wluk

"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).

haha true, when something goes wrong in big tech, your first thought is "what if we just double the instances and see if it goes away?"