AI 与机器学习2025arxiv.org
GLM-4.5：面向智能体、推理与编程（ARC）的基础模型
GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models
进阶 · 约 90 分钟读完前置知识：大模型后训练基础、强化学习（RLHF/RLVR）基础、Agent 系统基本概念
导读
GLM-4.5 是智谱与清华团队提出的面向智能体、推理与编码的基础模型（ARC Foundation Models）。论文详细披露了后训练方法：先分别训练推理、agent、chat 等专家模型，再将其能力蒸馏进统一模型；RL 实验还发现全 64K 上下文的单阶段训练优于渐进多阶段。据评论，其在 SWE-bench 上的表现可与更大或专有模型相比。
HN 关注点：这个开放模型在真实编码 agent 任务中接近 Claude 的表现，以及论文少见的后训练与 RL 工程细节披露。
HN 评论精选
这几条评论里：支持 1 质疑 2 亲历 1
亲历亲历使用称 GLM-4.5 编码能力接近 Claude 4，且更少篡改测试，稳定性优于 GPT-5。
我用 GLM-4.5 当编码模型有一阵子了，它真的非常好。在我做的编码 agent Octofriend 里，有时会把它和 Claude 4 搞混。主观体验：1）Claude 在需要跨大量上下文推理、考虑系统交互的整代码库任务上略好；2）GLM-4.5 更“诚实”——我很少见它像 Claude 那样改测试而不是修 bug 来让坏测试通过。两者都不错，GLM-4.5 还发现过 Claude 4 Sonnet 和 4.1 Opus 都没抓到的 bug。总体调试上 Claude 略胜，但很接近。和 GPT-5 比，Claude 和 GLM 更稳定；GPT-5 有时有长时间很出色的表现，代码质量主观更高，但一旦跑偏就很难拉回来。
reissbaker查看英文原文
I've been playing around with GLM-4.5 as a coding model for a while now and it's really, really good. In the coding agent I've been working on, Octofriend [1], I've sometimes had it on and confused it for Claude 4. Subjectively, my experience has been: 1. Claude is somewhat better at whole-codebase tasks, where you need to reason over a bunch of context and consider system interactions. 2. GLM-4.5 is somewhat better at being "honest" — i.e. I rarely see it doing the things Claude does like making broken tests pass by changing the test instead of fixing the bug. Both are quite good though, and GLM-4.5 has found bugs that both Claude 4 Sonnet and 4.1 Opus have failed to catch. In general I think Claude wins a little more frequently on debugging tasks than GLM-4.5, but it's close. Compared to GPT-5, both Claude and GLM feel like they're more consistent, although GPT-5 sometimes has long brilliant runs where it nails everything with subjectively higher code quality than either of the latter. However, once GPT-5 goes off the rails, it's hard to get it back on track, so it can be a bit frustrating to work with in comparison. 1: https://github.com/synthetic-lab/octofriend
质疑展示模型在简单计数任务上产生冗长、反复纠结的思考链。
我让它数“bluberry”里有几个 b，它的思考链非常冗长：先纠结这是不是“blueberry”的拼写错误，再逐字母数，又回到正确拼写重新数，绕了很久。
sriku查看英文原文
Ok so I asked "how many bs in bluberry" and it "thought" the following - ----- First, the question is: "How many bs in bluberry?" That seems like a typo. I think it's meant to be "blueberry," but it's spelled "bluberry" with one 'e' missing. Also, "bs" probably refers to the letter 'b'. So, rephrasing the question: How many times does the letter 'b' appear in the word "blueberry"? But the word given is "bluberry," which might be a misspelling of "blueberry." Let me confirm the correct spelling. The standard spelling is "blueberry," which has two 'e's: b-l-u-e-b-e-r-r-y. The question says "bluberry," which is missing the second 'e', so it's spelled b-l-u-b-e-r-r-y. That might be intentional or a mistake. I should answer based on the word as given. The question is: "How many bs in bluberry?" So, I need to count the 'b's in "bluberry." Let me write it out: B-L-U-B-E-R-R-Y. Now, I'll go through each letter: Position 1: B (that's one b) Position 2: L (not b) Position 3: U (not b) Position 4: B (that's the second b) Position 5: E (not b) Position 6: R (not b) Position 7: R (not b) Position 8: Y (not b) So, there are two 'b's in "bluberry." But is "bluberry" the correct word? It's probably a misspelling of "blueberry." In "blueberry," it's B-L-U-E-B-E-R-R-Y. Position 1: B (one) Position 2: L Position 3: U Position 4: E Position 5: B (two) Position 6: E Position 7: R Position 8: R Position 9: Y "Blueberry" has nine letters, and there are two 'b's: one at the start and one after the 'e'. In "bluberry," as given, it's eight letters: B-L-U-B-E-R-R-Y, with two 'b's. Perhaps the question is tricking me with the misspelling. But I should answer based on what's written. Another thought: "bs" might be plural for 'b', meaning how many 'b's, or it could be something else. But in context, it's likely about the letter. Maybe it's a riddle. "Bluberry" sounds like "blueberry," but spelled differently. Perhaps "bs" refers to something else, like the fruit or something, but that seems unlikely. The question says "bs," which is probably "b's" meaning the letter b. Let me read the question carefully: "How many bs in bluberry?" It might be a play on words. "Bluberry" could be intended as "blueberry," and "bs" might be misinterpreted. Another idea: in some contexts, "b" could stand for something, but here it's probably straightforward. Perhaps it's about the sound. But the question is written, so it's about spelling. I think I should just count the 'b's in the given word. So, in "bluberry": letters are B, L, U, B, E, R, R, Y. So, B appears twice. Therefore, there are two 'b's. But to be thorough, let's consider if "bluberry" is a word. It's not standard; it's a misspelling. However, for the purpose of this question, I'll use it as is. Perhaps the question has a typo, and it's meant to be "blueberry." But since it's written as "bluberry," I'll go with that. Final answer should be 2. The question says "bs," which might be lowercase, but in counting, case usually doesn't matter, and here it's all lowercase or mixed? The word is "bluberry," which is lowercase, and "bs" is lowercase, so it's fine. In English, when counting letters, we don't distinguish case unless specified. So, I think two is correct. But let me double-check the spelling. The user wrote "bluberry," which has one 'u' and no 'e' after 'u', but in "blueberry," it's "blue" with 'e'. In "bluberry," it's "blu" without 'e', then "berry." But in terms of letter count, it's clear. Perhaps it's "blueberry" misspelled, and I should correct it, but I think for accuracy, I'll use the given spelling. Another thought: the question might be "how many b's in blueberry?" but written with a typo. But since it's "bluberry," I'll answer based on that. So, response should be: There are two 'b's in "bluberry." To make it clear, I can say: The word "bluberry" contains the letter 'b' twice. But the questio…
质疑质疑论文在编码基准中唯独不列入 Qwen3 的选择性对比。
为什么编码基准里省略了 Qwen3，而其他基准没有？
lumost查看英文原文
Why was qwen3 omitted from the coding benchmark but not other benchmarks?
支持专业解读：专家模型蒸馏的统一后训练、单阶段 64K RL 优于多阶段、XML 函数调用模板等细节值得关注。
很欣赏这篇论文的深度，比常见的模型发布博客强多了。智谱/清华团队不仅讲了“是什么”，还讲了“怎么做”。后训练方法（第3节）最突出：为推理、agent、chat 分别训练专家模型，再把能力蒸馏进最终统一模型，像是让通才向专家委员会学习，比简单混数据更有结构。他们 RL 实验的两个发现很宝贵：全 64K 上下文的单阶段 RL 优于渐进多阶段（图6），这跟很多团队的假设相反；用类 XML 模板做函数调用以避开 JSON 转义地狱（图4）是实用的小而妙的工程决策。SWE-bench 表现也让它与更大或专有模型同列。
starchild3001查看英文原文
Really appreciate the depth of this paper; it's a welcome change from the usual model announcement blog posts. The Zhipu/Tsinghua team laid out not just the 'what' but the 'how,' which is where the most interesting details are for anyone trying to build with or on top of these models. The post-training methodology (Sec 3) is what really stands out to me. The idea of creating specialized 'expert models' for reasoning, agents, and chat, and then distilling their capabilities into a final unified model is a fascinating approach. It feels like a more structured way to solve the "jack of all trades, master of none" problem that can plague generalist models. Instead of just mixing all the data, they're essentially having a generalist learn from a committee of specialists. A couple of the findings from their RL experiments are pure gold for anyone working in this space. The counter-intuitive result that a single-stage RL process at the full 64K context length outperforms a progressive, multi-stage approach (Fig 6) is a fantastic lesson. I've seen teams assume the opposite would be true. Also, the pragmatic choice to use an XML-like template for function calls to avoid JSON escaping hell (Fig 4) may be a small but brilliant engineering decision that makes a huge difference in practice. Wrangling escaped code inside JSON turns out to be a mess. The performance on SWE-bench is impressive, putting it in the same league as much larger or proprietary models. What I’d love to see, and maybe others here have thoughts, is whether this hybrid training recipe holds up outside ARC-style evals. For example, do the agentic improvements transfer to messier, real-world workflows where APIs are undocumented, partial failures are common, and user input is full of ambiguity?
相关资源
从这篇论文所有 HN 讨论帖的全部评论里整理出来的链接。
历次讨论
- 2025Hacker News 讨论帖417 分 · 83 评
同领域还值得读
DeepSeek-R1：通过强化学习激励 LLM 的推理能力
DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via RL
研究如何用强化学习激励大语言模型的推理能力，提出 DeepSeek-R1，并探讨 RL 训练对模型推理表现的影响。
AI 与机器学习2025 · 1,351 分 · 1,056 评 · 引用 30进阶 · 约 80 分钟编辑精选1-bit LLM 时代：面向低成本计算的三值参数
The Era of 1-bit LLMs: ternary parameters for cost-effective computing
提出用三值参数 {-1, 0, 1} 替代 LLM 中的浮点权重，即所谓“1-bit LLM”。
AI 与机器学习2024 · 1,040 分 · 447 评 · 引用 19进阶 · 约 45 分钟编辑精选DeepSeek-V3.2：推进开放大语言模型的前沿
DeepSeek-v3.2: Pushing the frontier of open large language models
这是 DeepSeek-V3.2 的技术报告，延续了 DeepSeek 开源权重模型的路线，目标是让开放模型继续逼近甚至追平前沿闭源模型。
AI 与机器学习2025 · 982 分 · 465 评进阶 · 约 90 分钟编辑精选通过自我对弈与通用强化学习掌握国际象棋和将棋（AlphaZero）
Mastering Chess and Shogi by Self-Play with General Reinforcement Learning
DeepMind 的 AlphaZero 论文：用通用强化学习算法（自我博弈 + MCTS + 神经网络），从零开始仅数小时训练便在国际象棋和将棋上击败了世界顶级引擎 Stock…
AI 与机器学习2017 · 539 分 · 270 评 · 引用 13进阶 · 约 90 分钟编辑精选用强化学习发现更快的矩阵乘法算法
Discovering faster matrix multiplication algorithms with reinforcement learning
论文提出 AlphaTensor，用强化学习搜索矩阵乘法算法。
AI 与机器学习2022 · 518 分 · 111 评专业 · 约 90 分钟编辑精选