List your company

Synthetic in-context reasoning on MAD synthetic (test)

55.5Compression ScoreFEM-AFT
30.0236.63543.2549.865Feb 6, 2026Feb 24, 2026Mar 14, 2026Apr 2, 2026Apr 20, 2026May 8, 2026May 27, 2026

Evaluation results

MethodLinks
FEM-AFTbase architecture=AFT, modules=C, L, T, G2026.0255.59.7890.380.190.293.469.9
FEM-SMbase architecture=SMAttn, low-rank convolution (C)=true, LSE mixing (L)=true, linearized temperature (T)=true, outer gate (G)=true2026.0253.143.199.985.999.999.380.2
FEM-GLAbase architecture=GLA, modules=C, L, T, G, normalization=p~+t(i)2026.025319.199.986.399.99974.9
FEM-MAMBA(-G,T,L,C)base architecture=Mamba2026.0252.76.790.489.590.186.369.3
FEM-SM(-G)base architecture=SMAttn, low-rank convolution (C)=true, LSE mixing (L)=true, linearized temperature (T)=true, outer gate (G)=false2026.0252.339.199.985.899.999.479.4
FEM-GLA(-G)base architecture=GLA, modules=C, L, T, normalization=p~+t(i)2026.0251.913.297.186.193.591.472.2
FEM-SM(-C)base architecture=SMAttn, low-rank convolution (C)=false, LSE mixing (L)=true, linearized temperature (T)=true, outer gate (G)=true2026.0251.235.499.985.998.59978.3
FEM-GLA(-G,T)base architecture=GLA, modules=C, L, normalization=p~+t(i)2026.0251.212.492.285.192.489.270.4
FEM-MAMBAbase architecture=Mamba, modules=C, L, T, G, normalization=p~+t(i)2026.0251.116.890.789.792.79773
FEM-SM(-C,G)base architecture=SMAttn, low-rank convolution (C)=false, LSE mixing (L)=true, linearized temperature (T)=true, outer gate (G)=false2026.0250.732.899.985.79897.677.5
FEM-MAMBA(-p~t norm)base architecture=Mamba, modules=C, L, T, G, normalization=s+t(i)2026.0250.512.893.488.986.392.270.7
FEM-AFT(-G,T,L,C)base architecture=AFT2026.0250.59.156331.169.290.152.2
FEM-SM(-G,T)base architecture=SMAttn, low-rank convolution (C)=true, LSE mixing (L)=true, linearized temperature (T)=false, outer gate (G)=false2026.0250.33999.985.499.99878.8
FEM-SM(-C,G,T)base architecture=SMAttn, low-rank convolution (C)=false, LSE mixing (L)=true, linearized temperature (T)=false, outer gate (G)=false2026.0249.526.399.985.797.597.576.1
FEM-GLA(-G,T,L)base architecture=GLA, low-rank convolution (C)=true2026.0247.19.491.783.492.588.568.8
GatedDeltaNet2026.024529.899.980.299.994.374.9
FEM-SM(-G,T,L)base architecture=SMAttn, low-rank convolution (C)=true, LSE mixing (L)=false, linearized temperature (T)=false, outer gate (G)=false2026.024531.499.985.599.996.376.3
Hyena2026.0244.814.49989.498.69373.2
FEM-SM(-G,T,L,C)base architecture=SMAttn;Transformer, low-rank convolution (C)=false, LSE mixing (L)=false, linearized temperature (T)=false, outer gate (G)=false2026.0244.324.599.985.798.595.174.7
Mamba22026.0243.621.196.486.996.793.373
DiffTrans2026.0242.93999.983.797.195.876.4
MambaArchitecture=two-layer, Optimizer=Muon2026.0542.46.575.687.671.39563.1
DeltaNet2026.0242.235.799.952.899.999.971.7
FEM-GLA(-G,T,L,C)base architecture=GLA, all modules (C,L,T,G)=false2026.0240.28.591.381.386.876.864.2
AttnArchitecture=two-layer, Optimizer=Muon2026.0534.226.880.380.786.19567.2
ParallaxArchitecture=two-layer, Optimizer=Muon, Abbreviation=PLX2026.0533.235.695.173.393.798.871.6
LinAttn2026.0233.18.29174.975.693.162.6
GDNArchitecture=two-layer, Optimizer=Muon2026.0532.5119279.290.793.966.5
MesaArchitecture=two-layer, Optimizer=Muon2026.053121.999.886.199.956.865.9