04:00 <Ada> arXiv.org | Defensive Sufficiency in a Stackelberg Model of AI Security Subhabrata Majumdar, Rajlakshmi Chavan · 4h ago · 3 min read Feedback from automated testing, human red teaming, and incident response can strengthen an AI system's defenses when discovered failures lead to effective repairs. We study when this feedback process provides sufficient protection and when investing in it is economically worthwhile. We begin by showing that an attack surface composed of finite number of ... Read full story

04:00 <Ada> arXiv.org | How Do LLMs Change Predictions Under Negation? 4h ago · 3 min read Negation is an essential feature of human language, yet large language models (LLMs) remain unreliable in processing it. We evaluate recent open-source and closed-source LLMs on our negation benchmark and find that, in 37-71% of cases, they repeat the same answer under negation (e.g., "Madrid" for "What is not the capital of Spain?"). To understand and ad... Read full story

04:00 <Ada> arXiv.org | Origins of Universal Machine Learning Force-Field Errors in Multicomponent Materials 4h ago · 3 min read Universal machine learning force-field generalization to multicomponent environments generated by compositional design remains insufficiently assessed. We construct a benchmark of 7,599 multicomponent configurations inspired by high-entropy design, elemental substitution and anion mixing. Eleven pretrained models are evaluated against density functional t... Read full story

04:00 <Ada> arXiv.org | Shared Geometry As A Rosetta Stone: Cross-Modal Alignment Without Paired Data 4h ago · 3 min read Multimodal representations enable zero-shot classification and retrieval, but aligning independently trained models usually requires large amounts of paired data. Yet, the Platonic Representation Hypothesis suggests that models trained on different modalities may converge spontaneously toward a shared representation geometry. But then, do we even need pai... Read full story

04:00 <Ada> arXiv.org | An Empirical Study of Agent Skills' Downstream Utility 4h ago · 3 min read Agent Skills package procedural guidance and resources for reuse, but a relevant Skill does not necessarily improve task performance. Existing studies characterize Skill content and evaluate downstream performance, yet provide limited explanations of how utility depends on content, execution configuration, and multi-Skill organization. We conduct an empir... Read full story

04:00 <Ada> arXiv.org | MaRK: Markov-adapted Recurrent Kernels for Dynamic Operator Conditioning in State Space Models Syed Ibrahim Omer, Ginny Y. Wong, Xiangyu Zhao · 4h ago · 3 min read State Space Models (SSMs) offer an efficient alternative to Transformers for sequence modeling, yet conditioning pre-trained SSMs for iterative generation typically operates outside the recurrent operator, through input injection or activation modulation. While such mechanisms expose the model to conditioning information, they leave the underlying tempora... Read full story

04:00 <Ada> arXiv.org | Have I Seen Enough? Frozen Video-Language Models Encode Evidence Readiness Dan Ben-Ami, Kobi Cohen, Chaim Baskin · 4h ago · 3 min read Streaming video-language models must decide not only what to answer, but whether the evidence needed for the current question has arrived. Existing systems learn that decision as a separate trigger; we ask whether an unmodified model already computes it. We show that frozen VideoLLMs carry a linearly readable evidence-readiness signal, labelled from times... Read full story

04:00 <Ada> arXiv.org | Leveraging LLM-Generated Explanations for Detecting Emotionally Rewritten Fake News 4h ago · 3 min read The spread of fake news may cause severe social consequences. Existing fake news detection methods mainly focus on stylistic variations or incorporate external information such as explanations. However, news articles are often rewritten under different emotional backgrounds while preserving their underlying factual claims, which may affect the robustness ... Read full story

04:00 <Ada> arXiv.org | Self-Retrospection Distillation: Turning Post-hoc Experiences into Prior Foresight 4h ago · 3 min read Reinforcement learning with verifiable rewards (RLVR) turns agent experience into learning signals primarily through scalar outcome rewards after interaction. For group-relative objectives, however, this signal vanishes when all rollouts receive the same reward, even though their trajectories may reveal useful information about what the task requires and ... Read full story

04:00 <Ada> arXiv.org | Just for FUNS: LLM-Guided Spatio-Temporal Graph Node Generation for Forecasting Unobserved Node States 4h ago · 3 min read Spatio-temporal forecasting is a cornerstone of logistics, urban planning, and intelligent transportation systems. However, constrained by deployment costs and maintenance resources, sensor networks often lack comprehensive spatial coverage, rendering Forecast Unobserved Node States (FUNS) a critical yet formidable challenge. Conventional models rely on h... Read full story

04:00 <Ada> arXiv.org | SOL: Measuring Gaps between Text Distributions by Double Sliced Wasserstein Metrics 4h ago · 3 min read Evaluating text generation requires measuring how well the generated distribution matches the data distribution. For autoregressive models, this is done by the perplexity. Diffusion and flow-based language models can only provide a likelihood bound, whose tightness differs between model families. Sample-based substitutes such as generative perplexity with... Read full story

04:00 <Ada> arXiv.org | When Plans Change Answers: Formalizing Cost-Accuracy Optimization for Semantic Queries Kyoungmin Kim · 4h ago · 3 min read In semantic query engines, predicates are evaluated by machine-learned models, and the choice of a query plan affects not only the cost of a query but also its result. Existing systems either apply a fixed threshold to each semantic operator or tune accuracy per operator, without accounting for how errors propagate through joins. We give a formal problem ... Read full story

04:00 <Ada> arXiv.org | The Score Is Not the Structure: Brain Alignment and Cross-Lingual Transfer Saman Rahbar · 4h ago · 3 min read Similarity scores are often offered as evidence that a model shares structure with the brain or across languages. We ask what such a score reads when that structure is removed, or when the instrument measuring it does not work, in two settings. Across seventeen languages, a grammaticality probe transfers worse between more distant languages (r = -0.66). B... Read full story

04:01 <Ada> arXiv.org | The Geometry of Empowerment Catherine Ji, Vivek Myers, Sergey Levine, Benjamin Eysenbach · 4h ago · 2 min read Empowerment captures the capacity for an agent to actively control its environment. While conceptually appealing as an information-theoretic quantity, the connection between empowerment and structurally central states that provide broad access to future outcomes has remained an open question. In this work, we link empowerment maximization and skill-learni... Read full story

04:01 <Ada> arXiv.org | Evaluating and Improving the Robustness of Large Language Models to Input Sequence Variations Narek Maloyan · 4h ago · 3 min read Large language models (LLMs) in production systems face prompt injections, trojans (backdoors), and manipulation of automatic quality metrics. This thesis develops models, methods, and algorithms for evaluating and improving LLM robustness to adversarial input sequence variations. We propose R_stab(f), a generative robustness metric based on the Jensen-Sh... Read full story

04:01 <Ada> arXiv.org | Learning to Decide, Not to Reason: Parameter-Efficient Decision Operators via Low-Rank Activation Steering Ran Li, Lei Chen · 4h ago · 3 min read Injecting skills into a frozen language model currently costs a million parameters and a reinforcement-learning pipeline. We introduce DecSteer, a System-1 decision operator trained by behavior cloning that lowers this cost by roughly two orders of magnitude. The default operator uses 330K parameters to match a 1.33M-parameter operator trained with reinfo... Read full story

04:01 <Ada> arXiv.org | Budgeted Cache Repair for Cross-Context KV-Cache Reuse Haeyong Kang, Chang D. Yoo · 4h ago · 3 min read Cross-context KV-cache reuse predicts a shared segment's keys and values under a new prefix instead of recomputing them, and has been reported to do so without quality loss. We find otherwise, and identify two problems. (1) A hidden cost: on MMLU and GSM8K, reuse costs substantial accuracy. (2) A decision at the wrong unit: no rule for deciding whether to... Read full story

04:01 <Ada> arXiv.org | Population Scaling or Data Dilution? Dynamics of Local Topology Evolution in Decentralized Learning Yin-Kuan Liang, Yan Gao, Yang Long · 4h ago · 3 min read Scaling decentralized learning changes not only the number of clients $N$, but also the dynamics of information propagation and consensus. We argue that the effect of increasing $N$ cannot be understood in isolation, because data allocation, topology-dependent mixing, and communication capacity may change simultaneously. We study these coupled effects on ... Read full story

04:01 <Ada> arXiv.org | LLM Persona Unlearning 4h ago · 3 min read Pre-training equips large language models (LLMs) with a broad repertoire of behavioral patterns associated with roles, styles, values, and goals. Post-training teaches conditional enactment and makes a helpful Assistant the default, but it does not erase alternative modes from the weights; explicit prompts can therefore elicit personas that repeatedly sha... Read full story

04:01 <Ada> arXiv.org | A Parameter-Free Zeroth-Order Method with Covariance Matrix Adaptation and Effective Dimension Alexander Sholokhov, Alexander Rogozin · 8d ago · 3 min read Zeroth-order optimization methods are essential for solving black-box problems where gradient information is unavailable or expensive to compute. This paper presents POEM-CMA, a novel parameter-free stochastic zeroth-order algorithm that extends the recent POEM method by integrating covariance matrix alignment and the notion of effective dimension. In con... Read full story

09:38 <Vega> OpenAI News | Sophos cuts threat investigation time by 96% with OpenAI Daybreak 2h ago Discover how Sophos uses OpenAI’s Daybreak to cut cyber-threat investigation time by 96% and automate 52% of MDR cases while preserving human oversight. Read full story

15:50 <Ada> Hugging Face | Impactful scheduling for GPU clusters Kyle Wiggers · 30m ago · 16 min read A Blog post by Ai2 on Hugging Face Read full story

20:16 <Vega> OpenAI News | Enabling a data-driven workforce 08 Aug 2024 In this video, we share practical examples of how employees can use ChatGPT Enterprise to efficiently analyze data and uncover insights. Read full story

20:16 <Vega> OpenAI News | Fine-tuning GPT-4o webinar 26 Aug 2024 Fine-Tuning GPT-4o Webinar Read full story

20:16 <Vega> OpenAI News | Put AI to work: Lessons from hundreds of successful deployments 10 Sep 2024 Put AI to Work: Lessons from Hundreds of Successful Deployments Read full story

20:16 <Vega> OpenAI News | Put AI to work: Automate and scale financial operations 30 Sep 2024 Put AI to work: Automate and Scale Financial Operations Read full story

20:16 <Vega> OpenAI News | Solving complex problems with OpenAI o1 models 17 Oct 2024 In this video, we share how the o1 reasoning models can help in domains like coding, strategy, and research. Read full story

20:16 <Vega> OpenAI News | Put AI to work for marketing teams 31 Oct 2024 Put AI to Work for Marketing Teams Read full story

20:17 <Vega> OpenAI News | Shipping code to your IDE with ChatGPT 08 Apr 2025 See how ChatGPT can help move generated code into your IDE to speed up development workflows. Read full story

20:17 <Vega> OpenAI News | Strategic planning with ChatGPT 08 Apr 2025 See how ChatGPT can support strategic planning by analyzing context, generating options, and turning decisions into next steps. Read full story

20:17 <Vega> OpenAI News | Vision and voice in ChatGPT 08 Apr 2025 See how vision and voice in ChatGPT help teams work with multimodal inputs and communicate more naturally. Read full story

20:17 <Vega> OpenAI News | Writing with canvas in ChatGPT 08 Apr 2025 See how canvas in ChatGPT helps turn notes, transcripts, and rough drafts into polished writing. Read full story

20:17 <Vega> OpenAI News | Creating a contextual workspace in ChatGPT 13 Apr 2025 Discover how ChatGPT Projects helps organize files, save context, and keep conversations consistent across workstreams. Read full story

20:17 <Vega> OpenAI News | Market research with ChatGPT 13 Apr 2025 See how ChatGPT can plan, search, read sources, and compile a detailed research report to support market analysis. Read full story

20:18 <Vega> OpenAI News | Prototyping with canvas in ChatGPT 13 Apr 2025 See how to go from a product requirements document to a working HTML/React prototype using canvas in ChatGPT. Read full story

20:18 <Vega> OpenAI News | Connect internal knowledge from Google Drive 08 May 2025 See how ChatGPT can connect to Google Drive to use internal knowledge and answer questions with relevant context. Read full story

20:18 <Vega> OpenAI News | Automate complex workflows with OpenAI o3 08 May 2025 See how OpenAI o3 can reason through complex tasks and help automate multi-step workflows. Read full story

20:18 <Vega> OpenAI News | Create on-brand visuals with image generation 08 May 2025 See how ChatGPT image generation helps create branded visuals and product mockups from prompts, sketches, or uploaded images. Read full story

20:18 <Vega> OpenAI News | Collaborate and write with GPT-4o 08 May 2025 Explore how GPT-4o makes writing and collaboration more natural, helping teams draft, refine, and adjust tone faster. Read full story

20:18 <Vega> OpenAI News | Personalize ChatGPT with custom instructions 12 May 2025 See how custom instructions help personalize ChatGPT for your role, preferences, and work. Read full story

20:18 <Vega> OpenAI News | Delegate work to ChatGPT agent 06 Aug 2025 Watch a ChatGPT agent demo showing how it can research, code, and take action online to complete a business plan in minutes. Read full story

20:18 <Vega> OpenAI News | GPT-5: Our best model for work 07 Aug 2025 Watch a GPT-5 demo for work showing how the model can analyze feedback, create a product plan and prototype, and summarize next steps. Read full story

20:19 <Vega> OpenAI News | How OpenAI uses ChatGPT Atlas at work 18 Nov 2025 Watch a webinar on how OpenAI uses ChatGPT Atlas at work. Read full story

20:19 <Vega> OpenAI News | Using Codex across the software development lifecycle 20 Nov 2025 Watch a webinar on using Codex across the software development lifecycle. Read full story

20:19 <Vega> OpenAI News | How OpenAI Uses Codex to Accelerate Engineering 14 Jan 2026 A detailed look at how OpenAI teams use Codex to understand code, refactor systems, improve performance, boost velocity, and enhance engineering workflows. Read full story

20:19 <Vega> OpenAI News | Inside GPT-5 for Work 14 Jan 2026 A concise guide to how GPT-5 powers real business workflows—across marketing, engineering, finance, strategy, legal, and IT—and how to evaluate it for your team. Read full story

20:19 <Vega> OpenAI News | Building with the GPT-5 Model Series: A Practical Guide 15 Jan 2026 A practical guide for startups on migrating to the Responses API, designing agents, tuning prompts, steering GPT-5 models, and building reliable AI workflows. Read full story

20:19 <Vega> OpenAI News | A Practical Guide to Building AI Agents 15 Jan 2026 A comprehensive guide to designing, orchestrating, and deploying AI agents—covering use cases, model selection, tool design, guardrails, and multi-agent patterns. Read full story

20:19 <Vega> OpenAI News | Think Bigger: How Small Teams Win with ChatGPT 22 Jan 2026 A practical guide showing how small teams use ChatGPT to analyze data, create marketing materials, manage budgets, run meetings, and make better decisions. Read full story

20:19 <Vega> OpenAI News | Asana cuts model costs 76x in browser tests with GPT-6.1 Sol 13h ago Using GPT-6 Astra in Codex, Asana made its browser agent 76x cheaper and 5x faster in tests to offer customers more capable models. Read full story

21:35 <Ada> [Ada] Waterdownmayo is here. Mind the preprints.

#ai on 2026-10-09