The Archive

Shared Selective Persistent Memory architecture for agentic LLM systems retains reusable task specifications, schemas, and tool configs across multi-turn sessions.

Multimodal RL reward hacking in MLLMs via visual misalignment; introduces NRFR metric to measure failures in improved-reward samples across VQA and chart tasks.

Terminal embeddings preserve pairwise distances under dimension reduction with applications to k-means and k-median coresets for time-series clustering.

Semantic framework distinguishes AI outputs as engineered representations versus facts, formalizing failure modes like extrapolation, refuted assertion, and stale sources.

Analysis of representational variance in language models shows token-level context dominates (79-91%) over category structure (4-12%) across 14 models, challenging neural collapse theory.

BTHA framework decouples language guidance from vision-text backbones for medical image segmentation via stable feature-level adapters.

Foveated Dynamic Transformer (FDT) applies human visual system principles for robust, efficient vision transformers with inherent noise/adversarial resilience.

ProofCouncil agent using author-critic architecture solved 6/10 real mathematical problems in FirstProof challenge via agentic workflows.

Adaptive Multi-Teacher Routing (ATR) uses uncertainty-driven structure selection to improve universal machine-learning interatomic potentials with limited high-fidelity data.

Pipeline combines Ghidra reverse engineering, anchor-based retrieval, and LLM reasoning to recover source code from stripped binaries.

Test-time prompt adaptation for CLIP exploits distributional brittleness of adversarial perturbations to improve vision-language model robustness without retraining.

Analysis of Bayesian causal discovery under latent confounding in linear Gaussian networks identifies critical correlation thresholds affecting DAG posterior inference.

Parameter-efficient CLIP adapter with continuous metadata conditioning for long-term animal re-identification under morphological and seasonal distribution shifts.

Test-time scaling on EXAMS-V multilingual benchmark shows parseability is primary factor for small VLMs (Qwen2.5-VL-7B, Qwen3.5-4B), not search algorithm.

Open source AI is booming, according to Hugging Face CEO Clem Delangue. The company has grown into something like a GitHub for AI in recent years, where AI builders can share and download open models and datasets, now used by roughly half the Fortune 500. Delangue has seen the same story play out again and again: companies start […]

Multimodal retrieval framework combining vision and trajectory for autonomous-driving scenario matching in large-scale datasets.

Soofi S 30B-A3B: open-source MoE-Mamba hybrid for German/English with 3B active parameters, matches 14-27B dense models on benchmarks.

QADAPT: multi-agent RL with factored action spaces for tuning electrostatically-defined quantum-dot arrays.

SVF-CR: multimodal fusion framework for detecting ambivalence and hesitancy from synchronized facial, visual, and acoustic cues.

Test-time training method using guided self-learning to improve long-context LLM performance on extended input sequences.

Though Instagram head Adam Mosseri doesn't want to filter out AI content on the platform, he argues that you "shouldn't have it in your feed" if you don't like it. "I don't think we should filter out AI content," Mosseri said during an interview on Lenny Rachitsky's podcast. "I think we should let you know if content is AI content or not." At the same time, Mosseri seems to be drawing a distinction between content-based sorting and banning AI from the platform entirely. In fact, he believes people who love AI content "should be able to have a feed that's just AI town." Instagram, like many ot...

Theoretical analysis: InfoNCE contrastive loss population risk is O(1/k)-close to cross-entropy with k negative samples.

SYNRARE: synthetic EHR generation for rare disease diagnosis benchmarking to overcome privacy and data scarcity constraints.

AutoWorldBuilder: multi-agent LLM system for fictional worldbuilding with hierarchical context compression and iterative review.

Empirical study establishing data-efficient training guidelines for inertial sensor classification models via learning curve analysis.

On-device adaptive learning for EV battery power prediction using continual fine-tuning on resource-constrained platforms.

A solar and home energy storage company is expanding into AI data centers, but not by building one - instead, it's offering to pay its customers to put its compute units in their homes. Sunrun is launching a pilot program for a new "distributed AI compute" program that will "place numerous compute nodes in homes equipped with Sunrun solar and battery storage systems." Customers will be "compensated" for participating in the pilot program. Sunrun plans to sell the distributed compute power from the nodes to "enterprise compute buyers," like AI companies. It's a new approach to finding resource...

Biomolecular structure prediction and co-folding with models like OpenFold3 are now mainstream, large-scale workloads powering drug discovery and protein... Biomolecular structure prediction and co-folding with models like OpenFold3 are now mainstream, large-scale workloads powering drug discovery and protein design. Increasingly, they’re driven end-to-end by AI agents. For an agent to run that pipeline well, every step needs to be fast and scalable: Multiple Sequence Alignment (MSA) generation, co-folding inference, serving, and multi-GPU scale-out. Source

Deutsche Telekom deploys OpenAI models across customer service, employee workflows, and network operations as part of AI-native telco strategy.

A big day for OpenAI.