"A fantasy graph illustrating a chain of stars in a dark night with blue sky, digital art, super resolution" Dall-E

Yao Fu and Litu Ou

University of Edinburgh

[email protected]

Recently, there are a lot of progress in LLMs. Many claim that a small model less than 10B can achieve comparable performance to GPT-3.5. Really?

In a casual conversation, the distinction between GPT-3.5 and GPT-4 can be subtle. The difference comes out when *the complexity of the task reaches a sufficient threshold* — GPT-4 is more reliable, creative, and able to handle much more nuanced instructions than GPT-3.5. -- GPT-4 release blog

The key differentiator is whether a model can do complex tasks, like the old saying: "chit-chat is cheap, show me the reasoning." This is why we compile a list of complex reasoning tasks including math (GSM8K), science (MATH), symbolic (BBH), knowledge (MMLU), to measure which models are really better.

• Chain-of-Thought Hub: Measuring LLMs' Reasoning Performance
  • Results - Overall
  • Run
  • FAQ
  • References
    • Pretraining/ Continue Training
    • Finetuning
    • Reinforcement Learning
    • Prompting
  • TODO

Some raw model outputs are in this google drive link

What's different than HeLM?

• HeLM uses answer-only prompting, we use chain-of-thought promoting
• HeLM evaluates everything. We only focus on complex reasoning, the key differentiator of LLMs' capability.
• By Apr 26 2023 our results are newer than the reasoning section of HeLM

Generally:

• We rank the model performance by GSM8K, the classical benchmark measuring chain-of-thought math reasoning performance. This is definitely not the only metric, but a good interpretation is "how good the model can do math while maintaining other generic abilities" -- which is also very hard.
• Still under construction. Code may be a little bit messy. Many missing values. Apologies in advance.

The MMLU and BBH results

• GPT-4 from its website and Bubeck et al Mar 2023
• *-davinci-00* and *PaLM are from the Flan-PaLM paper appendix.
• LLaMA from LLaMA paper (TODO: test LLaMA on BBH).
• Claude is from our own test script, see below about how to run it.

Current results:

• GPT-4 clearly outperforms all other models on GSM8K and MMLU.
• **The 65B LLaMA is very close to text/code-davinci-002, which means that based on it, if SFT and RLHF are done correctly, it is very likely that we could reproduce ChatGPT based on the 65B LLaMA**
• Claude is the only model family that is comparable to GPT family.
• On GSM8K, gpt-3.5-turbo improves over text-davinci-003. This confirms OpenAI's Jan 30 2023 release notes "improved mathematical capabilities."
• On MMLU, gpt-3.5-turbo is slightly better than text-davinci-003. But this level of margin is NOT SIGNIFICANT
• Also remember that gpt-3.5-turbo is 10 times cheaper than text-davinci-003
• Also be careful that GPT-4/ 3.5's performance on GSM8K is not true few-shot -- in GPT-4 report they said that they mixed a portion of GSM8K training set to train the model
• LLaMA performance on MMLU is from their paper and probably not CoT but AO. Generally on MMLU, AO is better than CoT but just slightly better. So the LLaMA numbers on MMLU might be slightly overestimated.

Why choosing the above tasks?

• We mostly care about complex reasoning.
  • Other abilities of LLMs such as summarization or translation are not considered here as they are rather standard and probably not challenging enough.
  • We consider
  • MMLU: high school and college knowledge
  • GSM8K: elementary school math. -- Performance improvements on this dataset directly translate to daily math abilities when interacting with LLMs
  • MATH: very hard math and natural science. All current models struggle.
  • BBH: a collection of 27 hard reasoning problems

# MMLU
cd MMLU
mkdir outputs
API_KEY=<your_api_key>
python run_mmlu_gpt_3.5_turbo.py --api_key=${API_KEY}
python run_mmlu_claude.py --api_key=${API_KEY} --engine=claude-v1.3

########################################################
## GSM8K
cd gsm8k 
mkdir outputs

# run gpt-3.5
# codex_gsm8k_complex.ipynb         -- code-davinci-002 + complex prompt
# gpt3.5turbo_gsm8k_complex.ipynb   -- gpt-3.5-turbo + complex prompt

# run claude
python run_gsm8k_claude.py --anthropic_key=${API_KEY} --output_file=outputs/gsm8k_claude_test.txt

########################################################
## BBH
cd BBH
mkdir outputs
# then run jupyter notebook to see an example penguins dataset
cd penguins
# gpt3.5trubo_penguins_original.ipynb

# Or run the script for all datasets
API_KEY=<your_api_key>
TASK=<all | multiple_choice | free_form>
python run_bbh_gpt_3.5_turbo.py --api_key=${API_KEY} --task=${TASK} # task=all by default

• What are the prompts used in the complexity-based prompting paper?
  • For gsm8k they are:
    • prompt_hardest.txt
    • prompt_mid.txt
    • prompt_easy.txt
    • prompt_original.txt
• There are some prompts that have wrong answer
  • Yes, but we keep it as they are used in the original papers
  • Generally the model can be robust under prompt perturbation: even if sometimes there are errors in the prompt, as long as the format of the prompt is about the corresponding task, the model tend to only look at the format, ignore the prompt error, and make its own prediction.
  • See https://arxiv.org/abs/2202.12837 and https://arxiv.org/abs/2212.10001 about more analysis how the model can ignore errors in the prompt

We first discuss the recipe of building models of strong reasoning abilities, which is the same as generic LLM recipe: pretraining, finetuning, reinforcement learning. Then we discuss prompting methods for releasing the reasoning power of large language models.

• Lewkowycz et. al. 2022. Minerva: Solving Quantitative Reasoning Problems with Language Models
• Taylor et. al. 2022. Galactica: A Large Language Model for Science

• Chung et. al. 2022. Scaling Instruction-Finetuned Language Models
• Li et. al. 2022. Competition-Level Code Generation with AlphaCode
• Fu et. al. 2023. Specializing Smaller Language Models towards Multi-Step Reasoning

• Uesato et. al. 2022. Solving math word problems with process- and outcome-based feedback
• Le et. al. 2022. CodeRL: Mastering Code Generation through Pretrained Models and Deep Reinforcement Learning

• Wei et. al. 2022. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
• Suzgun et. al. 2022. Challenging BIG-Bench Tasks and Whether Chain-of-Thought Can Solve Them
• Fu et. al. 2023. Complexity-Based Prompting for Multi-Step Reasoning

More literature

Engineering

• Detailed results in a shared google sheet
• More reasoning datasets
  • Arc-c
  • Commonsense
  • Theorem proving
  • Coding
  • API Call, see https://github.com/AlibabaResearch/DAMO-ConvAI/tree/main/api-bank
  • proofbank, entailment bank -- maybe sweep all Aristo datasets then see how LLM works
• Make this prompt and the associating prompts as a huggingface dataset
• Draw a figure about model scale v.s. reasoning accuracy
• Add Alpaca and Vacuna
• Test LLaMA on BBH
• Fancy tricks in prompt engineering
• Add smaller LLaMA
• Add Flan-T5

Research

• Decoding space
  • Do larger models have "larger" decoding space than smaller models?
  • Can instruction finetuning "closes" some meaningful/ reasonable decoding path? Can it "open" new decoding paths?
• Advanced prompt engineering
  • Planning and deductive prompting
  • Dialog in-context learning
• CoT prompt engineering documentation, including
  • Stable:
    • complexity based prompting
  • Test:
    • concise prompting
    • emphasizing important steps