1. What the Eyes See, the LLMs Miss: Exploiting Human Perception for Adversarial Text Attacks
arXiv API published an update: What the Eyes See, the LLMs Miss: Exploiting Human Perception for Adversarial Text Attacks. Model availability, speed, and migration paths continue to change quickly across the AI stack. Verified releases are most valuable when they translate into adoption data, technical documentation, or broader customer rollout.
Aitoolsfi Summary:Perception Gap: LLM moderation systems fail to detect harmful content when adversarial text exploits the discrepancy between human visual processing and machine tokenization.
Tokenization Vulnerability: Attackers bypass safety filters by using character-level manipulations that remain legible to humans but appear as benign, unrecognizable noise to token-based models.
Moderation Risk: This fundamental blind spot forces a shift toward multimodal verification layers to prevent automated systems from being easily gamed by simple obfuscation techniques.
Source: arXiv API
2. Cranio-Diff: Diffusion-based Cross-domain Craniofacial Reconstruction with 2D X-ray Skull Guidance and Structural
arXiv API published an update: Cranio-Diff: Diffusion-based Cross-domain Craniofacial Reconstruction with 2D X-ray Skull Guidance and Structural. Model availability, speed, and migration paths continue to change quickly across the AI stack. Verified releases are most valuable when they translate into adoption data, technical documentation, or broader customer rollout.
Aitoolsfi Summary:Reconstruction Precision: Cranio-Diff overcomes the limitations of standard generative models by anchoring facial synthesis directly to patient-specific skeletal geometry.
Structural Guidance: The framework utilizes 2D X-ray inputs as a structural constraint to guide diffusion-based cross-domain mapping between bone and soft tissue.
Medical Imaging: This approach signals a shift toward physically-constrained generative models that prioritize anatomical accuracy over purely aesthetic visual synthesis.
Source: arXiv API
3. PsychoSafe: Eliciting Psychologically-Informed Refusals in Large Language Models
arXiv API published an update: PsychoSafe: Eliciting Psychologically-Informed Refusals in Large Language Models. Model availability, speed, and migration paths continue to change quickly across the AI stack. Verified releases are most valuable when they translate into adoption data, technical documentation, or broader customer rollout.
Aitoolsfi Summary:Refusal Strategy: PsychoSafe shifts LLM safety from blunt blocking to nuanced, psychologically-informed communication that maintains user rapport during denials.
Interaction Design: The framework integrates behavioral science principles directly into the model's response architecture to manage user expectations during refusal scenarios.
UX Evolution: This approach signals a move toward more sophisticated conversational interfaces that prioritize user experience without compromising essential safety boundaries.
Source: arXiv API
4. Observability for Delegated Execution in Agentic AI Systems
arXiv API published an update: Observability for Delegated Execution in Agentic AI Systems. Model availability, speed, and migration paths continue to change quickly across the AI stack. Verified releases are most valuable when they translate into adoption data, technical documentation, or broader customer rollout.
Aitoolsfi Summary:Execution Ambiguity: Current logging standards fail to distinguish between distinct delegation assignments, rendering standard audit trails functionally opaque.
Traceability Gap: The research highlights a structural flaw where identical execution traces mask the underlying logic of delegated task distribution.
Systemic Risk: Developers must implement granular metadata tagging to ensure verifiable accountability as automated delegation becomes a standard architectural pattern.
Source: arXiv API
5. An 84-Format Numeric Catalog with Bit-Exact Conformance Vectors: A Vendor-Neutral Reference for FP8, BF16, MXFP4, and
arXiv API published an update: An 84-Format Numeric Catalog with Bit-Exact Conformance Vectors: A Vendor-Neutral Reference for FP8, BF16, MXFP4, and. Model availability, speed, and migration paths continue to change quickly across the AI stack. Verified releases are most valuable when they translate into adoption data, technical documentation, or broader customer rollout.
Aitoolsfi Summary:Standardization Crisis: The explosion of specialized numeric formats has created a fragmented hardware landscape that threatens interoperability between training and inference silicon.
Conformance Reference: This catalog provides bit-exact vectors to harmonize 84 distinct formats, enabling developers to validate hardware precision across diverse micro-architectures.
Hardware Interoperability: Establishing vendor-neutral benchmarks will accelerate the adoption of low-precision formats like MXFP4, reducing reliance on proprietary, non-portable optimization stacks.
Source: arXiv API
6. GenEyePose: Patient-Free, Knowledge-Based Saccadic Eye Movement Modeling for Digital Neurophysiologic Biomarker
arXiv API published an update: Eye movements, including saccades, are widely regarded as highly sensitive and objective biomarkers of neurophysiologic states. Detecting saccadic signatures in neurologic diseases offers. Model availability, speed, and migration paths continue to change quickly across the AI stack. Verified releases are most valuable when they translate into adoption data, technical documentation, or broader customer rollout.
Aitoolsfi Summary:Diagnostic Automation: GenEyePose replaces human-subject data with synthetic knowledge models to identify neurological biomarkers through precise saccadic tracking.
Synthetic Modeling: The framework utilizes generative simulation to map eye movement patterns, bypassing the logistical constraints of traditional clinical patient data collection.
Clinical Scalability: This approach shifts neurophysiologic screening toward digital-first workflows, potentially lowering the barrier for early-stage disease detection in remote clinical settings.
Source: arXiv API
7. MeCo: One-Step MeanFlow-based Corrector for Multi-Channel Speech Separation
arXiv API published an update: MeCo: One-Step MeanFlow-based Corrector for Multi-Channel Speech Separation. Model availability, speed, and migration paths continue to change quickly across the AI stack. Verified releases are most valuable when they translate into adoption data, technical documentation, or broader customer rollout.
Aitoolsfi Summary:Perceptual Optimization: MeCo shifts speech separation focus from raw metric benchmarks to improving actual human listening quality.
MeanFlow Correction: The architecture utilizes a one-step MeanFlow-based corrector to refine multi-channel audio outputs post-separation.
Audio Fidelity: This approach signals a move toward generative correction methods that prioritize natural sound reproduction over traditional discriminative modeling.
Source: arXiv API
8. Correlation Is Not Enough: Embedding Human Metadata for Individual Causal Discovery
arXiv API published an update: Ask a pretrained biomedical language model whether "cortisol 28 ug/dL" and "stock-market volatility" are related, and it returns a cosine similarity of 0.83 on a scale where 1.0 means. Model availability, speed, and migration paths continue to change quickly across the AI stack. Verified releases are most valuable when they translate into adoption data, technical documentation, or broader customer rollout.
Aitoolsfi Summary:Spurious Correlation: Biomedical language models currently struggle to distinguish between semantic proximity and actual causal relationships in specialized datasets.
Metadata Integration: Researchers propose injecting human-derived causal metadata into embedding layers to force models beyond simple statistical association.
Scientific Reliability: This approach marks a shift toward grounding LLMs in structured domain logic to prevent dangerous misinterpretations in clinical research.
Source: arXiv API
Summary
Cognition, Qwen, Google, and Meta show a market moving past novelty and into operational pressure. The most important AI updates now sit around deployment boundaries: who can access a model, which tools an agent can call, how performance is measured in real tasks, and whether the business case is strong enough to justify production use.