I wonder if search engines linked to human-use-case engines (e.g. google/bing) start at a disadvantage because they have been historically incentivized to break themselves to support their business models? It seems reasonable to suppose that "good at selling ads" ≠ "good at finding results".
do you somehow control how non-trivial the queries are? The LLM generates them, right?
what if every engine returns garbage, or on the other hand, handles them too well?
building a benchmark like this in a genuinely fair way seems extremely hard to me. I’m very curious about the details, of course within what you can share.
It is hard to be fair I agree. We tried to be open about what we do here: github.com/keenableai/needle
The queries from what I can tell are not trivial. The actual github repo of the benchmark has a judgement/query browser where you can inspect the different query streams: https://keenableai.github.io/needle/
The judgements are available on huggingface so training is possible. But given that we evaluate on new queries daily there would need to be some generalization happening for this to show up in the benchmark.
One of the authors here. We have been seeing lots of benchmaxxing and leakage in standard web search benchmarks such as BrowseComp.
We developed this live benchmark with daily/hourly sampled fresh queries matching real agentic search traffic to estimate actual search performance of different AI search providers.
It seems that the agentic search benchmarks have fallen victim to reward hacking, just like the coding benchmarks. It's good to see mitigation efforts being made to address this problem.
What do you foresee in the future releases and improvements to this?
I wonder if search engines linked to human-use-case engines (e.g. google/bing) start at a disadvantage because they have been historically incentivized to break themselves to support their business models? It seems reasonable to suppose that "good at selling ads" ≠ "good at finding results".
do you somehow control how non-trivial the queries are? The LLM generates them, right?
what if every engine returns garbage, or on the other hand, handles them too well?
building a benchmark like this in a genuinely fair way seems extremely hard to me. I’m very curious about the details, of course within what you can share.
It is hard to be fair I agree. We tried to be open about what we do here: github.com/keenableai/needle
The queries from what I can tell are not trivial. The actual github repo of the benchmark has a judgement/query browser where you can inspect the different query streams: https://keenableai.github.io/needle/
Do you think it will be possible to train on this bench?
The judgements are available on huggingface so training is possible. But given that we evaluate on new queries daily there would need to be some generalization happening for this to show up in the benchmark.
One of the authors here. We have been seeing lots of benchmaxxing and leakage in standard web search benchmarks such as BrowseComp.
We developed this live benchmark with daily/hourly sampled fresh queries matching real agentic search traffic to estimate actual search performance of different AI search providers.
It seems that the agentic search benchmarks have fallen victim to reward hacking, just like the coding benchmarks. It's good to see mitigation efforts being made to address this problem.
What do you foresee in the future releases and improvements to this?