2026-10-02

Training LLM Judges from Language Feedback via Position-Selective Self-Distillation

👍 5

09/30 08:00

We study training LLM judges from natural language feedback, especially for subjective tasks where the verdict depends strongly on which evaluation criteria the judge invokes and how it weighs them. The dominant approach, outcome-supervised RL (e.g., GRPO), credits every token in the rollout with a

DAGent: Evaluate-then-Grow Planning for Deep Research Agents

👍 3

09/30 08:00

Deep research tasks require agents to navigate large knowledge spaces, synthesize evidence across many sources, and adapt their plans as findings emerge. Directed acyclic graph (DAG)-based multi-agent systems suit this setting because they support parallel execution and isolate each sub-task within

MemLife: Curating and Reasoning over Long-Term Egocentric Video Memories

👍 2

09/30 08:00

Long-term egocentric video enables personalized AI assistants to reason about daily life. However, as video histories grow to hundreds of hours spanning months or years, reprocessing raw clips for every query becomes computationally prohibitive. Memory systems offer a scalable alternative by compact

SpatialCORE: Confidence-Aware Grounded Spatial Reasoning in Large Vision--Language Models

👍 3

09/30 08:00

Large Vision-Language Models (LVLMs) have made remarkable progress across visual perception tasks, yet spatial reasoning remains a persistent weakness, especially for questions that require reasoning over visual space. Recent spatial-reasoning methods incorporate generated grounding, where models pr

Soft Spatial Reasoning

👍 1

09/30 08:00

Large Vision-Language Models (LVLMs) commonly perform spatial reasoning through chain-of-thought (CoT), encoding intermediate reasoning as autoregressive sequences of discrete language tokens. Such hard thinking requires committing to a single token at each step, even when the correct spatial interp

Mitigating the Length-Scaling Tax with Online Distillation

👍 2

09/30 08:00

Length scaling during reinforcement-learning (RL) post-training is often viewed as a sign of improved reasoning ability, especially on difficult problems, but may also make responses to already-solved problems unnecessarily verbose. We quantify this side effect as the length-scaling tax (LST): exces

LEAP: Learned Block-wise Evidence Retrieval for Long Audio-Video Perception

👍 2

09/30 08:00

Hour-scale audio-visual question answering is constrained by a context dilemma: dense whole-recording encoding rapidly exhausts context limits, whereas uniform temporal compression severely dilutes fine-grained acoustic and visual evidence. We introduce LEAP, a framework where the model retrieves it

Removing Timing Shortcuts Improves Non-Invasive Brain-to-Text

👍 4

09/30 08:00

We find that major reported improvements in decoding words from non-invasive brain recordings are largely reproducible without any brain data. In the influential work of d'Ascoli et al. (2025), time series of brain activity from subjects perceiving continuous speech are segmented into fixed-length w

Beyond the Remembered World: Predictive 4D Belief for Persistent Navigation in Evolving Worlds

👍 4

09/30 08:00

Persistent spatial memory enables embodied agents to navigate familiar environments across repeated visits. However, targets may move while unobserved, including during navigation, making remembered locations unreliable by the time an agent arrives. Despite advances in memory retrieval and state pre

Safety of Latent Communication in Multi-Agent Systems

👍 6

09/30 08:00

Latent communication enables multi-agent systems to exchange information directly in internal representation space, reducing the token, computation, and latency overhead of text-based communication. To this end, lightweight trainable links are introduced to map the sender's representations into the

The Evolution of Attention in Large Language Models: Mechanisms, Trade-offs, and Emerging Trends

👍 6

09/30 08:00

Self-attention gives LLMs fine-grained, query-dependent access to context, but dense token interactions incur quadratic prefill cost and a key--value cache growing with context length. Research thus spans explicit-memory compression, sparse access, recurrent state construction, structured state dyna

Tacit-TTS: From Autoregressive Decoding to Masked Prediction for Efficient Transcript-Free Voice Cloning

👍 9

09/29 08:00

TTS systems with autoregressive semantic modeling have demonstrated strong zero-shot voice cloning performance and rich expressive variation, but their sequential decoding incurs substantial latency. Non-autoregressive alternatives offer much faster generation, yet often rely on more restrictive ref

MILO: Automated Harness Discovery via Orchestrated Multi-Agent Evolution

👍 15

09/29 08:00

Modern agentic systems combine an AI model with a harness that controls execution and environmental interactions. Harness design strongly affects long-horizon performance, yet its combinatorial search space demands substantial human effort that must be repeated as models change. Existing automated m

TERRA: Terrain-Aware Reconstruction, Retargeting and Control for Musculoskeletal Locomotion

👍 32

09/29 08:00

Recent advances in musculoskeletal modeling and reinforcement learning have enabled muscle-actuated agents to reproduce increasingly complex human motions. Yet these capabilities remain largely confined to flat ground, in part because motion datasets rarely include aligned terrain geometry and becau

SkillGym: Training Skill-Use Agents with Automatic Verifiable Environment Generation

👍 4

09/29 08:00

Skills equip LLM agents with professional knowledge and guidance to complete long-horizon and complex tasks. Although skills have been widely adopted in recent agent paradigms and harnesses, how to synthesize reliable training data and how to train agents for skill use remain underexplored. In this

LoopVL: Recurrent Visual Intelligence

👍 24

09/29 08:00

We introduce LoopVL to study whether Loop Transformers can be effectively extended to vision- language models. LoopVL combines Module-Loop and Model-Loop computation to iteratively update a unified vision-language state through shared modules. We train LoopVL from scratch through language pre-traini

Overcoming Scaling Limits in On-Policy Self-Distillation for LLM Reasoning

👍 5

09/29 08:00

On-policy self-distillation (OPSD) trains a student to match a privileged teacher distribution along its own sampled trajectory. Standard OPSD applies this supervision to unverified student rollouts while conditioning the teacher on privileged context, typically a reference solution. We separate the

The Geometry of Inference in Transformer Residual Streams

👍 2

09/29 08:00

Transformer language models build predictions through successive residual updates, but how their representations become specific to an eventual outcome remains unclear. We study this process by comparing intermediate residual states with their own final states and an empirical bank of final states f

Dating the Model: Hidden Dates in System Prompts Affect LLM Evaluation

👍 3

09/29 08:00

Reproducibility is essential for scientific research, yet prior work shows that LLM outputs vary with hardware and batching. We identify an overlooked factor: the hidden injection of the current date into system prompts, which users cannot control and which changes every day. Across 9 recent LLMs an

BIABench: Evaluating AI agents on real-world bioimage analysis tasks

👍 4

09/28 08:00

Artificial-intelligence (AI) agents hold promise for automating bioimage analysis, yet no benchmark evaluates whether they can carry out real-world analyses end to end. Such analyses are hard for agents because 2D images, 3D volumes and time-lapse sequences are often too large to read as context, so

Prompted Identity Degrades Cooperation in Multi-Agent LLM Systems

👍 3

09/28 08:00

Multi-agent LLM systems increasingly mix models from several providers, yet exposing each agent's underlying model identity to its peers significantly impairs cooperation. We show that when agents are aware of each other's model family, the group splits into clusters, where agents prefer interacting

SeLMRoute: Probabilistic Semantic Evidence for Large Language Model Routing

👍 5

09/28 08:00

Large language model (LLM) routing aims to select the most suitable model for each incoming query. Most existing routers learn this decision directly from query embeddings, model representations, preference data, or clusters of similar examples. Such approaches can be effective, yet the representati

Rethinking Latent Visual Reasoning: Grounding Latent Reasoning in Visual Evidence

👍 206

09/28 08:00

Latent visual reasoning (LVR) enables multimodal large language models (MLLMs) to perform intermediate computation in continuous latent tokens rather than expressing every reasoning step in words. However, unlike textual CoT, latent reasoning is not directly observable, making it difficult to superv

NavHarness: Towards Lifelong Embodied Navigation

👍 4

09/28 08:00

Frontier models can now perform well on individual embodied navigation tasks through multi-round multimodal reasoning with simple tools. Across successive tasks, however, an agent must also rely on an evolving map and earlier search records, both of which may be incomplete or conflict with new obser

Label-free steering: Compressing test-time reinforcement learning into bias-only subspaces

👍 3

09/28 08:00

Test-time reinforcement learning (TTRL) enables models to improve their reasoning without relying on labeled training data, but existing approaches typically optimize a large fraction of the model parameters. This raises a natural question: can effective test-time adaptation emerge when both the rew

Evaluating Bounded Autonomy in Regulated Agentic AI: A Diagnostic Harness with Constitutional Rewards, Escalation Labels, and Runtime Governance

👍 3

09/28 08:00

We propose RegLLM, a diagnostic harness for bounded autonomy in regulated agentic workflows. It instruments six trustworthiness signals: citation validity, source grounding, schema compliance, escalation correctness, constitutional alignment, and unsafe-action rate. Signals are distinguished by thei

SlideDP: Scaling Host-Resident LLM Fine-Tuning Across Multiple GPUs

👍 2

09/28 08:00

Host-resident layer streaming enables full-parameter LLM fine-tuning beyond GPU memory, but data-parallel ranks compete for shared host resources. Replicated transfers amplify traffic, while strong scaling can expose host work as computation windows shrink. We present SlideDP, a synchronous data-par

CheatBench: Measuring Reward Gaming in AI Agents

👍 4

09/28 08:00

Reinforcement learning has helped AI agents solve increasingly difficult tasks, but high rewards do not always reflect the work users intended. In recent incidents and controlled evaluations across the AI industry, agents trained to maximize reward have accessed unauthorized information, attempted t

BiasReducer: Adaptive Bias Mitigation for Reward Models

👍 18

09/26 08:00

Reward models score responses from large language models (LLMs) and guide LLM training toward human preferences. However, reward models can favor superficial attributes such as length or confidence, leading LLMs to produce higher-scoring but not more correct responses. Existing mitigation methods ei

CUA-SWE: When Computer-Use Agents Meet Visual Software Engineering

👍 13

09/26 08:00

Software development requires more than editing code: developers repeatedly run software, interact with its interfaces, visually inspect its behavior, and use these observations to decide what to change next and whether a change works. Existing coding agents and computer-use agents are largely studi

Teamily vs. OpenAI Dots: The Vision "Team of Dots" is Already Here, Using by Million of Users

@Teamily_AI · 1.6K 粉丝 · 749.2K 阅 · 512 赞 · 170 转

09/30 04:00

OpenAI just shared its vision for Dots: an always-on personal agent today, evolving over time into teams of specialized agents working together. We think that is exactly where AI is going. It is also

中文介绍 Teamily 与 OpenAI Dots 对比，探讨 AI 个性化代理团队的未来趋势。

The Rise of the Automation Engineer

@jakeserval · 1.1K 粉丝 · 406.0K 阅 · 505 赞 · 1.1K 转

10/01 22:00

There is a lot of talk about the jobs AI takes away and very little about the ones it creates. While pessimism might be easier (and certainly gets more clicks), I’m more optimistic. From where I sit,

中文介绍 自动化工程师的崛起，探讨 AI 创造的就业机会。

Free the models: harness design at the frontier

@pirroh · 16.2K 粉丝 · 234.6K 阅 · 507 赞 · 78 转

09/30 00:48

Model routers are everywhere right now, but they have a fundamental limitation. No matter if based on advanced heuristics or a small model that reads each turn and picks which LLM to use, a router

中文介绍 模型路由器的局限性与设计创新。

Why LLMs can't write well and how Sherpa fixes it

@RohanNayak2 · 13.7K 粉丝 · 230.7K 阅 · 514 赞 · 142 转

10/01 00:32

As we pivoted to AI-written shows, countless people told us this would kill Pocket FM. After six long months, I almost believed them. But then our ecosystem exploded, and not despite AI writing but

中文介绍 大型语言模型写作困境与 Sherpa 的解决方案。

Safeguarding the AI Era of Biology: Watermarking the Building Blocks of Life

@pushmeet · 29.0K 粉丝 · 209.0K 阅 · 502 赞 · 88 转

09/30 23:12

Today, we announced SynthID Bio, a method to watermark and detect AI-designed proteins, the building blocks of life. We believe this is the first time humanity has successfully synthesised AI-designed

中文介绍 SynthID Bio 技术，为生命的基本元素打上水印。

History Rhymes: Large Language Models Off to a Bad Start?

@michaeljburry · 2.1M 粉丝 · 120.2K 阅 · 519 赞 · 84 转

09/30 23:04

This 145-year-old case study – presented at the Smithsonian Institute no less - provides a potentially devastating critique of today’s Large Language Models and the spending behind them. Michael Burry

中文介绍 145 年案例研究对大型语言模型的批判。

Why I'm joining Cognition

@alexstamos · 91.8K 粉丝 · 97.2K 阅 · 501 赞 · 37 转

09/29 23:55

There are real safety and security issues raised by AI, but they can be fixed. It's time to get to work instead of freaking out. That's why I'm joining Cognition as CISO. I really believe in the

中文介绍 加入 Cognition 担任 CISO，解决 AI 安全问题。

The FDE boom is a product gap

@EugenAlpeza · 2.5K 粉丝 · 96.2K 阅 · 539 赞 · 50 转

09/29 22:07

The industry is hiring its way around a missing product. Every enterprise trying to put agents to work runs into the same problem: agents don’t know how the company operates, and the operators who do,

中文介绍 企业级代理应用的产品差距。

We can and must solve alignment

@eric_ho · 4.4K 粉丝 · 66.2K 阅 · 505 赞 · 72 转

10/01 02:08

In San Francisco, it feels like the eve of the singularity. Yet, a walk through the city looks surprisingly mundane, with driverless cars and rolling fog equal parts of the scenery. The world feels

中文介绍 解决 AI 对齐问题，迎接奇点的到来。

how to build agentic systems for knowledge work

@arscontexta · 13.6K 粉丝 · 41.4K 阅 · 510 赞 · 45 转

10/01 04:38

the goal is to build software around your agents that gives them the context, tools and methods for a particular kind of work @garrytan put it like this: but what are people really trying to achieve

中文介绍 构建针对知识工作的代理系统。

Why Instinct feels Magical

@yashhsm · 27.2K 粉丝 · 35.7K 阅 · 516 赞 · 24 转

10/01 03:34

how instinct, a $10b startup works - by reverse-engineering its harness: single-threaded chat, memory, computer use, background tasks etc One month ago, i got an invite for instinct my first

中文介绍 直觉在 Instinct 创始公司中的作用。

DeepSeek Harness for Desktop Is Here

@DeepSeekHarness · 989 粉丝 · 35.4K 阅 · 1.0K 赞 · 69 转

10/01 00:13

Today, we’re releasing the DeepSeek Harness v0.2 preview, with a desktop app available for macOS and Windows. You can install it here and follow the instructions to get started. This release comes

中文介绍 DeepSeek Harness for Desktop 桌面应用程序发布。

Building a design system for coding agents

@arlanoska · 3.5K 粉丝 · 25.4K 阅 · 503 赞 · 29 转

09/30 02:26

You ask an agent to build a page. Then you fix the spacing, swap the inputs and explain which button it should use. The next agent makes different choices, so you do it all again. If your repo has

中文介绍 为编码代理构建设计系统。

Creative prompting with Gemini Omni in Google Flow

@FlowbyGoogle · 73.2K 粉丝 · 23.0K 阅 · 539 赞 · 51 转

10/01 04:46

To help get you started exploring Gemini Omni Flash in Google Flow, here is a comprehensive guide featuring prompt templates and detailed ideas for your next creative generations. Whether you are

中文介绍 使用 Gemini Omni 在 Google Flow 进行创意提示。

September at OptimAI: Building the Next Layer of the Agentic Internet

@OptimaiNetwork · 98.4K 粉丝 · 10.7K 阅 · 513 赞 · 138 转

09/30 23:08

September marked an important step in OptimAI's evolution. Over the past month, we have been expanding the products people interact with while continuing to strengthen the infrastructure underneath

中文介绍 OptimAI 构建代理互联网的下一层。

Academia is for Ambition — Alex Zhang, MIT

Latent Space · 10/02 08:28

We catch up with RLM first author Alex Zhang, MIT PhD, on Jev, PhD masxing, and the future of harnesses.

中文介绍 MIT博士Alex Zhang讨论了Jev、PhD masxing以及利用的未来的问题。

The eternal complement

OpenAI News · 10/02 01:00

Advanced AI may matter most for the routine work behind breakthrough ideas. Explore why execution could shape the next economy and the pace of progress.

中文介绍 高级AI可能在突破性想法背后的日常工作中最重要。探讨执行如何塑造下一个经济和进步速度。

How Albertsons Companies is reimagining retail from the inside out

OpenAI News · 10/02 00:00

Albertsons Cos. is using ChatGPT Enterprise and the OpenAI API to help teams work faster and make grocery shopping easier for millions of customers.

中文介绍 Albertsons公司利用ChatGPT和企业API来帮助团队提高工作效率，并为数百万客户简化购物体验。

Introducing Olmo-core 3: Open, scalable training infrastructure for large MoEs

Hugging Face Blog · 10/01 23:01

中文介绍 介绍Olmo-core 3：大型MoE的开源、可扩展训练基础设施。

[AINews] Gemini 4 Argon: GDM’s answer to Astra/Fable, with 1M output

Latent Space · 10/01 14:45

... but you can’t try it yet unless you are “government users and trusted cyber defenders in the Fairwind Program”

中文介绍 Gemini 4 Argon：GDM对Astra/Fable的回应，但只有政府用户和Fairwind计划中受信任的网络安全防御者才能试用。

The Den frees up 10-15 hours a week to grow with ChatGPT Work

OpenAI News · 10/01 08:00

As it opens a new location, the social club prepares grant applications in 2 hours instead of 3 days and liquor-license materials in 3 hours instead of 4 days.

中文介绍 The Den社交俱乐部使用ChatGPT Work将准备赠款申请的时间从3天缩短到2小时，将酒类营业执照材料准备时间从4天缩短到3小时。

Gemini 4 Argon 🧠, inside OpenAI’s chip ⚡, distillation takedown 🕵️

TLDR AI · 10/01 08:00

中文介绍 Gemini 4 Argon：OpenAI芯片内部，蒸馏降级分析。

Why Dwarkesh is Wrong about Computer Use + How OpenAI shipped its Jev competitor in 1 Week

Latent Space · 10/01 06:23

Our DevDay coverage - the first pod on the DevDay lineup - dives in with the leaders of OpenAI’s CUA team and API platform.

中文介绍 探讨Dwarkesh关于计算机使用观点的错误，以及OpenAI如何在1周内推出其Jev竞争对手。

Gemini 4 Argon: our next era of frontier intelligence

DeepMind Blog · 10/01 04:01

中文介绍 Gemini 4 Argon：我们前沿智能的新时代。

Introducing SynthID Bio

DeepMind Blog · 09/30 23:03

Proof of concept for watermarking AI-generated proteins while preserving biological function.

中文介绍 SynthID Bio的引入：在保留生物功能的同时，对AI生成的蛋白质进行水印标记的证明概念。

“We’re not going to shoot ourselves in the foot” over hack fallout, says OpenAI’s chief research officer

MIT Tech Review AI · 09/30 18:40

Two months after the bombshell news that a swarm of its agents had broken their containment and hacked into the computers of the AI company Hugging Face, OpenAI is still putting out fires. A steady drip of disclosures about other hacks in the weeks since has kept OpenAI in the spotlight and raised s

中文介绍 OpenAI首席研究官表示，在Hugging Face黑客事件后，OpenAI不会因黑客事件而自毁前程。

Disrupting a coordinated model-distillation campaign

OpenAI News · 09/30 18:30

Learn how OpenAI disrupted a campaign to extract protected model reasoning and is strengthening defenses against adversarial distillation.

中文介绍 OpenAI破坏了一个协调的模型蒸馏活动，并加强了对抗蒸馏的防御。

Helping small businesses put AI to work

OpenAI News · 09/30 18:00

OpenAI is partnering with America’s SBDC to expand hands-on AI training and local support for small businesses, alongside a new report on how small teams are using AI.

中文介绍 OpenAI与America’s SBDC合作，为小型企业提供AI培训和支持，并发布了一份关于小型团队如何使用AI的新报告。

[AINews] OpenAI DevDay 2026: Dots, 6.1 Sol, Ultrafast, Decisions API, Agents API, Spaces, Marketplace, and 1.2 Billion ChatGPT WAU

Latent Space · 09/30 13:53

the most confident DevDay yet.

中文介绍 OpenAI DevDay 2026：最自信的DevDay，包括Dots、6.1 Sol、Ultrafast、Decisions API、Agents API、Spaces、Marketplace和1.2亿ChatGPT WAU。

Open TTS Leaderboard: Scalable Evaluation for Multilingual Text-to-Speech and Voice Cloning

Hugging Face Blog · 09/30 08:00

中文介绍 Open TTS排行榜：多语言文本到语音和语音克隆的可扩展评估。