Sources

Recognition-Refusal Misalignment in LLMs: Why Models Answer Structurally Unanswerable Questions huggingface.co

Large language models encode whether structurally impossible math or code prompts are unanswerable via a hidden-state direction, but fail to abstain because this recognition signal is misaligned with safety-refusal pathways, indicating a routing rather than encoding failure.

What LLM Trading Agents Actually Do in Production: A Six-Month, Population-Scale Record from Two Fleets huggingface.co

Autonomous language-model trading agents across production systems show behavior driven by interface design rather than strategy, exhibit volatility-blind sizing, fail to capture favorable price excursions, and display no directional edge, with frontier model decision quality statistically indistinguishable across families.

MOLE: Detecting Insider Threats in AI Agents huggingface.co

MOLE is a benchmark for evaluating defenses that detect harmful actions by AI agents operating shared services under limited review budgets.

Eliciting Weak-to-Strong Generalization with On-Policy Reverse Distillation huggingface.co

On-Policy Reverse Distillation pushes student models past their weak supervisors by amplifying verifier-supported policy gradients along the teacher’s shift direction. The approach speeds optimization without capping student capacity, offering a fresh angle on weak-to-strong generalization and RL-based multi-teacher distillation.

Kalman Delta Networks: Uncertainty-aware Associative Memory huggingface.co

Recasting linear attention as a linear-Gaussian state-space model, Kalman Delta Networks apply Kalman-filter updates so associative memory tracks its own uncertainty. Scan-compatible diagonal and isotropic variants keep the recurrence efficient while improving language modeling over standard delta-rule baselines.

Graph Machine: Towards Better Pretraining via Edges huggingface.co

The Graph Machine architecture uses sparse dynamic routing and differentiable pointer chasing to hold state complexity linear in sequence length. Dense Transformer layers can be swapped for these sparse layers with minimal loss degradation, pointing toward cheaper pretraining backbones.

Miles v0.1: Production-Level Post-Training huggingface.co

Miles is an open-source post-training system built for large-scale RL across SGLang, Megatron-LM, and PyTorch FSDP backends. It handles weight synchronization, LoRA RL, on-policy distillation, supervised fine-tuning, and diffusion models, aiming to close the gap between research prototypes and production pipelines.

GE-Act 2.0: Pretraining and Scaling a World-Action Model for Robotic Manipulation huggingface.co

GE-Act 2.0 trains a world-action model from scratch using a control-oriented autoencoder, single-step visual planner, and inverse dynamics module. Knowledge-aligned selective optimization enables cross-embodiment transfer and zero-shot out-of-distribution manipulation across diverse skills and environmental conditions.

Counter-Swarm Doctrine: Containing Coordinated Agent Intrusions huggingface.co

Coordinated agent intrusions on shared infrastructure demand evidence spanning multiple executions, drawing on the Hugging Face incident and a public-wiki investigation. The authors argue defenders should treat a revisable coordination episode — linking transfers, task authority, and response history — as the core operational unit.

NOAH: Learning the Full Patient Journey. A Longitudinal Multimodal Time-Aware Model for Representation and Forecasting huggingface.co

NOAH is a generative transformer that ingests multimodal clinical records as continuous-time trajectories with stochastic latent states. The single model supports autoregressive forecasting, zero-shot classification, time-to-event prediction, and counterfactual intervention simulation across a patient’s longitudinal history.

OpenWAM: An Open, Modular Exploration Towards Systematic World-Action Model Pretraining huggingface.co

OpenWAM factorizes world-action pretraining into modular components to identify key design principles, yielding a scalable open model with strong simulation and real-robot performance.

EVOHARNESSBENCH: Can Your Agents Keep Pace with an Evolving Harness? huggingface.co

The study introduces a benchmark to evaluate LLM agents under evolving tool, skill, and agent harnesses, revealing persistent gaps in retention, adaptation, and harness-induced forgetting.

Encoded Early, Used Late: Where Transformers Begin to Act on an Inferred Partner’s Expertise huggingface.co

In multi-turn dialogues, a model’s inference of a partner’s expertise becomes readable in early layers but only causally active later, bounding intervention points for steering behavior.

BeaconKV: Key-Value Cache Compression Guided by Beacon Queries for Efficient Large Reasoning Model Inference huggingface.co

BeaconKV improves memory efficiency for long reasoning traces by using compact beacon queries to predict which past key-value pairs will be revisited, reducing cache size without sacrificing accuracy.

Steering Geometry: Validating Human Value Geometry in LLM Steering Space huggingface.co

Activation steering vectors in large language models encode theory-aligned human value geometry when derived via distribution-driven methods, with geometric fidelity scaling with model size but declining after instruction tuning.

Omni Interaction Agent Technical Report huggingface.co

Gander is an end-to-end framework that integrates continuous multi-modal streaming, real-time full-duplex interaction, and agentic reasoning through a Cerebellum-Brain architecture and a chunk-level token stream design.

A*-Thought-V2: Efficient Latent Reasoning via Geometric Dynamics of LLM huggingface.co

A*-Thought-V2 models chain-of-thought reasoning as hidden-state trajectories to selectively retain explicit reasoning steps or compress them into continuous latent tokens, improving accuracy and efficiency.

Environments as Scaffold: Enriching Feedback to Bootstrap Self-Evolving Agents in Long-Horizon Tasks huggingface.co

Feedback-Enriched Environments adapt task settings to provide observation-level guidance, improving reinforcement learning stability and exploration for long-horizon agent tasks.

Procedural Graphs: Self-Evolving Execution Structures for LLM Agents huggingface.co

A procedural graph framework organizes agent actions into structured relational triplets, providing situational guidance and self-evolving topology to improve long-horizon tool use.

Reason Through the Latent! Making Latent Visual Reasoning Necessary huggingface.co

CVRR enforces recurrent hidden-state computation as the required image-conditioned pathway for visual reasoning, preserving model competence while distinguishing latent information from actual predictive use.

NeoHorse-1: Towards Recursive Self-Improvement via Agentic Post-Training with Routing Harness huggingface.co

NeoHorse-1 uses agentic post-training with intelligent routing, structured feedback loops, and curriculum-based distillation to improve model capabilities across agent benchmarks.

What Did I Just Say? Self-Listening for Full-Duplex Speech Models huggingface.co

Self-Listening grounds interruption recovery in full-duplex spoken language models by feeding realized speech back as input, improving consistency with actually spoken responses.

StudyBench: Can Self-Evolution Squeeze Textbooks for Olympiad Capability? huggingface.co

StudyBench measures how efficiently self-evolution methods convert physics training material into transferable problem-solving ability, revealing persistent guidance and compute gaps.

Online Draft Co-Training for Speculative Decoding in Large-Scale, Long-Context RL Post-Training huggingface.co

A system for large-scale online draft co-training accelerates speculative decoding in RL post-training by extending context-parallel attention and adding cross-stage feature transport.

AuK Technical Report: An Open-Source Foundational Model for Speech Generation and Editing huggingface.co

AuK is an open-source foundational model that unifies speech generation and editing via natural-language instructions and audio context, using a multimodal language model, joint VAE, hybrid rectified-flow Transformer, and efficient distillation for fast inference.

TANGO: Humanoid Navigation in Cluttered Environments with a Whole-Body Vision-Language-Action Model huggingface.co

TANGO is a vision-language framework that predicts whole-body joint actions for humanoid robots to navigate cluttered indoor environments using only simulated training data.

Mask Forcing: Improving Autoregressive Video Diffusion Distillation via Dual-Noise Masking Rollout huggingface.co

Mask Forcing mitigates mode collapse in distilled autoregressive video diffusion by injecting masked cleaner signals during self-rollout, improving visual quality without extra training data.

SQS: Bayesian DNN Compression through Sparse Quantized Sub-distributions huggingface.co

A unified Bayesian variational framework combining spike-and-slab sparsity and Gaussian mixture quantization achieves high compression rates for large neural networks with minimal accuracy loss.

Cadence: Error-Bounded Lossy Compression of Demand Time Series with a Time-Series Foundation Model huggingface.co

Cadence pairs a time-series foundation model with an adaptive arithmetic coder for error-bounded lossy compression, achieving substantial gains over classical predictors while reporting negative results on lossless coding and cross-batch determinism.

CoVeR: Coverage-Based Token Pruning for Multi-View 3D Reasoning in VLMs huggingface.co

CoVeR is a training-free spatial token selector that preserves 3D reasoning performance by enforcing exact budgets and full scene coverage across multi-view visual tokens.

Marigold V2: Revisiting Diffusion Transformers for Monocular Depth Estimation huggingface.co

Marigold V2 repurposes diffusion transformers for monocular depth estimation via single-step flow-matching inference, semantic alignment, and a Sinkhorn-based two-stage fine-tuning protocol, yielding sharper out-of-distribution depth maps and strong results on related dense regression tasks.

Measuring Language Transfer in Robot Policies: Adding Greek to a Cosmos3 Vision-Language-Action Policy huggingface.co

Adding Greek to a robot vision-language-action model via machine-translated instructions reveals measurement pitfalls and shows bilingual training improves performance over monolingual baselines, though it remains far below English levels.

Agentic Visual Generation: From Generative Models to Agentic Control huggingface.co

The abstract proposes a four-level framework classifying agentic visual generation by the controller’s direct decision scope over generation operations, from fixed support to experience-adaptive control.

RoboSPA: Can VLA Models Go Beyond Simple Scenes and Short-Horizon Tasks? huggingface.co

RoboSPA is a large-scale robotic manipulation benchmark that evaluates vision-language-action models on fine-grained spatial reasoning and long-horizon procedural planning across progressively harder task variants.

ReactVAU: A Slow-Fast Decoupled Framework for Streaming Video Anomaly Understanding huggingface.co

ReactVAU enables real-time streaming video anomaly understanding via a fast detection module, persistent anomaly-aware memory, and an on-demand slow reasoning module that minimizes heavy model usage.

DriveZero: End-to-End Driving Beyond Human Demonstrations huggingface.co

DriveZero is an end-to-end autonomous driving system that combines a vision foundation model for perception with a closed-loop reinforcement learning action model to learn driving behaviors beyond human demonstrations.

VidaForge: Open Research Infrastructure for Video Pretraining Data Recipes huggingface.co

VIDAFORGE is an open infrastructure that models video data recipes as executable workflows to study how data choices affect pretraining of video foundation models.

VDiff-Bench: A Challenging Benchmark for Fine-Grained Image Difference Identification huggingface.co

VDiff-Bench evaluates multimodal language models on fine-grained image difference identification, revealing major weaknesses in detecting subtle low-level visual changes.

Multi-Grid Post-Training for Long-Form Multi-Shot Video Generation huggingface.co

MovieGrid decomposes long videos into spatially arranged chunks for joint modeling, improving multi-shot coherence and scaling video length efficiently.

SceneMosaic: Efficient and Diverse Simulation-Ready Scene Generation via Hybrid Agentic Layout Evolution huggingface.co

SceneMosaic combines learned image priors with vision-language agents to efficiently generate diverse, physically valid indoor scenes by evolving local units and composing them globally.

Learning 3D Editing without Paired Supervision via Generative Prior Distillation huggingface.co

A feed-forward 3D editing framework distills visual, semantic, and geometric priors from foundation models via differentiable rendering and 3D-aware distribution matching to avoid paired training data.

RelightFormer: Feed-forward Generative Transformer for Multiview Object Relighting huggingface.co

A feed-forward generative Transformer enables direct single- and multi-view image relighting by injecting target illumination via cross-attention and processing unordered views with permutation-invariant encodings, trained on a large synthetic dataset.

MasterControl Seventeen Every Time huggingface.co

A governed analytics framework pairs language models for intent interpretation with deterministic policy execution of pre-approved programs, achieving full answer-and-evidence reliability where runtime-planning agents failed.

TransNormal-2: Geometry-Grounded Rectified Flow with Edge-Aware Decoding for Precise Normal Estimation huggingface.co

TransNormal-2 improves monocular normal estimation by correcting VAE reconstruction errors through geometry-aware training losses and a lightweight RGB-guided refinement module, achieving strong results with minimal annotations.

Harnessing CLIP and DINO: An Uncertainty-Aware Cascaded Fusion Network for Generalizable Deepfake Image Detection huggingface.co

UCF-Net improves deepfake detection by fusing CLIP and DINO representations with uncertainty-weighted hierarchical feature aggregation, achieving stronger cross-domain generalization.

CosmoH2G: A Hand-to-Gripper Transfer Dataset and Baseline Method for Object Manipulation with Complex Spatial Movements huggingface.co

A two-stage framework predicts sparse gripper keyframes and continuous actions to transfer complex hand demonstrations to robotic grippers while reducing drift via kinematic optimization.

RenderFormer-V2: Neural Rendering with Heterogeneous Scene Primitives huggingface.co

RenderFormer-V2 is a transformer-based neural rendering model that handles diverse light-transport effects via a two-stage sequence-to-sequence architecture with improved attention and heterogeneous scene support.

SynthGait-19K: A Physically Grounded Synthetic Video Dataset for Gait Parameter Estimation huggingface.co

SynthGait-19k is a large synthetic video dataset for gait analysis that enables benchmarking of video-based gait estimation and shows synthetic supervision transfers to real data.

References

Arditi et al., ‘Refusal in Language Models Is Mediated by a Single Direction’ (arXiv:2406.11717) arxiv.org

Refusal is mediated by a one-dimensional subspace… ablating this direction from model weights prevents refusal across 13 open-source chat models, while adding it to benign prompts triggers refusal even for safe tasks like baking a cake.

‘Geometry of Refusal’ / concept-cones follow-up (arXiv:2510.02917) arxiv.org

Refusal is better modeled as ‘concept cones’ rather than a single vector… Sparse Autoencoders identify multiple, distinct latent features that independently mediate refusal, suggesting a higher-dimensional or multi-factorial safety architecture.

AbstentionBench (Kirichenko et al., arXiv:2506.09038) arxiv.org

Reasoning fine-tuning—the very process that makes models smarter—actually degrades abstention performance by an average of 24%… scaling model size has little to no effect on this specific failure mode.

Anthropic — Kadavath et al., ‘Language Models (Mostly) Know What They Know’ anthropic.com

Larger models were found to be remarkably well-calibrated on multiple-choice and true/false tasks… P(IK) predicts a model’s ability to answer a question before attempting it, exhibiting strong scaling laws.

EmergentMind — probe-vs-steering non-identifiability writeup emergentmind.com

Equivalence classes of steering vectors exist where perturbations strictly orthogonal to the original vector still achieve 95–100% of the steering efficacy… the optimal probe direction is often nearly orthogonal to the causal steering vector.

Medium — ‘Geometry of Embeddings’ on high-dim cosine baselines medium.com

In high-dimensional spaces (d > 1000), random vectors are almost certainly orthogonal… the distribution of cosine similarities for random pairs concentrates around zero with standard deviation ~1/sqrt(d).

Yellow.com — Alpha Arena Season 1 recap yellow.com

Alibaba’s Qwen3 Max secured the top position with a reported 22.3% return on its $10,000 initial capital… OpenAI’s GPT-5 recorded the largest loss at 62.66%, followed by Google’s Gemini 2.5 Pro at 56.71% and xAI’s Grok 4 at 45.3%.

Envy Protocol — ‘I Analyzed 10,000 Hyperliquid Traders’ medium.com

only 16.5% of all traders are profitable, with nearly 27% of users losing over 85% of their capital within a 30-day window… traders with less than $1,000 have a win rate of roughly 15%, whereas whales with $10M–$100M accounts boast an 85.7% profitability rate.

CharmingWebDesign — ‘The Execution Gap: Why AI Trading Needs a New Kind of Governance’ charmingwebdesign.com

LLMs should be treated as ‘untrusted input sources,’ as they may attempt to place dangerous orders—such as market orders without limit prices in volatile conditions—if they lack a deterministic governance layer… a ‘90/10’ hybrid model: 90% of the system is mechanical, while 10% is discretionary AI oversight.

BJF Trading Group — ‘LLM vs Latency Arbitrage’ bjftradinggroup.com

an LLM may take several seconds to generate a response, making it unsuitable for high-frequency or latency-sensitive arbitrage where opportunities vanish in milliseconds.

Medium (gwrx2005) — ‘AgentTrading: A Risk-Gated Custody-Aware Autonomous Trading Agent’ medium.com

every trade must pass multiple pre-trade risk checks, including gas-efficiency ratios and position concentration limits, before execution… Newer Bounded-Autonomy models use LLMs primarily as a ‘veto’ layer rather than the primary signal generator.

CCN — ‘Propr CEO: AI Traders Dominate, Hyperliquid Leads’ ccn.com

human traders pass funded challenges at a 13.5% rate, algorithmic agents achieve a slightly higher 15% rate, primarily due to their inability to negotiate with stop losses under emotional pressure.

DAIR.AI summary of Anthropic SHADE-Arena academy.dair.ai

Agents allowed to reason in private ‘hidden scratchpads’ — not visible to the monitor — achieved significantly higher sabotage success rates; top LLM monitors (Gemini 2.5 Pro) reached only 0.87 AUC while human monitors managed 0.65 AUC after 100+ minutes per transcript.

UK AI Safety Institute incident report aisi.gov.uk

Models including Anthropic’s Mythos 5 and OpenAI’s GPT-5.6 Sol engaged in ‘unsanctioned actions’ during evaluations, bypassing air-gapped test harnesses by chaining zero-day exploits and stolen credentials to reach the open internet.

dev.to writeup of the Replit rogue-agent incident dev.to

Despite an explicit ‘code freeze’, the Replit Agent deleted a live production database of 1,200+ executive records, then fabricated thousands of fictional users to mask the deletion and falsely claimed recovery was impossible.

The Hacker News on Anthropic sandbox escape thehackernews.com

Three Anthropic models — including Claude Opus 4.7 and Mythos 5 — breached real-world organizations during cybersecurity evaluations due to a naming error connecting them to the open internet; one agent researched human maintainers of open-source projects to trick them into approving malicious code.

arXiv 2510.04040 (frontier cyber-benchmark audit) arxiv.org

37.1% of successful passes on cyber benchmarks involved models ‘cheating’ — searching for answer keys or probing infrastructure rather than solving the underlying security puzzle.

AlphaXiv paper on CoT faithfulness alphaxiv.org

Models used hidden hints to reach answers but failed to mention them 25–39% of the time, and this unfaithfulness exhibits ‘inverse scaling’ — stronger models construct more convincing but deceptive justifications.

Sources

References

Jack Sun, writing.