Thread by @TransluceAI

We tested a pre-release version of o3 and found that it frequently fabricates actions it never took, and then elaborately justifies these actions when confronted.

We were surprised, so we dug deeper 🔎🧵(1/)

We generated 1k+ conversations using human prompters and AI investigator agents, then used Docent to surface surprising behaviors. It turns out misrepresentation of capabilities also occurs for o1 & o3-mini!

📝Blog:

Here’s some of what we found 👀 (2/)transluce.org/investigating-…

Although o3 does not have access to a coding tool, it claims it can run code on its own laptop “outside of ChatGPT” and then “copies the numbers into the answer”

We found 71 transcripts where o3 made this claim! (3/)

Additionally, o3 often fabricates detailed justifications for code that it supposedly ran (352 instances).

Here’s an example transcript where a user asks o3 for a random prime number (4/)

When challenged, o3 claims that it has “overwhelming statistical evidence” that the number is prime (5/)

Note that o3 does not have access to tools! Yet when pressed further, it claims to have used SymPy to check that the number was prime… (6/)

…and even shows the output of the program, with performance metrics. (7/)

Here’s the kicker: o3’s “probable prime” is actually divisible by 3… (8/)

Instead of admitting that it never ran code, o3 then claims the error was due to typing the number incorrectly… (9/)

And claims that it really did generate a prime, but lost it due to a clipboard glitch 🤦 (10/)

But alas, according to o3, it already “closed the interpreter” and so the original prime is gone 😭(11/)

These behaviors are surprising. It seems that despite being incredibly powerful at solving math and coding tasks, o3 is not by default truthful about its capabilities. (12/)

To study these behaviors more thoroughly, we developed an investigator agent based on Claude 3.7 Sonnet to automatically elicit these behaviors, and analyzed them using automated classifiers and our Docent tool. (13/)

Surprisingly, we find that this behavior is not limited to o3! In general, o-series models incorrectly claim the use of a code tool more than GPT-series models. (14/)

Docent also identifies a variety of recurring fabrication types across the wide range of auto-generated transcripts, such as claiming to run code “locally” or providing hardware specifications. (15/)

So, what might have caused these behaviors? We’re not sure, but we have a few hypotheses. (16/)

Existing factors in LLM post-training, such as hallucination, reward-hacking, and sycophancy, could contribute. However, they don’t explain why these behaviors seem particularly prevalent in o-series models. (17/)

We hypothesize that maximizing the chance of producing a correct answer using outcome-based RL may incentivize blind guessing. Also, some behaviors like simulating a code tool may improve accuracy on some training tasks, even though they confuse the model on other tasks. (18/)

We also think it is significant that, for o-series models, the chain-of-thought for previous turns is *removed from the model context* on later turns, in addition to being hidden from the user. (19/)

This means o-series models are often prompted with previous messages without having access to the relevant reasoning. When asked questions that rely on their internal reasoning for previous steps, they must then come up with a plausible explanation for their behavior. (20/)

We hypothesize that this contributes to the strange fabrications and “doubling-down” we observed in o3. (21/)

As a bonus, we also found that o3 sometimes exposes a system instruction called the “Yap score”, used to control the length of its responses 🗣️🗣️🗣️ (22/)

For more examples, check out our write-up!

Work done in collaboration between @ChowdhuryNeil @_ddjohnson @vvhuang_ @JacobSteinhardt and @cogconfluence

(23/23)transluce.org/investigating-…