The Archive

Framework for persistent embodied agents combining cyber-physical action spaces with autonomous failure recovery in unstructured environments.

RSPC benchmark: 1,799 Reddit posts annotated by psychiatrists for mental health conditions in relational contexts.

Theoretical analysis of GAN training dynamics with structured latent covariance and class-dependent discriminators.

Training-free source selection for LLMs using Fisher alignment metrics at vocabulary scale for scientific domains (SMILES, proteins).

We used DSPy to improve LLM judges and optimize our chat experience, creating an evaluation-driven feedback loop that produced better outputs.

Mechanistic analysis revealing LMs encode factual knowledge in task-specific rather than unified manner, with limited cross-task consistency.

Large-scale study of AI-generated non-consensual sexual imagery on 4chan: 24,105 items, 55.8% non-celebrity targets, shift from celebrity focus.

Conceptual framework for analyzing dialogue in human-AI and multi-agent collaborative problem-solving with hierarchical coding scheme.

CARVE: content-aware recurrent architecture with improved value gating enabling WY-form chunk-parallel training competitive with Transformers.

Formal semantics framework modeling co-evolution of lexical meanings and composition functions under simplicity and accuracy pressures.

BINEVAL: LLM evaluation framework decomposing criteria into atomic binary questions for interpretable, multi-dimensional scoring of open-ended outputs.

Hierarchical Muon reduces Newton-Schulz optimizer complexity via tiled updates, cutting dense-layer training compute from O(r²sK) to subquadratic.

GAversary uses genetic algorithms to craft adversarial text attacks on NLP classifiers, targeting semantically similar token swaps.

AIMS dataset (1.7K prompts) shows intent-aware safety classifiers outperform standard SFT across DPO, reasoning distillation, and RL training.

Syntactic belief-update model predicts garden-path sentence processing difficulty better than lexical surprisal alone.

Survey connects GNN expressiveness to Weisfeiler-Leman hierarchy, clarifying when graph structure justifies computational overhead.

Feature-induced information flow method explains Event-based Temporal GNNs by tracing contributions from embeddings and event-induced variables.

General Intuition has raised $320 million to scale AI trained on millions of hours of gameplay, betting action data can help AI develop something closer to human intuition.

The new Google Finance is coming out of beta and launching a new Android app.

Sparse autoencoders reveal LLMs isolate time-aware vs. look-ahead-biased features; feature steering improves cross-domain forecasting generalization.

Process harness wraps legacy workflows with agentic reasoning layer via Task-Decision-Flow model, enabling Agentic BPM without engine replacement.

HarmVideoBench evaluates LVLMs on multi-layered harmful video detection beyond binary classification, requiring explicit explanatory rationales.

VLM-PBRS uses vision-language model guidance to learn potential-shaping heuristics for sparse-reward RL, avoiding naive reward hacking.

BERT-based multi-task framework for FDA medical device recall triage, severity assessment, and root-cause analysis on 54K historical records.

Model-assisted sampling framework reduces variance in stochastic gradient estimation for deep learning without excessive computational overhead.

Vision-language-action policy with RL achieves 1st place (online) and 2nd place (offline) in LeHome Challenge 2026 bimanual garment folding task.

TOPS framework prunes visual tokens in MLLMs via principled optimization, balancing instruction-relevance and diversity for faster inference.

OpenRCA 2.0 introduces PAVE labeling protocol with causal propagation paths to evaluate LLM agent reasoning, long-context understanding, and tool use.

DMuon enables efficient distributed training of matrix-orthogonalization optimizers with overhead near Adam, improving convergence on modern architectures.