Isn't that precisely the reason why we introduced the term hallucination? Because llms have historically always made up bullshit of they cannot answer directly... If they now nailed this to maybe the model not respond instead of responding incorrectly, then a lot of previously unusable usecases would become feasible.
So I feel like that's exactly the right metric and the way to track it wrt hallucinations.
The point is that it's not a useful metric on its own. For example, redirecting from /dev/null also achieves a zero hallucination rate.
We want the hallucination rate to decrease while the overall answer rate of queries remains sufficiently high. For more specifics, look into ROC and AUC.
Models can answer "I don't know". Hallucination benchmarks, including this, give the models the option to "not attempt". It's just that the metric linked doesn't take into account the rate of correct answers at all. It has its uses in analyzing incorrect vs not attempted answers, but gives a very partial picture.
Its very annoying this has been in the capability of models since the very beginning. It could check how probable its token values are and if those fall below a certain threshold either say "I don't know", or output the most probable (well, more like least improbable) tokens but give a very clear, very strong warning that it is a shot in the dark and likely to contain hallucinations.
But no, Google and OpenAI would rather always have an answer ready and tell you to mix glue into your pizza toppings :)
Hallucination detection is an open problem. If it were that simple, people would indeed "just" do it.
Basically the problem is that LLMs aren't trained on things they don't know; an alternative way of saying this is that they're not trained on things they're not trained on, which is obviously true.
When you RL a model and it answers incorrectly, you don't teach it to answer "I don't know", you teach it to answer correctly instead. This makes it very hard for it to realize when it doesn't know things.
Models tend to default to their training data even when they lack sufficient context, they've never been trained to recognize their own uncertainty, so they hallucinate confidently instead.
I don't have much to add other than this observation that we seem to have moved away from eating one small rock per day for nutritional value, and adding gasoline in spaghetti.
The glue on pizza reference brought back memories :)
The probability of tokens is unfortunately a poor proxy for confidence because it is entirely possible for "mixing glue" to appear in a sentence about making pizza depending on context. It might even be likely if the user has asked the model to lie.
the observation goes beyond garbage in garbage out. Mainly that we're always operating from some prior and limited understanding. That what may look like a hallucination could be closer to the truth than our current frameworks of understanding allow us to admit. The hermeneutic circle.
Was there something about this specific model and submission that made you feel compelled to write this self-evident observation?
Or would you describe your methodology as more like picking a random sentence fragment as an input value then generating completions from your existing corpus without any post-input "learning" process related to the rest of the source material?
The big question for me having used a lot of these SOTA chinese models is: what is its token efficiency like?
Running Step 3.5 Flash locally for example, it's an amazingly capable model all things considered, but it's token efficiency is so bad that it gets out performed by most others wall-clock time (even with my MTP-support for it hacked in to llama.cpp: despite being trained on three heads, MTP 2 is the sweet spot, and only gets it from 20tk/s to 30tk/s on my Spark)
The DeepSeek models and Qwen 3.5 Plus are also good examples of this: compared to Opus, and especially GPT 5.5 they use many more tokens to get to the same answers.
I'm really hoping that Qwen 3.7 is better in this regard, can't wait to try it out
(ps. running DeepSeek v4 Flash on my Spark is absolutely wild, thanks antirez if you see this haha)
> Gemma 4 31B is notably token efficient, using 39M output tokens to run the Intelligence Index vs 98M for Qwen3.5 27B (Reasoning). This is ~2.5x fewer output tokens for a model scoring 3 points lower. For context, the other models at the 42-point intelligence level also use significantly more tokens: MiniMax-M2.5 (56M), DeepSeek V3.2 (Reasoning, 61M), and GLM-4.7 (Reasoning, 167M)
Right and all of my own evals back this up for Gemma 4...
...except its notably worse at coding in an agent context even with a harness setup to do exactly what Google says it should do (wrt. to sending summarised thinking back and so on)
So despite it being far better token efficiency wise, it's just worse for what I need to use it for compared to DSv4 Flash or Qwen 3.6 27B
