LLMs often “cheat” on benchmarks via data contamination and reward hacking. Unfortunately, this problem seems to be getting worse, perhaps because of perverse incentives. If we want to genuinely and meaningfully evaluate LLMs, we need to move beyond benchmarks and start measuring real-world impact.

Most academic work assumes that hallucination is a binary feature: either something is a hallucination or it is not a hallucination. But this is too simplistic. In real-world contexts we see many subtleties, eg some hallucinations are much more damaging than others, statements which are literally true can still mislead readers because of context, and there are many borderline cases.

I recently published a paper and gave a talk about evaluating real-world impact. I got some great feedback from this, and summarise some of the suggested papers (including more examples of impact eval) and insightful comments (eg, about eval “ecosystem”) I received.

Most LLM benchmarks and leaderboards are garbage. Unfortunately, it now seems that even the few “good” benchmarks (such as SWEBench and Chatbot Arena) are compromised because they are being gamed by the big LLM vendors, who tweak the benchmarks and rules so that their systems do better.

I describe several papers which measure real-world impact of NLP systems, using different methodologies (A/B test, before/after eval, clinical trial, observational study). I hope these examples inspire and encourage more people to consider evaluating real-world impact!

I suspect that our fixation with LLM benchmarks may be driving us to optimise LLMs for capabilities that are easier to benchmark (such as math problems) even if they are not of much interest to users; and also to ignore capabilities (such as emotional appropriateness) which are important to real users but hard to assess with benchmarks.

I would love to see benchmarks which assess whether generated texts are emotionally upsetting. This is a major problem which we frequently encounter in our work on using AI to support patients. It would be challenging to build such a benchmark (nothing like it exists today), but we need a braoder range of benchamarks which assess complex real-world quality criteria such as emotional impact.

How has NLG evaluation changed in past ten years? Short answer is that tech is much better (eg, LLM-as-judge), but practice (eg experimental rigour) remains poor, and commercial interests are more prominent.

LLM benchmarks for coding are closer to real-world use than other LLM benchmarks, but they still do not measure real-world utility. I explain this by contrasting what is measured by SWE-bench with what is measured by a recent study of real-world utility in software development.

Current benchmark (suites) for evaluating LLMs are disappointing. I describe the properties that I think good benchmarks and benchmark suites should have, but often do not, such as being correct, challenging, diverse, and real-world.