The Archive

MaLoRA and MaLOR introduce dynamic, recurrent state-space adapters for task and token-level LLM adaptation, improving reasoning via selective modulation.

GAMUT benchmark introduces two-level meta-rubrics to measure factual completeness in long-form LLM generation, addressing precision-recall gap in factuality eval.

ResearchArena framework evaluates AI control and monitoring for detecting sabotage in automated AI R&D agents across safety/capability post-training and optimization tasks.

CircuitKIT open-source library unifies circuit discovery, evaluation, and intervention workflows for mechanistic interpretability with automated contrastive prompts.

Anthropic blocks authors from opting out of $1.5B settlement at last minute.

Off-Context GRPO uses privileged training information (solution prefixes) to enable RL with verifiable rewards on hard reasoning problems avoiding zero-gradient plateaus.

Staypoint detection benchmark provides ground-truth annotations for semantic trajectory analysis from noisy GPS data, addressing lack of standardized evaluation.

Real-time SDF-based mapping and motion planning for autonomous UAVs unifies occupancy mapping and trajectory optimization using signed distance representations.

Framework generalizes batch normalization and neural modules to Riemannian manifolds, Lie groups, and gyrogroups for manifold-valued deep learning.

SHRED-ROM reduced-order modeling with shallow recurrent decoders enables real-time optimal control synthesis for high-dimensional dynamical systems.

Study models how imperfect LLM detectors distort user incentives and downstream metrics, showing counterintuitive effects of detection as behavioral intervention.

LangGraph practitioner guide with three executable recipes for stateful multi-step agent workflows in business processes.

Safety analysis showing distributed, normalized failures in deployed AI systems are harder to instrument than obvious errors.

Anthropic donates additional $20M to Public First Action, totaling $40M commitment to AI policy advocacy.

OpenAI launches ChatGPT for Small Businesses program offering training and automation tools via ChatGPT Work.

There are new 3.6 and 3.5 models today, but Google is already training Gemini 4.

A federal judge has signed off on Anthropic's $1.5 billion class action settlement with authors who accused the company of training its AI models on copyrighted books, as reported earlier by Reuters. In an order on Monday, Judge Araceli Martínez-Olguín writes that the…

RL framework for predicting lean blowout in gas turbine combustors via optimal reactor clustering.

Network-agnostic GNN initialization layer to improve spatial generalization in traffic assignment models.

Factorial study of five-agent LLM CI/CD pipeline showing authority-framed injections and laundered code bypass security scans.

Auditable fraud detection pipeline combining gradient boosting, graph features, TreeSHAP, and bounded LLM investigation agent.

BioSecBench-Surveillance: 100-task verifiable benchmark for AI agents inferring pathogen genomic analysis pipelines from raw data.

PathAgentBench: benchmark for vision-language agents on gigapixel whole-slide pathology images evaluating multi-scale evidence-seeking.

Robust financial statement fraud detection framework using LLMs on structured+textual data with temporal generalization evaluation.

Controlled study of prompt design (format, instruction count, context length) effects on instruction adherence and hallucination across five LLMs.

Multi-relational GCNs for user modeling in personalized systems; applies existing GNN techniques to recommendation tasks.

Inference-time steering mitigates cross-lingual factual inconsistency in LLMs through contextual intervention strategies.

Neural network surrogates for thermodynamic property prediction in supercritical combustion; domain-specific physics simulation.

DBMol leverages AlphaFold-3 and Boltz-2 structure prediction models for de novo small molecule drug design.

MeetingToM benchmark evaluates multimodal LLMs on theory-of-mind reasoning in multi-party meeting scenarios.