The Archive

CATCH applies continuous adversarial training to reduce dishonesty in LLMs while preserving capabilities on knowledge-grounded tasks.

EliteReplay combines GRPO with quality-diversity search to preserve solution diversity in RL post-training rather than converging to single policies.

Scaffolding in coding agents improves performance on ARC-AGI-3 and Terminal-Bench even with stronger models; environment-agnostic design persists.

Rational synthesis framework for multi-agent systems with probabilistic temporal objectives and Nash equilibrium constraints on strategy profiles.

Active preference learning method shifts burden from humans to robots by identifying under-supervised task features and requesting targeted clarification.

LEASH pairs weak agents with frontier models using uncertainty signals for cheap escalation decisions, with oversight mechanisms for strong agent monitoring.

Librarian agents proactively curate and reorganize long document corpora during inference-time preparation to anticipate diverse queries.

X-RIC framework integrates agentic AI with RL and domain-specific language model for explainable RAN intelligent controller optimization.

Policy gradient RL for LLM post-training loses reward invariance under regularized importance sampling; proposes baseline centering fix.

Topology-aware loss functions for wind-field super-resolution favor different model architectures than pointwise metrics.

BitSparse losslessly compresses sparse ReLU activations during LLM training via custom GPU kernels, reducing memory footprint.

PRISMA algorithm evaluates K policies simultaneously in contextual bandits with sample complexity scaling in log(K).

Mandarin LLM exhibits overuse of corrective construction 'not X, but Y'; proposes decision tree to distinguish substantive vs. defective uses.

Anthropic grants METR independent evaluator access to Claude systems; exploring partnerships with Hugging Face and others for safety assessment.

Disagreement-based routing for weak-to-strong model cascades avoids costly ground-truth labels in agentic tasks.

LinguaGuardBench reveals agent safety guards trained on English fail to generalize across German, French, Turkish, Swahili, Amharic.

Three-level shortcut audit framework identifies spurious correlations in agent safety evaluation datasets (LinuxArena, TS-Bench).

SurvICL tabular foundation model trained on synthetic data outperforms Cox proportional-hazards model on survival analysis benchmarks.

Single-Increment Rejection Sampler (SIRS) estimates log acceptance probability for parallel masked diffusion with one query instead of per-token.

Training-free layer-wise consolidation fuses specialist models into one student using only forward passes, eliminating distillation's training cost.

OS-Helix is a trained multimodal agent unifying GUI interactions and code execution in a single action space for desktop automation.

Meta details hardware-model co-design principles for optimizing AI training on AMD GPUs and MTIA accelerators versus NVIDIA baseline.

Audio-vision cross-modal probing reveals vision encoders learn material damping properties via acoustic signal analysis.

SCOPED trains capability-specific experts then distills explicit skills into VLMs via reward-guided reflection to avoid gradient conflicts.

FutureLift refines multimodal motion forecasts by representing proposals as heterogeneous future graphs for autonomous driving.

GNN-based local refinement optimizes layout decomposition for directed self-assembly lithography patterning integration.

High-probability generalization bounds for randomized block coordinate descent in nonconvex machine learning optimization.

TRACE uses topology-aware prediction-adaptive objectives to enforce connectivity in thin-structure medical image segmentation.

FEB-30 benchmark separates memory correctness from retrieval/generation quality in agent evaluation across 30 fact-evolution trajectories.

There's a growing trend in the consumer tech world: gadgets designed to spy on everything and everyone around you, regardless of whether or not anyone else actually consented to it. Now, with the advent of "AI personal assistants" like Meta's Muse, you're being cheerfully asked not just to give up your own data, but also the information of your friends, family, the person you went on a Hinge date with five years ago - and virtually anyone else you've ever come in contact with. Since Muse was released last month, the agent has already been downloaded millions of times and hit No. 1 on Apple's ...