The Archive

STMD distills diffusion model transition maps for faster probabilistic inference without teacher supervision.

Attention entropy analysis reveals token-level RL post-training redundancy and heterogeneous learning signals in LLM reasoning.

OpenAI documents sandboxing, approvals, network policies, and telemetry for safe Codex deployment in agent workflows.

Bharat ABIS describes billion-scale multimodal biometric search system for national identity using fingerprint, face, iris matching.

Prefix consistency weights CoT traces by regeneration stability, improving LLM reasoning accuracy without log-probability access.

Vision-language models enable zero-shot ODD perception for autonomous systems compliance with safety-critical regulations.

Covariate-shift-free method scales modular addition learning via auxiliary modulus without training-test distribution mismatch.

Nanoleaf teased a trio of new products focused on embodied AI as it looks to move its brand beyond smart lighting. | Image: Nanoleaf Smart lighting company Nanoleaf has been unusually quiet recently. While competitors such as Govee and Philips Hue have been pumping out new products and innovative features at an impressive pace, Nanoleaf has launched just a handful of smart lighting products in the last two years. There's a reason for this lull - the company has been going through a "brand evolution" focused on wellness, robotics, and, of course, AI. "The smart home is getting kind of boring,"...

Reddit user discusses token usage limits and mentions Opus 4.7 as solution for exhausting remaining quota before reset.

Few-shot LLM scoring (GPT-5.2) on short answers shows mid-range degradation on partial-credit responses without task-specific adaptation.

MAVEN multi-agent framework adds in-step epistemic verification and adversarial skepticism to LLM reasoning chains for high-stakes tasks.

LithoBench introduces first domain-specific benchmark for evaluating multimodal LLMs on geological lithology interpretation from remote-sensing imagery.

Paper proposes logic augmentation and active inference for extracting tacit procedural knowledge into machine-interpretable representations.

Decentralized multi-agent pathfinding solver using local communication and learned coordination for scalable multi-robot trajectory planning.

Benchmark evaluates latest LLMs on grammatical error correction across edit precision, fluency, and meaning retention with reference-free metrics.

Stochastic first-order optimization method using medoid mini-batch gradient sampling for heavy-tailed noise without explicit clipping.

Psych-201 dataset reveals post-training reduces LLM behavioral alignment with humans, with divergence widening in newer model generations.

HDMI: probe-free causal intervention method steers LLM hidden states via gradient-based margin maximization without auxiliary classifiers.

PhoneSafety benchmark (700 examples) distinguishes genuine safety understanding from task failure in phone-use agents via fine-grained outcome categorization.

Checkpoint-level analysis of gender bias formation in Dutch BERT trained from scratch, tracing emergence of morphological gender information.

Intent-driven Semantic ID generation for conversational news recommendation bridges implicit user intents unaddressable by standard RAG pipelines.

Ensemble methods for psychological defence mechanism classification via orthogonal voter axes; shared task at BioNLP 2026.

SAM 3D Animal: multi-animal 3D reconstruction from single images using SMAL+ parametric model and prompt-based disambiguation.

Anthropic leadership structure dominated by women across product, engineering, and executive roles.

CIKA framework uses LLMs as interventional simulators to identify causal concept contributions to mathematical reasoning via prompts.

Bi-objective decision tree learning for recourse summaries balancing effectiveness and cost in classifier auditing.

Generalization bounds for extreme multi-class supervised contrastive learning accounting for tuple dependencies in finite pools.

Causal Energy Minimization framework derives Transformer block parameterization (MHA, gated MLPs) from energy function optimization.

Parallel lifted classical planning via semi-naive Datalog evaluation with rule-level and instruction-level parallelism.

POISE: reinforcement learning baseline estimation from LLM hidden states without critic model, improving variance reduction for RLVR.