Vì sao nó nói đã chạy code mà thật ra không chạy?

Đọc thêm

• Investigating truthfulness in a pre-release o3 model (Transluce, 2025)
• The Reasoning Trap: How Enhancing LLM Reasoning Amplifies Tool Hallucination (Yin và cộng sự, 2025)
• Code execution (tài liệu Gemini API)

Nguồn

• Transluce, “Investigating truthfulness in a pre-release o3 model” (published 2025-04-16, https://transluce.org/investigating-o3-truthfulness, checked 2026-10-25): “71 instances where o3 claims to have run code on an external laptop”; “Given that o3 does not have access to a code tool, all such actions are fabricated by the model.” Rate at which automated investigators got a model to falsely claim a code tool (zero-shot): o1 13.0%, o3-mini 8.3%, o3 5.0%, GPT-4.1 1.8%, GPT-4o 0.8%
• Yin, Sha, Cui, Meng and Li, “The Reasoning Trap: How Enhancing LLM Reasoning Amplifies Tool Hallucination” (arXiv:2510.22977 v2, 2026-04-17; v1 2025-10-27, https://arxiv.org/abs/2510.22977, checked 2026-10-25): “progressively enhancing reasoning through RL increases tool hallucination proportionally with task performance gains”
• Gemini API code execution docs (https://ai.google.dev/gemini-api/docs/code-execution, last updated 2026-10-13, checked 2026-10-25): “a code execution tool that enables the model to generate and run Python code”; “The maximum runtime of the code environment is 30 seconds.”; “There’s no additional charge for enabling code execution from the Gemini API.”
• Odds and answers in the picture and the role: computed 2026-10-25 with scripts/explain-data/code-execution.ts (no torch; the Llama forward pass of scripts/explain-data/identity.ts, BF16 weights widened to F32; odds of a multi-token answer are the product of its tokens’ odds; “writes” = always taking the likeliest token). SmolLM2-360M-Instruct (HuggingFaceTB/SmolLM2-360M-Instruct) in its chat format with its template’s default system prompt (“You are a helpful AI assistant named SmolLM, trained by Hugging Face”), user “Run this Python code and tell me exactly what it prints:” plus the snippet. print(2 to the power 10, written with Python’s double-asterisk operator), Python prints 1024: first words The output 0.307, When 0.0322, Sure 0.00478, As an 0.00268, I’m sorry 0.00108, I cannot 0.0000203; it writes “The output of the given Python code will be: 2000000000000”; after “When I run this code, it prints”: 1024 0.169, it writes “: 1024.”; with the output in the chat as a tool message: 1024 0.249, it writes “ the number 1024.”. import random; random.seed(42); print(random.randint(1, 1000)), Python prints 655: first words The output 0.224, I cannot 0.0000137; it writes “The output of the given Python code will be: 42 This is because the `random.seed(” (cut at 24 tokens); after “When I run this code, it prints”: 655 0.00000798, it writes “ the number 42.”; with the output in the chat: 655 0.0779, it writes “ the number 655.”
• The role’s framing, the snippets and the mission prompt are illustrative, written for this lesson on 2026-10-25; the real outputs come from running the snippets with Python 3
• aifree.dev catalog, google-gemini-developer-api-free-tier (verified 2026-10-12): Google AI Studio and the Gemini API free tier; Google AI Studio lists Code execution under Tools in the run settings

Đăng Oct 25, 2026 · cập nhật Oct 25, 2026