The Archive

Gradus synthesizes graded training tasks for LLM agents in policy-constrained environments by decoupling task design from solution difficulty.

RITE benchmark evaluates LLM agent resilience to silent tool errors in e-commerce systems where incorrect intermediate outputs lack failure signals.

I2R-Pose estimates 3D human poses from radar in smart homes by initializing pose predictions and refining joint localization.

Analysis of adversarial L1 attacks shows concentrated perturbations are not necessary; distributed budgets retain 97% attack success.

Engram Memory Network improves prototype-based explainability by using neuroscience insights to encode task-relevant information like human memory.

Foundation-style pretraining framework transfers solar PV forecasting across geographically dispersed stations without per-site retraining.

CrystalAR applies autoregressive modeling with space-group symmetry constraints to generate novel crystalline structures.

AC/DC on a Budget reduces dense training phase cost in sparsification by zero-initializing masked parameters to optimize FLOPs-performance tradeoff.

Vision-language models improve counting accuracy on large sets through agent-based epistemic actions (marking, pointing) without retraining.

Proposes thermodynamic singularity criterion to detect memorization vs. proper learning in diffusion models across architectures.

Analyzes Muon spectral optimizers on convex quadratics, characterizing gain and length regimes for constant step sizes.

S²-MAS framework simulates 500M household agents with compact FinL-mlx decision model trained on financial scenarios.

Measures how LLM probability estimates shift from single new information with varying framing using δ-SNLI dataset.

2D CNN classifies hypnotic susceptibility from dense-array EEG during guided imagery on 62 participants.

OpenAI's Terminal X eval agent achieves 88.6% agreement with senior investor judgment using GPT-6.1, detecting 25% more mandate deviations.

RSI Control Arena testbed explores misaligned frontier models sabotaging successor training via recursive self-improvement.

Anytime-valid runtime monitor detects deterioration of learned barrier certificates in stochastic systems via trajectory auditing.

World2HOSI orchestrates multi-expert models to generate long-horizon human-object-scene interactions from 3DGS scenes.

Measures task readability sparsity (k*) across 19 LMs, finding base models allocate consistent feature counts for 9 core tasks.

Drill-down evaluation framework decomposes Text-to-SQL generation into sub-queries to trace hallucination patterns across complexity levels.

DeepSeek V4.1 Flash 1017, MiniMax M3.1 Flash, and Mercury 3 released; limited immediate adoption expected.

Mayor Zohran Mamdani said Monday night that New Yorkers must resist the narrative that an AI takeover is inevitable and demand regulations that protect the public and put power back in their hands.

CALIBER benchmark measures confidence calibration in reasoning models, reducing Expected Calibration Error on BigMathDigits.

Cohere discusses enterprise AI governance challenges including visibility, accountability, and control at scale.

OpenAI acquires Orchestra Bio to apply AI to life sciences R&D, targeting drug discovery and development acceleration.

Simon Willison reflects on how coding agents may disrupt traditional programmer employment by automating logic-puzzle work that defined the profession's appeal and compensation.

In August, Nscale and Anthropic announced a $45 billion compute deal. The two companies are reportedly discussing yet another deal of the same magnitude.