Datasets:
Navigation questions ambiguity
Hi! Not sure if this is the place for comments on the benchmarks themselves, but I feel like some of the navigation questions still have too much ambiguity. Several of them just give numbers of steps without specifying you should face forward, or that each block of steps need to be in the same direction, and don't mention any turns. I'm assuming each question is asked to the LLM with a fresh context, so there isn't a lot of clues to how it should be interpreted.
Like this one that R1 gets wrong: "If you follow these instructions, do you return to the starting location? Take 2 steps. Take 10 steps. Take 8 steps." If you take 2 steps back, 10 forward, 8 back, you'll end up where you started. Or this one that Sonnet 3.5 (Oct) gets wrong: "If you follow these instructions, do you return to the starting location? Take 3 steps. Take 6 steps. Take 9 steps." That could easily be interpreted as 3 steps forward, 6 forward, 9 back.
Even this question, while more defensible, is a bit wiggly: "If you follow these instructions, do you return to the starting location? Take 5 steps. Take 4 steps. Take 3 steps." Since you can have a right triangle of those lengths and it doesn't say you can only walk in cardinal directions.
I guess the defense for all of these might be that it says "If you follow these instructions, do you return" vs "can you return". The latter would imply more strongly that it is a puzzle where you're supposed to figure out how possible it is to return based on any set of directions. Then again, the fact that the steps are split into separate blocks to begin with, does imply there is there's something different about them. I haven't read the original paper, but it seems odd especially considering how unambiguous some of the other questions are. Especially the first two questions don't seem to really be testing navigation per se, but rather reading between the lines of how the question is worded.
Edit: I hadn't actually read R1's reasoning until now and man it is a bit painful with it going back and forth trying to decide if it should interpret the 2 steps, 10 steps, 8 steps as a simple straight line forward or if it has directional flexibility since it isn't mentioned. I was a being a bit cheeky with the triangle comment above, but it even considered triangles and if it could form a closed path at one point. I'm pretty sure if I got this question I'd overthink it the same way, so maybe I'm biased, but it doesn't seem like a good question.