Daily Papers · 2026-09-09
📄 论文日报 | 来源：Hugging Face Daily Papers
🗓 Episode：2026.09.08 | 离散扩散无损加速；流平衡优化自改进
1. 通过离散扩散实现大语言模型的无损加速
- 英文标题：Unlocking Lossless Speedups in LLMs via Discrete Diffusion
- arXiv：2609.04010v1
💡 一句话解读：这篇研究要解决大语言模型生成速度慢的问题。传统模型一个字一个字地生成，很慢。他们提出一种新方法，让模型能同时生成多个字，但又不牺牲质量。就像把流水线改成并行处理，速度能提升3倍，而且不需要额外的辅助模型。
大语言模型（LLMs）的成功很大程度上归功于下一个词预测（NTP），但其自回归（AR）结构需要缓慢的顺序词生成。为了克服这一瓶颈，我们引入了扩散增强的大语言模型，这是一类新模型，它定义了AR模型分布，同时使用扩散从该分布中并行抽取多个词。我们将这些模型的参数解耦为两组：使用标准NTP目标训练的AR权重，以及用于同时生成多个词的轻量级扩散权重。扩散权重通过简单的扩散蒸馏阶段学习，该阶段对现有LLM训练流程的额外开销可忽略不计。我们还引入了Ψ-Spec，一个采样器家族，能够在固定上下文长度下实现无损加速和推理时扩展。与推测解码不同，我们的方法不需要单独的草稿模型。与扩散LLM（d-LLMs）不同，它在加速生成的同时不牺牲底层AR模型的质量。由此产生的模型称为Uno，可以从头训练或通过增强现有的开放权重AR LLM来构建。Uno在评估的每个批处理大小上都比领先的推测解码方法实现更高的吞吐量，并且相对于基础AR模型提供高达3倍的加速，包括在设备支持的最大批处理大小下。值得注意的是，我们的8B Uno模型在智能体工具使用、编码和长上下文推理的所有评估基准上，都优于领先的开放d-LLM——26B DiffusionGemma，以及专有的Mercury 2。我们在https://s-sahoo.github.io/uno/发布代码和检查点。
📄 English Abstract
Large Language Models (LLMs) owe much of their success to next-token prediction (NTP), but their autoregressive (AR) structure requires slow, sequential token generation. To overcome this bottleneck, we introduce diffusion-augmented LLMs, a new class of models that defines an AR model distribution while using diffusion to draw multiple tokens in parallel from that distribution. We decouple the parameters of these models into two sets: AR weights, trained using the standard NTP objective, and lightweight diffusion weights, trained to generate multiple tokens simultaneously. The diffusion weights are learned through a simple Diffusion Distillation phase that adds negligible overhead to existing LLM training pipelines. We also introduce $Ψ$-Spec, a family of samplers that enables lossless acceleration and inference-time scaling at a fixed context length. Unlike speculative decoding, our method requires no separate draft model. Unlike diffusion LLMs (d-LLMs), it accelerates generation without sacrificing the quality of the underlying AR model. The resulting models, called Uno, can be trained from scratch or built by augmenting existing open-weight AR LLMs. Uno achieves higher throughput than leading speculative-decoding methods at every evaluated batch size and delivers up to $3\times$ speedups over the base AR model, including at the largest batch size supported by the device. Notably, our 8B Uno model outperforms the leading open d-LLM, the 26B DiffusionGemma, and the proprietary Mercury 2 across all evaluated benchmarks in agentic tool use, coding, and long-context reasoning. We release code and checkpoints at: https://s-sahoo.github.io/uno/
2. 流平衡：基于验证器的在线策略推理经验自我改进方法
- 英文标题：FlowBalance: Verifier-Grounded Self-Improvement from On-Policy Reasoning Experience
- arXiv：2609.03241v1
💡 一句话解读：这篇研究要解决AI推理模型自我改进时容易‘自欺欺人’或‘钻牛角尖’的问题。他们提出一种叫FlowBalance的方法，像老师批改作业一样，用外部验证器打分，同时参考模型自己的判断，但只在外部打分说‘对’时才听自己的，说‘错’时反而反向调整，从而让AI学得更稳、更好。
推理模型可以从自身的在线策略经验中改进，但这种内部循环是脆弱的：终端验证器提供可靠但稀疏的监督，而密集的同模型指导可能强化错误自信或将学习过度集中于狭窄的解决模式。我们提出FlowBalance，一种基于验证器的自我改进方法，它学习完整响应的归一化分布。对于每条在线策略轨迹，同一策略的冻结训练时视图利用特权上下文产生词级对数概率增益，这些增益聚合成轨迹级自我指导分数。FlowBalance用验证器导出的组优势校准该分数：在正优势轨迹上保留指导，在负优势轨迹上反转指导，当滚动组不提供结果偏好时禁用指导。由此产生的能量指数性地重新加权参考策略，而轮廓轨迹平衡通过每个滚动组的一次对数配分估计拟合归一化目标。这实现了通过轨迹平衡进行结果校准的自我指导，无需单独的词级模仿损失。我们的分析建立了组内对比保持、最小变化反向KL特征、目标奖励的单调验证器控制，以及对被拒绝响应上假阳性自我指导的精确校正。在数学推理上，FlowBalance在Qwen3-4B和Qwen3-8B上均优于FlowRL的平均性能，同时提高了训练速度和稳定性，避免了直接OPSD的响应长度崩溃，并在受控AIME24诊断中表现出更高的正确策略多样性。
📄 English Abstract
A reasoning model can improve from its own on-policy experience, but this inner loop is fragile: terminal verifiers provide reliable yet sparse supervision, while dense same-model guidance can reinforce false confidence or overconcentrate learning on a narrow solution mode. We introduce FlowBalance, a verifier-grounded self-improvement method that learns a normalized distribution over complete responses. For each on-policy trajectory, a frozen training-time view of the same policy uses privileged context to produce token-level log-probability gains, which are aggregated into a trajectory-level self-guidance score. FlowBalance calibrates this score with the verifier-derived group advantage: guidance is retained on positive-advantage trajectories, reversed on negative-advantage trajectories, and disabled when the rollout group provides no outcome preference. The resulting energy exponentially reweights a reference policy, and profiled trajectory balance fits the normalized target with one log-partition estimate per rollout group. This realizes outcome-calibrated self-guidance via trajectory balance, without a separate token-level imitation loss. Our analysis establishes within-group contrast preservation, a minimum-change reverse-KL characterization, monotonic verifier control of target reward, and an exact correction against false-positive self-guidance on rejected responses. On mathematical reasoning, FlowBalance improves average performance over FlowRL on both Qwen3-4B and Qwen3-8B, while also improving training speed and stability, avoiding direct OPSD’s response-length collapse, and exhibiting higher correct-strategy diversity in a controlled AIME24 diagnostic.
3. ENEAS：嵌入引导的神经集成自适应分割
- 英文标题：ENEAS: Embedding-guided Neural Ensemble for Adaptive Segmentation
- arXiv：2609.03756v1
💡 一句话解读：视频追踪和分割AI常把雕像、画作等当成真人，或物体离开画面后仍“幻觉”存在。ENEAS提出一个统一方法：既能精准追踪单个目标，又能按文字描述发现所有同类物体，并通过语义验证层区分“看起来像”和“真的是”，避免误判。
我们提出ENEAS，一种统一的、可通过文本提示的实例追踪与语义发现方法。文本提示分割模型，包括SAM 3等最新基础模型，仍存在时间幻觉、空间碎片化和语义误分类问题：当物体离开视野时未能报告目标缺失，在极端特写时分割局部纹理而非完整物体，并优先考虑视觉特征而非本体论现实，导致雕像、绘画或反射等视觉上相似的伪影被分割为目标实体。ENEAS通过单一方法双向工作：对唯一实例进行精确追踪和高质量分割，以及通过语义验证层解析文本查询所命名的每个实例的开放概念发现。对于追踪，我们将几何稳健的SeC架构（先前仅限于点交互）扩展了文本提示适配器，并利用其时间记忆，使目标在消失时保持而不漂移到干扰物，即使填满整个视野也能保持完整。对于发现，验证层将高速视觉嵌入匹配与条件VLM细化相结合，仅对模糊候选进行语义推理，这过滤了纯视觉模型无法区分的本体论错误，同时保持低延迟。ENEAS专为3D重建设计，其中单个误分类干扰物会破坏资产，它解锁了视频、广泛库以及时间或空间无序数据集合的高质量语义追踪和分割，并具备区分真实实例与其“替身”（看起来相似但并非相同的事物）的能力。代码和模型可在https://github.com/speridlabs/eneas获取。
📄 English Abstract
We present ENEAS, a unified, text-promptable method for instance tracking and semantic discovery. Text-promptable segmentation models, including the latest foundation models such as SAM 3, still suffer from temporal hallucinations, spatial fragmentation, and semantic misclassification: they fail to report target absence when an object leaves the field of view, segment local textures instead of the complete object during extreme close-ups, and prioritize visual features over ontological reality, so that visually similar artifacts such as statues, paintings, or reflections are segmented as target entities. ENEAS works two ways from a single method: precise tracking and high-quality segmentation of a unique instance, and open-concept discovery of every instance a text query names, resolved by a semantic verification layer. For tracking, we extend the geometrically robust SeC architecture, previously limited to point interactions, with a text-prompting adapter and leverage its temporal memory, so that the target is held through disappearance without drifting to distractors and kept whole even when it fills the entire view. For discovery, the verification layer combines high-speed visual embedding matching with conditional VLM refinement, invoking semantic reasoning only for ambiguous candidates, which filters out the ontological errors that visual-only models cannot distinguish while keeping latency low. Designed with 3D reconstruction in mind, where a single misclassified distractor corrupts the asset, ENEAS unlocks high-quality semantic tracking and segmentation of video, of broad libraries, and of collections of temporally or spatially unordered data, together with the discrimination to tell true instances from their doppelgangers: things that look alike but are not the same. The code and models are available at https://github.com/speridlabs/eneas
4. 具身技能（EmbodiedSkills）：编排、训练与部署VLA智能体的统一框架
- 英文标题：EmbodiedSkills: A Unified Framework for Orchestrating, Training, and Deploying VLA Agents
- arXiv：2609.01281v1
💡 一句话解读：这篇研究要解决机器人执行复杂长任务时，光会预测动作不够的问题。它提出了一个叫EmbodiedSkills的统一框架，把每个技能决策当成一个执行提案：执行前检查条件，执行后验证结果。这个框架有个固定接口，可以方便地更换底层模型，还能记录过程数据用于训练和优化。
视觉-语言-动作（VLA）模型将视觉观察和语言指令直接映射为机器人动作，但长时程任务需要的不仅仅是动作预测。智能体必须在物理状态演变过程中协调感知、规划、执行、进度验证和恢复。动作预测或模型生成的技能决策本身并不能保证提议的操作在当前状态下有效，也不能保证其结果会被验证。我们提出EmbodiedSkills，一个统一框架，将每个技能决策视为一个执行提案：运行时在执行前检查其前提条件，并在执行后验证结果。一个共享的可执行技能接口将高层技能选择、受限的低层VLA执行和动作后验证连接在单个智能体循环中。由于该接口保持不变，低层VLA策略可以被替换或调整，而无需改变智能体循环。该接口还将规划、执行、验证和恢复事件记录为结构化轨迹，为各个组件提供监督，并在有交互反馈时支持可选的在线适应。我们在RoboTwin 2.0和LIBERO上使用Qwen3-VL和OpenPI/pi0.5实例化了EmbodiedSkills。任务适应的低层VLA策略在50个RoboTwin 2.0任务上平均成功率达到86.20%，在四个LIBERO套件上达到97.40%。这些结果确立了EmbodiedSkills中使用的任务适应低层VLA策略的执行性能。在四个记忆依赖的RMBench任务上，相同的任务适应执行方法平均成功率为12.5%。该框架提供了一个可训练和可检查的智能体层，用于将这些策略转化为闭环具身系统。
📄 English Abstract
Vision-language-action (VLA) models map visual observations and language instructions directly to robot actions, but long-horizon tasks require more than action prediction. An agent must coordinate perception, planning, execution, progress verification, and recovery as the physical state evolves. An action prediction or a model-generated skill decision does not, by itself, guarantee that the proposed operation is valid in the current state or that its outcome will be verified. We propose EmbodiedSkills, a unified framework that treats each skill decision as an execution proposal: the runtime checks its prerequisites before execution and verifies the outcome afterward. A shared executable-skill interface connects high-level skill selection, bounded low-level VLA execution, and post-action verification within a single agent loop. Because this interface remains fixed, low-level VLA policies can be replaced or adapted without changing the agent loop. The interface also records planning, execution, verification, and recovery events as structured trajectories, which provide supervision for individual components and can support optional online adaptation when interactive feedback is available. We instantiate EmbodiedSkills with Qwen3-VL and OpenPI/pi0.5 on RoboTwin 2.0 and LIBERO. Task-adapted low-level VLA policies achieve an average success rate of 86.20% across 50 RoboTwin 2.0 tasks and 97.40% across the four LIBERO suites. These results establish the execution performance of the task-adapted low-level VLA policies used in EmbodiedSkills. On four memory-dependent RMBench tasks, the same task-adapted execution approach achieves 12.5% average success. The framework provides a trainable and inspectable agent layer for turning these policies into closed-loop embodied systems.
5. 一种症状，三个杠杆：同策略自蒸馏的批判性综述
- 英文标题：One Symptom, Three Levers: A Critical Review of On-Policy Self-Distillation
- arXiv：2608.25936v1
💡 一句话解读：这篇论文要解决的问题是：一种叫“同策略自蒸馏”的AI训练方法虽然省钱省力，但会导致模型“崩溃”——只会走老路，失去创造力。作者提出用三个“杠杆”来理解和控制这个问题：信号用在哪儿、给模型看什么信息、以及信号何时改变。文章没有新实验，主要是梳理和统一已有研究，帮助大家看清哪些是共识，哪些还有争议。
同策略蒸馏（on-policy distillation）在模型自身生成的内容上训练语言模型，同时由教师模型逐词打分。它结合了模仿学习的密集监督与强化学习的同策略采样，但需要另一个更大的模型作为教师。同策略自蒸馏（On-Policy Self-Distillation, OPSD）消除了这一成本：教师即模型本身，但以测试时学生无法获得的特权信息（如参考答案、计划或环境反馈）为条件。教师并不比学生更强，只是信息更充分。早期结果令人鼓舞，其准确性与强化学习相当，而生成的token数仅为后者的一小部分。然而，产生信号的不对称性也导致了偏差。目前一个主要失败模式主导了该领域：崩溃（collapse），即模型能产生的推理路径逐渐收窄。崩溃并非OPSD特有，但特权信息会加剧它。本综述将崩溃视为一种由三个杠杆控制的症状：(i) 信号施加的位置，即token如何加权；(ii) 教师所看到的内容，即特权信息的性质；(iii) 信号变化的时间，即教师的动态和指导的衰减。我们将范围限定在数学推理领域，因为该方法起源于此，且其失败模式记录最为详尽。我们未报告新实验。本文的贡献在于结构性：为不同论文中命名各异的现象提供统一词汇，并明确区分了已有定论与仍有争议的部分。
📄 English Abstract
On-policy distillation trains a language model on its own generations while a teacher scores them token by token. It combines the dense supervision of imitation learning with the on-policy sampling of reinforcement learning. But it requires a second, larger model to act as teacher. On-Policy Self-Distillation (OPSD) removes that cost. The teacher is the model itself, conditioned on privileged information the student will not have at test time, such as a reference solution, a plan, or environment feedback. The teacher is no stronger than the student, only better informed. Early results were promising, with accuracy comparable to reinforcement learning at a fraction of the generated tokens. But the same asymmetry that produces the signal also biases it. One failure mode now dominates the field: collapse, the progressive narrowing of the set of reasoning paths the model can produce. Collapse is not specific to OPSD, though privileged information aggravates it. This review treats collapse as a symptom governed by three levers: (i) where the signal is applied, that is, how tokens are weighted; (ii) what the teacher is shown, that is, the nature of the privileged information; and (iii) when the signal changes, that is, the teacher’s dynamics and the decay of guidance. We restrict our scope to mathematical reasoning, where the method originated and where its failure modes are best documented. We report no new experiments. The contribution is structural: a shared vocabulary for phenomena named differently across papers, and a clear line between what is settled and what is still disputed.
6. 先验证再蒸馏：在策略蒸馏中的提示级教师门控
- 英文标题：Verify Before You Distill: Prompt-Level Teacher Gating for On-Policy Distillation
- arXiv：2609.02998v1
💡 一句话解读：这篇研究要解决的是：在AI模型训练中，用老师模型教学生模型时，老师可能对某些问题自己也不懂，却硬教，导致学生学坏。提出的方法是：先小范围测试老师，确认它在这类问题上靠谱，才让它教；不靠谱就换用其他更稳妥的方式。这样既教得好，又省算力。
在策略蒸馏（On-policy distillation, OPD）中，冻结的教师模型在学生自身的轨迹上提供密集的令牌级监督，以加速后训练。然而，标准的OPD对所有提示统一应用这种监督，不检查教师对每个提示是否可靠。由于反向KL是模式寻求的，一个自信但错误的教师可能引发强烈且误导性的更新。分布代理（如熵或师生似然一致性）衡量不确定性或一致性，但并未直接验证结果正确性。我们提出了教师门控在策略蒸馏（Teacher-Gated On-Policy Distillation, TGOPD），其原则是：在允许密集监督之前，应在提示级别验证教师可靠性。TGOPD通过少量由验证器评分的教师探针来估计可靠性，当可靠性检查通过时，将每个提示专门路由到密集OPD，否则路由到验证器基础的GRPO。在数学、代码和指令遵循的4B和35B学生模型中，TGOPD在所有六个单域设置中均优于标准OPD，并在多域训练下于两个规模均取得更高的七基准平均值。通过利用原本闲置的教师容量进行可靠性估计，TGOPD还减少了异步OPD中教师端的计算浪费，在实测的4B单域运行中，教师节点GPU利用率从9.8%提升至78.9%。
📄 English Abstract
On-policy distillation (OPD) accelerates post-training by providing dense token-level supervision from a frozen teacher on the student’s own rollouts. Vanilla OPD applies this supervision uniformly across prompts, without checking whether the teacher is reliable for each prompt. Because reverse KL is mode-seeking, a confidently wrong teacher can induce a strong yet misleading update. Distributional proxies, such as entropy or teacher-student likelihood agreement, measure uncertainty or agreement but do not directly verify outcome correctness. We introduce Teacher-Gated On-Policy Distillation (TGOPD), built on the principle that teacher reliability should be verified at the prompt level before dense supervision is admitted. TGOPD estimates reliability from a small set of verifier-scored teacher probes and routes each prompt exclusively to dense OPD when the reliability check passes or to verifier-grounded GRPO otherwise. Across 4B and 35B students in mathematics, code, and instruction following, TGOPD outperforms Vanilla OPD in all six single-domain settings and achieves higher seven-benchmark averages at both scales under multi-domain training. By using otherwise-idle teacher capacity for reliability estimation, TGOPD also reduces teacher-side compute waste in asynchronous OPD, increasing teacher-node GPU utilization from 9.8% to 78.9% in the measured 4B single-domain run.
7. 还有哪些需要修复？探索经对话生成的产物中修订传播的高性价比测试时计算
- 英文标题：What Else Needs Fixing? Exploring Cost-Effective Test-Time Compute for Revision Propagation in Artifacts Generated Through Conversation
- arXiv：2609.03254v1
💡 一句话解读：这篇研究要解决的是：当用户让AI在对话中修改一个作品（如文章）的某一部分时，AI常常只改了那一处，而没把相关的地方也一起改对。他们提出了一个测试方法，并比较了多种AI修改策略，发现用几个AI同时给出修改方案再选最好的，既省钱效果又好。
大型语言模型（LLMs）常通过对话中的迭代生成和修订来帮助用户创建产物。这里的挑战是，当用户仅指定局部修改时，LLMs必须识别相关依赖并将修订传播到产物的所有受影响部分。本文研究了LLMs在对话生成产物上的这一能力，其中产物上下文及其依赖可能嵌入在对话历史中。为实用起见，我们还探索了该新场景下性价比高的测试时计算。具体而言，我们为此场景引入了一个新基准，并评估了九种修订方法，包括顺序反思和并行采样变体，使用gpt-oss-20b/120b、gpt-5.4-mini和qwen3.5-9b/27b/122b在该基准上进行测试。结果显示，基线方法的准确率为68.3%–93%，而最具性价比的方法是从三个并行样本中通过基于LLM或中位数选择来挑选，这可将准确率提高2.2%–9.7%。我们的代码和数据集可在https://github.com/ntt-dkiku/llm-revision-propagation获取。
📄 English Abstract
Large Language Models (LLMs) often help users generate artifacts through iterative cycles of generation and revision in conversation. A challenge here is that, when users specify only a local change during revision, LLMs must instead identify the relevant dependencies and propagate the revision to all affected parts of the artifact. This paper studies this ability of LLMs on conversationally generated artifacts, where the artifact context and its dependencies may be embedded in the conversation history. Toward practical use, we also explore cost-effective test-time compute for this new setting. Specifically, we introduce a new benchmark for this setting, and evaluate nine revision methods, including sequential reflection and parallel sampling variants, using gpt-oss-20b/120b, gpt-5.4-mini, and qwen3.5-9b/27b/122b on the benchmark. The results show that baselines achieve accuracies of 68.3–93%, and the most cost-effective method is selecting from three parallel samples using either LLM-based or medoid selection, which improves accuracy by 2.2–9.7%. Our code and dataset are available at https://github.com/ntt-dkiku/llm-revision-propagation.