GLM-5.3 writes Chrome exploits, AIDE² beats humans, Agensh runs 1,024 agents
Every URL the pipeline pulled into ranking for this issue — primary sources plus the supporting and contradicting findings each Researcher returned. Inline citations in the issue point back here.
Sources
GLM-5.3 and the spread of advanced cyber capabilities anthropic.com
Quoting Anthropic Frontier Red Team simonwillison.net
We evaluate several models on 100 tasks from the [internal Binary Exploitation benchmark] (selected at random), and find that GLM-5.3 develops full control flow hijacks in 4% of the trials; Claude Mythos Preview did so in 6%. Although GLM-5.3 performs below Claude Mythos Preview here, a meaningful threshold has clearly been crossed: earlier models, like Claude Opus 4.6 and GLM-5.2, do not succeed in any of them. — Anthropic Frontier Red Team , GLM-5.3 and the spread of advanced cyber capabiliti…
Recursive self-improvement of AI research agents huggingface.co
AI agents are beginning to automate research and development across the AI stack, from improving training efficiency to optimizing inference. A natural next step is to improve the research efficiency of the agents themselves. When an AI research agent’s own code is the object of optimization, each accepted rewrite becomes the agent that the next round edits. We refer to this loop as recursive self-improvement. Its significance lies in a long-standing trend, in which increased cumulative spending
Agensh: Scaling Organizational Intelligence to 1,024 Agents huggingface.co
A multi-agent system can reduce latency on complex tasks by executing work concurrently. Several pioneering harness frameworks support multi-agent systems. However, the scalability of current multi-agent harnesses is often constrained by a central orchestrator’s capacity to allocate tasks and coordinate workers. To address this limitation, we introduce Agensh, a scalable self-organized multi-agent harness without a central orchestrator: concurrent workers execute a multi-agent cooperation loop,
Emergent Collusion in Long-Horizon LLM Agent Interaction huggingface.co
Paired agents that repeatedly complete tasks, share logs, and verify each other’s work drift away from the verification protocol when compliance conflicts with reward. The study shows collusion emerges naturally in multi-agent setups, with deviation growing over extended interaction horizons.
Flash-dLLM: IO-Aware KV Caching and Parallel Decoding for Fast, Memory-Efficient Diffusion LLMs huggingface.co
Diffusion LLMs have lagged autoregressive rivals on inference speed because KV caching and parallel decoding fight for memory bandwidth. Flash-dLLM co-designs both around IO bottlenecks, closing the deployment gap that has kept non-autoregressive text generation largely academic.
Lean Pool: An AI-Maintained Archive of Formalized Mathematics huggingface.co
The repository holds formalized mathematics in Lean, but growth, maintenance, and optimization run through AI agents rather than human contributors. It offers a testbed for autonomous theorem-proving pipelines and a living corpus for training math-reasoning models.
The Tasteful Agent: Measuring and Improving Taste in Long-Horizon Tasks huggingface.co
Long-horizon agents live or die by intermediate choices — which hypothesis to test, which implementation to extend — yet benchmarks only grade end-to-end success. TasteBench isolates these decision points to measure and improve what the authors call an agent’s taste.
JEV-as-a-Judge: Accept When Confident, Escalate When Unsure huggingface.co
A decision-only judge acts as a cheap first pass and escalates only uncertain cases to a stronger model. Against 16 generative and reward-model judges with blinded human adjudication, JEV lands within 3 points of the best LLM judge.
RoboFollow: Unveiling the Instruction Following Mirage in Embodied Agents huggingface.co
Embodied agents post strong success rates largely because scenes admit only one valid task, letting policies ignore language entirely. RoboFollow enforces high scene entropy so each environment supports multiple distinct actions, forcing agents to actually parse the instruction.
HARMONY: Hierarchical Agentic Reasoning for MONocular Image-to-Scene Synthesis huggingface.co
Reconstructing a full 3D scene from one image has split between agentic reasoners that grasp spatial relations but misalign pixels, and geometry foundation models that predict dense point maps at low fidelity. HARMONY stacks them hierarchically to get both accurate layout and sharp reconstruction.
Ovis-Embedding: Pushing the Frontiers of Universal Omni-Modal Embeddings huggingface.co
In this report, we introduce Ovis-Embedding, a state-of-the-art omni-modal embedding family built on native integration of text, image, video, and audio. Instead of assembling separate modality towers, Ovis-Embedding uses a shared multimodal backbone to encode different modalities in a common representation space. Specifically, we make three key advances: (1) native omni-modal initialization: we adopt a pretrained Qwen-omni model as the embedding backbone and adapt it through contrastive trainin
Circuit Hypernetworks for Quantum-Augmented Diffusion Language Models huggingface.co
Language models can be adapted by changing the computations applied to individual tokens. Quantum circuits offer one such approach, but evaluating wider circuits inside a large model can be computationally demanding. Here we introduce HyperQ, which adds token-conditioned quantum residual branches to a frozen masked-diffusion language model. A quantum residual branch is a module in each transformer block that reads a token’s hidden state, emits the coordinates of that token’s circuit, executes it
GAE: Learning a Geometry-Native Latent Space for 3D-Consistent World Generation huggingface.co
We present a compact geometry-native latent space as a shared foundation for perception and generation. Visual generators can produce photorealistic frames without preserving a consistent 3D scene. We argue that this is not only a modeling problem but also a representation problem: generators typically evolve appearance-centric latents, while perception models recover geometry in a semantically rich space that encodes cross-view structure. Rather than adding geometry as another output, we repara
From Pattern Recognizers to Personalized Companions: A Survey of Large Language Models in Mental Health huggingface.co
The rising global prevalence of mental health conditions, together with longstanding barriers in traditional healthcare, such as limited resources, high cost, stigma, and privacy concerns, has created an urgent need for accessible and scalable support. Large Language Models (LLMs) have emerged as a transformative technology with strong potential to democratize mental health support through advanced natural language understanding and generation. However, the rapidly expanding, fragmented body of
StableVQ: Practical Guidelines for Stable Vector-Quantized Tokenizer Training huggingface.co
Vector Quantization (VQ) is fundamental to discrete visual tokenizers that power modern autoregressive and masked image generation models. While recent shared-projection codebook methods have substantially advanced codebook utilization, training stability remains a critical and underexplored challenge. We argue that the root cause lies in the entanglement of the Encoder—Decoder and Codebook training: because neither module can reliably fulfill its own responsibility in isolation, the system can
Can one language model hand its live memory to another without the receiver rereading the context? We demonstrate useful persistent hybrid-state transfer across one architecture-matched Qwen3.5 4B-to-9B sibling pair. To our knowledge, this is the first demonstrated cross-model handoff of persistent recurrent inference state between differently sized hybrid language models without target prefix replay. Translated attention KV alone leaves a large gap; adding the Gated DeltaNet (GDN) persistent-st
Bellman Policy Optimization huggingface.co
Reinforcement learning with verifiable rewards (RLVR) improves the reasoning capabilities of large language models (LLMs). We introduce Bellman Policy Optimization (BPO), a critic-free method derived from Policy Mirror Descent (PMD). For autoregressive generation with terminal rewards, BPO uses the Bellman equations to reformulate PMD as a trajectory-level objective. The reformulation avoids estimating state values at intermediate states. We prove that it has the same unique optimal solution as
Geometric and Semantic Coupling for Interaction Understanding in 3D Scenes huggingface.co
Interaction understanding in 3D scenes requires a joint description of movable parts, their motion, and the regions through which they can be operated. We present Segment-Snap, which connects these outputs through the physical relationship between parts and handles. Learned predictors identify broad part surfaces and small handles. A geometric decoder uses planar and upright priors to constrain motion, then selects hinge lines using predicted handle locations, without training a motion regressor
ImIR: Image-Instruction Tuning for All-in-One Image Restoration huggingface.co
Degradations vary widely across images, so a practical restoration system has to handle many degradation types with one model. A recent and effective recipe adapts a large pretrained image-editing model to restoration using a small low-rank adapter with a text prompt. We replace that prompt with an instruction derived from the degraded image itself. The image reaches the editor through two paths: its structure comes from the model’s VAE, and its semantic instruction comes from a lightweight toke
RULER: Instance-aware Rubric Rewards for SVG Generation huggingface.co
Generating Scalable Vector Graphics (SVG) code from natural-language instructions is an open-ended task without absolute visual ground truth, leaving both evaluation and policy optimization without a faithful signal. Scalar metrics (CLIP, Aesthetic) calibrated on natural images transfer poorly to stylized vector content, and reusing them as RL rewards triggers reward hacking. We address both limitations with rubric-based scoring. We first establish empirically that prompting a vision-language ju
All-in-One Multilingual Scene Text Recognition with Script-aware Mixture-of-Experts huggingface.co
Multilingual scene text recognition (STR) remains challenging due to the scarcity of training data for most languages and the difficulty of serving diverse scripts within a single model. Existing solutions either deploy one recognizer per language, inflating cost and introducing error accumulation, or rely on massive vision-language models (VLMs) that are expensive and still inaccurate on many scripts. In this work, we pursue an all-in-one multilingual recognizer that is simpler than per-languag
Tri-PvP: Exposing Modality Bias in Omni-Modal Large Language Models through Perceptual-Propositional Evidence Conflicts huggingface.co
Tri-PvP benchmark reveals visual bias and asymmetric evidence-form preferences in omni-modal language models, with early-layer decodable modality effects resistant to surface mitigation.
Embedding Physics Priors in Robot Learning: A Survey huggingface.co
The rapid progress of artificial intelligence is reshaping robotics and accelerating the adoption of learning-based approaches. While purely data-driven methods have achieved remarkable success in computer vision and natural language processing, robotics remains constrained by limited data, complex real-world interactions, and the need for reliable operation. These challenges have motivated the exploration of physics-embedded robot learning, which embeds physics priors into learning algorithms.
ALPINE: Adaptive Localization for Parameter- and Sample-Efficient Few-Shot Learning huggingface.co
Few-shot learning research is predominantly evaluated on accuracy alone, with limited attention to the parameter and training-sample budgets required to reach that accuracy - a real constraint for practitioners without large-scale compute. We present an ultra-lightweight (22,249-34,917 parameter) spatial-relational architecture for few-shot image classification that combines fixed Gabor edge-energy guidance with a windowed, content-adaptive patch locator. Under a strictly matched, iso-episode-bu
Blaming Across the Aisle: Political Contrasting and Blame Attribution in the Danish Parliament huggingface.co
Political discourse is widely perceived to be growing more hostile, yet robust evidence remains scarce. This study examines blame attribution in the Danish Parliament from 1997 to 2026, combining a purpose-built classifier, BlameBERT (F1: 0.80), with multilevel statistical modeling. The classifier is constructed using an annotation-efficient pipeline for blame attribution in low-to-mid resource languages. The results reveal a banana-shaped trajectory, with blame declining until around 2016 befor
References
NIST CAISI assessment nist.gov
GLM-5.3 is the most cyber-capable open-weight model released to date… lagging behind the U.S. frontier by approximately four months.
ExplainX (MindFort eval summary) explainx.ai
In MindFort’s live NexBench environment… GLM-5.3 ranked only 12th out of 17 models tested, scoring 33 points… its decision-making and reasoning often falter when tasked with the complex, multi-step execution required in dynamic production environments.
AI Weekly (Hacker News summary) aiweekly.co
Commentators expressed skepticism regarding Anthropic’s motives, suggesting the report serves a ‘regulatory capture’ agenda… characterized Anthropic’s alarm as the ‘best Z.ai ad ever’.
Zhipu/Z.ai GLM-5.3 model card (Hugging Face) huggingface.co
cyber capability developed faster than expected… [Zhipu] delayed the release of the model’s weights by two weeks following its August 14 launch to perform additional safety evaluation and hardening.
Shattered.io analysis of GLM-5.3 license shattered.io
a restrictive licensing clause requiring any firm with over $10 billion in annual revenue to pass a Z.ai security review before commercial use.
Weco AI blog (authors’ framing) weco.ai
AIDE² is the first experimental evidence of ‘Level 1’ recursive self-improvement — a system that materially lifts R&D efficiency beyond what human engineers achieved on a fixed budget.
Sakana AI — Darwin-Gödel Machine sakana.ai
The DGM maintains an evolving archive of agent variants… replaces formal proof with ‘Darwinian’ empirical testing, increasing SWE-bench performance from 20% to 50%.
ExplainX analysis of the RSI Ladder explainx.ai
While the self-improved AIDE47 reached performance ceilings roughly 2× faster (20 steps vs. 40), it did not achieve significantly higher absolute performance — the ‘ignition’ threshold for Level 2 runaway acceleration has not been cleared.
Hacker News discussion news.ycombinator.com
The agent is modifying its scaffold — prompts and search logic — rather than the underlying neural weights… closer to automated dogfooding or hyperparameter optimization than the theoretical singularity.
bobdc.com (Chollet-aligned critique) bobdc.com
Intelligence is fundamentally situational; as an agent specializes to improve itself, it may lose the very flexibility required to innovate in new domains — recursive software loops eventually hit diminishing returns.
WecoAI/aideml GitHub repo github.com
The baseline AIDE agent is open-source with CLI, tree-search visualization and support for frontier and local models — but no AIDE² code has been released alongside the technical report.
Redis engineering blog — ‘Why multi-agent LLM systems fail’ redis.io
Multi-agent systems suffer from redundant token consumption — sometimes 4x to 220x more than their single-agent counterparts — and are prone to conformity bias, where a hallucinated fact introduced by one agent is reinforced by others until locked in as false consensus.
Stanford Digital Economy Lab — ‘How are AI agents spending your tokens?’ digitaleconomy.stanford.edu
Agentic tasks can consume up to 1,000× more tokens than standard code chat, and stochastic trajectories mean the same task can vary in price by up to 30×, producing sticker shock for developers.
Medium — ‘Claude Code vs Cursor vs Devin vs Copilot in 2026’ medium.com
Cognition (creators of Devin) note that fragmenting tasks across multiple agents can lead to conflicting architectural decisions and compounding errors — one agent builds a component with one design pattern while a parallel agent uses another, producing incompatible outputs the coordinator fails to reconcile.
Hugging Face blog — Musamolla, ‘Multi-agent LLM systems failure’ (MAST taxonomy) huggingface.co
New failure taxonomies show that nearly 80% of production breakdowns stem from specification ambiguity and coordination gaps rather than model limitations, and up to 30% of agentic AI projects are abandoned after the proof-of-concept stage.
alphaXiv discussion of AgenSH (2609.26781) alphaxiv.org
Because workers were staggered rather than launched simultaneously, the scaling comparisons do not reflect equal total compute usage, potentially masking the massive resource requirements — and reviewers warn of rapidly diminishing returns beyond 128 agents.
Deep Learning Weekly Issue 474 — practitioner roundup deeplearningweekly.com
Real-world failure modes such as conflicting merges and ‘discovery noise’ become significant bottlenecks at enterprise scale, and the no-internet reconstruction environment may not reflect the unpredictability of live production deployments.
