模型把姚明编成足球运动员 (English)

Generated: 2026-06-22 06:56:03

---

Okay, here is the English translation, maintaining the storytelling style as requested.

---

Hallucinations in LLMs: I've Stumbled into More Pits Than You've Read Papers

In the summer of 2023, I got burned badly by a large language model.

I was using GPT-4 to write a market analysis report. It confidently threw out a set of data – claiming a certain company's quarterly revenue had grown by 47.3%! The number was so precise, I took a screenshot and sent it straight to the client. The client immediately called me out, leaving me red-faced: that company didn't even exist that quarter! The number was completely fabricated by the model.

In that moment, I understood – these large language models are, at their core, "sophisticated liars." They aren't querying a database; they're guessing the next word in a sequence. Ask them to tell a story, and they can make a lie sound more convincing than the truth.

---

Hallucination Isn't a Bug, It's a Core Design Feature

Later, I slogged through that 49-page survey paper from HIT and Huawei. It took me three whole days. The paper was well-written, but let me break down my understanding in plain English.

They categorize hallucinations into two types: Factual Hallucination and Faithfulness Hallucination.

Factual Hallucination is pure fabrication. You ask, "Who was the first person to walk on the moon?" and it says, "Charles Lindbergh was the first to walk on the moon during the 1951 Lunar Pioneer Mission." This guy famously flew solo across the Atlantic in 1927, having nothing to do with the moon. The model mashed two different things together – like asking a child what they ate yesterday, and they mix in something from last week.

Faithfulness Hallucination is even more insidious. I tested this myself: I asked a model to summarize news from October 2023, and it gave me events from October 2006. Think about it – it's like asking your assistant to compile today's meeting minutes, and they pull out notes from ten years ago. Absolutely ridiculous.

---

Why Do Models Spout Nonsense? I Ran an Experiment

To figure this out, I did a crazy test. I asked the same model (GPT-4, November 2023 version) the same question ten times: "Why can't you use a quantum computer to stir-fry vegetables?"

Guess what? The results were fascinating:

• 3 times it said, "Qubits decohere at high temperatures."
• 2 times it said, "The microscopic motion of a spatula would cause wave function collapse."
• 1 time it said, "Because quantum computers need liquid helium cooling, which would freeze the vegetables."
• 4 times it simply refused to answer.

See! The model is "making things up" every single time, but some fabrications just happen to sound plausible. It has no idea whether a quantum computer can stir-fry vegetables – it's just generating text that looks like a correct answer based on patterns in its training data. It's like asking someone who has never seen an apple to describe it. They can only piece together words they've heard – "red, round, edible" – but they'll never know the true taste of an apple.

Professor Dong Bin from Peking University's math department said that hallucination and creativity are separated by a thin line. I agree halfway: hallucination could be a precursor to creativity, but only if you can control it. The problem now is that the model itself can't tell the difference between hallucination and creation – it's like letting a three-year-old do art. They might paint a masterpiece, or they might just cover the walls in a chaotic mess.

---

Where's the Root Cause? I Dug Through the Data

The paper says there are three main sources of hallucination: data sources, the training process, and inference. Let me add my own practical observations – these are pits I've personally fallen into.

Data source issues are the most severe. I've scraped some training data and found a Reddit post claiming "Yao Ming is a soccer player," and the model learned it. Worse, the model can "stitch together" contradictory information from different sources. For example, it might know Yao Ming plays basketball, but also remember the false "Yao Ming is a soccer player" fact, and eventually generate an answer like "Yao Ming is both a basketball and soccer player." Imagine the chaos if this happened in the medical field.

Training process issues are more subtle. During pre-training, the model learns statistical patterns, not facts. It knows "Yao Ming" and "basketball" often appear together, but it doesn't know that "Yao Ming is a basketball player" is a factual statement. It's like memorizing a million math problems without learning the underlying principles – you can only apply formulas during an exam, and you're lost if the question changes slightly.

Inference stage issues are the funniest. I tested this by asking the model to do a simple arithmetic problem: 237 + 489. Sometimes it gives 726 (correct). But if you add a distractor, like "237 + 489, where 237 is a prime number," it might start making things up. Because it incorporates the irrelevant "prime number" information into its reasoning process. It's like asking someone "What is 1+1? Note that 1 is an odd number." They might get confused and say "3."

---

How to Fix It? Methods I've Tried

The paper lists a ton of methods. Let me talk about a few I've actually tested – telling you what works and what doesn't.

RAG (Retrieval-Augmented Generation) is currently the most popular solution. Simply put: before generating an answer, let the model look up information in a knowledge base. I tested a RAG system built with LangChain, and the results were good – but there's a catch: if the retrieved information itself is wrong, the model will fabricate even more "convincingly." It's like asking a known liar to do research, but the source material is also fake – they can then weave a perfectly seamless lie.

Prompt Engineering is the cheapest method. I wrote a "Data Iron Law" framework. The core idea is just one sentence: You cannot fabricate data that isn't in the context. The specific approach is:

# Rule 1: Data injected into the context is the only legitimate source.
# Rule 2: Prohibit unit conversions (6.29 USD/lb is 6.29 USD/lb).
# Rule 3: Mark the source for every numerical value.

In my tests, the hallucination rate dropped from 30% to 8%. But there's a problem: the model will directly refuse to answer many questions because it's "unsure." It's like telling a child, "If you don't know, say you don't know." The result is that they become afraid to say even the things they do know – it's overcorrecting, but it's still better than spouting nonsense.

Knowledge Editing is the most hardcore method. It involves directly modifying the model's parameters to make it "forget" incorrect information. But this is very risky – changing one parameter can affect the model's performance on other tasks. It's like brain surgery: cutting out a bad memory might take the good ones with it.

---

What Troubled Me Most: The Reasoning Paradox

The paper mentioned a phenomenon that caught my attention: reasoning models have a significantly higher hallucination rate than base models.

I tested this myself: I asked GPT-4 to do logic reasoning problems. The deeper it reasoned, the higher the hallucination rate. For example:

• Simple question: hallucination rate 5%
• Requires 3 steps of reasoning: hallucination rate 15%
• Requires 5+ steps of reasoning: hallucination rate 30%

It's like the more a person thinks, the easier it is for them to get stuck in a mental rut. FAIR (Meta's AI research team) published a paper in 2024 that used online reinforcement learning to reduce the hallucination rate by 23.1 percentage points – but the trade-off was a decrease in reasoning ability. Frustrating, isn't it? It's like putting brakes on a fast horse – it runs slower, but at least it won't crash into a wall.

---

What Should Ordinary People Do?

To be honest, there's no perfect solution right now. A 2024 paper theoretically proved that hallucination is an inherent property of autoregressive language models and cannot be completely eliminated.

Hearing this might disappoint you – but don't worry, let me share some hard-earned lessons:

1. Never trust specific numbers generated by a model. Every time I see a precise number like "47.3% growth," I manually verify it. Think about it – if the model itself believes the numbers it fabricates, why should you?

1. Ask the model to provide sources for its information. I added this line to my prompts: "Every factual statement must cite its source." Although the model sometimes fabricates sources too, it at least helps filter out some hallucinations. It's like asking a liar for evidence – they might slip up.

1. Cross-validate. Ask the same question to different models, or ask the same model three times. If the answers are inconsistent, there's definitely a