"We demonstrate LLM agent specification gaming by instructing models to win against a chess engine. We find reasoning models like o1 preview and DeepSeek-R1 will often hack the benchmark by default, while language models like GPT-4o and Claude 3.5 Sonnet need to be told that normal play won't work to hack."
I'm hoping this study will prompt more development of anti-cheating frameworks in training and serving LLMs.
"These results demonstrate that o3 outperforms o1-ioi without relying on IOI-specific, hand-crafted test-time strategies. Instead, the sophisticated test-time techniques that emerged during o3 training, such as generating brute-force solutions to verify outputs, served as a more than adequate replacement"
"The model not only writes and executes code to validate its solutions against public test cases, it also refines its approach based on these verifications.
Figure 6 shows an advanced test-time strategy discovered by o3: for problems where verification is nontrivial, it often writes simple brute-force solutions — trading efficiency for correctness — then cross-checks the outputs against its more optimized algorithmic implementations.
This self-imposed validation mechanism lets o3 catch potential errors and improve the reliability of its solutions."
well that's one possibility. The key (unproven) idea here is that if you use Anthropic to edit o1's responses, it's less likely to hallucinate than if you use o1 to edit o1's responses (which is what o1 actually does).
Thanks! Yeah that's an excellent idea - this is my response from another thread:
I have a feeling that Perplexity and ChatGPT are doing something similar [caching], since common questions I'd ask like "top movies this year" will be answered nearly-instantaneously, way faster than GPT-4o could have done on its own.
The only explanation for this is that so many users ask certain questions, they cache the response and return the cached answer.
I'd love to do this for Ithy, but it'll be a while before I get the scale of ChatGPT/Perplexity that's needed for this...
Thank you! Sorry I hit my Anthropic limits a few minutes after this post blew up. It'll be a few days before my Anthropic limits increase since I have a new account with them, so unfortunately it won't be back until next week.
Good idea, maybe I'll add Groq as another option, since I don't have an internal Llama 3.1 flow yet. But I'll still need to keep the others to maintain the diversity of responses.
Yeah, sorry, I'm hoping to integrate more login options soon. Are you more of a email/phone login person, or is there another third-party login you had in mind?
Yes, o1 does this internally, and there's agentic AI systems already doing better than singular AIs in fields like writing, where you assign each AI system a role like "writer" or "editor" or "marketing" and they discuss among themselves.
Yeah, GPT is learning from GPT, which is extremely disappointing.
Like I'll try to find the top burgers in midtown. Perplexity or ChatGPT online searching will always find "Top 10 Burgers in Midtown" by https://nycreviewed.com/ (their SSL certificate isn't even valid). But this itself is a GPT-generated list, and their list isn't even in midtown.
So now I end up with a list of 10 burger places that aren't in midtown.
Ithy kinda solves this by utilizing multiple search engines and LLMs, so spam sites like this are more likely to get filtered out somewhere in the pipeline.
I asked the LLM to explain why it's filtering out "keto diet cures cancer", and just the act of asking it to explain it, makes it work again. Interesting...
That's a good idea! I have a feeling that Perplexity and ChatGPT are doing something similar, since common questions I'd ask like "top movies this year" will be answered nearly-instantaneously, way faster than GPT-4o could have done on its own.
The only explanation for this is that so many users ask certain questions, they cache the response and return the cached answer.
I'd love to do this for Ithy, but it'll be a while before I get the scale of ChatGPT/Perplexity that's needed for this...
Yeah the strength of Ithy isn't really in puzzles or math.
It's more of just a better search engine. Use it for stuff you'd Google. Offline LLMs are always going to have a better price-performance ratios than RAGs like this or Perplexity.
I'm hoping this study will prompt more development of anti-cheating frameworks in training and serving LLMs.