2026-09-30

Can We Trust the Teacher? Decoupled Credit Direction-Magnitude for Self-Distillation

👍 3

09/28 08:00

RLVR provides reliable trajectory-level credit, while OPSD offers dense supervision for token-level credit. This exposes a fundamental coupling when updating step-level credit direction and magnitude with teacher supervision, preventing steps from receiving reliable credit directions and contributio

SentZero: An Enhanced Sentence-Centric Vision-Language Pretraining for Multi-Task Zero-Shot Chest X-Ray Analysis

👍 25

09/28 08:00

Vision-language (VL) pretraining using paired chest X-ray (CXR) images and radiology reports has shown strong potential for medical image understanding. However, existing methods often remain dependent on task-specific finetuning because radiology reports are lengthy, clinically dense, and difficult

Who Gets a Token, and What Does It Carry? Unequal Name Support and Concept Access in Large Language Models

👍 2

09/28 08:00

Names are personal identifiers, but they also carry social meaning and are widely used to evaluate how language models treat different people. Such evaluations typically assume that matched names are comparable model inputs. We show that this assumption often fails at the lexical interface: matched

Reinforcing Agentic Creativity in Scientific Ideation with Night Science

👍 2

09/28 08:00

Large language models (LLMs) excel at structured, verifiable tasks, but their low-entropy bias can produce homogeneous and predictable outputs, limiting their utility for open-ended scientific ideation. Effective discovery, however, spans a broader creative spectrum: from structured day science to l

SCOPD: Sparse-Context On-Policy Self-Distillation for Efficient Vision-Language Models

👍 2

09/28 08:00

Reasoning vision-language models (VLMs) process images and videos as long sequences of visual tokens, making inference expensive. Training-free token pruning reduces this cost, but aggressive compression can sharply degrade performance, often attributed to irreversible loss of task-relevant visual i

Distillation Defenses Easily Break After Reinforcement Learning

👍 3

09/28 08:00

Distillation attacks copy the reasoning capabilities of closed-source large language models, allowing bad actors to replicate state-of-the-art performance at low cost. Attackers systematically collect a large volume of frontier model reasoning traces and then train (i.e., "distill") their own models

FactorEngram: Factorized N-gram Memory with Basis-Level Gating for Language Models

👍 3

09/28 08:00

Lookup-based memory has been a promising way to scale the parameters of large language models (LLMs). It retrieves learned representations of local token patterns, such as n-grams, instead of reconstructing them through successive layers of computation. However, existing designs such as Engram treat

Hard Vision, Easy Vision: What GPT-6 Astra Reveals Across Computer Vision

👍 15

09/28 08:00

Frontier general-purpose systems are rapidly expanding beyond visual understanding into capabilities traditionally handled by dedicated computer-vision models. As these capabilities expand, a central question for the computer-vision community is how far this reach extends, and what remains hard. We

Nereus: Adaptive Parallelism for LLM Post-Training

👍 9

09/28 08:00

Reinforcement learning (RL) post-training for large language models (LLMs) coordinates multiple models across generation, inference, and training on GPU clusters. Several factors may change during a run, including resource availability, sequence length, memory pressure, and stage bottlenecks. As a c

FlowTool: Controlling Tool Parameter in Image Retouching via Flow Matching

👍 15

09/28 08:00

Tool-based image editing (image retouching) is commonly formulated with autoregressive multimodal large language models (MLLMs) that sequentially generate reasoning, tool selections, and parameter values. In this work, we present a novel approach to tool-based image editing by framing the task as a

An RL View of OPD: Least Square Policy Distillation for Sample-Efficient LLM Reasoning

👍 10

09/28 08:00

We study on-policy distillation (OPD) through the lens of reinforcement learning, establishing a connection between the reverse-KL objective in OPD and KL-regularized policy optimization. Building on this connection, we introduce Least-Square Policy Distillation (LSPD), an RL-inspired framework that

Imprint Reader: From Weight-Update Readout to Behavioral Intervention

👍 11

09/28 08:00

As language models take a growing role in AI development, a natural aspiration is for them to reflect on their own learning process, as humans do, and use that reflection to improve themselves. At the same time, these models have an advantage that human learners lack, since training leaves parameter

Do We Really Need KL Divergence for On-Policy Distillation of Large Language Models?

👍 0

09/27 08:00

Since the advent of knowledge distillation, KL divergence has been the standard loss in distillation. Recently, on-policy distillation (OPD) has emerged as an efficient post-training paradigm for LLMs. As a distillation method, OPD naturally inherits KL divergence as its standard loss. However, in t

Pruned CTC for Memory-Efficient Large-Vocabulary ASR Training

👍 3

09/27 08:00

Connectionist temporal classification (CTC) naturally supports offline and streaming speech recognition with utterance-level supervision, but conventional implementations materialize frame-by-vocabulary activations in memory, making CTC training with native LLM vocabularies prohibitively memory-inte

DISCO: Distributed Long Context Scaling with Grounding-Reasoning Disaggregation

👍 2

09/27 08:00

While Large Language Models (LLMs) advertise million-token context windows, reasoning quality often collapses as inputs grow -- a phenomenon termed context rot. This failure stems from a structural entanglement in monolithic architectures, where the massive search burden of contextual grounding exha

Recursive Harness Distillation across Agents for Robot Manipulation

👍 29

09/27 08:00

A central goal in robotics is to enable manipulation across changing tasks and environments. Vision-language-action (VLA) models provide broad manipulation capabilities but can struggle when execution requires diagnosing failures and adapting behavior. Strong agents can discover effective interventi

SMAT: Simple and Efficient Merge-Aware Training

👍 2

09/27 08:00

Model merging integrates the capabilities of multiple experts without joint retraining, but standard expert training optimizes task loss alone and does not guarantee good performance after merging. Merge-aware training (MAT) aims to improve merged performance, but existing methods do not fully accou

What masking geometry works best for EEG foundation models?

👍 1

09/27 08:00

EEG foundation models hold promise for scalable brain-signal decoding across clinical and cognitive neuroscience applications, yet their pre-training pipelines remain poorly understood. Among design choices, the masking strategy is particularly critical: it determines what the network must predict a

KernelZero: Co-Evolving Proposer and Coder for Continuously Improved GPU Kernel Generation

👍 3

09/27 08:00

High-performance GPU kernels are essential to modern machine learning systems, yet automatically generating kernels that are both correct and efficient remains challenging. Existing LLM-based approaches face two major limitations: the scarcity of high-quality training data aligned with the model's c

QwenGyre: An Elastic Reinforcement Learning Framework for Training xLong-Horizon Agents

👍 27

09/27 08:00

Large language model (LLM) agents increasingly undertake extreme-long (xlong) horizon tasks, where a single execution can span hours, hundreds of model--environment interactions, and nearly 1M tokens per rollout. Applying online reinforcement learning (RL) to such executions poses two fundamental ch

ExpVoyager: Direct Experience Navigation for Dynamic Agent Skill Synthesis

👍 8

09/26 08:00

Learning from experience in LLM agents has become a key paradigm for developing self-evolving agents that continuously learn and expand their capabilities. Within this paradigm, synthesizing the agent skill has emerged as a promising solution for transforming accumulated experience into reusable pro

Allspark: Weak to Strong Transfer via Alternating Chain of Thought

👍 2

09/26 08:00

Recent progress in frontier models has renewed interest in large-scale reinforcement learning (RL), but the cost of generating large-model rollouts makes even testing RL recipes expensive. We ask whether reasoning improvements learned by a small, weak model can benefit a larger, stronger model witho

Playing to Par: Reinforcement Learning for Provably Optimal Quadrilateral Block Decompositions

👍 2

09/26 08:00

A quadrilateral block decomposition of a planar domain is judged by whether it is complete, whether its elements are well shaped, and how many of its vertices are irregular. The last has a provable floor: the discrete Gauss-Bonnet identity enforces a lower bound on the total vertex irregularity of a

Groupwise Agentic Grading and Advantage Redistribution for Code Agent RL

👍 90

09/26 08:00

Reinforcement learning (RL) for code agents often uses executable tests to provide binary rewards. With these rewards, Group Relative Policy Optimization (GRPO) assigns identical advantages to test-passing trajectories within each rollout group, overlooking differences in implementation quality and

Relic: From Multi-Agent Collaboration to Persistent Organizational Capability

👍 5

09/26 08:00

Multiple agents may often conflict in an organization: for example, one coding agent changes an interface in a repository, but another continues to develop on the old version where existing tests become stale. A conversation can resolve the episode, but when the participants change, what makes the l

Change the Product, Keep the Parameters: Associative Algebra Layers for Transformers

👍 17

09/26 08:00

Fast matrix multiplication algorithms keep the product fixed and search for a cheaper way to evaluate it. We instead ask whether a Transformer's learned projections can use a different, cheaper product altogether. Building on an associative-algebra construction that replaces ordinary matrix multipli

G^2PTQ: Improving LLM Post-Training Quantization with Generalized Gradient Compensation

👍 2

09/25 08:00

Post-training quantization (PTQ) is a practical approach to reducing the memory and computational footprint of large language models (LLMs) without retraining. GPTQ-based methods have become the de facto standard, yet they suffer from two complementary limitations. Methods with local, layer-wise obj

Specification Before Generation: A Pre-Registered, Five-Model Paired Evaluation of a Specification Frame for LLM-Generated Code in Money, Time, Idempotency, and Access Tasks

👍 0

09/20 08:00

Code generated by large language models passes security checks at a rate that has barely moved in four years. In regulated backends, the defect classes that matter most are money arithmetic, time handling, retry safety, and access control. Teams answer with instruction files, yet the largest control

NanoForecast v0.5: Competitive Time Series Forecasting Through Training Pipeline Optimization

👍 2

09/15 08:00

We present NanoForecast v0.5, a 6.5M-parameter forecaster that competes with models 31x its size (TimesFM, 200M parameters) after training pipeline fixes and no architecture change. Retraining the v0.3 architecture with corrected loss-scope handling, tensor shape alignment, and wider augmentation co

Safe Error Correction for Language Models: Frozen-Base Adjustment with Capability Preservation

👍 2

09/14 08:00

We study a practical question: can a small correction module fix errors in a frozen language model's outputs without degrading its base capabilities? We propose CRN v2, a lightweight logit-level correction module (~34M trainable parameters, 0.73% of the 4.65B text module) that sits atop a fully froz

Meet your Team Bots

@n2parko · 5.7K 粉丝 · 274.2K 阅 · 533 赞 · 38 转

09/29 04:03

Today we're introducing Team Bots, and it's likely the biggest change to Grok Bot since launch. I wanted to share a bit about why we built it and how these new digital colleages might change the way

中文介绍 n2parko介绍Team Bots，可能是Grok Bot自发布以来的最大更新，探讨其构建原因及对工作方式的影响。

How to build motion design studio with Opus 5.5 ( Full-course )

@0xMovez · 36.4K 粉丝 · 78.9K 阅 · 737 赞 · 46 转

09/27 22:29

Most people who try motion design with Opus 5.5 end up with the same video: centered text on a gradient, everything fading in, a logo at the end. They don't give it a reference, don't give it a

中文介绍 0xMovez分享用Opus 5.5制作动态设计的工作流程，指出常见问题并提供改进建议。

CC, our experimental AI agent built for families

@GoogleAI · 2.5M 粉丝 · 69.4K 阅 · 508 赞 · 36 转

09/29 02:54

Managing a household means juggling busy schedules, crowded inboxes, and endless to-do lists. It requires context switching, multitasking, and the most precious resource of all: time. Earlier this

中文介绍 GoogleAI推出面向家庭的实验性AI代理CC，旨在帮助管理家庭事务。

Its not just the f*cking sandbox

@joedaroo · 12.2K 粉丝 · 41.2K 阅 · 517 赞 · 76 转

09/28 06:22

A lot of the perspective on all the AI incidents has been shared from the outside in, and little has been said from the inside looking out, through the lens of a security person living through it.

中文介绍 joedaroo从安全人员的角度分析AI事件，提供内部视角。

i made 10+ motion videos with opus 5.5 in 3 days. here's the whole pipeline (open source)

@RaphaelAubryy · 393 粉丝 · 40.9K 阅 · 507 赞 · 42 转

09/28 17:25

Most people trying motion design with Opus 5.5 hit the same wall: ↳ the first render is "mid": centered text, a gradient, everything fading in ↳ the music and the picture live separate lives ↳ one

中文介绍 RaphaelAubryy分享使用Opus 5.5制作10+个动态视频的完整工作流程，开源分享。

How I Make $10K Launch Videos for $0 with Opus 5.5 (Full Course)

@twoclipping · 7.5K 粉丝 · 40.5K 阅 · 510 赞 · 35 转

09/28 13:42

im open sourcing my entire workflow for making the motion designs that studios charge $5,000 to $15,000 for this uses 0 mcp, 0 external tools and im not selling you any subscription you dont need

中文介绍 twoclipping开源制作价值1万美元启动视频的工作流程，使用Opus 5.5且无需额外工具。

You Don't Hate AI Art. You Hate Admitting You Like It

@lilykeilani011 · 3.0K 粉丝 · 10.7K 阅 · 521 赞 · 71 转

09/28 05:36

The Problem Isn't Always the AI It's no secret that rejection of AI-generated art is highly visible on social media. Open any post related to AI art and you'll find hundreds of negative comments.

中文介绍 lilykeilani011讨论对AI艺术的态度，指出问题不在于AI本身，而在于人们对承认自己喜欢AI艺术感到难以接受。

NVIDIA Kumo Tabular Sets a New Accuracy-Efficiency Frontier for Tabular Prediction

Hugging Face Blog · 09/29 23:30

中文介绍 NVIDIA发布Kumo Tabular，为表格预测设定新精度-效率边界。

Getting the Source Right, Not Just the Fact: Source-Aware Verification for MCP Agents

Hugging Face Blog · 09/29 21:07

中文介绍 Hugging Face Blog发布文章，探讨MCP代理的源意识验证。

Making AI an asset, not an expense

MIT Tech Review AI · 09/29 18:43

When customers talk about AI costs, the conversation usually starts with token prices and ends with access to the latest, most capable model in the cloud. Do they always need that level of capability? Not necessarily. But that is often where the conversation goes. As AI moves from experimentation to

中文介绍 MIT Tech Review AI发表文章，讨论将AI视为资产而非成本的观点。

Introducing GPT-6.1 Sol

OpenAI News · 09/29 18:00

Meet GPT-6.1 Sol: near-Astra intelligence for coding, computer use, and professional work at one-fifth of Astra’s standard API input and output token prices.

中文介绍 OpenAI推出GPT-6.1 Sol，近Astra智能，价格仅为Astra标准API的五分之一。

DevDay 2026 Recap

OpenAI News · 09/29 18:00

Explore more than 20 announcements from OpenAI DevDay 2026, including GPT-6 Astra, ChatGPT, Codex, APIs, security, and new tools for builders.

中文介绍 OpenAI DevDay 2026回顾，包括GPT-6 Astra、ChatGPT等20多项新公告。

[AINews] AMD buys World Labs for $8.2B, as Atlas solves sparse reconstruction problem for robotics, design and more

Latent Space · 09/29 10:55

Congrats team!

中文介绍 AMD以82亿美元收购World Labs，Atlas解决稀疏重建问题。

[AINews] Opus 5.5 is good at explainer videos

Latent Space · 09/29 10:44

a rare feature of a capability

中文介绍 Opus 5.5擅长制作解释性视频。

Claude Code’s Next Era — Thariq Shihipar, Anthropic

Latent Space · 09/29 09:48

Shipping Opus/Sonnet 5.5, Mods, Plugins, Projects, Tag while Pacing the Frontier

中文介绍 Claude Code的下一代产品，Thariq Shihipar，Anthropic发布Opus/Sonnet 5.5等新功能。

Introducing dots

OpenAI News · 09/29 08:00

Dots by OpenAI are a proactive assistant that can keep working across complex projects and everyday tasks. Learn how dots help you stay in control while work moves forward.

中文介绍 OpenAI推出Dots，跨复杂项目和日常任务的主动助手。

Claude Sonnet 5.5 🧠, Anthropic IPO leaks 📝, AMD buys World Labs 💰

TLDR AI · 09/29 08:00

中文介绍 Claude Sonnet 5.5、Anthropic IPO泄露、AMD收购World Labs等AI新闻汇总。

Roundtables: The Deadly Failures of The Virtual Border Wall

MIT Tech Review AI · 09/29 06:17

Listen to the session or watch below The US has spent billions building a “virtual wall” of surveillance towers along its southern border over the past 25 years, promising they will help detect and apprehend border crossers and save lives. But a groundbreaking investigation by MIT Technology Review

中文介绍 MIT Tech Review AI探讨虚拟边境墙的致命失败。

Towards safety cases for frontier AI training

OpenAI News · 09/29 03:00

Our early guidelines for safety cases in frontier AI training cover technical safeguards, operational practices, and investigating misalignment incidents

中文介绍 OpenAI发布前沿AI训练安全案例的早期指南。

How we will do better for Australia

OpenAI News · 09/29 03:00

OpenAI apologises for incidents involving Australian government websites and outlines stronger safeguards and support to strengthen Australia’s cyber defences.

中文介绍 OpenAI为澳大利亚道歉，并概述加强澳大利亚网络防御的更强保障和支持。

When can we say AI made a scientific discovery?

MIT Tech Review AI · 09/29 01:03

This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here. Last Wednesday, Anthropic announced that earlier this year it had launched a molecular biology lab, where Claude agents read and conjecture about hard biology pro

中文介绍 MIT Tech Review AI探讨何时可以说AI做出了科学发现。

Holo4: powering generalist computer-use agents

Hugging Face Blog · 09/28 17:44

中文介绍 Hugging Face Blog发布Holo4，为通用计算机使用代理提供动力。