Benchmark Almanac

CaLM

GPT-4 (snapshot unspecified) 56.8% tops its leaderboard Claude 2 45.3% leads the models we track
Not in model cards

Models answer causal discovery, intervention, and counterfactual reasoning tasks in English and Chinese. Evaluation compares answers with reference labels to measure accuracy and robustness across causal targets and prompting methods.

The full suite withholds some ground-truth labels. Open-ended explanation tasks also use manual understandability judgments. CaLM Lite is a smaller configuration with public labels.

Released
May 2024
Kind
Language reasoning
Level
University
Scored
Checked automatically
Answers
Short answer
Tasks
Public, plus a private test set
Measures
Capability
Subject
Mathematics
Language
Chinese, English
In and out
Text → Text
Interaction
Single turn
Refresh
One fixed set
Run by
A host
Strongest model tested
Claude 2 #192 on the capability index
Last model added
September 2023
Scores refreshed
November 18th, 2026

In the authors’ words

Causal reasoning is viewed as crucial for achieving human-level machine intelligence. Recent advances in language models have expanded the horizons of artificial intelligence across various domains, sparking inquiries into their potential for causal reasoning. In this work, we introduce Causal evaluation of Language Models (CaLM), which, to the best of our knowledge, is the first comprehensive benchmark for evaluating the causal reasoning capabilities of language models. First, we propose the CaLM framework, which establishes a foundational taxonomy consisting of four modules: causal target (i.e., what to evaluate), adaptation (i.e., how to obtain the results), metric (i.e., how to measure the results), and error (i.e., how to analyze the bad results). This taxonomy defines a broad evaluation design space while systematically selecting criteria and priorities. Second, we compose the CaLM dataset, comprising 126,334 data samples, to provide curated sets of causal targets, adaptations, metrics, and errors, offering extensive coverage for diverse research pursuits. Third, we conduct an extensive evaluation of 28 leading language models on a core set of 92 causal targets, 9 adaptations, 7 metrics, and 12 error types. Fourth, we perform detailed analyses of the evaluation results across various dimensions (e.g., adaptation, scale). Fifth, we present 50 high-level empirical findings across 9 dimensions (e.g., model), providing valuable guidance for future language model development. Finally, we develop a multifaceted platform, including a website, leaderboards, datasets, and toolkits, to support scalable and adaptable assessments. We envision CaLM as an ever-evolving benchmark for the community, systematically updated with new causal targets, adaptations, models, metrics, and error types to reflect ongoing research advancements. Project website is at https://opencausalab.github.io/CaLM.

Abstract of Causal Evaluation of Language Models (arXiv:2405.00622, CC0)
28
Entries on its own leaderboard, 8 models we track

GPT-4 (snapshot unspecified) tops the leaderboard with 56.8% Claude 2 leads with 45.3% among models released by September 2023

Best score per entry from the benchmark's own leaderboard, read on November 18th, 2026. Where a model ran with several settings or harnesses, we keep its best. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. The best model we track is Claude 2, at 45.3%. This view shows only models we track, each at its best.

The newest model with a score here came out in September 2023, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

  1. 1 GPT-4 (snapshot unspecified) 56.8%
  2. 2 GPT-3.5-Turbo (snapshot unspecified) 48.2%
  3. 3 text-davinci-003 47.3%
  4. 4 Claude 2 45.3%
  5. 5 text-davinci-002 42.3%
  6. 6 internlm-chat-20b 36.4%
  7. 7 Qwen-14B 34.5%
  8. 8 InternLM-chat (7B) 34.2%
  9. 9 Baichuan2-chat (7B) 33.9%
  10. 10 text-davinci-001 33.3%
  11. 11 Qwen-7B 32.9%
  12. 12 text-curie-001 31.7%
  13. 13 Baichuan1-7B 30.8%
  14. 14 Baichuan2-chat (13B) 30.1%
  15. 15 davinci (175B) 29.9%
  16. 16 Vicuna-v1.3 (33B) 29.8%
  17. 17 Wizardcoder (15B) 28.9%
  18. 18 Llama2-chat (70B) 28.6%
  19. 19 Llama 2-13B 28.6%
  20. 20 curie (6.7B) 28.4%
  21. 21 Koala (13B) 28.3%
  22. 22 text-babbage-001 28.3%
  23. 23 Llama 2-70B 28.1%
  24. 24 Llama 2-7B 27.6%
  25. 25 text-ada-001 25.3%
  26. 26 babbage (1.3B) 23.4%
  27. 27 ada (0.35B) 22.9%
  28. 28 Baichuan1-chat (13B) 21.7%

Source: Official CaLM leaderboard

#EntryScoreSettingDateSource
1 GPT-4 (snapshot unspecified) 56.8% Official leaderboard
2 GPT-3.5-Turbo (snapshot unspecified) 48.2% Official leaderboard
3 text-davinci-003 47.3% Official leaderboard
4 Claude 2 45.3% Jul 2023 Official leaderboard
5 text-davinci-002 42.3% Official leaderboard
6 internlm-chat-20b 36.4% Sep 2023 Official leaderboard
7 Qwen-14B 34.5% Sep 2023 Official leaderboard
8 InternLM-chat (7B) 34.2% Official leaderboard
9 Baichuan2-chat (7B) 33.9% Official leaderboard
10 text-davinci-001 33.3% Official leaderboard
11 Qwen-7B 32.9% Sep 2023 Official leaderboard
12 text-curie-001 31.7% Official leaderboard
13 Baichuan1-7B 30.8% Jun 2023 Official leaderboard
14 Baichuan2-chat (13B) 30.1% Official leaderboard
15 davinci (175B) 29.9% Official leaderboard
16 Vicuna-v1.3 (33B) 29.8% Official leaderboard
17 Wizardcoder (15B) 28.9% Official leaderboard
18 Llama2-chat (70B) 28.6% Official leaderboard
19 Llama 2-13B 28.6% Jul 2023 Official leaderboard
20 curie (6.7B) 28.4% Official leaderboard
21 Koala (13B) 28.3% Official leaderboard
22 text-babbage-001 28.3% Official leaderboard
23 Llama 2-70B 28.1% Jul 2023 Official leaderboard
24 Llama 2-7B 27.6% Jul 2023 Official leaderboard
25 text-ada-001 25.3% Official leaderboard
26 babbage (1.3B) 23.4% Official leaderboard
27 ada (0.35B) 22.9% Official leaderboard
28 Baichuan1-chat (13B) 21.7% Official leaderboard
#ModelScoreSettingDateSource
1 Claude 2 45.3% Jul 2023 Official leaderboard
2 internlm-chat-20b 36.4% Sep 2023 Official leaderboard
3 Qwen-14B 34.5% Sep 2023 Official leaderboard
4 Qwen-7B 32.9% Sep 2023 Official leaderboard
5 Baichuan1-7B 30.8% Jun 2023 Official leaderboard
6 Llama 2-13B 28.6% Jul 2023 Official leaderboard
7 Llama 2-70B 28.1% Jul 2023 Official leaderboard
8 Llama 2-7B 27.6% Jul 2023 Official leaderboard

Rankings by part

We read 2 parts from the benchmark’s own leaderboard: English and Chinese. Each ranks models on one part of the test. Claude 2 leads both. Arrows show how far a model moved from its overall rank among the models we track.

English

8 models we track have a score here.

  1. 1 Claude 2 49.4%
  2. 2 internlm-chat-20b 38%
  3. 3 Qwen-14B 34.8%
  4. 4 Qwen-7B 32.4%
  5. 5 Baichuan1-7B 30.4%
  6. 6 Llama 2-70B ↑1 28.5%
  7. 7 Llama 2-13B ↓1 27%
  8. 8 Llama 2-7B 26.4%

Number one over time

Lab claims count from the day the card came out, independent runs from the model’s release.

Anthropic
Own leaderboard Not tested since May 2024
Claude 2 45.3%

Source: Official CaLM leaderboard

Scores over time

All the scores we have, one dot each, so a model tested at several settings appears more than once.

Plot with other benchmarks
On the benchmark’s own leaderboard Best so far
0% 20% 40% 60% 80% 100% 2023 Near the ceiling (85%) Claude 2 Baichuan1-7B 0% 20% 40% 60% 80% 100% 2023 Near the ceiling (85%) Claude 2 Baichuan1-7B

Source: Official CaLM leaderboard

No lab model card reports it yet

We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.

More in Knowledge

All 332 in Knowledge
  • MMLU-Pro In use Harder 10-option version of MMLU with reasoning-focused questions across 14 subjects. Intern-S2-397B 89.8%
  • MMLU In use 57-subject multiple-choice exam questions from elementary to professional level. Scored by accuracy. Seed1.8 92.3%
  • SuperGPQA In use 26,529 graduate-level multiple-choice questions across 285 disciplines. Accuracy. Qwen 3.6 Max (Preview) 73.9%
  • WideSearch In use 200 bilingual broad info-seeking tasks: agent collects many atomic facts into a table. Scored by success rate and row/item F1. Tencent Hy4 preview 83.9%
  • BBEH micro average In use Models solve harder replacements for each of the 23 BIG-Bench Hard tasks, 4,520 examples in total. The score is the share of all examples answered correctly. Gemma 4 31B IT 74.4%
  • HLE-Diamond In use Models answer 1,000 cleaned expert questions from Humanity's Last Exam, 500 on reasoning and 500 on knowledge, closed book. Graded on exact answers. Claude Fable 5.2 77.9%

Data audit

9 of 10 checks pass, 1 warning. Look first at the Newest score check. The newest came 3 years ago, in September 2023, for Qwen-7B. Every benchmark

Catalog record 2 pass
  • Pass Task access Public tasks, plus a private test set.
  • Pass Confidence Checked against a primary source.
Scores 1 warning 3 pass 1 note
  • Pass Independent runs Scores from the official leaderboard (28 entries).
  • Pass Models scored 8 models we track, plus 20 other entries on its leaderboard.
  • Warning Newest score The newest came 3 years ago, in September 2023, for Qwen-7B.
  • Note Frontier models None of the top 10 has a score. The strongest tested is Claude 2 (#192).
  • Pass Headroom The best is 45.3%, under 85%, so it still separates models.
Leaderboard import 3 pass
  • Pass Importer We read the board on opencausalab.github.io: 8 rows for models we track and 20 other entries at the last read.
  • Pass Last read November 18th, 2026.
  • Pass Read worked The last read worked.
Consistency 1 pass 1 note
  • Note Lab claims No model has both a lab number and an independent run.
  • Pass Known issues No published audit found problems.

How we check