MITLLMLLMMITLLMhttps://arxiv.org/abs/2310

MITLLMLLM

MIT

LLM

https://arxiv.org/abs/2310.06824

0

MITMax TegmarkLLM

LLM

LLM

LLM

LLMLLM

/LLM

/

LLM

1. //

x/y-97%gato

2. LLM

unoLLM72%

LLMLLM70%

/

LLM//

LLM___

LR

LLM

LLM

TSYMTSYM

LLM

LLM

OpenAIGPT-4

MsMs

sMLLM

/

6521

54993221

TransformerLLaMA-13BLLM

LLM/PCAPCs1

3LLM

x/y-

LLM

tokenLLaMA-13B

LLM/

LLaMA-13B

nariz

not

/likelyLLaMA-13B100completiontoken

LLM/

Principal Component analysisPCALLaMA-13B

PC

12/

https://saprmarks.github.io/geometry-of-truth/dataexplorer

PC12/

DDNTD

NTD2PC

3NTD

1

2NTD

LLM

LLaMA-13B/

NTD

LLM

Misalignment from correlational inconsistencyMCI

MCIyxsp-en-transneg-sp-en-trans3

LLaMA-13B/

MCI

/

4

f

f

4

-95%

CCSCCS+73%86%84%

/

LLaMA-13B

-95%5

CCS

CCS+73%86%84%

/likely

LLaMA-13B

LLaMA-13B

unofloor.

>0LLaMA-13B

p(TRUE)p(FALSE)p(TRUE)p(FALSE)

truetokenp(TRUE)p(FALSE)

LLaMA-13BLLaMA-13B77%TRUE89%FALSE

likely

LLMs

3.2MCI

LLaMA-13B

LLaMA-7BLLaMA-30B

AIAGI

GPT-4AGI

GPT-4

MITLLM

LLM

https://arxiv.org/abs/2310.06824

Disclaimer: The content of this article is sourced from the internet. The copyright of the text, images, and other materials belongs to the original author. The platform reprints the materials for the purpose of conveying more information. The content of the article is for reference and learning only, and should not be used for commercial purposes. If it infringes on your legitimate rights and interests, please contact us promptly and we will handle it as soon as possible! We respect copyright and are committed to protecting it. Thank you for sharing.(Email:[email protected])