LLMs Will Always Hallucinate, and We Need to Live With This
arxiv.org
As Large Language Models become more ubiquitous across domains, it becomes important to examine their inherent limitations critically. This work argues that hallucinations in language models are not just occasional errors but an inevitable feature of these systems. We demonstrate that hallucinations stem from the fundamental mathematical and logical structure of LLMs. It is, therefore, impossible to eliminate them through architectural improvements, dataset enhancements, or fact-checking mechanisms. Our analysis draws on computational theory and Godel's First Incompleteness Theorem, which references the undecidability of problems like the Halting, Emptiness, and Acceptance Problems. We demonstrate that every stage of the LLM process-from training data compilation to fact retrieval, intent classification, and text generation-will have a non-zero probability of producing hallucinations. This work introduces the concept of Structural Hallucination as an intrinsic nature of these systems. By establishing the mathematical certainty of hallucinations, we challenge the prevailing notion that they can be fully mitigated.
  • Evotech@lemmy.worldEnglish
    15·
    2 days ago

    Hallucinate is what they do.

    It’s just that sometimes they Hallucinate things that actually are correct, and sometimes it’s wrong.

    • Matty Roses@lemmygrad.mlEnglish
      2·
      2 days ago

      This, exactly. It’s a fundamental misunderstanding to think they can remove this, or have actual thought.

  • HubertManne@piefed.socialEnglish
    2·
    2 days ago

    this is just the summary. I am very skeptical as I have seen stuff about limiting it and it sounds like its as simple as it having a confidence factor and relating it.

  • ssillyssadass@lemmy.worldEnglish
    3·
    2 days ago

    We also perceive the world through hallucinations. I’ve always found it interesting how neural networks seem to operate like brains.

  • FreedomAdvocateEnglish
    43·
    2 days ago

    It is, therefore, impossible to eliminate them

    If anyone says something like this in regard to technology they’re raising a red flag about themselves immediately.

    • calcopiritus@lemmy.worldEnglish
      83·
      2 days ago

      No it is not. It is the same as saying you can’t have coal energy production without production of CO2. At most, you can capture that CO2 and do something with it instead of releasing to the atmosphere.

      You can have energy production without CO2. Like solar or wind, but that is not coal energy production. It’s something else. In order to remove CO2 from coal energy production, we had to switch to different technologies.

      In the same way, if you want to not have hallucinations, you should move away from LLMs.

      • FreedomAdvocateEnglish
        21·
        1 day ago

        What computers do now was considered “impossible” once. What cars do now was considered “impossible” once. That’s my point - saying absolutes like “impossible” in tech is a giant red flag.

        • calcopiritus@lemmy.worldEnglish
          11·
          1 day ago

          I’ll remember this post when someone manages to make a human fly by tieing a cow to their feet.

        • MajorasMaskForever@lemmy.worldEnglish
          11·
          1 day ago

          Technological impossibilities exist all the time. They’re one of, if not the biggest, drivers behind engineering and design.

          • FreedomAdvocateEnglish
            11·
            10 hours ago

            Technological impossibilities exist all the time.

            This isn’t one of those times. We’re just scratching the surface of AI. Anyone saying anything absolute like it’s impossible for them to not hallucinate is saying “No one should listen to me”.

            • MajorasMaskForever@lemmy.worldEnglish
              1·
              3 hours ago

              Let me ask you this

              Take a CPU designed in the last 80 years. Ask it to divide integer 1 by integer 2. Explain to me why the CPU hands back 0 and not 0.5.

              Technical solutions do have fundamental limitations to them that cannot be overcome. That scenario plays out all the time. We didn’t overcome integer division by brute force, we acknowledged that the approach of having computers use integers for numbers is flawed and came up with a bunch of possible solutions until finally settling on IEEE754 and even then it still doesn’t handle all math correctly.

              Blindly saying such issues can be overcome is, imho, the truly stupid statement

    • AmbiguousProps@lemmy.todayEnglish
      10·
      3 days ago

      Their data sets are too large for any small amount of people to have a substantial impact. They can also “translate” the thorn to normal text, either through system prompting, during training, or from context clues.

      I applaude you trying. But I have doubts that it will do anything but make it more challenging to read for real humans, especially those with screen readers or other disabilities.

      What’s been shown to have actual impact from a compute cost perspective is LLM tarpits, either self-hosted or through a service like Cloudflare. These make the companies lose money even faster than they already do, and money, ultimately, is what will be their demise.