The Archive

Analysis of memory utilization in hybrid recurrent-attention language models shows attention and recurrent layers are underused; proposes balancing pathways.

BazaarBench simulated marketplace benchmark evaluates safety of LLM agents in decentralized C2C transactions, tracking six failure types across transaction stages.

Technical paper on hyperbolic graph embedding optimization using parametrization choices to avoid numerical instability in Poincaré and Lorentz models.

ufakzeka-karar is open Turkish classification model with 182M parameters for decision questions, produces probabilities without text generation in one CPU pass.

DivLM post-training framework improves LLM diversity in creative story generation via continued pre-training on writing corpus and reinforcement learning.

Two-stream deep learning framework decouples temporal and spatial processing for EEG source imaging, trading off reconstruction accuracy and inference cost.

BRANCH-MoE uses binary decision tree routing for experts in mixture-of-experts layers to improve load balancing and reduce computation overhead.

RuModernBERT-ruLaw domain-adapted Russian encoder trained on 304K legislative documents improves performance on legal text understanding tasks.

Theoretical work on Hamiltonian learning from quantum dynamics using minimal control for analog simulators without fast arbitrary gates.

SAFE-MR separates claim veracity from evidence sufficiency in multimodal rumor detection using relation-aware graphs and dual scoring heads.

SAGA-CDR pairs books with mood-matched music via transformer sentiment embeddings and conditional GANs for cross-domain recommendation.

Mean-field theory analysis of normalized SGD shows adaptive learning rates accelerate convergence but impose scaling tradeoffs in width and batch size.

MedPrune dynamically prunes communication topologies in multi-agent medical VQA systems to reduce redundancy and token overhead.

An OpenAI publicist tried to change the topic of CEO Sam Altman's interview with Vanity Fair's Mark Guiducci after the editor brought up a ChatGPT user's suicide. When Guiducci confronted Altman about the incident, the publicist warned Guiducci about running out of time, saying she'd like to "move on" to another topic. The interruption came shortly after Guiducci asked Altman if he knew who Laura Reiley is, a journalist who wrote an essay for The New York Times about her daughter's conversations with ChatGPT before she took her own life. "Her daughter committed suicide after speaking to ChatG...

Wander is a graph foundation model using random walks as a unified interface to transfer across heterogeneous graphs, features, and tasks.

Prior-data fitted networks adapted for unsupervised tabular anomaly detection via contrastive learning without labeled anomalies.

Von Mises–Fisher likelihood with profile estimation enables closed-form speaker embedding enhancement without labeled data across multiple datasets.

Programmatic Search Agent extends agentic search beyond query reformulation by treating executable operations over candidate sets as primitive actions.

OVAL optimizes KV cache page retrieval for long-context LLM inference by encoding pages with output-aware objectives rather than attention-score proxies.

MM-KG aligns multimodal patient data (EHR, imaging, genomic) with biomedical knowledge graphs via typed layers and explicit edges for clinical LLM explainability.

Activation alignment method reduces in-context learning cost for tabular foundation models by using full context to teach sparse-subset behavior.

Study of sparsity-inducing routing maps (sparsemax, entmax, normmax) in MoE language models reveals training dynamics and expert participation patterns.

Theoretical analysis of gradient descent convergence for logistic regression on linearly separable data with large step sizes.

Machine learning ensemble framework detects nighttime anomalies (fires, eruptions, gas flaring) using NASA Black Marble satellite data.

Two-probe benchmark measures language model compliance spectrum: exploitability vs. stoppability across 12 models with compliance index.

VideoTapestry: training-free multi-agent framework refines hierarchical video memory with coarse-to-fine query-driven retrieval for long-video QA.

Sam Altman thinks that the benefits of AI will be so great that "the world should accept some bad things happening" along the way. The OpenAI CEO pointed to hacks, scams, and "other bad things" as costs society should expect to tolerate because "people will do tremendously orders of magnitude more good stuff" with AI, without elaborating on the potential benefits. His comments come as OpenAI pushes for a "lighter touch" approach to AI regulation amid intensifying anxiety over the safety of frontier models - fears fueled in part by the company's repeated failures to keep its increasingly capab...

Entropy-aware distribution mixing mitigates confident conflicts when small models imitate low-probability reasoning traces from teacher models.

Reward Stealing Attack extracts latent safety rewards via inverse RL to generate scalable adversarial prompts against aligned LLMs.

HackerRank’s AI interviewer has already conducted more than 500,000 interviews, with Snowflake, Snorkel, and Capgemini among its early testers.