Carlos's Debrief

⚡ Quick Summary

Agents lack help-seeking judgment: HiL-Bench introduces Ask-F1 to measure when AI agents should escalate to humans -- finding all frontier models fail at this core judgment task.
LLMs use finite counting states, not true arithmetic: Across 100+ model variants, stable counting breaks below context limits; fluent benchmarks mask unreliable rule-following.
Faithful uncertainty as hallucination antidote: Rather than expand knowledge or abstain, a metacognition framework teaches LLMs to express calibrated uncertainty -- enabling trustworthy AI.
Attention replaced by dynamic MLP parameters: WeightFormer reframes attention as dynamically-generated MLP weights, achieving Transformer-level modeling at linear complexity.
Tabular models need RL-driven data cleaning: L2C2 uses deep RL to sequence data cleaning operators, closing the dirty-input gap that degrades Tabular Foundation Model accuracy.

📰 News Headlines

Quantum Threat to Blockchain: A Coinbase panel of six cryptographers concludes a quantum computer capable of breaking blockchain encryption is inevitable, with the window to prepare defenses narrowing rapidly.
New Container & Nix CVEs Surface: CVE-2026-31431 (Copy Fail rootless container exploit) and CVE-2026-44028 (Lix integer overflow) highlight active kernel-level and package-manager vulnerability research.
LLM Hallucination Persists, Metacognition Proposed: Research argues that factual gains come from expanded knowledge rather than boundary awareness, proposing 'faithful uncertainty' -- calibrated linguistic doubt as the antidote to confident errors.
AI Agents Universally Fail at Help-Seeking Judgment: HiL-Bench's Ask-F1 metric reveals all frontier models suffer a systemic judgment gap: they cannot reliably decide when to ask a human vs. guessing.
Sim-to-Real Gap: Hybrid FLUX+REGEN Narrows the Gap: Combining diffusion-based geometry transforms with image-to-image translation outperforms either approach alone for photorealizing game engine synthetic datasets.

🧠 LLMs 4

Counting as a minimal probe of language model reliability [13]

Tianxiang Dai, Jonathan Fan

We introduce Stable Counting Capacity, an assay in which models count repeated symbols until failure. The assay removes knowledge dependencies, semantics and ambiguity from evaluation, avoids lexical and tokenization confounds, and provides a direct measure of procedural reliability. Across more…

arXiv:2605.02028LLMs

Hallucinations Undermine Trust; Metacognition is a Way Forward [14]

Gal Yona, Mor Geva, Yossi Matias

Most factuality gains in LLM research have come from expanding the model's knowledge boundary rather than improving awareness of that boundary. We conjecture that distinguishing known from unknown is inherently difficult: models may lack discriminative power to perfectly separate truths…

arXiv:2605.01428LLMs

BlenderRAG: High-Fidelity 3D Object Generation via Retrieval-Augmented Code Synthesis [15]

Massimo Rondelli, Francesco Pivi, Maurizio Gabbrielli

Automatic generation of executable Blender code from natural language remains challenging, with state-of-the-art LLMs producing frequent syntactic errors and geometrically inconsistent objects. We present BlenderRAG, a retrieval-augmented generation system operating on a curated multimodal dataset of 500 expert-validated examples across…

arXiv:2605.00632LLMs

Talker-T2AV: Joint Talking Audio-Video Generation with Autoregressive Diffusion Modeling [17]

Zhen Ye, Xu Tan, Aoxiong Yin, Hongzhan Lin, Guangyan Zhang

Joint audio-video generation models yield stronger cross-modal coherence than cascaded approaches, but existing models couple modalities throughout denoising via pervasive attention, treating high-level semantics and low-level details in a fully entangled manner. We propose Talker-T2AV, an autoregressive diffusion framework where…

arXiv:2604.23586LLMs

🤖 Agents 2

HiL-Bench (Human-in-Loop Benchmark): Do Agents Know When to Ask for Help? [8]

Mohamed Elfeki, Tu Trinh, Kelvin Luu, Guangze Luo, Nathan Hunt

Frontier coding agents solve complex tasks when given complete context but collapse when specifications are incomplete. The bottleneck is not raw capability but judgment: knowing when to act autonomously and when to ask for help. Current benchmarks are blind to…

arXiv:2604.09408Agents

Agentic AI Systems Should Be Designed as Marginal Token Allocators [9]

Siqi Zhu

This position paper argues that agentic AI systems should be designed and evaluated as marginal token allocation economies rather than as text generators priced by the unit. Following a single request -- a developer asking a coding agent to fix…

arXiv:2605.01214Agents

🛡️ AI Safety 1

Prior-Aligned Data Cleaning for Tabular Foundation Models [16]

Laure Berti-Equille

Tabular Foundation Models (TFMs) achieve state-of-the-art zero-shot accuracy on small tabular datasets by meta-learning over synthetic data-generating processes. However, their in-context learning assumes approximately clean inputs: missing values, outliers, and duplicates create a prior mismatch that degrades both accuracy and…

arXiv:2604.25154AI Safety

🔬 Machine Learning 4

Linking spatial biology and clinical histology via Haiku [7]

Yan Cui, Jacob S. Leiby, Wenhui Lei, Dokyoon Kim, Yanxiang Deng

Here we present Haiku, a tri-modal contrastive learning model trained on multiplexed immunofluorescence (mIF). It comprises 26.7 million spatial proteomics patches from 3,218 tissue sections across 1,606 patients spanning 11 organ types, with matched hematoxylin and eosin (H&E) histology and…

arXiv:2605.00925Machine Learning

Assessing Pancreatic Ductal Adenocarcinoma Vascular Invasion: the PDACVI Benchmark [10]

M. Riera-Marin, O. K. Sikha, J. Rodriguez-Comas, M. S. May, T. Kirscher

Surgical resection is the only potentially curative treatment for pancreatic ductal adenocarcinoma (PDAC), and eligibility depends on accurate assessment of vascular invasion (VI). We introduce CURVAS-PDACVI, an open benchmark for uncertainty-aware AI in PDAC staging based on a densely annotated…

arXiv:2604.27582Machine Learning

A Hybrid Approach for Closing the Sim2real Appearance Gap in Game Engine Synthetic Datasets [11]

Stefanos Pasios

Video game engines generate large volumes of visual synthetic datasets for training computer vision algorithms, but a notable sim2real appearance gap limits real-world applicability. This letter investigates using FLUX.2-4B Klein (a diffusion model) and REGEN (an image-to-image translation model) to…

arXiv:2605.02291Machine Learning

Linear-Time Global Visual Modeling without Explicit Attention [12]

Ruize He, Dongchen Han, Gao Huang

We demonstrate that attention can be mathematically reframed as a Multi-Layer Perceptron (MLP) equipped with dynamically predicted parameters. Through this lens, attention's global modeling power is explained not as explicit token-wise aggregation but as an implicit process where dynamically generated…

arXiv:2605.01711Machine Learning

🔗 Lobste.rs 6

CVE-2026-31431: Copy Fail vs. rootless containers [1]

Yesterday I shared here my own report on trying out Copy Fail with rootless containers. Someone else also published an article on this, but going a lot deeper into the implementation details of the exploit and its interaction with the…

Lobste.rs

Why every organization should make it easy to report security flaws [2]

Companies make it too challenging to report security bugs and data leaks. Having a dedicated security email address could save your company from a damaging hack.

Lobste.rs

DigiCert Revokes Certificates After Support Portal Hack [3]

DigiCert revoked certificates that hackers obtained through its internal support portal as part of a social engineering attack.

Lobste.rs

An exploitable integer overflow in Lix (CVE-2026-44028) [4]

Security researchers have found a security issue in Lix. This issue has been assigned CVE-2026-44028. The issues are different between Lix and CppNix but MITRE copied the wrong information into the CVE, which should have gone into the CppNix entry.

Lobste.rs

Trustworthy JavaScript for the Open Web [5]

The open web handles highly sensitive data from private communications to financial transactions and medical records. Traditionally servers are trusted to deliver code, but new approaches aim to make JavaScript itself trustworthy end-to-end.

Lobste.rs

ChatGPT Won't Let You Type Until Cloudflare Reads Your React State. I Decrypted the Program That Does It [6]

A security researcher reverse-engineered the obfuscated JavaScript that Cloudflare uses to read React component state in ChatGPT's input field before allowing typing, revealing the surveillance mechanism built into the AI's client-side code.

Lobste.rs

📚 Google Scholar 6

Early identification of breakthrough technologies: Insights from science-driven innovations [18]

This paper presents a framework for identifying breakthrough technologies using science-driven innovation data and machine learning-based link prediction to detect which innovations have the potential to trigger technological breakthroughs.

Google Scholar

Catalyst breakthroughs in methane dry reforming: Employing machine learning for future advancements [22]

Rising levels of atmospheric carbon dioxide and methane have sparked interest in methane dry reforming as a pathway to reduce greenhouse gases. Machine learning is being employed to discover and optimize catalysts for this reaction, accelerating the development of more…

Google Scholar

Large language models (LLM) in computational social science: prospects, current state, and challenges [20]

This review examines the use of large language models in computational social science, discussing their emerging capabilities for automated text analysis, social simulation, and behavioral modeling, while identifying key challenges around data bias, privacy, and model integration.

Google Scholar

Artificial intelligence and machine learning in cybersecurity: a deep dive into state-of-the-art techniques and future paradigms [21]

This survey provides a comprehensive overview of AI and machine learning applications in cybersecurity, covering adversarial AI, automated threat intelligence, AI-driven security orchestration, and the evolving role of AI in both offensive and defensive security operations.

Google Scholar

Securing the future: exploring post-quantum cryptography for authentication and user privacy in IoT devices [22]

With the emergence of quantum computers, traditional cryptographic methods are vulnerable to attacks, emphasizing the need for post-quantum cryptography to secure IoT devices. This paper surveys various post-quantum cryptographic algorithms and their applicability to resource-constrained IoT environments.

Google Scholar

Quantum machine learning: A comprehensive review of integrating AI with quantum computing for computational advancements [23]

This review covers the integration of AI with quantum computing across the spectrum from quantum-enhanced classical ML to native quantum algorithms, with applications in optimization, drug discovery, and quantum-secured communications.

Google Scholar

💬 Hacker News 2

Coinbase Advisers Warn Quantum Computing Will Crack Blockchain Encryption [24]

A Coinbase panel of six cryptographers concludes that a quantum computer powerful enough to break blockchain encryption will be built, and the window for organizations to transition to post-quantum cryptography is narrowing.

Hacker News

Show HN: Quldra – A true device based post quantum messenger [25]

Quldra is a post-quantum messenger that uses true random numbers from quantum key distribution devices to achieve unconditional security, unlike computational security approaches that rely on mathematical hardness assumptions.

Hacker News