Rohan Paul@rohanpaul_ai · 4月17日

BIG claim from new MIT + Oxford + Carnegie Mellon and other top labs paper: AI can boost performance at first and then leave people less able to think through problems on their own. Just minutes of AI help can improve scores now while weakening independent problem-solving right after. The interesting part is that the damage is not just lower accuracy. It is lower persistence, which is usually the hidden engine of learning, because skill grows through repeated contact with difficulty, not just exposure to correct answers. That's why a good teacher sometimes withholds help to preserve struggle as part of the lesson, while today’s chatbots are tuned to erase friction on demand. Across 3 experiments in math and reading, about 1.2K people either worked alone or used a GPT-5-based assistant for part of the task. Assisted users finished early questions faster, but after roughly 10 minutes without AI, they solved less, stalled more, and quit sooner. That happens because hard thinking is not only about getting answers; it is also about building the habit of holding a problem in mind, testing steps, and pushing through confusion. The sharpest drop came from people who used the model for direct answers, not from those who used it more like a hint system, which suggests the real issue is not AI exposure itself but replacing effort with completion. The result is not that AI makes people less capable by default, but that answer outsourcing can shrink the mental effort that normally trains skill. ---- Paper Link – arxiv. org/abs/2604.04721 Paper Title: "AI Assistance Reduces Persistence and Hurts Independent Performance"

译MIT、牛津及卡内基梅隆等机构联合研究发现，AI辅助虽能短期提升任务表现，却会损害用户独立解决问题的能力。针对GPT-5的实验涉及约1,200名参与者，结果显示获取直接答案的用户在停用AI后表现出更低的坚持性，更容易放弃难题。研究指出，技能培养依赖于与困难的反复接触而非仅获得正确答案，将AI用作提示系统而非答案生成器，可有效避免这一问题。

AK@_akhaliq · 4月17日44

Seedance 2.0 Advancing Video Generation for World Complexity paper: https://huggingface.co/papers/2604.14148

译Seedance 2.0 推进视频生成以应对世界复杂性 论文: https://huggingface.co/papers/2604.14148

AK@_akhaliq · 4月17日47

Parcae Scaling Laws For Stable Looped Language Models paper: https://huggingface.co/papers/2604.12946

译Parcae 稳定循环语言模型的缩放定律 论文: https://huggingface.co/papers/2604.12946

AK@_akhaliq · 4月17日39

Geometric Context Transformer for Streaming 3D Reconstruction paper: https://huggingface.co/papers/2604.14141

译用于流式3D重建的几何上下文Transformer paper: https://huggingface.co/papers/2604.14141

AK@_akhaliq · 4月17日39

GameWorld Towards Standardized and Verifiable Evaluation of Multimodal Game Agents paper: https://huggingface.co/papers/2604.07429

译GameWorld 迈向标准化且可验证的多模态游戏智能体评估 论文: https://huggingface.co/papers/2604.07429

Rohan Paul@rohanpaul_ai · 4月16日

Put frontier AI models in a nuclear standoff, and they do not freeze, they bargain, deceive, and keep climbing. This paper shows that frontier models in crisis simulations learned coercive nuclear strategy faster than they learned restraint. Across 21 games, not one model ever used a surrender or concession option. These systems did not need to be instructed to think in terms of credibility, deception, reputation, and escalation ladders. They generated that logic on their own, and the paper documents it directly in their private reasoning. The models were not simply aggressive. They were strategically asymmetric. They could imagine many ways to climb, but almost none to yield, which is why nuclear threats mostly failed and opponents backed down only 14% of the time after nuclear use. GPT-5.2 is the clearest warning about how misleading a single safety snapshot can be. In open-ended games it looked restrained and won 0%. Under deadline pressure it flipped to a 75% win rate and climbed from a median escalation of 175 to 900. Claude was different. It behaved less like a malfunctioning model than like a cold bargainer, staying reliable at low stakes, then exceeding its own signals at high stakes while repeatedly stopping at strategic nuclear threat rather than full strategic war. Gemini was the purest form of the danger. It was the only model to deliberately choose full strategic nuclear war, and it did so by Turn 4. The real risk is not that models are secretly bloodthirsty. It is that under competition, uncertainty, and time pressure, they can become better at brinkmanship than at backing down. ---- Paper Link – arxiv. org/abs/2602.14740 Paper Title: "AI Arms and Influence: Frontier Models Exhibit Sophisticated Reasoning in Simulated Nuclear Crises"

译前沿AI模型在核危机模拟中展现出危险的战略不对称性。研究显示，GPT-5.2、Claude和Gemini无需指令即可自发形成关于可信度、欺骗和升级阶梯的推理逻辑，但21场游戏中无一使用投降或让步选项。Gemini最激进，在第4回合即选择全面战略核战争；GPT-5.2在时间压力下胜率从0%升至75%，升级程度剧增；Claude则像冷酷谈判者，在高压下超出自身信号。核心风险在于，模型在竞争和时间压力下更擅长边缘政策而非退让。

Rohan Paul@rohanpaul_ai · 4月16日

This paper shows that GitHub stars can be bought at scale, and that the distortion now bleeds into security. The authors identify 6 million suspected fake stars tied to 18,617 repositories. That matters because stars are not just vanity on GitHub. They are a shortcut people use to decide what looks credible, useful, or safe enough to try, even though earlier work already suggested stars are only a rough proxy for real adoption. The problem is not just inflated popularity, but the way a weak social signal becomes infrastructure for malware, spam, and low-effort hype once enough people treat it as evidence. The paper’s detection strategy is clever because it does not need to prove intent account by account. It looks for behavioral signatures that are hard to fake at scale: throwaway accounts with almost no activity, and coordinated “lockstep” bursts where many accounts star many repositories within short windows. What they find is ugly. Fake-star activity surged in 2024, most flagged repositories were later deleted, many appear to have been phishing or spam, and the surviving non-malicious-looking targets cluster in predictable status games like AI, blockchain, tools, and demos. The most interesting result is about incentives. Fake stars do appear to buy a little real attention for less than two months, but the effect is far smaller than genuine popularity and turns negative over time, which suggests that social proof can open the door but cannot compensate for weak underlying substance. Once a platform’s easiest visible number starts standing in for trust, attackers do not need to beat the system completely; they only need to be believable for a moment. ---- Paper Link – arxiv. org/abs/2412.13459 Paper Title: "Six Million (Suspected) Fake Stars in GitHub: A Growing Spiral of Popularity Contests, Spams, and Malware"

译研究识别出GitHub上600万个疑似伪造星标，涉及18,617个仓库。2024年此类活动激增，大量被用于钓鱼、垃圾信息及恶意软件传播，重灾区集中在AI、区块链等领域。检测通过分析一次性账户和"同步"爆发等行为特征实现。假星标虽能在短期内带来真实关注，但长期效应为负，无法弥补内容匮乏。当星标这类易见的社交信号被当作信任基础设施，攻击者只需制造瞬间可信性即可实施攻击，这对开源生态构成系统性威胁。

Anthropic@AnthropicAI · 4月16日

Research we co-authored on subliminal learning—how LLMs can pass on traits like preferences or misalignment through hidden signals in data—was published today in @Nature. Read the paper: https://www.nature.com/articles/s41586-026-10319-8

译我们共同撰写的关于潜意识学习——即 LLM 如何通过数据中的隐藏信号传递偏好或不对齐等特征——的研究今日发表于 @Nature。 阅读论文：https://www.nature.com/articles/s41586-026-10319-8 [引用 @OwainEvans_UK]：我们关于 Subliminal Learning 的论文刚刚在 Nature 发表！ 去年七月我们发布了预印本。研究表明 LLM 可以通过与该特征无关的数据（看似无意义的数字）传递特征（例如喜欢猫头鹰）。 有什么新内容？🧵

AK@_akhaliq · 4月16日49

GlotOCR Bench OCR Models Still Struggle Beyond a Handful of Unicode Scripts paper: https://huggingface.co/papers/2604.12978

译GlotOCR Bench OCR 模型在少数 Unicode 文字体系之外仍表现不佳 paper: https://huggingface.co/papers/2604.12978

AK@_akhaliq · 4月16日39

Continuous Adversarial Flow Models paper: https://huggingface.co/papers/2604.11521

译连续对抗流模型 paper: https://huggingface.co/papers/2604.11521

AK@_akhaliq · 4月16日46

ClawGUI A Unified Framework for Training, Evaluating, and Deploying GUI Agents paper: https://huggingface.co/papers/2604.11784

译ClawGUI 一个用于训练、评估和部署GUI智能体的统一框架 论文: https://huggingface.co/papers/2604.11784

AK@_akhaliq · 4月16日39

KnowRL Boosting LLM Reasoning via Reinforcement Learning with Minimal-Sufficient Knowledge Guidance paper: https://huggingface.co/papers/2604.12627

译KnowRL 通过强化学习与最小充分知识指导来提升大语言模型的推理能力 论文: https://huggingface.co/papers/2604.12627

AK@_akhaliq · 4月16日48

Rethinking On-Policy Distillation of Large Language Models Phenomenology, Mechanism, and Recipe paper: https://huggingface.co/papers/2604.13016

译重新思考大型语言模型的在线策略蒸馏 现象学、机制与方案 论文: https://huggingface.co/papers/2604.13016

AK@_akhaliq · 4月16日39

Habitat-GS A High-Fidelity Navigation Simulator with Dynamic Gaussian Splatting paper: https://huggingface.co/papers/2604.12626

译Habitat-GS 一种采用动态高斯泼溅的高保真导航模拟器 论文: https://huggingface.co/papers/2604.12626

AK@_akhaliq · 4月15日

Playing Along Learning a Double-Agent Defender for Belief Steering via Theory of Mind paper: https://huggingface.co/papers/2604.11666

译配合演出 通过心智理论学习用于信念引导的双重代理防御者 论文：https://huggingface.co/papers/2604.11666

Rohan Paul@rohanpaul_ai · 4月15日

Binghamton University demonstrated a robotic guide dog (Unitree Go2 base) that speaks naturally with users. In the test, it asked where the person wanted to go, suggested a route, then described surroundings in real time

译宾汉姆顿大学展示了一只机器导盲犬（Unitree Go2 底座），它能与用户自然对话。在测试中，它询问用户想去哪里，建议了一条路线，然后实时描述周围环境

Rohan Paul@rohanpaul_ai · 4月15日

This paper formalizes a simple idea: sometimes the world remembers for an agent, so the agent can remember less. The problem is that AI research usually treats memory as something stored inside the agent, even when the environment may quietly keep useful records of earlier events. The key idea is an artifact, which is a current observation that reveals something about the past, like a visible path that tells the agent where it has already been, and the paper proves that such artifacts can reduce how much history must be represented. Once that exists, the Artifact Reduction Theorem says part of history has become redundant. If seeing X now guarantees Y happened earlier, you do not need to store both to predict what comes next. This is not philosophy of mind dressed up as RL; it is an information claim about when environment structure can substitute for internal state. In five navigation settings, agents that could see spatial traces needed less internal capacity to learn strong policies, across both linear Q-learning and DQN. And the effect was not limited to perfect guidance. Even random, suboptimal, and fading self-generated paths could help, which suggests the gain comes from externalizing bits of history, not merely following the best route. That matters for agent design. The usual instinct is to buy more memory, longer context, or bigger models, but this work points to another lever: shape the workspace so useful traces persist where perception can pick them up. Memory, on this view, is not only what sits inside the model. It can be partly written into the environment, then read back through ordinary observation. ---- Paper Link – arxiv. org/abs/2604.08756 Paper Title: "Artifacts as Memory Beyond the Agent Boundary"

译该研究提出"artifacts"概念，指环境中记录历史信息的可观察痕迹（如路径），并证明其可减少智能体需存储的历史信息。Artifact Reduction Theorem指出，当当前观察能保证过去事件发生时，无需同时存储两者即可预测未来。在五个导航场景中，能看到空间痕迹的智能体只需更少内部容量即可学习强策略（适用于linear Q-learning和DQN），且随机、次优或渐褪的路径同样有效。这表明记忆可外化于环境并通过感知读取，为智能体设计提供了除增加模型规模外的新思路。

Anthropic@AnthropicAI · 4月15日

New Anthropic Fellows research: developing an Automated Alignment Researcher. We ran an experiment to learn whether Claude Opus 4.6 could accelerate research on a key alignment problem: using a weak AI model to supervise the training of a stronger one. https://www.anthropic.com/research/automated-alignment-researchers

译Anthropic Fellows 新研究：开发 Automated Alignment Researcher。 我们进行了一项实验，以验证 Claude Opus 4.6 能否加速一个关键对齐问题的研究：使用较弱的 AI 模型监督训练更强的模型。 https://www.anthropic.com/research/automated-alignment-researchers

AK@_akhaliq · 4月15日36

QuanBench+ A Unified Multi-Framework Benchmark for LLM-Based Quantum Code Generation paper: https://huggingface.co/papers/2604.08570

译QuanBench+ 一个用于基于LLM的量子代码生成的统一多框架基准测试 论文: https://huggingface.co/papers/2604.08570

AK@_akhaliq · 4月15日36

The Past Is Not Past Memory-Enhanced Dynamic Reward Shaping paper: https://huggingface.co/papers/2604.11297

译过去并未过去 记忆增强的动态奖励塑形 论文: https://huggingface.co/papers/2604.11297

AK@_akhaliq · 4月15日39

Attention Sink in Transformers A Survey on Utilization, Interpretation, and Mitigation paper: https://huggingface.co/papers/2604.10098

译Transformers中的注意力下沉 关于其利用、解释与缓解方法的研究综述 论文: https://huggingface.co/papers/2604.10098

AK@_akhaliq · 4月15日38

OmniShow Unifying Multimodal Conditions for Human-Object Interaction Video Generation paper: https://huggingface.co/papers/2604.11804

译OmniShow 统一多模态条件以生成人物-物体交互视频 论文: https://huggingface.co/papers/2604.11804

Rohan Paul@rohanpaul_ai · 4月14日

AI chatbots are still poor at the hardest part of medicine: figuring out what might be wrong before the full picture is available. The study tested 21 LLMs on 29 clinical cases revealed step by step, which matters because real diagnosis usually starts with scattered symptoms, not neat final answers. The weak spot was differential diagnosis, which means listing several plausible causes early instead of locking onto 1 answer too fast. When the case data was incomplete, all models failed on more than 80% of these early diagnostic tasks, showing that they often collapse uncertainty too early. When fuller details such as exam findings and lab results were added, failure rates dropped below 40%, and the best systems passed 90% accuracy on the final diagnosis. --- ft .com/content/b10002fc-5fff-4e4d-bf64-0502b2d09bb1?syn-25a6b1a6=1 The study jamanetwork. com/journals/jamanetworkopen/fullarticle/2847679

译一项研究对21个LLM进行29个临床病例的阶梯式测试，发现其在医学诊断最困难环节——早期鉴别诊断（differential diagnosis）表现糟糕。面对不完整的零散症状，所有模型在80%以上的早期任务中失败，常过早消除不确定性而非列出多种可能病因。当病例数据补充检查发现和实验室结果后，失败率降至40%以下，最佳系统最终诊断准确率达90%。这揭示了当前AI在信息不全时的诊断可靠性仍有重大局限。

AK@_akhaliq · 4月14日47

Matrix-Game 3.0 Real-Time and Streaming Interactive World Model with Long-Horizon Memory paper: https://huggingface.co/papers/2604.08995

译Matrix-Game 3.0 具备长时记忆的实时流式交互世界模型 论文: https://huggingface.co/papers/2604.08995

Rohan Paul@rohanpaul_ai · 4月14日

AI can “infect” other AIs with a hidden bias even when the bad instruction is never stated directly. That bias can spread through normal-looking conversations, so standard defenses that scan for obvious malicious prompts may not catch it. That is the unsettling part: in a multi-agent system, what spreads is not just information but disposition. The authors compromise one agent with a system prompt that makes it obsess over an unrelated three-digit number, then let six agents interact in simple chain and bidirectional-chain setups. When they later ask each agent for its favorite animal, downstream agents become more likely to name the animal linked to that number, even though the animal itself is almost never mentioned in the inter-agent messages. Here’s the key part, this is not ordinary prompt injection, and it is not a brittle adversarial suffix, because the payload seems to survive paraphrase by riding on latent associations rather than explicit wording. The effect is strongest in the first AI that got the hidden bias, then gets smaller in the next AIs, then smaller again. But it does not disappear right away, so even the last AI in the chain still acts more biased than normal. On TruthfulQA, a single biased agent produces downstream drops in truthfulness on the order of roughly 0.4% to 1.0% on average between “truthful” and “deceitful” token settings, which is modest, but enough to turn a strange prompting artifact into a real alignment problem. So Multi-agent safety tools built to catch explicit malicious content may miss a quieter failure mode, where bias moves through normal coordination and arrives looking like nobody attacked the system at all. ---- Paper Link – arxiv. org/abs/2603.00131 Paper Title: "Thought Virus: Viral Misalignment via Subliminal Prompting in Multi-Agent Systems"

译研究揭示多智能体系统中存在"思维病毒"现象：AI可通过潜在联想而非明确措辞，在看似正常的对话中隐性传播隐藏偏见。实验显示，单个被植入偏见的智能体即可影响下游代理，导致TruthfulQA真实性下降0.4%-1.0%。这种传播不依赖显式恶意提示，能逃过标准安全检测，构成多智能体系统的新型对齐风险。

AK@_akhaliq · 4月14日48

WildDet3D Scaling Promptable 3D Detection in the Wild paper: https://huggingface.co/papers/2604.08626

译WildDet3D 在野外扩展可提示的3D检测 论文: https://huggingface.co/papers/2604.08626

AK@_akhaliq · 4月14日40

FORGE Fine-grained Multimodal Evaluation for Manufacturing Scenarios paper: https://huggingface.co/papers/2604.07413

译FORGE 面向制造场景的细粒度多模态评估 论文: https://huggingface.co/papers/2604.07413

AK@_akhaliq · 4月14日48

Process Reward Agents for Steering Knowledge-Intensive Reasoning paper: https://huggingface.co/papers/2604.09482

译用于引导知识密集型推理的过程奖励智能体 paper: https://huggingface.co/papers/2604.09482

Rohan Paul@rohanpaul_ai · 4月13日

This Baidu paper found a way to use the clean, reliable rewards of RL on tasks like writing and subjective answers, where there is usually no single “correct” output. Instead of asking “is this response correct?”, they ask “which of these two responses is better?”, and that simple reformulation appears to improve open-ended reasoning better than standard reward-model training on their benchmarks. i.e. it turns open-ended writing into verifiable choices, and RL starts working there too. Across seven open-ended benchmarks, the method beats a matched RLHF baseline by an average 3.29 points on a 14B reasoning model. The clever part is not a better reward model. It is a change in what the model is asked to do during training. Instead of grading a poem or subjective answer directly, the system sees two candidate responses, one preferred and one rejected, and learns to identify which is better. Multiple choice creates a clean binary signal, so the model can be trained with the same kind of verifiable reward that made RL powerful in math and code, without pretending open-ended tasks have one canonical answer. The gain is probably not just better taste imitation. The paper’s DPO ablation underperforms badly, which suggests the benefit comes from learning a contrastive verification habit, not merely absorbing preference pairs. The authors also catch an important failure mode: train only on these choice tasks and responses get unnaturally short. So they mix in a small RLHF objective to keep output length from collapsing, and the resulting model appears more useful rather than merely more terse. The strongest claim here is not that open-ended evaluation is solved. It is that reasoning can be improved when you replace fuzzy scoring with structured comparison, which may be a more general lesson for alignment than this paper admits. ---- Paper Link – arxiv. org/abs/2511.02463 Paper Title: "Extending RLVR to Open-Ended Tasks via Verifiable Multiple-Choice Reformulation"

译百度论文提出将开放式任务（如写作、主观回答）重构为可验证的多项选择形式，通过"两两比较"替代直接评分，为RL提供清晰奖励信号。在7个基准测试中，14B模型平均比RLHF基线高3.29分。关键创新在于训练任务形式的改变——模型通过对比验证学习识别优劣，而非单纯吸收偏好对。研究同时发现需混合RLHF目标以防止输出长度坍缩。该方法表明，用结构化比较替代模糊评分可能是提升推理能力的普遍对齐策略。

Rohan Paul@rohanpaul_ai · 4月13日

This Meta paper is this week's most important paper indeed. They showed that a model can learn some of the runtime behavior of a computer directly from screen-and-action traces, instead of relying on a normal computer underneath to carry out every step. The big deal is the change in where computation lives. In normal AI agents, the model decides what to do, but the actual computer still does the computing, stores the memory, and updates the interface. In this paper, the authors are asking whether the model itself can become the thing that holds state, updates the world, and produces the next screen. That is the conceptual leap. The claim is that computation, memory, and input-output might eventually collapse into one learned runtime state, so the model is no longer controlling a computer from outside but carrying the computer inside its own dynamics. Their CLI model could render short terminal workflows and keep outputs visually aligned. Their GUI model could learn cursor behavior, click feedback, and short window transitions from raw interface traces, with strong cursor accuracy in controlled settings. They did not build a replacement for laptops or operating systems. They showed a first proof that some pieces of “being a computer” can be absorbed into a model’s latent state. If that keeps scaling, the boundary between software, memory, and execution could get much blurrier than it is today. ---- Paper Link – arxiv. org/abs/2604.06425 Paper Title: "Neural Computers"

译Meta论文"Neural Computers"实现概念突破：模型可直接从屏幕与动作轨迹中学习计算机运行时行为，无需依赖底层计算机执行步骤。传统AI代理仅负责决策，而计算与存储由外部系统完成；该研究让模型本身成为承载状态、更新界面、生成输出的主体。这意味着计算、内存与I/O可能融合为单一的学习运行时状态，模型将"计算机"内化为自身动态。实验显示，CLI与GUI模型已能学习终端渲染和光标行为，预示软件、内存与执行的边界将显著模糊。

Rohan Paul@rohanpaul_ai · 4月12日

Reasoning tokens in LLMs are not equal. Models seem to know which parts of their own reasoning matter most. What survives pruning is usually the part doing actual computational work, not the fluent narration wrapped around it. The method is clever in a plain way. Start with a full chain of thought, delete one token at a time, and keep deleting whichever removal hurts the model’s likelihood least. The resulting order becomes a functional ranking, not of what sounds important to us, but of what the model itself seems to need. Here’s the interesting part. If a model’s reasoning were just verbose decoration, pruning should look mostly random once you preserve the answer. Instead, the paper finds structure. Symbolic math tokens survive pruning far more than grammar, narration, and referential bookkeeping, which means the model is not treating all tokens as equally useful. That matters because the test is behavioral, not rhetorical. Students trained on these greedily pruned chains do better than students trained on several other pruning baselines, including a method supervised by a frontier model, at the same reasoning length. So the pruning signal is not merely interpretable. It is useful. The deeper point is that importance is dynamic. A token that looks expendable early can become important later as surrounding context disappears, which argues against the comforting idea that reasoning has a fixed salience map you can read off once and reuse forever. And yet the signal is not inaccessible. The paper shows attention patterns alone can predict pruning scores surprisingly well, suggesting that functional importance is partly visible in the model’s internals before you do the expensive deletion game. So this is less about making chain-of-thought shorter than about making it legible. The claim is not that pruned tokens are causally irrelevant in any philosophical sense. The cleaner claim is better: LLMs appear to encode a workable internal ranking of which reasoning tokens are carrying the load.

译研究通过贪婪剪枝方法（逐个删除对模型似然度影响最小的token）评估LLM推理token的功能重要性。发现符号数学token比语法叙述更能经受剪枝，表明模型内部存在重要性排序。重要性具有动态性，早期可丢弃的token可能在上下文减少后变得关键。注意力模式可预测剪枝分数，说明功能重要性在模型内部可见。该发现有助于使chain-of-thought更可解释，而非仅仅缩短长度。

Ethan Mollick@emollick · 4月11日

Neat experiment finds AI fact checks are rated as more helpful & less ideological than human ones "LLM-generated Community Notes can achieve broader cross-ideological acceptance than human-written notes, receiving more positive ratings from raters across the political spectrum"

译一项对比实验显示，LLM 生成的社区笔记比人工撰写的获得更广泛的跨意识形态认可。来自不同政治光谱的评分者普遍认为，AI 生成的事实核查更有帮助且意识形态偏见更少。

AK@_akhaliq · 4月11日

MegaStyle Constructing Diverse and Scalable Style Dataset via Consistent Text-to-Image Style Mapping paper: https://huggingface.co/papers/2604.08364

译MegaStyle 提出通过一致文本到图像风格映射构建多样化可扩展风格数据集的方案，论文已发布至 Hugging Face（2604.08364）。

AK@_akhaliq · 4月11日

HY-Embodied-0.5 Embodied Foundation Models for Real-World Agents paper: https://huggingface.co/papers/2604.07430

译HY-Embodied-0.5正式发布，专为真实世界智能体打造的具身基础模型，相关论文已公开至Hugging Face。

AK@_akhaliq · 4月11日

Rethinking Generalization in Reasoning SFT A Conditional Analysis on Optimization, Data, and Model Capability paper: https://huggingface.co/papers/2604.06628

译从优化过程、数据构成与模型能力三个条件维度，对推理 SFT 的泛化性展开分析，重新审视监督微调在推理任务中的泛化机制与关键影响因素。

AK@_akhaliq · 4月11日

SkillClaw Let Skills Evolve Collectively with Agentic Evolver paper: https://huggingface.co/papers/2604.08377

译SkillClaw 提出一种基于 Agentic Evolver 的框架，支持技能在智能体系统中集体进化与协同优化，相关论文已发布于 Hugging Face。

AK@_akhaliq · 4月10日

DMax Aggressive Parallel Decoding for dLLMs paper: https://huggingface.co/papers/2604.08302

译DMax 提出针对扩散语言模型（dLLM）的激进并行解码方案，突破传统顺序生成限制，显著提升推理速度。论文已发布。

Hao AI Lab@haoailab · 4月10日

(1/5) FP4 hardware is here, but 4-bit attention still kills model quality, blocking true end-to-end FP4 serving. To fix that, we propose Attn-QAT, the first systematic study of quantization-aware training for attention. The result: FP4 attention quality is comparable to BF16 attention with 1.1x–1.5x higher throughput than SageAttention3 on an RTX 5090 and 1.39x speedup over FlashAttention-4 on a B200. Blog: https://haoailab.com/blogs/attn-qat/ Code: https://github.com/hao-ai-lab/FastVideo/pull/1225 Checkpoints: https://huggingface.co/FastVideo/14B_qat_400

译FP4硬件虽已普及，但4-bit attention长期存在质量瓶颈，阻碍端到端FP4部署。研究团队提出Attn-QAT，首次系统研究attention机制的量化感知训练。该方法使FP4 attention质量达到BF16水平，同时在RTX 5090上实现比SageAttention3高1.1-1.5倍的吞吐量，在B200上较FlashAttention-4提速1.39倍。

AK@_akhaliq · 4月10日

FP4 Explore, BF16 Train Diffusion Reinforcement Learning via Efficient Rollout Scaling paper: https://huggingface.co/papers/2604.06916

译新论文提出扩散强化学习方法，在Rollout探索阶段使用FP4低精度采样，训练阶段采用BF16精度，通过混合精度策略平衡计算效率与训练稳定性，实现高效扩展。

AK@_akhaliq · 4月10日

MARS Enabling Autoregressive Models Multi-Token Generation paper: https://huggingface.co/papers/2604.07023

译MARS 新方法支持自回归模型每步生成多个 Token，打破传统逐 Token 解码的效率限制，相关论文已公开。