The Archive

SHAP-based explainability algorithm for time series foundation models enables transparent forecasting in critical infrastructure applications.

Methodological analysis of unit definitions in surprisal theory and language model probability assignments for cognitive modeling.

Penalty regularization approach achieves global optimality for constrained exploration in safety-bounded reinforcement learning.

So I started the morning with 1 message to summarize everything after I woke up on a session, and immediately got hit with usage limit exceeded (Im on max 5x plan). So I thought maybe it was my cron session (checked it and there were no tasks done at all over night). I have nothing else running.. After 5 hours, I started running a session again to continue working, 17 minutes later (I know its 17 minutes exact because I had a youtube video playing at the same time). Just went to 37% used. How is this even possible? The task I did was to create a simple .ps1 script. I've used claude code sin...

TACHIOM system improves multivector retrieval efficiency via token-aware clustering and hierarchical indexing for dense passage retrieval.

Claw-Eval-Live benchmark separates refreshable workflow signals from reproducible snapshots to evaluate evolving LLM agent task performance.

Crab runtime enables semantics-aware checkpoint/restore for autonomous agent sandboxes, bridging agent-OS semantic gap for fault tolerance and RL.

GPT-5.5 completed a multi-step cyber-attack simulation in 11 min ($1.73) vs. 12 hrs for human expert; UK AI Security Institute benchmark.

Latent adversarial detection uses residual stream activation analysis to identify multi-turn prompt injection attacks with 93.8% accuracy.

Link lets users connect cards, banks, and subscriptions, then authorize AI agents to spend securely via approval flows.

Critical analysis of AI sign language translation systems from disability justice perspective, highlighting bias and excluded deaf community input.

PRISM mitigates distributional drift in multimodal model post-training via three-stage black-box distillation before RL, addressing SFT-induced capability degradation.

S²VAE improves 3D geometry preservation in visual world models by learning latent representations that encode scene structure over appearance alone.

Theoretical framework shows sparse autoencoders can capture concept manifolds rather than assuming independent linear directions, with implications for interpretability.

DEFault++ hierarchically detects, classifies, and diagnoses faults in transformer attention mechanisms and components without runtime errors.

1X showcases NEO factory as part of humanoid robotics competition, with brief mention of factory artwork.

Novel splitting techniques for bipolar set-based argumentation frameworks incorporating collective attacks and supports.

Today, game developers can begin integrating NVIDIA DLSS 4.5 with Dynamic Multi Frame Generation, Multi Frame Generation 6X, and the second-generation... Today, game developers can begin integrating NVIDIA DLSS 4.5 with Dynamic Multi Frame Generation, Multi Frame Generation 6X, and the second-generation transformer model for NVIDIA Super Resolution. In this post, we’ll go over new technologies and resources to share with our game-developer community, including: At CES 2026, we introduced DLSS 4.5, extending its AI-driven… Source

Neural network techniques are increasingly used in computer graphics to boost image quality, improve performance, and streamline content creation. Approaches... Neural network techniques are increasingly used in computer graphics to boost image quality, improve performance, and streamline content creation. Approaches like super resolution, denoising, and neural rendering help real-time engines work more efficiently, offering new creative possibilities while keeping performance in mind. Unreal Engine 5 (UE5) has taken several steps in this direction… Source

Auto-FlexSwitch reduces storage overhead in dynamic model merging via learnable compression of task-specific weight increments.

Hybrid neural-Kalman filtering improves UAV state estimation in degraded sensing by combining learned nonlinear dynamics with principled uncertainty quantification.

Court filing alleges xAI distilled OpenAI models; significant if verified but requires legal context confirmation.

Manus, an AI company Meta acquired for $2 billion last year is running ads promising quick, easy money with AI: Find local businesses without websites or with bad websites, have AI build them one, then call them up and sell it to them. As part of the campaign, Manus was paying content creators to build out Instagram, YouTube, and TikTok accounts that promote its AI product as an easy, lucrative gig. (The creators' TikTok accounts were taken down after The Verge inquired about them.) Some of these videos would also appear as official ads for Manus, but the posts on the paid creator accounts th...

FiLMMeD applies feature-wise linear modulation for cross-problem generalization in multi-depot vehicle routing via neural combinatorial optimization.

Framework mapping methodological dimensions of classroom interaction research (scale, duration, modality) in AI-enabled educational contexts.

Guidelines for adversarial benchmark task design in terminal-agent evals, distinguishing verification logic rigor from prompt-based task writing.

Neuro-symbolic framework integrating first-order logic, causal models, and RL for explainable, verifiable adaptations in safety-critical rule-based systems.

AI Security Institute benchmarks GPT-5.5 against Mythos on cyber-exploitation tasks; GPT-5.5 achieves 71.4% on expert-level tasks, performing comparably to Mythos.

Study characterizes consistency of emergent misalignment personas across fine-tuning domains in Qwen 2.5 32B, measuring correlation between harmful behavior and self-assessment.

TopBench benchmark (779 samples) evaluates LLMs on implicit predictive reasoning over tabular data, addressing latent intent recognition beyond retrieval.