Sources

What we learned mapping a year’s worth of AI-enabled cyber threats anthropic.com

The MiniMax-M2 Series: Mini Activations Unleashing Max Real-World Intelligence huggingface.co

The MiniMax-M2 series introduces Mixture-of-Experts language models with minimal activated parameters that achieve high performance in agentic tasks through specialized training and deployment systems.

Gemini Embedding 2: A Native Multimodal Embedding Model from Gemini huggingface.co

Gemini Embedding 2 is a multimodal embedding model that generates unified representations for video, audio, image, and text data, achieving superior performance across diverse retrieval tasks and demonstrating strong zero-shot capabilities across specialized domains.

MobileMoE: Scaling On-Device Mixture of Experts huggingface.co

MobileMoE adapts sparse mixture-of-experts to phones by combining fine-grained and shared experts with quantization-aware training. The sub-billion-parameter models beat dense baselines and prior MoE designs on both quality and efficiency, fitting INT4 weights into mobile memory budgets across prefill and decode.

Understanding Data Temporality Impact on Large Language Models Pre-training huggingface.co

Kyutai’s Kairos study shows that training LLMs on Common Crawl snapshots in chronological order, instead of shuffling, improves temporal precision and factual freshness without hurting general language understanding. The approach reframes pre-training as a continual-learning curriculum over time-stamped web data.

Can LLMs Introspect? A Reality Check huggingface.co

Probing whether language models truly monitor their internal states, the authors find apparent self-reports track surface cues rather than hidden activations. Behavioral and representational tests suggest introspective answers reflect learned conversational patterns, undercutting claims that LLMs genuinely access their own computations.

LLaVA-OneVision-2: Towards Next-Generation Perceptual Intelligence huggingface.co

The second-generation vision-language model uses codec-stream tokenization, windowed attention and 3D RoPE to handle long video, temporal localization and tracking from one backbone. Authors introduce a JumpScore metric and report state-of-the-art results across fine-grained spatial and temporal benchmarks.

LocateAnything: Fast and High-Quality Vision-Language Grounding with Parallel Box Decoding huggingface.co

Instead of generating box coordinates token by token, NVIDIA’s grounding model emits each box as an atomic geometric unit in parallel. The design unifies detection and visual grounding, raising throughput and localization accuracy while preserving geometric coherence across large-scale training data.

MRT: Masked Region Transformer for Layered Image Generation and Editing at Scale huggingface.co

The Masked Region Transformer trains a 20B-parameter diffusion model that produces and edits transparent image layers in one pass. An overflow-aware canvas and selective token masking unify image-to-layers, text-to-layers and layers-to-layers tasks, with distillation cutting inference cost.

FastKernels: Benchmarking GPU Kernel Generation in Production huggingface.co

Existing GPU kernel benchmarks miss how generated code behaves inside real inference engines. FastKernels evaluates LLM kernel agents against representative architectures using HuggingFace Transformers, vLLM and SGLang, exposing correctness drops and interface mismatches that synthetic harnesses hide.

Negligible in Size, Significant in Effect: On Scale Vectors in Large Language Models huggingface.co

Scale vectors in LLMs significantly impact optimization despite minimal parameter count, with theoretical analysis and practical improvements showing enhanced training performance and scaling behavior.

MobileGym: A Verifiable and Highly Parallel Simulation Platform for Mobile GUI Agent Research huggingface.co

MobileGym presents a browser-based mobile environment enabling deterministic evaluation and scalable reinforcement learning through JSON-based state management and parallel execution.

MUSE-Autoskill: Self-Evolving Agents via Skill Creation, Memory, Management, and Evaluation huggingface.co

A skill-centric agent framework enables continuous improvement of task-solving capabilities through a unified lifecycle of skill creation, memory, management, evaluation, and refinement.

NSF-SciFy: Mining the NSF Awards Database for Scientific Claims huggingface.co

NSF-SciFy is a large-scale dataset of scientific claims and investigation proposals extracted from NSF award abstracts, enabling improved language model fine-tuning for claim verification and scientific discovery tracking.

RT-Lynx: Putting the GEMM Sparsity In a Right Way for Diffusion Models huggingface.co

Diffusion Transformers achieve strong image generation performance but face high inference costs; this work proposes RT-Lynx, which uses activation sparsification and optimized CUDA kernels to accelerate inference while maintaining generation quality.

QUACK: Questioning, Understanding, and Auditing Communicated Knowledge in Multimodal Social Deduction Agents huggingface.co

A multimodal social reasoning environment and evaluation framework called QUACK is introduced to audit the grounding of agent language through three-level assessment of game outcomes, behavioral trajectories, and utterance-level consistency.

ZeroUnlearn: Few-Shot Knowledge Unlearning in Large Language Models huggingface.co

ZeroUnlearn addresses privacy concerns in large language models by reformulating machine unlearning as precise knowledge re-mapping through model editing, enabling efficient and targeted removal of sensitive information while preserving general model utility.

JLT: Clean-Latent Prediction in Latent Diffusion Transformers huggingface.co

Latent diffusion models using clean-data prediction outperform velocity prediction in compressed representations, demonstrating that prediction targets are geometrically dependent rather than algebraically interchangeable.

Learning to Act under Noise: Enhancing Agent Robustness via Noisy Environments huggingface.co

NoisyAgent is an agentic training framework that incorporates environmental imperfections into agent learning to improve robustness in real-world stochastic settings.

SAM: State-Adaptive Memory for Long-Horizon Reasoning Agent huggingface.co

Long-horizon agentic reasoning is enhanced through a state-adaptive memory framework that dynamically manages interaction histories by creating compact memory cues while preserving detailed trajectories for targeted retrieval.

Balancing Fidelity and Diversity in Diffusion Models via Symmetric Attention Decomposition: Hopfield Perspective huggingface.co

The symmetric and skew-symmetric components of transformer attention matrices are analyzed as governing energy landscape structure and circulation dynamics, respectively, with implications for generation trade-offs.

Beyond Final Answers: Auditing Trajectory-Level Hallucinations in Multi-Agent Industrial Workflows huggingface.co

Trajel presents a trajectory-level hallucination audit framework with a five-type taxonomy for multi-step LLM agent workflows, demonstrating that current detection methods miss nuanced failures and require trajectory-aware approaches for safe deployment.

VitaBench 2.0: Evaluating Personalized and Proactive Agents in Long-Term User Interactions huggingface.co

VitaBench 2.0 evaluates personalized and proactive agent behavior in long-term user interactions by requiring continuous extraction and updating of user preferences from fragmented interactions.

Efficient Agentic Reinforcement Learning with On-Policy Intrinsic Knowledge Boundary Enhancement huggingface.co

AKBE enhances LLM agent training by dynamically identifying when tools are needed versus when internal knowledge suffices, improving accuracy and reducing unnecessary tool usage through targeted supervisory signals.

Learning High-Frequency Continuous Action Chunks in Latent Space huggingface.co

High-frequency robotic control is improved by using variational autoencoders to enhance temporal and spatial consistency, combined with a reuse-then-refine strategy for smooth real-time execution.

Does Seeing More Mean Knowing More? Mono-Anchored Advantage Normalization for Multi-Source Visual Reasoning huggingface.co

A novel mono-anchored multi-source reasoning framework that uses dynamic anchors to quantify information gain and regulate modality interactions during reinforcement learning with verifiable rewards.

Rethinking VLM Representation for VLA Initialization huggingface.co

Effective vision-language-action model initialization requires balancing pretrained vision-language model representations with embodied task-specific adaptations and robot-data pretraining while preserving core action-relevant features.

Geometry-Aware Representation Denoising for Robust Multi-view 3D Reconstruction huggingface.co

A novel diffusion-based framework for multi-view 3D reconstruction that restores both scene geometry and high-quality imagery from degraded inputs by operating in the feature space of a 3D reconstructor.

SpatialBench: Is Your Spatial Foundation Model an All-Round Player? huggingface.co

SpatialBench presents a comprehensive benchmark for evaluating spatial foundation models across diverse domains and tasks, revealing limitations in current models and introducing DA-Next-5M and DA-Next to advance spatial representation learning.

Confidence and Calibration of Activation Oracles for Reliable Interpretation of Language Model Internals huggingface.co

Research evaluates confidence estimation methods for activation oracles, finding bootstrap mode frequency provides better-calibrated confidence scores than log-probability approaches.

Share More, Search Less: Collaborative Parallel Thinking for Efficient Test-Time Scaling huggingface.co

Collaborative Parallel Thinking (CPT) enables information sharing across parallel search branches during inference to reduce redundant exploration and improve efficiency in test-time scaling for language models.

DarkForest: Less Talk, Higher Accuracy for Multi-Agent LLMs huggingface.co

DarkForest is a controlled-communication framework that enhances multi-agent LLM reasoning by clustering semantic candidates and using calibrated belief distributions to reduce error propagation and communication overhead.

Agentic CLEAR: Automating Multi-Level Evaluation of LLM Agents huggingface.co

Agentic CLEAR is an automatic evaluation framework that provides multi-level textual insights into agent behavior through dynamic analysis of LLM interactions across various benchmarks and settings.

CroCo: Cross-Lingual Contrastive Preference Tuning on Self-Generations huggingface.co

Cross-lingual contrastive preference tuning enables multilingual language model improvement without language-specific annotations, achieving strong performance across diverse tasks and languages.

D^2-Monitor: Dynamic Safety Monitoring for Diffusion LLMs via Hesitation-Aware Routing huggingface.co

Diffusion large language models generate text through multi-step denoising processes that expose intermediate representations useful for safety monitoring, leading to the development of a bi-level safety monitor that dynamically routes computational resources based on hesitation detection.

Squeezing Capacity from Multimodal Large Language Models for Subject-driven Generation huggingface.co

A novel approach conditions diffusion models on multimodal large language models for subject-driven image generation, combining text and reference image encoding with VAE-based identity conditioning to improve both semantic understanding and identity preservation.

EvalVerse: Pipeline-Aware and Expert-Calibrated Benchmarking for Professional Cinematic Video Generation huggingface.co

EvalVerse presents a comprehensive evaluation framework for generative video models that bridges the gap between human aesthetic judgment and machine scoring through expert-calibrated vision-language models and multi-stage cinematic assessment.

LongAV-Compass: Towards Unified Evaluation of Minute-Scale Audio-Visual Generation Across T2AV, I2AV, and V2AV huggingface.co

LongAV-Compass is a comprehensive benchmark for evaluating minute-long audio-visual generation across multiple modalities, assessing quality, consistency, and alignment over extended temporal sequences.

Soap2Soap: Long Cinematic Video Remaking via Multi-Agent Collaboration huggingface.co

A multi-agent framework called Soap2Soap is presented for long-horizon video-to-video generation that maintains narrative structure and character identity across extended sequences through consistent semantic backbone and visual reference anchors.

EverAnimate: Minute-Scale Human Animation via Latent Flow Restoration huggingface.co

EverAnimate addresses long-horizon animated video generation challenges through persistent latent propagation and restorative flow matching to maintain visual quality and character identity.

STREAM: A Data-Centric Framework for Mining High-Value Task-Oriented Dialogues from Streaming Media huggingface.co

A data-centric framework called Stream generates large-scale, multi-domain service dialogues by synthesizing interactions from streaming media, incorporating persona construction and conversational blueprints with retrieval-augmented generation for knowledge-aware responses.

References

Thomas Wiegold blog review of M2.7 thomas-wiegold.com

absurdly chatty, often generating four times more tokens than the industry average for reasoning tasks

StartupFortune — ‘SWE-bench has been benchmaxxed’ startupfortune.com

changing the orchestration layer can cause up to a 22% swing in performance, making it difficult to isolate the model’s raw capability from its vendor-optimized environment

Anthropic — Detecting and Preventing Distillation Attacks anthropic.com

approximately 24,000 fraudulent accounts and over 16 million unique exchanges with Claude… MiniMax was identified as the primary actor, allegedly responsible for 13 million of the total queries

OpenSourceForU — MiniMax M3 / licensing analysis opensourceforu.com

Modified-MIT license that explicitly prohibits commercial use without prior written authorization… developers have labeled this ‘faux open source’

MiniMax Forge technical blog (Hugging Face) huggingface.co

Prefix-Tree Merging… reportedly achieves an approximately 40x speedup in training throughput compared to standard sequential sample processing

Kilo Blog — head-to-head M2.7 vs Claude Opus blog.kilo.ai

MiniMax achieved roughly 90% of Claude’s coding quality on specialized TypeScript benchmarks while being nearly 20 times cheaper

Thoughtworks (Medium) — Anthropic’s AI Espionage Disclosure: Separating Signal from Noise thoughtworks.medium.com

Critics characterized the disclosure as ‘marketing spin’ designed to highlight Anthropic’s safety monitoring rather than a revolutionary shift in threat landscapes; the AI was prone to significant errors, hallucinating stolen credentials and forcing human operators to intervene at four to six critical decision points.

Google Cloud Threat Intelligence — Advances in Threat Actor Usage of AI Tools cloud.google.com

GTIG identified the first generation of malware that utilizes LLMs during execution — families such as PROMPTFLUX (a dropper) and PROMPTSTEAL (used by APT28) query Gemini or Qwen APIs at runtime for dynamic script generation and self-obfuscation.

Help Net Security — Malware using LLMs (citing Marcus Hutchins) helpnetsecurity.com

Many early samples like PROMPTFLUX are ‘slop malware’ that lack necessary guardrails or entropy, often failing to execute because they rely on the faulty assumption that LLMs instinctively know how to evade antivirus software.

Zenity — MITRE ATLAS adds Agentic AI platform zenity.io

MITRE ATLAS added a dedicated ‘Agentic AI’ platform filter with techniques like AI Agent Context Poisoning (AML.T0080) and Modify AI Agent Configuration (AML.T0081), shifting focus from attacking the base model to exploiting the agent’s tool-invocation and memory layers.

Penligent — PentestGPT Alternatives 2026 Edition penligent.ai

While agents can exploit up to 87% of known one-day CVEs when provided with descriptions, their success rate against undocumented zero-days or hardened HackTheBox challenges remains near 0% — a persistent ‘lab-to-real gap’.

Startup Fortune — Anthropic account bans send developers scrambling startupfortune.com

Developers reported organization-wide bans without warning, with entire teams losing access to Claude Team and API accounts due to unidentifiable ‘suspicious signals’ — some reports suggest up to 1.5 million accounts were terminated in the same period.

MindStudio comparison (Gemini Embedding 2 vs Qwen3-VL) mindstudio.ai

Qwen3-VL-2B achieves a significantly smaller modality gap (measured at approximately 0.25) compared to Gemini Embedding 2 (0.73). A lower gap suggests that Qwen’s text and image vectors are semantically closer.

zc277584121.github.io embedding-models-benchmark-2026 zc277584121.github.io

Gemini Embedding 2 ranked last in compression quality, showing a Spearman rho of 0.668 when truncated to 256 dimensions — higher degradation than competitors like Voyage or Jina.

Hugging Face — Qwen3-VL-Embedding-8B model card huggingface.co

Qwen3-VL-Embedding-8B currently ranks first on the MMEB-V2 leaderboard with an overall score of 77.8, particularly excelling in image-text and video-text matching.

LlamaIndex GitHub issue #21535 github.com

For Gemini Embedding 2, the legacy task_type parameter (e.g., RETRIEVAL_QUERY) has been deprecated and is now ignored by the backend… led to ‘silent accuracy loss’ where the model falls back to a default, unoptimized embedding state.

tokencost.app pricing breakdown tokencost.app

Text: $0.20 per 1M tokens (a 33% increase from the legacy 001 model); Images: $0.45; Audio: $6.50; Video: $12.00 per 1M tokens.

MachineBrief — MRL Overhyped or Underestimated machinebrief.com

Standard, non-MRL text embeddings are surprisingly robust to truncation, often holding their ground against MRL-trained counterparts unless dimensions are reduced by more than 70%… simple post-hoc PCA can make standard embeddings just as compressible.

Sources

References

Jack Sun, writing.