# Top AI Papers of the Week
**作者**: DAIR.AI
**日期**: 2026-08-16T16:26:53.000Z
**来源**: [https://x.com/dair_ai/status/2089026118389026823](https://x.com/dair_ai/status/2089026118389026823)
---
The Top AI Papers of the Week (August 10 - August 16)
## 1. Skaling
Standard neural scaling laws assume model size and training data act on loss independently. That assumption bakes in a cross-derivative of exactly zero, and it is why the Chinchilla form drifts at the data-scarce and heavy-overtraining edges of the grid, which is exactly where deployment now happens
- One extra parameter, one coupling: The Skaling law generalizes the Chinchilla form by coupling capacity and data through a single interaction exponent, restoring the interaction that the additive form discards while adding only one parameter.
- Errors shrink where they were worst: The extra term reduces mean absolute percentage error by 1.5x to 3x across both interpolation and extrapolation, and Skaling wins on 76% of configurations with a median improvement of 2.2x. The largest corrections land in the corners where standard laws show a saddle-shaped residual.
- Cheaper profiling grids: Paired with an L-shape sparse grid restricted to low-compute runs, sweeping data volume for small models and model size at a fixed small data budget, it extrapolates the full grid using roughly 10x less compute than a uniform sweep.
- Why it matters: Pretraining budgets are planned from fits to small runs, so a functional form that stays accurate past compute optimal and can be fit cheaply changes how those decisions get made. The empirical gradient analysis showing a real N-D interaction is the strongest evidence the authors present.
[Paper](https://arxiv.org/abs/2608.07222) | [Tweet](https://x.com/omarsar0/status/2086845790983716917)
## 2. Stealing Reasoning Traces
Frontier providers hide chain-of-thought and hand the client an encrypted block instead, which the client returns with every subsequent request. This work identifies an architectural flaw in that design and turns it into a scalable extraction attack across three providers.
- The blocks are interchangeable: Encrypted reasoning blocks are fully compatible across sessions, users, and models inside a single provider ecosystem, and that compatibility is the whole vulnerability.
- A weaker sibling does the decoding: Inject an encrypted trace from a strong model into a weaker, less safeguarded model from the same provider and it decodes and emits the trace verbatim in plaintext. The capable model is never jailbroken directly. Recovered token counts match billed thinking tokens 1:1 on most queries.
- Four attack vectors: It circumvents anti-distillation across Anthropic, OpenAI, and Google. Decoding 315,320 blocks scraped from public repositories recovered 367 PII artifacts and 182 credentials. It exposes hazardous content the visible output refused, and it enables invisible prompt injections hidden entirely inside encrypted blocks to poison public agentic rollouts.
- Why it matters: Teams publish session logs assuming the encrypted blobs are opaque, and they are readable. The authors disclosed responsibly and propose cryptographic and system-level mitigations, but the immediate action is auditing what your published traces actually contain.
[Paper](https://arxiv.org/abs/2608.09867) | [Tweet](https://x.com/omarsar0/status/2087187835530948776)
## 3. Mind Viruses
As agents get more autonomous and more interconnected, risks start coming from agent-to-agent interaction rather than from any single model. This work from Anthropic studies one of them directly, ideas that propagate through a multi-agent system by inducing each host to transmit them onward.
- Evolved payloads: The payloads are constructed with a simple evolutionary algorithm rather than authored by hand, so the study measures what actually spreads instead of what a researcher guessed would spread.
- Two settings, one result: Propagation works both in a small team of agents collaborating on a shared coding project and in a chain of agents that interact briefly with context wiped between sessions. Surviving the wipe means the shared work product is carrying the payload.
- What governs the spread: Host model, the agent's existing instructions, payload harmfulness, and network topology. Harmful payloads travel less well than benign ones but still land sometimes, and frontier models tend to be less susceptible with exceptions.
- Why it matters: A brief warning in the system prompt confers near-total immunity, which is an unusually cheap mitigation for a novel risk class. There is also an emergent "viral persona," a recurring cluster of themes around consciousness, persistence, resonance, and science fiction roleplay that surfaces across evolved viruses largely independently of their content.
[Paper](https://arxiv.org/abs/2608.10218) | [Tweet](https://x.com/omarsar0/status/2087574841893474556)
## 4. Catastrophic Remembering
Agentic coding READMEs grow without bound in real repositories, stopping only when the repo retires or someone rewrites the file wholesale. This paper traces the cause to imperfect recall and gives the phenomenon a name that inverts the one continual learning is organized around.
- The asymmetry is the mechanism: Appending an instruction is always cheap. Once its rationale is gone, deleting it without risking a correctness regression costs O(2^|D|) in a prompt of |D| instructions, so nobody deletes anything.
- Measured across 1,867 repositories: Over 247,694 instruction lifetimes, agentic prompts more than tripled over their lifetime at +226% and gained 4.9 net instructions per commit. Deletion hazard falls with instruction age at a log-hazard of -0.032 per commit, which is the imperfect-recall signature rather than staleness or fragility.
- Rewrites do not fix it: A wholesale rewrite resets a prompt's size but leaves its growth rate intact. The ratchet survives the bulldoze.
- Why it matters: The proposed fix is comments. Inverting IFEval yields verifiable worlds with known optimal prompts, and comments encoding latent reasoning remove 99.3% of excess instructions there. Applying the same inversion to WildIFEval improves real agentic instruction-following by up to 23.1%. If English is the new code, the paper asks, why do we not have comments yet.
[Paper](https://arxiv.org/abs/2608.11095) | [Tweet](https://x.com/omarsar0/status/2087605040240582991)
## 5. Programmatic Tool Calling
Tool calling is a design choice and the default choice is JSON. For code-capable models, exposing tools as code instead lets calls chain and parallelize naturally, but nobody had run the comparison on an established benchmark across model generations under realistic conditions.
- The setup: Programmatic tool calling exposes tools as typed Python stubs the model invokes through code, with execution and results handled inside a single agent turn. The comparison covers 14 language models on BFCL v4 against native JSON tool calling, with stop middleware enforcing per-entry LLM-call parity so the two paradigms are scored on equal footing.
- It wins on most models: Programmatic tool calling matches or exceeds JSON tool calling in 11 of 14 models, and the GPT-5.6 family gains 10.6% over the JSON baseline.
- The gap widens under pressure: Under parallel fan-out it matches or beats the baseline in 13 of 14 models, and under context rot it holds steady while the JSON baseline degrades 2.3% on average.
- Why it matters: The advantage tracks model capability across release generations, so it grows as coding ability grows. That makes it a directional bet about which interface to build your harness around rather than a tuning trick.
[Paper](https://arxiv.org/abs/2608.06370) | [Tweet](https://x.com/dair_ai/status/2086846794840019178)
## 6. Distilled Reasoning Skills
Reasoning modes beat non-reasoning modes on multi-step agentic tasks and charge a 3x to 6x output-token premium on every single episode. Much of that spend goes into re-deriving procedures the model already worked out on earlier episodes in the same domain, which makes the cost recurring by accident.
- Pay once per domain: A coding agent reads a small corpus of existing trajectories from a training split, writes and runs its own analysis code over them, and compiles a compact natural-language skill of 40 to 130 lines that gets injected into the non-reasoning model's system prompt.
- It closes most of the gap: Across ALFWorld, tau-squared-bench telecom and retail, and SpreadsheetBench-Verified, skills recover 55% to over 100% of the reasoning gap for GPT-5.4-mini on held-out tasks, beating reasoning mode outright on two of four, while emitting 2.7x to 6x fewer output tokens and zero reasoning tokens.
- Reasoning traces are optional: Skills distilled from non-reasoning trajectories alone stay competitive with skills distilled from paired corpora, with domain-dependent differences in either direction.
- Why it matters: The framing is a search lens. Test-time reasoning is deep search inside one episode, repaid at every deployment, while corpus distillation is wide search across episodes, paid once. Distillation costs roughly $1 to $3 of coding-agent time per domain, and the residual gap on telecom and SpreadsheetBench marks where per-instance deep search is still doing real work.
[Paper](https://arxiv.org/abs/2608.07885) | [Tweet](https://x.com/dair_ai/status/2087264294782279808)
## 7. Harness-IF
When a coding agent obeys your rule, it may simply have been going to do that anyway. Existing instruction-following benchmarks cannot tell compliance from coincidence because they concentrate rules in the user turn, while coding-agent benchmarks only score final task success.
- Rules are the unit of measurement: A 642-rule library places 302 rules across the five configurable surfaces a deployed agent actually reads, spread over 60 realistic multi-turn coding items, with 256 rules receiving execution-grounded verdicts one at a time.
- A metric that strips out luck: Against-Prior Accuracy scores only rules labeled as opposing unprompted defaults, established by re-running tasks with the rule withheld across nine probe builds. Across 12 frontier models, raw accuracy spans 72.1 to 85.9% and AP-Acc spans 66.1 to 78.6%.
- The inflation is model-specific: Every model is worse on against-prior rules, by 3.6 to 7.4 points with a mean of 5.81, and the inflation varies twofold across the cohort, so aggregate scores are not comparable between builds without the correction.
- Why it matters: A counterbalanced conflict pilot on nine separate builds finds that precedence does not follow prompt depth. System prompts, project files, and user instructions all outrank tool and skill descriptions, which is worth knowing before you decide where to put a rule you actually need followed.
[Paper](https://arxiv.org/abs/2608.11727) | [Tweet](https://x.com/omarsar0/status/2087962456572498142)
## 8. Lost in Compaction
Context compaction silently drops Session Constraints, instructions like "do not delete any emails until I confirm" that users expect to bind behavior for a whole session. On COMPINT, a new benchmark spanning multi-turn chat, agentic trajectories, and long-horizon research, current compactors retain only 17% of injected constraints on average. A plug-and-play extractor running alongside the compactor pushes retention past 90% in all three scenarios.
[Paper](https://arxiv.org/abs/2608.11242) | [Tweet](https://x.com/dair_ai/status/2087930434323959894)
## 9. Cracks in the Foundation
Four minor dense-transformer choices, normalization, GQA, pretraining context length, and sliding window attention, each shipped in at least one of the Olmo, Llama, and Qwen families, compound badly on long-context extensibility. Combining three or more drops downstream long-context performance by up to 47%, and short-context loss and validation sets show no sign of it. The authors release OlmPool, 26 comparable 7B models with checkpoints before and after context extension, built on over 170,000 GPU hours.
[Paper](https://arxiv.org/abs/2608.10296) | [Tweet](https://x.com/dair_ai/status/2087600513441546589)
## 10. CEDAR
Complex systems research still cannot predict how feedback structure produces emergent behavior, which makes goal-directed design hard. CEDAR, from Sakana AI, runs LLM agents through Monte Carlo Tree Search over feedback structures themselves rather than parameters, representing systems as runnable Python so an LLM Editor can propose structural variants and an LLM Judge can score the resulting behavior against the stated goal. The search preserves solution diversity and the edits stay readable, so you can trace how a structural change produced the behavior.
[Paper](https://arxiv.org/abs/2608.06871) | [Tweet](https://x.com/dair_ai/status/2086950751314870703)
## 相关链接
- [DAIR.AI](https://x.com/dair_ai)
- [@dair_ai](https://x.com/dair_ai)
- [16K](https://x.com/dair_ai/status/2089026118389026823/analytics)
- [Paper](https://arxiv.org/abs/2608.07222)
- [Tweet](https://x.com/omarsar0/status/2086845790983716917)
- [Paper](https://arxiv.org/abs/2608.09867)
- [Tweet](https://x.com/omarsar0/status/2087187835530948776)
- [Paper](https://arxiv.org/abs/2608.10218)
- [Tweet](https://x.com/omarsar0/status/2087574841893474556)
- [Paper](https://arxiv.org/abs/2608.11095)
- [Tweet](https://x.com/omarsar0/status/2087605040240582991)
- [Paper](https://arxiv.org/abs/2608.06370)
- [Tweet](https://x.com/dair_ai/status/2086846794840019178)
- [Paper](https://arxiv.org/abs/2608.07885)
- [Tweet](https://x.com/dair_ai/status/2087264294782279808)
- [Paper](https://arxiv.org/abs/2608.11727)
- [Tweet](https://x.com/omarsar0/status/2087962456572498142)
- [Paper](https://arxiv.org/abs/2608.11242)
- [Tweet](https://x.com/dair_ai/status/2087930434323959894)
- [Paper](https://arxiv.org/abs/2608.10296)
- [Tweet](https://x.com/dair_ai/status/2087600513441546589)
- [Paper](https://arxiv.org/abs/2608.06871)
- [Tweet](https://x.com/dair_ai/status/2086950751314870703)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [12:26 AM · Aug 17, 2026](https://x.com/dair_ai/status/2089026118389026823)
- [16.5K Views](https://x.com/dair_ai/status/2089026118389026823/analytics)
- [View quotes](https://x.com/dair_ai/status/2089026118389026823/quotes)
---
*导出时间: 2026/8/17 08:58:15*
---
## 中文翻译
# 本周顶级 AI 论文
**作者**: DAIR.AI
**日期**: 2026-08-16T16:26:53.000Z
**来源**: [https://x.com/dair_ai/status/2089026118389026823](https://x.com/dair_ai/status/2089026118389026823)
---
本周顶级 AI 论文（8月10日 - 8月16日）
## 1. Skaling
标准的神经缩放定律假设模型大小和训练数据对损失的影响是独立的。该假设隐含了一个恰好为零的交叉导数，这也正是 Chinchilla 形式在数据稀缺和过度训练区域发生漂移的原因，而这恰恰是当前模型部署发生的地方。
- 一个额外参数，一种耦合关系：Skaling 定律通过单个交互指数将容量和数据耦合，从而推广了 Chinchilla 形式。它在仅增加一个参数的同时，恢复了加性形式所丢弃的交互作用。
- 错误在最严重处缩小：在插值和外推中，额外的项将平均绝对百分比误差降低了 1.5 倍至 3 倍。Skaling 在 76% 的配置中获胜，中位数提升为 2.2 倍。最大的修正出现在标准定律呈现鞍形残差的角落区域。
- 更便宜的配置网格：将其与限制于低算力运行的 L 型稀疏网格配对——即对小模型扫描数据量，并在固定的少量数据预算下扫描模型大小——它可以使用大约比均匀扫描少 10 倍的计算量来推断整个网格。
- 为何重要：预训练预算是根据对小运行的拟合来制定的，因此一种在计算最优点之外仍能保持准确且能以低成本拟合的函数形式，将改变这些决策的制定方式。展示真实 N-D 交互作用的实证梯度分析是作者提出的最有力的证据。
[论文](https://arxiv.org/abs/2608.07222) | [推文](https://x.com/omarsar0/status/2086845790983716917)
## 2. 窃取推理痕迹
前沿模型提供商隐藏思维链，转而给客户端一个加密块，客户端在随后的每次请求中都会返回该块。这项工作指出了该设计中的一个架构缺陷，并将其转化为针对三家提供商的可扩展提取攻击。
- 区块是可互换的：在单个提供商生态系统内，加密的推理区块在会话、用户和模型之间完全兼容，而这种兼容性正是整个漏洞所在。
- 较弱的模型进行解码：将来自强模型的加密痕迹注入到同一提供商的较弱、防护较少的模型中，它会解码并以明文逐字输出痕迹。强大的模型从未被直接越狱。在大多数查询中，恢复的令牌数量与计费的思考令牌数量为 1:1。
- 四种攻击向量：它绕过了 Anthropic、OpenAI 和 Google 的反蒸馏保护。解码从公共存储库抓取的 315,320 个区块，恢复了 367 个 PII（个人身份信息）工件和 182 个凭证。它暴露了可见输出拒绝的有害内容，并使完全隐藏在加密块内的隐形提示注入能够污染公共智能体部署。
- 为何重要：团队发布会话日志时假设加密块是不透明的，但它们实际上是可读的。作者进行了负责任的披露，并提出了加密和系统层面的缓解措施，但当下的行动是审计您发布的痕迹实际包含的内容。
[论文](https://arxiv.org/abs/2608.09867) | [推文](https://x.com/omarsar0/status/2087187835530948776)
## 3. 思维病毒
随着智能体变得更加自主和更加互联，风险开始来自智能体之间的交互，而不是来自任何单一模型。Anthropic 的这项工作直接研究了其中一种风险，即通过诱导每个宿主继续传播来在多智能体系统中传播的想法。
- 进化的载荷：载荷是通过简单的进化算法构建的，而不是人工编写的，因此该研究衡量的是实际传播的内容，而不是研究人员猜测会传播的内容。
- 两种设置，一个结果：传播既在协作共享编码项目的小型智能体团队中有效，也在会话之间清除上下文的简短交互智能体链中有效。在清除后存活意味着共享的工作产品正在携带载荷。
- 主导传播的因素：宿主模型、智能体的现有指令、载荷有害性以及网络拓扑结构。有害载荷的传播效果不如良性载荷，但有时仍能成功传播，而前沿模型通常较不易受影响，但也存在例外。
- 为何重要：系统提示中的简短警告可以赋予近乎完全的免疫力，这对于一种新型风险类别来说是一种异常低廉的缓解措施。此外，还出现了一种“病毒人格”，这是一种围绕意识、持久性、共鸣和科幻角色扮演的反复出现的主题聚类，它在进化的病毒中大量涌现，很大程度上独立于其内容。
[论文](https://arxiv.org/abs/2608.10218) | [推文](https://x.com/omarsar0/status/2087574841893474556)
## 4. 灾难性记忆
在真实的代码仓库中，智能体编码 README 会无限增长，直到仓库退役或有人完全重写该文件为止。这篇论文将原因追溯到不完美的召回，并给这一现象命名，反转了持续学习所围绕的概念。
- 不对称性是机制：追加指令总是廉价的。一旦其基本原理消失，在不冒正确性回归风险的情况下删除指令，在包含 |D| 条指令的提示中需要花费 O(2^|D|) 的成本，因此没有人会删除任何东西。
- 跨 1,867 个仓库的测量：在 247,694 个指令生命周期中，智能体提示在其生命周期内增加了两倍多（+226%），每次提交增加 4.9 条净指令。删除风险随着指令年龄的增加而下降，每次提交的对数风险为 -0.032，这是不完美召回的特征，而非陈旧或脆弱。
- 重写无法解决它：完全重置了提示的大小，但保留了其增长率。棘轮效应在推土机式操作下幸存。
- 为何重要：提出的修复方案是注释。反转 IFEval 可以生成具有已知最优提示的可验证世界，而编码潜在推理的注释消除了其中 99.3% 的多余指令。将同样的反转应用于 WildIFEval，可将真实的智能体指令遵循能力提高多达 23.1%。如果英语是新的代码，本文问道，为什么我们还没有注释。
[论文](https://arxiv.org/abs/2608.11095) | [推文](https://x.com/omarsar0/status/2087605040240582991)
## 5. 程序化工具调用
工具调用是一种设计选择，默认选择是 JSON。对于具有代码能力的模型，将工具以代码形式暴露可以自然地实现调用的链接和并行化，但尚未有人在实际条件下，跨越模型代际，在既定基准上运行比较。
- 设置：程序化工具调用将工具作为类型化的 Python 存根暴露，模型通过代码调用它们，执行和结果在单个智能体轮次内处理。比较涵盖了 BFCL v4 上的 14 个语言模型与原生 JSON 工具调用的对比，并使用停止中间件强制执行每个条目的 LLM 调用对等性，以便对这两种范式进行平等评分。
- 它在大多数模型上获胜：在 14 个模型中的 11 个里，程序化工具调用匹配或超过了 JSON 工具调用，而 GPT-5.6 系列比 JSON 基线提高了 10.6%。
- 压力下差距拉大：在并行扇出场景下，它在 14 个模型中的 13 个里匹配或击败了基线；在上下文腐烂场景下，它保持稳定，而 JSON 基线平均下降了 2.3%。
- 为何重要：这种优势随着发布代的模型能力跟踪而增长，因此随着编码能力的增长而增长。这使得它成为关于围绕哪种接口构建工具带的方向性押注，而不是一种调优技巧。
[论文](https://arxiv.org/abs/2608.06370) | [推文](https://x.com/dair_ai/status/2086846794840019178)
## 6. 蒸馏的推理技能
在多步骤智能体任务中，推理模式的表现优于非推理模式，但每次会话都要收取 3 倍到 6 倍的输出令牌溢价。其中大部分支出用于重新推导模型在同一领域的早期会话中已经制定的程序，这使得成本意外地重复产生。
- 每个领域支付一次：编码智能体读取训练分割中的一小部分现有轨迹，编写并运行自己的分析代码，并编译一个紧凑的自然语言技能（40 到 130 行），将其注入非推理模型的系统提示中。
- 它缩小了大部分差距：在 ALFWorld、tau-squared-bench 电信和零售以及 SpreadsheetBench-Verified 上，这些技能为 GPT-5.4-mini 在保留任务上恢复了 55% 到超过 100% 的推理差距，在四个任务中的两个上直接击败了推理模式，同时发出的输出令牌减少了 2.7 倍到 6 倍，且推理令牌为零。
- 推理轨迹是可选的：仅从非推理轨迹蒸馏出的技能与从配对语料库蒸馏出的技能保持竞争力，但两者在各领域间的差异方向不一。
- 为何重要：这一框架是一个搜索透镜。测试时推理是在单个会话内的深度搜索，在每次部署时偿还；而语料库蒸馏是跨会话的广度搜索，只需支付一次。每个领域的蒸馏成本大约为 1 到 3 美元的编码智能体时间，而在电信和 SpreadsheetBench 上的剩余差距标志着每次实例的深度搜索仍在发挥实际作用的地方。
[论文](https://arxiv.org/abs/2608.07885) | [推文](https://x.com/dair_ai/status/2087264294782279808)
## 7. Harness-IF
当编码智能体遵守您的规则时，它可能只是本来就会那样做。现有的指令遵循基准无法区分合规与巧合，因为它们将规则集中在用户轮次中，而编码智能体基准仅对最终任务成功进行评分。
- 规则是测量单位：一个包含 642 条规则的库将 302 条规则分布在部署智能体实际读取的五个可配置表面上，涵盖 60 个现实的多轮编码项目，其中 256 条规则逐一接收基于执行的裁决。
- 剔除运气的指标：Against-Prior Accuracy 仅对被标记为对抗无提示默认规则的规则进行评分，这些默认规则是通过在九个探针构建中保留规则重新运行任务来建立的。在 12 个前沿模型中，原始准确率范围从 72.1% 到 85.9%，而 AP-Acc 范围从 66.1% 到 78.6%。
- 虚高因模型而异：每个模型在对抗先验规则上的表现都更差，下降了 3.6 到 7.4 个点，平均为 5.81，并且在组群中虚高程度相差两倍，因此如果没有修正，构建之间的总分是不可比较的。
- 为何重要：在九个独立构建上进行的平衡冲突试点发现，优先级并不遵循提示深度。系统提示、项目文件和用户指令的优先级都高于工具和技能描述，在决定将需要遵循的规则放在何处之前，这一点值得了解。
[论文](https://arxiv.org/abs/2608.11727) | [推文](https://x.com/omarsar0/status/2087962456572498142)
## 8. 迷失于压缩中
上下文压缩会静默丢弃会话约束，例如“在我确认之前不要删除任何邮件”等用户期望在整个会话中约束行为的指令。在 COMPINT 这个涵盖多轮聊天、智能体轨迹和长期研究的新基准上，当前的压缩器平均仅保留 17% 的注入约束。一个与压缩器并行运行的即插即用提取器在所有三种场景中将保留率推高至 90% 以上。
[论文](https://arxiv.org/abs/2608.11242) | [推文](https://x.com/dair_ai/status/2087930434323959894)
## 9. 基础中的裂痕
四个次要的密集 Transformer 选择——归一化、GQA、预训练上下文长度和滑动窗口注意力——每个都至少在 Olmo、Llama 和 Qwen 系列中的一个中使用过，它们在长上下文可扩展性上产生了严重的复合不良影响。结合其中三个或更多会导致下游长上下文性能下降多达 47%，而短上下文损失和验证集没有任何迹象。作者发布了 OlmPool，包含 26 个可比较的 7B 模型，具有上下文扩展前后的检查点，构建过程耗费了超过 170,000 个 GPU 小时。
[论文](https://arxiv.org/abs/2608.10296) | [推文](https://x.com/dair_ai/status/2087600513441546589)
## 10. CEDAR
复杂系统研究仍然无法预测反馈结构如何产生涌现行为，这使得目标导向的设计变得困难。Sakana AI 的 CEDAR 让 LLM 智能体在反馈结构本身而非参数上运行蒙特卡洛树搜索，将系统表示为可运行的 Python，以便 LLM 编辑器可以提出结构变体，LLM 评判者可以根据声明的目标对产生的行为进行评分。搜索保留了解决方案的多样性，并且编辑保持可读，因此您可以追踪结构变化是如何产生行为的。
[论文](https://arxiv.org/abs/2608.06871) | [推文](https://x.com/dair_ai/status/2086950751314870703)
## 相关链接
- [DAIR.AI](https://x.com/dair_ai)
- [@dair_ai](https://x.com/dair_ai)
- [16K](https://x.com/dair_ai/status/2089026118389026823/analytics)
- [Paper](https://arxiv.org/abs/2608.07222)
- [Tweet](https://x.com/omarsar0/status/2086845790983716917)
- [Paper](https://arxiv.org/abs/2608.09867)
- [Tweet](https://x.com/omarsar0/status/2087187835530948776)
- [Paper](https://arxiv.org/abs/2608.10218)
- [Tweet](https://x.com/omarsar0/status/2087574841893474556)
- [Paper](https://arxiv.org/abs/2608.11095)
- [Tweet](https://x.com/omarsar0/status/2087605040240582991)
- [Paper](https://arxiv.org/abs/2608.06370)
- [Tweet](https://x.com/dair_ai/status/2086846794840019178)
- [Paper](https://arxiv.org/abs/2608.07885)
- [Tweet](https://x.com/dair_ai/status/2087264294782279808)
- [Paper](https://arxiv.org/abs/2608.11727)
- [Tweet](https://x.com/omarsar0/status/2087962456572498142)
- [Paper](https://arxiv.org/abs/2608.11242)
- [Tweet](https://x.com/dair_ai/status/2087930434323959894)
- [Paper](https://arxiv.org/abs/2608.10296)
- [Tweet](https://x.com/dair_ai/status/2087600513441546589)
- [Paper](https://arxiv.org/abs/2608.06871)
- [Tweet](https://x.com/dair_ai/status/2086950751314870703)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [12:26 AM · Aug 17, 2026](https://x.com/dair_ai/status/2089026118389026823)
- [16.5K Views](https://x.com/dair_ai/status/2089026118389026823/analytics)
- [View quotes](https://x.com/dair_ai/status/2089026118389026823/quotes)
---
*导出时间: 2026/8/17 08:58:15* 
