EVL

MuSR

Multistep soft reasoning

benchmarkReasoningAccuracy on multi-step reasoning scenarios756 items

What this measures

Multistep Soft Reasoning — long, natural-language scenarios (murder mysteries, team-allocation puzzles) requiring you to chain many clues together. Tests reasoning over messy free text.

Sample

A 1,000-word whodunit: weigh means, motive, and opportunity across suspects to name the murderer.

Leaderboard · 100 models

Accuracy on multi-step reasoning scenarios — higher is better

1T3Q-qwen2.5-14b-v1.0-e3Open LLM LB · 2026-05

38.7

2T3Q-Qwen2.5-14B-Instruct-1M-e3Open LLM LB · 2026-05

38.7

3calme-3.2-instruct-78bOpen LLM LB · 2026-05

38.5

4calme-3.1-instruct-78bOpen LLM LB · 2026-05

36.5

5CalmeRys-78B-Orpo-v0.1Open LLM LB · 2026-05

36.4

6calme-2.4-rys-78bOpen LLM LB · 2026-05

34.6

7DS-R1-Distill-Q2.5-14B-Harmony_V0.1Open LLM LB · 2026-05

31.9

8miscii-14b-1M-0128Open LLM LB · 2026-05

28.8

9DeepSeek-R1-Distill-Qwen-14BOpen LLM LB · 2026-05

28.7

10Galactic-Qwen-14B-Exp2Open LLM LB · 2026-05

28.5

11Qwen2.5-14B-Vimarckoso-v3-IF-VariantOpen LLM LB · 2026-05

28.4

12o-distil-qwenOpen LLM LB · 2026-05

28.3

13Coma-II-14BOpen LLM LB · 2026-05

28.1

14Gauss-Opus-14B-R999Open LLM LB · 2026-05

27.8

15Lion-Lamarck-v.1.1.0Open LLM LB · 2026-05

27.5

16Lion-Lamarck-v.1.0.9Open LLM LB · 2026-05

27.4

17Sombrero-Opus-14B-Sm1Open LLM LB · 2026-05

27.2

18Viper-Coder-v1.1Open LLM LB · 2026-05

26.2

19Viper-Coder-v1.6-r999Open LLM LB · 2026-05

26.2

20IF-reasoning-experiment-40Open LLM LB · 2026-05

25.9

21Sombrero-Opus-14B-Sm4Open LLM LB · 2026-05

25.8

22Qwen2.5-14B-ReasoningMergeOpen LLM LB · 2026-05

25.6

23Lamarck-14B-v0.6-002-model_stockOpen LLM LB · 2026-05

25.5

24Condor-Opus-14B-ExpOpen LLM LB · 2026-05

25.4

25DeepSeek-R1-Distill-Qwen-25.5B-BrainstormOpen LLM LB · 2026-05

25.2

26Qwenvergence-14B-v9Open LLM LB · 2026-05

25.2

27Qwentinuum-14B-v013Open LLM LB · 2026-05

25.0

28Qwenvergence-14B-v12-Prose-DSOpen LLM LB · 2026-05

24.8

29Hermes-3-Llama-3.1-70B-lorablatedOpen LLM LB · 2026-05

24.7

30Rombos-LLM-V2.5-Qwen-32bOpen LLM LB · 2026-05

24.7

31Sombrero-Opus-14B-Sm2Open LLM LB · 2026-05

24.5

32novablast-previewOpen LLM LB · 2026-05

24.5

33TheBeagle-v2beta-32B-MGSOpen LLM LB · 2026-05

24.5

34Llama-3.1-Tulu-3-70B-SFTOpen LLM LB · 2026-05

24.5

35RombosBeagle-v2beta-MGS-32BOpen LLM LB · 2026-05

24.5

36Phi-4-AbliteratedRPOpen LLM LB · 2026-05

24.4

37QwQ-Buddy-32B-AlphaOpen LLM LB · 2026-05

24.4

38Phi4-Slerp4-14BOpen LLM LB · 2026-05

24.4

39Llama-3.1-Tulu-3-70BOpen LLM LB · 2026-05

24.3

40Oxyge1-33BOpen LLM LB · 2026-05

24.3

41Barcenas-14b-phi-4Open LLM LB · 2026-05

24.2

42orca_mini_v7_72bOpen LLM LB · 2026-05

24.2

43ultiima-32BOpen LLM LB · 2026-05

24.1

44DRT-o1-7BOpen LLM LB · 2026-05

24.1

45phi-4-14bOpen LLM LB · 2026-05

24.0

46NQLSG-Qwen2.5-14B-OriginalFusionOpen LLM LB · 2026-05

24.0

47deepseek-llm-67b-chatOpen LLM LB · 2026-05

23.9

48Chocolatine-2-14B-Instruct-v2.0Open LLM LB · 2026-05

23.9

49Apollo_v2-32BOpen LLM LB · 2026-05

23.8

50Arcee-BlitzOpen LLM LB · 2026-05

23.8

51Microsoft_Phi-4Open LLM LB · 2026-05

23.8

52Phi-4-Megatron-EmpatheticOpen LLM LB · 2026-05

23.7

53RYS-XLargeOpen LLM LB · 2026-05

23.7

54phi-4-300stepsOpen LLM LB · 2026-05

23.7

55phi-4Open LLM LB · 2026-05

23.7

56Mahou-1.5-llama3.1-70BOpen LLM LB · 2026-05

23.7

57phi-4-1stepsOpen LLM LB · 2026-05

23.6

58Chocolatine-2-14B-Instruct-v2.0.1Open LLM LB · 2026-05

23.5

59Hermes-3-Llama-3.1-70BOpen LLM LB · 2026-05

23.4

60Llama-3.1-Tulu-3-70B-DPOOpen LLM LB · 2026-05

23.4

61Llama-3-70B-Synthia-v3.5Open LLM LB · 2026-05

23.4

62Qwenvergence-14B-v0.6-004-model_stockOpen LLM LB · 2026-05

23.3

63Phi4-SuperOpen LLM LB · 2026-05

23.3

64Phi-4-SuperOpen LLM LB · 2026-05

23.3

65phi-4-abliteratedOpen LLM LB · 2026-05

23.2

66jessi-v0.4-falcon3-7b-instructOpen LLM LB · 2026-05

23.2

67Apollo-70BOpen LLM LB · 2026-05

23.1

68Phi-4-Super-o1Open LLM LB · 2026-05

22.9

69Phi-4-Super-1Open LLM LB · 2026-05

22.9

70Phi-4-EmpatheticOpen LLM LB · 2026-05

22.7

71Qwen2.5-32BOpen LLM LB · 2026-05

22.7

72Lamarck-14B-v0.6-model_stockOpen LLM LB · 2026-05

22.7

73Llama3.1-Gutenberg-Doppel-70BOpen LLM LB · 2026-05

22.7

74RYS-XLarge-baseOpen LLM LB · 2026-05

22.4

75Qwen2.5-7B-nerd-uncensored-v1.5Open LLM LB · 2026-05

22.4

76LamarckInfusion-14B-v2Open LLM LB · 2026-05

22.4

77Llama3.1-Flammades-70BOpen LLM LB · 2026-05

22.3

78Qwenvergence-14B-v10Open LLM LB · 2026-05

22.3

79openbuddy-llama3.3-70b-v24.1-131kOpen LLM LB · 2026-05

22.3

80Qwenvergence-14B-v12-ProseOpen LLM LB · 2026-05

22.3

81Calcium-Opus-14B-Elite2Open LLM LB · 2026-05

22.2

82jessi-v0.6-falcon3-7b-instructOpen LLM LB · 2026-05

22.2

83Phi-4-o1Open LLM LB · 2026-05

22.2

84Lamarck-14B-v0.7-FusionOpen LLM LB · 2026-05

22.1

85Calcium-Opus-20B-v1Open LLM LB · 2026-05

22.1

86NQLSG-Qwen2.5-14B-MegaFusion-v8.6Open LLM LB · 2026-05

22.1

87Qwen2.5-14B-Ultimav2Open LLM LB · 2026-05

22.0

88LamarckInfusion-14B-v2-loOpen LLM LB · 2026-05

22.0

89WomboCombo-R1-Coder-14B-PreviewOpen LLM LB · 2026-05

22.0

90Chocolatine-Fusion-14BOpen LLM LB · 2026-05

22.0

91Megatron-Opus-14B-2.1Open LLM LB · 2026-05

21.9

92Equuleus-Opus-14B-ExpOpen LLM LB · 2026-05

21.9

93GWQ-9B-PreviewOpen LLM LB · 2026-05

21.8

94SuperThoughts-CoT-14B-16k-o1-QwQOpen LLM LB · 2026-05

21.8

95Qwenvergence-14B-v3-ProseOpen LLM LB · 2026-05

21.8

96RomboUltima-32BOpen LLM LB · 2026-05

21.7

97Qwen-14B-ProseStock-v4Open LLM LB · 2026-05

21.7

98Magellanic-Opus-14B-ExpOpen LLM LB · 2026-05

21.6

99Falcon3-MoE-2x7B-InsructOpen LLM LB · 2026-05

21.6

100Qwenvergence-14B-v13-Prose-DSOpen LLM LB · 2026-05

21.5

Benchmark health

headroom

Plenty of room at the top — this benchmark still separates frontier models clearly.

34.1

top-10 mean

38.7

best

100

models

What improves this score

Datasets & environments shown to raise it.

No data yet.

Research using this benchmark

all →

K-Quantization and its Impact on Output Performance Distributional Energy-Based Models for Uncertainty-Aware Structured LLM Reasoning

Matched by name in title/abstract.