Long-context language modeling on HELMET benchmark leaderboard · SOTA2 Research

Long-context language modeling on HELMET

247Summarization ScoreQwen2.5-14B + RLVR

Evaluation Results

Method	Links
Qwen2.5-14B + RLVR2025.09		247	—	420	—	46.7	25.5
Qwen2.5-14B + SPELL2025.09		230	—	392	—	52.3	31.6
Qwen2.5-32B + SPELL2025.09		212	—	564	—	61.4	35
Qwen2.5-32B + RLVR2025.09		210	—	167	—	52.7	25.8
Qwen2.5-32B2025.09		162	—	428	—	59	28.1
Qwen2.5-7B + RLVR2025.09		143	—	26	—	49.3	18.5
Qwen2.5-7B + SPELL2025.09		137	—	104	—	54.2	24.2
Qwen2.5-7B2025.09		41	—	35	—	50	11.5
Qwen2.5-14B2025.09		37	—	15	—	42.4	15.2
CoPE2026.02		32.78	21.02	85.5	99	67.83	61.23
CoPE2026.02		32.37	19.06	—	34	—	28.48
CoPE2026.02		30.88	15.07	86.4	97.88	63.17	58.68
CoPE2026.02		30.81	16.89	83.7	98.88	67.72	59.6
CoPE2026.02		29.76	13.1	79.4	99.63	68.67	58.11
RoPE2026.02		29.18	6.46	74.6	99.75	68.38	55.74
RoPE2026.02		28.46	8.39	80.2	99.13	67.44	56.86
CoPE2026.02		27.89	18.23	84.7	76	56.78	52.72
HardClip2026.02		25.7	9.28	77	98.5	67.5	55.6
HardClip2026.02		25.68	7.44	73.1	99.75	68.06	54.81
RoPE2026.02		21.76	8.52	83.4	98.13	66.67	55.7
HardClip2026.02		18.55	10.16	79.8	98.5	66.11	54.62
RoPE2026.02		11.1	7.67	85.5	97.63	62.78	52.94
HardClip2026.02		9.33	9.31	77.3	82.13	62.05	48.02
RoPE2026.02		9.06	7.93	—	26.13	—	14.37
HardClip2026.02		8.6	9.24	—	36.86	—	18.23
HardClip2026.02		6.93	10.31	82.2	94.38	59.72	50.71
RoPE2026.02		6.31	8.21	82.1	71.38	53.44	44.29