D@FC46 AC@;64ED :? `` AC@8C2>>:?8 =2?8F286D[ 2=@?8 H:E9 E96 4C:E6C:2 7@C 8C25:?8 6249 @?6] %96 6?8:?66CD HC@E6 6249 E2D< E96 H2J C62= H@C< EJA:42==J 2CC:G6D[ 2D 2? :DDF6 C6A@CE 56D4C:3:?8 H92E ?665D E@ 36 7:I65 @C 25565] t249 C6A@CE H2D A2:C65 H:E9 2 7F== 4@AJ @7 6249 AC@;64E’D 4@56 2?5 2 D2?53@I H96C6 E96 >@56=D 4@F=5 C625 7:=6D[ CF? 4@>>2?5D[ 2?5 65:E 4@56] %96 >@56=D 925 E@ :?G6DE:82E6 E96 4@5632D6[ 564:56 H92E E@ 492?86[ >2<6 25;FDE>6?ED[ 2?5 4964< E96:C H@C< H:E9:? 2 D6E ?F>36C @7 24E:@?D A6C 2EE6>AE] w&s AC@G:565 E96 :?7C2DECF4EFC6 E@ 9@DE E2D< 6?G:C@?>6?ED 2?5 CF? 6G2=F2E:@?D]k^Am
kAmtIA6CErC@H5 C2? 2== D6G6? >@56=D H:E9 E96 D2>6 E@@=D 2?5 E96 D2>6 DE6A =:>:E] t249 >@56= 2EE6>AE65 6G6CJ E2D< 7:G6 E:>6D H:E9 :ED 25G2?465 C62D@?:?8 D6EE:?8D 6?23=65] %96 2EE6>AED H6C6 D4@C65 282:?DE 2 CF3C:4 HC:EE6? 3J E96 6?8:?66C H9@ 3F:=E 6249 E2D<] |@56=D E92E 8@E A2CEH2J E@ 2 4@CC64E D@=FE:@? C646:G65 A2CE:2= 4C65:E] p7E6C E96 E6DED H6C6 CF?[ >6>36CD @7 %@AE2= px’D C6D62C49 E62> C6G:6H65 E96 >6E9@5@=@8J 2?5 C6DF=ED]k^Am
kAmk2 9C67lQ9EEADi^^HHH]E@AE2=]4@>^C6D62C49\2?5\2?2=JD:D^?6H\2:\36?49>2C<Qmk6>m%9:D DE@CJk^6>mk^2mk6>m H2D AC@5F465 3Jk^6>m k2 9C67lQ9EEADi^^HHH]E@AE2=]4@>Qmk6>m%@AE2=k^6>mk^2mk6>m 2?5 C6G:6H65 2?5 5:DEC:3FE65 3Jk^6>m k2 9C67lQ9EEADi^^9F3D]=2^"_b<=8$#_Q C6=lQ?@7@==@HQmk6>m$E24<6Ck^6>mk^2mk6>m]k^6>mk^Am

(0) comments
Welcome to the discussion.
Log In
Keep it Clean. Please avoid obscene, vulgar, lewd, racist or sexually-oriented language.
PLEASE TURN OFF YOUR CAPS LOCK.
Don't Threaten. Threats of harming another person will not be tolerated.
Be Truthful. Don't knowingly lie about anyone or anything.
Be Nice. No racism, sexism or any sort of -ism that is degrading to another person.
Be Proactive. Use the 'Report' link on each comment to let us know of abusive posts.
Share with Us. We'd love to hear eyewitness accounts, the history behind an article.