Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

Long-context language modeling on RULER 16K context

83Accuracy (RULER 16K)

SINK + TAIL + TOP-K + ϕ

-0.09621.47743.0564.623Apr 7, 2026
Updated 10d ago

Evaluation Results

MethodLinks
2026.04
83
2026.04
82.9
2026.04
82.8
2026.04
82.8
2026.04
82.8
2026.04
82.7
2026.04
82.5
2026.04
82.5
2026.04
82.4
2026.04
82.4
2026.04
81.8
2026.04
77.5
2026.04
77.2
2026.04
75.5
2026.04
72.1
2026.04
66.5
2026.04
65.5
2026.04
65.3
2026.04
64.7
2026.04
64.6
2026.04
64.3
2026.04
64.1
2026.04
63.7
2026.04
63.6
2026.04
63
2026.04
62.7
2026.04
60.6
2026.04
59.2
2026.04
58.3
2026.04
55.5
2026.04
54.5
2026.04
53.6
2026.04
47.5
2026.04
46.8
2026.04
46.6
2026.04
44
2026.04
43.3
2026.04
42.9
2026.04
41.2
2026.04
40.8
2026.04
40.6
2026.04
40.4
2026.04
39.9
2026.04
37.9
2026.04
37
2026.04
36.8
2026.04
36.6
2026.04
35.5
2026.04
33.8
2026.04
33.5
2026.04
33.4
2026.04
31.1
2026.04
31
2026.04
30.5
2026.04
29.9
2026.04
29.4
2026.04
26
2026.04
24.3
2026.04
24
2026.04
22.1
2026.04
16.7
2026.04
9.1
2026.04
7.9
2026.04
6.8
2026.04
6.5
2026.04
6.4
2026.04
5.3
2026.04
4.9
2026.04
4.4
2026.04
4.2
2026.04
3.6
2026.04
3.1