对于1，可以看我们论文里面的Appendix E, 详细地讲述了如何在sglang里面相对高效的支持这样的需求。走裸transformer慢得无法接受，走vllm只支持offline extract，但我们这套逻辑要求online serving —— sglang 改起来更简单一点，相对来说。（但也有相当 多的patch，瘫，这个feature没什么人用，很多模型适配/和其他的feature的适配都一点没做）
对于2，我们的动机就是Agent层零适配。但其实坦白讲，也不能做到完全零适配 —— 我们这个头总要给Agent留一个开关，总不能所有工具调用都进行裁剪 —— 假如把 echo 裁剪掉了是相当“令模疑惑”的。现在的实现就是模型可以直接输出一个特定的标识符“SKIP PRUNE IN NEXT X TURNS”，Agent框架检测到标识符之后bypass接下来的X轮。
Refactor本质上怎么judge的交代不清。我觉得refactor应该是用类似ProgramBench的黑盒judge方法，但粗看下来本文章还是白盒黑盒混一起，用类SWE-bench的方式judge，也没有刻意讨论说refactor judge的过程中对于主要行为和边界行为的一致性的检查的保证。感觉上是yet another SWE-bench，虽然文章批评了SWE-bench系列的issue-test不匹配的错漏，但自己对于这个问题的保证方法也就是agent核查+人工检查，和SWE-bench并没有任何的区别，为什么要相信这里的人工比SWE-bench OpenAI的人工更好呢
部分结论已经过时。例如mini swe agent不如openhands、harness对模型性能表现很重要——首先这是非常well known的结论，完全不知道这个bench带来了什么新知；其次论文里面的表格细看都有明显的反这个结论的。我们知道前沿模型对于harness的依赖在逐渐减弱，如果使用opus5或者k3之类的模型，在不同的harness和数据集上，SWE表现都相当稳定。而论文里面的最大论据是GPT 5.2的从20到40的性能变化，但同一张表格里面就有Gemini-3-Pro的反向结论，以及甚至可以发现越强的模型性能变化越少，这也符合现在的认知
Refactor真的是一个非常重要的问题吗？ 或者说，这种issue driven而不是spec driven的refactor符合真实场景吗？我们不妨回看可能甚至不太成功的rewrite Bun in Rust运动，也是一个totally spec driven的工作，并且refactor的另一个judge维度是所谓的代码整洁性和设计架构，而文章避过了这个问题，并没有给出除了测试之外更好的judge维度
from selector.rules import Quad, SoftCap, UnaryPreference, SparseBinaryPreference from selector.solver import Selector defgroup_by_repo(ds): out = defaultdict(list) for i, r inenumerate(ds["repo"]): out[r].append(i) return{k: np.asarray(v, np.int64)for k, v in out.items()} deftag_boost(tag, delta): deffn(ds): return np.where(np.array([tag in(t or[])for t in ds["tags"]]), delta,0.0).astype(np.float64) return fn res = Selector( ds, k=700, scores=scores, preferences=[ UnaryPreference(name="testing_boost", transform=tag_boost("testing",0.3)), Quad(name="repo_var", grouping=group_by_repo, alpha=0.05), ], ).solve(max_iter=30)
Sample Loss Norm: 代码片段的长短有着很大的差别，更关键的是，与文本领域大部分的信息都是稀疏不同，保留的代码行既有稀疏的片段（大部分行被删除）、又有稠密的片段（大部分行被保留），且长度上也有比较大的区别。如果简单在token level计算，模型训练相对不稳定，将每一个样本的loss使用样本的长度归一化能更好的平衡长短样本
还有一个有趣的发现是，以Qwen3-Reranker-0.6B这种LLM based reranker作为基底，它的自定义instruction能力让下游任务的泛化更加轻松了。例如代码补全任务的query是上一行代码行而不是自然语言文字，这对于大部分模型其实是难以理解的，但我们可以在instruction中描述补全任务，使得模型在训练相对少的补全任务上性能也相当不错。
To determine the optimal ingest order, we need a way to tell how similar one repository is to another (similar in terms of their content), so we invented a new probabilistic data structure to do this in the same class of data structures as MinHash and HyperLogLog. This data structure, which we call a geometric filter, allows computing set similarity and the symmetric difference between sets with logarithmic space. In this case, the sets we’re comparing are the contents of each repository as represented by (path, blob_sha) tuples. Armed with that knowledge, we can construct a graph where the vertices are repositories and edges are weighted with this similarity metric. Calculating a minimum spanning tree of this graph (with similarity as cost) and then doing a level order traversal of the tree gives us an ingest order where we can make best use of delta encoding. Really though, this graph is enormous (millions of nodes, trillions of edges), so our MST algorithm computes an approximation that only takes a few minutes to calculate and provides 90% of the delta compression benefits we’re going for.
❯ tldr grep grep Find patterns in files using regular expressions.More information: https://www.gnu.org/software/grep/manual/grep.html. - Search for a pattern within a file: grep "{{search_pattern}}" {{path/to/file}} - Search for an exact string (disables regular expressions): grep {{[-F|--fixed-strings]}} "{{exact_string}}" {{path/to/file}} - Search for a pattern in all files recursively in a directory, showing line numbers of matches, ignoring binary files: grep {{[-r|--recursive]}} {{[-n|--line-number]}} --binary-files {{without-match}} "{{search_pattern}}" {{path/to/directory}} - Use extended regular expressions (supports ?, +, {}, (), and |), in case-insensitive mode: grep {{[-E|--extended-regexp]}} {{[-i|--ignore-case]}} "{{search_pattern}}" {{path/to/file}} - Print 3 lines of [C]ontext around, [B]efore or [A]fter each match: grep --{{context|before-context|after-context}} 3 "{{search_pattern}}" {{path/to/file}} - Print file name and line number for each match with color output: grep {{[-H|--with-filename]}} {{[-n|--line-number]}} --color=always "{{search_pattern}}" {{path/to/file}} - Search for lines matching a pattern, printing only the matched text: grep {{[-o|--only-matching]}} "{{search_pattern}}" {{path/to/file}} - Search stdin for lines that do not match a pattern: cat {{path/to/file}} | grep {{[-v|--invert-match]}} "{{search_pattern}}"
Effi-code: Unleashing code efficiency in language modelsSWIFTCODER: Enhancing Code Generation in Large Language Models through Efficiency-Aware Fine-tuning
Full Sampling: 每组采 样的k个rollout计算奖励，之后整组的奖励由每个rollout的最大值给出
Bootstrap Sampling: Full Sampling虽然提高了性能，但计算量太大。为了减少推理次数，同时保持组数不变，采用bootstrap采样，先生成一个候选池，再从这个池子里面抽取k个答案，就形成了一组（这样允许某个答案被分到多个组里面重用），论文中候选池大小就和正常top1大小相同，也就是平均而言每个样本被重用k次
文章的另一个重要的实验结论是，微调student llm(θd)是必要的，之前的研究中没有相关模块，会导致性能的大幅度下降。这细想其实是一个很有趣的事情，可以注意到，压缩的时候是没有接触到query信息的（这也是为什么称为离线的原因），可以理解为某种意义上的LLM as an embedder，而加入了query和embedding再训练的时候，θd一边学会了如何理解自己产生的embedding，另一方面学会了如何根据query去选择embedding，整体上类似于ColBERT架构的Reranker（前面是multi-vec embed, 后面是maxsim）
An especially egregious form of this issue can take place with “context poisoning” – where many parts of the context (goals, summary) are “poisoned” with misinformation about the game state, which can often take a very long time to undo. As a result, the model can become fixated on achieving impossible or irrelevant goals.
The Berkeley Function-Calling Leaderboard is a tool-use benchmark that evaluates the ability of models to effectively use tools to respond to prompts. Now on its 3rd version, the leaderboard shows that every model performs worse when provided with more than one tool4. Further, the Berkeley team, “designed scenarios where none of the provided functions are relevant…we expect the model’s output to be no function call.” Yet, all models will occasionally call tools that aren’t relevant.
上下文冲突
A Microsoft and Salesforce team documented this brilliantly in a recent paper.
分阶段提供信息，模型的表现严重下降
We find that LLMs often make assumptions in early turns and prematurely attempt to generate final solutions, on which they overly rely. In simpler terms, we discover that when LLMs take a wrong turn in a conversation, they get lost and do not recover. 我们发现，LLM 们经常在早期阶段做出假设，并过早地尝试得出最终解决方案，而他们过度依赖这些解决方案。简而言之，我们发现，当 LLM 们在对话中走错方向时，他们会迷失方向，无法恢复。
Andrew Karpathy依然擅长炒作，他的观点是LLM as a new OS. Context is RAM
上下文工程：在上下文窗口中为下一步填充恰到好处的信息的科学
有哪些呢？
Instructions
Knowledge
Tools
上下文工程策略：写入上下文，选择上下文，压缩上下文，隔离上下文
写入上下文：
临时笔记板，可以是会话的状态对象，也可以是简单的工具调用写文件
Anthropic 的研究表明，将“笔记板”工具与特定领域的提示配对使用可以带来显著的收益，与专业代理的基准相比，最高可提高 54%。这也称作上下文卸载（Context Offload）, 参考 The "think" tool: Enabling Claude to stop and think \ Anthropic
Anthropic identified three scenarios where the context offloading pattern is useful:
Anthropic 确定了上下文卸载模式有用的三种场景：
Tool output analysis. When Claude needs to carefully process the output of previous tool calls before acting and might need to backtrack in its approach;
Policy-heavy environments. When Claude needs to follow detailed guidelines and verify compliance; and
Sequential decision making. When each action builds on previous ones and mistakes are costly (often found in multi-step domains).
To our knowledge, no benchmark evaluates document retrieval systems in practical settings; in an end-to-end manner, across several document types and topics, and by evaluating the use of both textual and visual document features.
classCheckCitationFaithfulness(dspy.Signature): """Verify that the text is based on the provided context.""" context:str= dspy.InputField(desc="facts here are assumed to be true") text:str= dspy.InputField() faithfulness:bool= dspy.OutputField() evidence:dict[str,list[str]]= dspy.OutputField(desc="Supporting evidence for claims") context ="The 21-year-old made seven appearances for the Hammers and netted his only goal for them in a Europa League qualification round match against Andorran side FC Lustrains last season. Lee had two loan spells in League One last term, with Blackpool and then Colchester United. He scored twice for the U's but was unable to save them from relegation. The length of Lee's contract with the promoted Tykes has not been revealed. Find all the latest football transfers on our dedicated page." text ="Lee scored 3 goals for Colchester United." faithfulness = dspy.ChainOfThought(CheckCitationFaithfulness) faithfulness(context=context, text=text)
However, traditional RAG has limitations: it usually queries a single data source and only performs one retrieval pass, so if the initial results are poor or the query is phrased oddly, the answer will suffer
但是，传统的 RAG 存在局限性：它通常查询单个数据源，并且只执行一次检索传递，因此如果初始结果不佳或查询措辞奇怪，答案将受到影响
There’s no built-in mechanism for the system to reason about how to retrieve better information or to use additional tools if needed.
系统没有内置机制来推理如何检索更好的信息或在需要时使用其他工具。
Indexing throughput depends on corpus size for Annoy and HNSW, where throughput is halved when corpus size is increased by 10x.
对于 Annoy 和 HNSW，索引吞吐量取决于语料库大小，当语料库大小增加 10 倍时，吞吐量就会减半。
Indexing throughput for RPLSH is independent of corpus size.
RPLSH 的索引吞吐量与语料库大小无关。
Annoy is 4.5 to 5 times faster than HNSW.
Annoy 比 HNSW 快 4.5 到 5 倍 。
RPLSH is 23 to 24 times faster than HNSW at 1M documents.
对于 1M 文档，RPLSH 的速度比 HNSW 快 23 到 24 倍 。
查询吞吐
Observations: 观察结果：
HNSW outperforms Annoy and RPLSH. At corpus size 1M the QPS is 9 times as high as Annoy, and 16 times as high as RPLSH at comparable quality. Similar observations between hnswlib and Annoy are found in ANN Benchmarks, where the QPS of hnswlib is 5-10 times higher at the same quality on all tested datasets.
HNSW 的表现优于 Annoy 和 RPLSH。在 1M 语料库规模下，其每秒查询速度 (QPS) 是 Annoy 的 9 倍 ，在同等质量下是 RPLSH 的 16 倍 。在 ANN 基准测试中也发现了 hnswlib 与 Annoy 之间的类似现象：在所有测试数据集上，相同质量下 hnswlib 的每秒查询速度 (QPS) 比 Annoy 高 5-10 倍。
The Annoy index is almost 3 times larger than the HNSW index, which results in ~40% more total memory usage in the 1M SIFT dataset.
Annoy 索引几乎比 HNSW 索引大 3 倍，这导致 1M SIFT 数据集的总内存使用量增加约 40%。
Both indexes are independent of dimension size, but max points in a leaf node (Annoy) and max links per level (HNSW) might need adjustments with higher dimensionality to get decent quality.
这两个索引都与维度大小无关，但叶节点中的最大点数（Annoy）和每级的最大链接数（HNSW）可能需要使用更高的维度进行调整才能获得不错的质量。
docker compose ps # 得到结果应该如下 NAME IMAGE COMMAND SERVICE CREATED STATUS PORTS ollama ollama/ollama "/bin/ollama serve" ollama About a minute ago Up About a minute 0.0.0.0:11434->11434/tcp, :::11434->11434/tcp
❯ docker exec -it ollama ollama list NAME ID SIZE MODIFIED qwen2.5:7b 845dbda0ea48 4.7 GB 2 hours ago What's next: Try Docker Debug for seamless, persistent debugging tools in any container or image → docker debug ollama Learn more at https://docs.docker.com/go/debug-cli/
from hypothesis import given, strategies as st @given(st.lists(st.integers())) deftest_sort_is_sorted(xs): result =sorted(xs) assert result ==sorted(result) assertset(result)==set(xs)
// EncodeMulti encodes a cookie value using a group of codecs. // // The codecs are tried in order. Multiple codecs are accepted to allow // key rotation. // // On error, may return a MultiError. funcEncodeMulti(name string, value interface{}, codecs ...Codec)(string,error){ iflen(codecs)==0{ return"", errNoCodecs } var errors MultiError for_, codec :=range codecs { encoded, err := codec.Encode(name, value) if err ==nil{ return encoded,nil } errors =append(errors, err) } return"", errors }
type Session struct{ // The ID of the session, generated by stores. It should not be used for // user data. ID string // Values contains the user-data for the session. Values map[interface{}]interface{} Options *Options IsNew bool store Store name string } // A Cookie represents an HTTP cookie as sent in the Set-Cookie header of an // HTTP response or the Cookie header of an HTTP request. // // See https://tools.ietf.org/html/rfc6265 for details. type Cookie struct{ Name string Value string Path string// optional Domain string// optional Expires time.Time // optional RawExpires string// for reading cookies only // MaxAge=0 means no 'Max-Age' attribute specified. // MaxAge<0 means delete cookie now, equivalently 'Max-Age: 0' // MaxAge>0 means Max-Age attribute present and given in seconds MaxAge int Secure bool HttpOnly bool SameSite SameSite Raw string Unparsed []string// Raw text of unparsed attribute-value pairs }
// Save adds a single session to the response. func(s *CookieStore)Save(r *http.Request, w http.ResponseWriter, session *Session)error{ encoded, err := securecookie.EncodeMulti(session.Name(), session.Values, s.Codecs...) if err !=nil{ return err } http.SetCookie(w,NewCookie(session.Name(), encoded, session.Options)) returnnil }
// WaitWithTimeout blocks until the WaitGroup counter is zero or timeout // returns true if timeout func(w *Wait)WaitWithTimeout(timeout time.Duration)bool{ c :=make(chanstruct{},1) gofunc(){ deferclose(c)
因此，开蒸！总之也是训了XLM-RoBERTa-large & mBERT的模型替代LLM（可以发现现在的RAG基本就是 
