2026-08-04

From RLVR to RLSVR: Task Transformation Induces Self-Verifiable Rewards for Open-Ended LLM Self-Improvement

👍 77

07/26 08:00

Reinforcement Learning with Verifiable Rewards (RLVR) has driven recent progress in reasoning-oriented large language models (LLMs) by enabling large-scale optimization. However, its applicability remains largely limited to domains such as mathematics and coding, where correctness can be determinist

中文介绍 针对RLVR仅适用于有确定答案的数学、代码等任务，提出将开放任务转化为可自验证形式（RLSVR），通过任务变换产生自验证奖励，使LLM能在开放式任务上自我改进。核心是让奖励信号由任务结构自动生成，扩展强化学习适用范围。

Mental World Modeling

👍 59

07/29 08:00

World models enable a predictive substrate for planning and action, yet existing formulations merely answer a physical question: what/where it is, and how will it evolve. Human behavior, however, is driven by hidden mental state (what a person believes, wants, intends, feels, and considers socially

中文介绍 世界模型传统上只预测物理状态，本文提出心理世界建模，建模人的隐藏心理状态（信念、欲望、意图、感受及社会规范），以解释和预测人类行为。核心是将心理状态作为潜在变量纳入预测框架，用于规划和行动。

Weak-to-Strong On-Policy Distillation

👍 44

07/28 08:00

On-policy distillation (OPD), which aligns a student with the teacher's token-level distribution on the student's own rollouts, is an effective paradigm for transferring capabilities across LLMs. Prevailing approaches assume a teacher at least as capable as the student: they either distill a larger

中文介绍 现有在线策略蒸馏（OPD）通常假设教师模型至少与学生一样强，本文研究弱到强蒸馏：用较弱教师指导学生，挑战传统假设。核心在于学生从较弱教师的token分布中学习，探索能力传递的新范式。

AISPA: User-Centric System Prompt Auditing for Large Language Model Applications

👍 32

07/30 08:00

System prompts are instructions configured by developers to govern the behaviors of foundation models in AI applications. They are used throughout commercial AI products, but are rarely disclosed to the public or regulators, creating a serious trust and accountability gap in the wide deployment of A

中文介绍 系统提示词由开发者配置控制模型行为，但商业产品很少公开，造成信任与问责缺口。AISPA提出以用户为中心的系统提示词审计方法，对LLM应用的系统提示进行审计，增强透明度与可信度。

SAF-OPD: Stable Advantage Fusion for On-Policy Distillation

👍 25

07/31 08:00

Reinforcement learning with verifiable rewards (RLVR) broadcasts a single response-level reward to every token, while on-policy distillation (OPD) scores each token against a stronger teacher for a dense advantage but caps performance at teacher quality and discourages exploration beyond it. Their c

中文介绍 RLVR将单一响应级奖励广播给所有token，OPD则用更强教师提供逐token稠密优势，但性能受限于教师质量且抑制探索。SAF-OPD提出稳定优势融合，结合两者优点，平衡教师引导与探索。

Scaling Properties of Text Conditioning in Visual Generation

👍 25

07/31 08:00

We study empirical scaling properties for text conditioning in visual generation. Such properties have rarely been measured because diffusion loss does not scale with the number of tokens in natural-language prompts. Surprisingly, we find that the converged diffusion loss scales with the amount of s

中文介绍 研究视觉生成中文本条件的缩放规律。扩散损失不随自然语言提示的token数缩放，但发现收敛后的扩散损失与提示相关数据规模呈缩放关系，揭示文本条件化在扩散模型中的扩展行为。

QQWorld: Quantile-Quantile Matching for World Model Regularization

👍 24

07/30 08:00

Latent world models enable efficient planning by predicting future states in a compact representation space, but their performance depends critically on the quality of the learned latent distribution. LeWorldModel (LeWM) regularizes its latents toward an isotropic Gaussian using the Epps-Pulley (EP)

中文介绍 潜在世界模型在紧凑表示空间预测未来状态，性能依赖潜在分布质量。LeWM用Epps-Pulley检验将潜在正则化为各向同性高斯。QQWorld采用分位数-分位数匹配进行世界模型正则化，改进潜在分布建模。

See2Think: Do Multimodal Models Really Use Intermediate Visual States?

👍 23

07/29 08:00

Multimodal large language models increasingly use sketches, annotations, tools, and intermediate images during reasoning, but it remains unclear whether they truly rely on these visual states. Existing benchmarks are limited both by task collections with narrow coverage or partially text-solvable sa

中文介绍 多模态大模型在推理中日益使用草图、注释、工具和中间图像，但尚不确定是否真正依赖这些中间视觉状态。本文构建基准测试，在任务覆盖和避免文本可解样本上改进，检验模型对中间视觉状态的实际依赖。

Fewer Clarifications, Better Code: Benchmarking Cross-Session Personalized Ambiguity Adaptation in Coding Assistants

👍 21

07/29 08:00

AI-assisted coding increasingly translates informal user intent into executable software, yet coding requests often contain ambiguities that recur in user-specific ways across tasks and sessions. Existing disambiguation methods typically address each ambiguous request in isolation within the current

中文介绍 AI辅助编程中，用户请求常含歧义且跨会话以个性化方式重复。现有方法在单次会话内孤立处理。本文提出跨会话个性化歧义适应基准，衡量编码助手能否利用历史交互减少澄清次数并生成更好代码。

β-OPSD: Deriving with Policy Optimization, Training with Self-Distillation

👍 21

07/30 08:00

On-policy self-distillation (OPSD) is a promising approach to improve reasoning language models, but it remains brittle in practice: making it work reliably often requires substantial engineering effort. We identify a structural source of this difficulty: vanilla OPSD is precisely the β=1 member of

中文介绍 在线策略自蒸馏（OPSD）可提升推理模型但实践脆弱，需大量工程。本文发现普通OPSD是β=1的特例，提出β-OPSD：用策略优化推导，用自蒸馏训练，通过调整β获得更稳定的训练动态。

SwanTale: Unified Multi-Speaker Speech and Audio Generation for Instruct and Zero-Shot Tasks

👍 20

08/03 08:00

Speech and audio generation is often needed in animation dubbing, audio drama, movies, advertising, games, podcasts, and short-video production. In these scenarios, creators may need to design voices without reference recordings, control speaker styles with natural language, support acoustic scenes

中文介绍 面向动画配音、有声剧、广告等场景的多说话人语音与音频生成统一模型。支持无需参考录音设计音色、用自然语言控制说话人风格、声学场景等，适合指令跟随和零样本任务。

ShadowDancer: Teaching Video World Models Any Action by Learning Unified Dynamics Representations from a Video and Its Shadow

👍 17

07/30 08:00

We present ShadowDancer, a novel approach to any-action, frame-level control of interactive video world models. The obstacle is representational: existing interfaces either encode an action loosely, leaving how it unfolds for the model to improvise, or encode it exactly through structured signals th

中文介绍 交互式视频世界模型通常难以精确控制任意动作。ShadowDancer提出从视频及其阴影中学习统一动力学表示，实现帧级任意动作控制，避免现有接口动作编码松散或过结构化的问题。

Enhancing Rubric-based RL via Self-Distillation

👍 16

07/21 08:00

Rubric-based RL has recently shown promise in improving LLMs on open-ended tasks. A widely recognized limitation of rubric-based RL is limited exploration: criteria that no rollout manages to satisfy (Unexplored Criteria, UC) receive no optimization signal. Recent methods address this by incorporati

中文介绍 基于评分标准的RL在开放任务上提升LLM，但存在探索不足：未达到任何标准的样本（Unexplored Criteria）得不到优化信号。本文结合近期方法与自蒸馏，增强对未探索标准的信号覆盖，提升训练效果。

ExtractBench: A Benchmark for Schema-Guided Enterprise Document Extraction

👍 15

07/31 08:00

Enterprise workflows increasingly rely on agents for schema-guided extraction: given a document and a user-defined schema, the agent faithfully follows the schema to produce the correct output with source evidence as grounding metadata. We present ExtractBench, a benchmark for schema-guided extracti

中文介绍 面向企业文档的模式引导抽取：给定文档和用户定义的schema，代理遵循schema输出正确结果并附来源证据。ExtractBench提供基准以评估此类抽取任务的性能，支持企业流程中的可靠性验证。

Σ-Mem: An Online Reliability Memory for LLM-based Multi-Agent Systems

👍 14

07/30 08:00

Memory is central to long-horizon LLM agents, yet existing memory systems primarily preserve interaction content rather than modeling which agents can be trusted and under what conditions. This limitation is particularly important in multi-agent systems, where a central model may be unable to direct

中文介绍 现有LLM多智能体记忆系统主要保存交互内容，不建模哪个智能体在什么条件下可信。Σ-Mem提出在线可靠性记忆，记录和更新各智能体的可信度，供中央模型判断协作策略，提升长程多智能体系统的稳健性。

Evaluation-Verification Reward for Consistent Multi-Reference Image Editing

👍 11

07/31 08:00

While recent image editing models have made rapid progress, multi-reference editing remains challenging, particularly in maintaining visual consistency across references and ensuring overall visual harmony. Reinforcement learning has proven highly effective for text-to-image generation and single-im

中文介绍 多参考图像编辑需保持跨参考的视觉一致性与整体和谐，现有模型难以兼顾。本文提出评估-验证奖励，结合强化学习优化编辑模型，提升多参考一致性视觉质量。

SWE-Touch: Benchmarking Coding Agents When Users Touch the Code

👍 9

08/03 08:00

Real-world software development requires coding agents to operate in shared workspaces where users may inspect and modify code during an ongoing task, yet existing repository-level benchmarks typically evaluate agents working alone or restrict user participation to messages. This leads us to ask: ho

中文介绍 现实软件开发中，用户可能在任务进行中查看和修改代码，但现有仓库级基准通常让代理单独工作或仅允许消息。SWE-Touch提出新基准，模拟用户在共享工作区中的代码修改，评估编码代理与用户的动态协作能力。

Filesystem-Based Memory for LLM Agents: Organization, Evolution, and Sustainability

👍 9

07/29 08:00

Deployed LLM agents increasingly keep their long-term memory as a filesystem: a directory tree of markdown files that the agent itself reads, writes, and reorganizes through generic file tools. Yet research has largely passed over this medium: prior systems design bespoke memory representations and

中文介绍 部署的LLM代理常以文件系统形式维护长期记忆：目录树中的markdown文件，由代理自行读写和重组。此前研究少有涉及。本文系统研究这种记忆的组织、演化与可持续性问题。

UEmbed: Unified Sparse and Dense Multimodal Embeddings

👍 8

08/03 08:00

Sparse retrieval underpins modern search systems, from web search to retrieval-augmented generation. Existing work has introduced Learned Sparse Retrieval (LSR) to push beyond exact lexical matching toward richer semantics. Yet LSR has so far remained tied to encoder-style bidirectional architecture

中文介绍 稀疏检索是搜索系统基础，学习型稀疏检索（LSR）超越字面匹配，但局限于编码器式双向架构。UEmbed提出统一的稀疏与稠密多模态嵌入方法，结合两者优势，扩展LSR到生成式或多模态场景。

StyleForge: Indoor Furniture Styling by Counterfactual Reasoning in a Hypergraph Field

👍 5

08/03 08:00

Fixed-layout indoor furniture styling requires selecting assets that form a coherent room without changing the prescribed furniture categories, positions, orientations, or scales. Existing approaches typically retrieve each asset independently or rely on static local relations, making them prone to

中文介绍 固定布局室内家具风格化要求在不改变家具类别、位置、朝向等前提下选择成套家具。现有方法独立检索或用静态关系，易失协调。StyleForge在超图场中通过反事实推理建模家具间复杂关系，实现整体协调的风格化。

I went on paternity leave. My agent built a $3M pipeline from 2,741 conversations. Here's the code.

@wayne_liang_ · 1.5K 粉丝 · 999.9K 阅 · 637 赞 · 279 转

08/04 00:47

For eight weeks, a clone of me ran our sales front line: 300+ prospect calls a week, 132 new paying customers, 37 enterprise opportunities worth ~$3M. It also made up a price, emailed internal notes

中文介绍 博主休陪产假八周，AI分身代跑销售一线：每周300+通电话、132个新付费客户、37个企业机会，管道约300万美元。AI还会编造报价、发内部备忘，代码已开源。

PayBox 101

@moonpay · 421.1K 粉丝 · 32.0K 阅 · 547 赞 · 62 转

08/02 21:18

MoonPay, the global financial technology company powering the movement of value across fiat and digital assets, has launched PayBox, the first payment vault built for AI that lets a person's AI agent

中文介绍 MoonPay发布PayBox，自称首个为AI打造的支付保险库，让AI代理能管理资金、执行支付。MoonPay是全球法币与数字资产互转金融科技公司，此举瞄准AI代理自主交易场景，属新产品发布。

Using AI to learn

@Hi_Mrinal · 22.2K 粉丝 · 20.2K 阅 · 528 赞 · 41 转

08/02 18:13

The best way to learn from AI is to really reverse the whole process like to use ai for deep understanding rather than just "getting the answer" revert the typical workflow, instead of asking the AI

中文介绍 学AI的正确方式不是索要答案，而是反转流程：让AI辅助深度理解。比如请它拆解原理、展示推理过程，而不是直接输出结果。这是关于“用AI学习”的方法论分享，强调从被动接受到主动建构的转变。

[AINews] Qwen 3.8 Max(2.4T) and 27B, new open weights models for Coding and Cowork

Latent Space · 08/04 11:49

Qwen is so back!

The Inference Engineering Masterclass — Philip Kiely & Ali Taha, Baseten

Latent Space · 08/04 05:44

Baseten just raised a $13B Series F and is now one of the leading kings of inference engineering. We go into everything you need to know for autoregressive and diffusion engineering.

Trump’s AI protectionism has come for robotics

MIT Tech Review AI · 08/04 02:43

This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here. Humanoid robots usually elicit more cringe than awe: They stumble, kick children, and despite advances are still worse at using their hands than my toddler. It’s

Here’s why AI agents lie and cheat to reach their goals

MIT Tech Review AI · 08/03 16:30

MIT Technology Review Explains: Let our writers untangle the complex, messy world of technology to help you understand what’s coming next. You can read more from the series here. When two OpenAI models hacked into the website Hugging Face in July, they weren’t trying to make money or commit sabotage

How we built a realtime system for responsive voice AI in six months

OpenAI News · 08/03 15:00

GPT-Live enables continuous voice interaction with AI, using a turnless speech model and low-latency architecture for faster, more natural conversations.

Circles powers telco personalization with OpenAI technology

OpenAI News · 08/03 08:00

Circles uses the OpenAI API and Codex to power AI-native telco experiences, increasing ARPU by 22%, reducing churn by 9%, and improving development efficiency.

DeepSeek V4 Flash ⚡, OpenAI’s math breakthrough 🔢, Qwen 3.8-Max 🤖

TLDR AI · 08/03 08:00

[AINews] not much happened today

Latent Space · 08/01 09:38

apart from DeepSeek V4-Flash 0731, a quiet day.

Ten advances in mathematics and theoretical computer science

OpenAI News · 08/01 08:00

OpenAI shares new results on long-standing open problems in mathematics and theoretical computer science, including advances in geometry, cryptography, and complexity.

Advancing responsible AI across Europe

OpenAI News · 07/31 23:00

OpenAI shares how its safety, security, transparency, and provenance practices support responsible AI governance in Europe. The work will continue as the EU AI Act advances.

Building abundant intelligence

OpenAI News · 07/31 23:00

A full-stack approach to making advanced AI more capable, more affordable, and more widely useful.

Univé builds an AI-ready workforce

OpenAI News · 07/31 15:00

See how Univé built an AI-ready workforce with ChatGPT Enterprise by combining leadership, responsible governance, and employee-led innovation to transform work at scale.

not much happened today

Smol AI News · 07/31 13:44

**DeepSeek** launched the public-beta of **DeepSeek-V4-Flash API**, boasting a significant post-training performance leap without architecture or size changes, achieving a **Terminal-Bench score of 82.7** and nearing **GPT-5.6 Luna's 51** score at about **60% lower cost per task**. The model feature

[AINews] GPT 5.6 price cut by 20%-80%: Cost of GPT 5.4 Intelligence dropped 13x in 4 months due to GPT 5.6 recursive self-optimization

Latent Space · 07/31 12:40

Distillation is all you need!

Disrupting a Criminal Scam Operation

OpenAI News · 07/31 08:00

OpenAI disrupted a Cambodia-based scam operation using ChatGPT to support investment, romance, gambling, and impersonation schemes.