CompactionRL: Reinforcement Learning with Context Compaction for Long-Horizon Agents
CompactionRL: RL training for long-horizon agents with context compaction, jointly optimizing task and summary generation.
Search the full wire by company, model, lab, or keyword. Every story we have ever aggregated.
CompactionRL: RL training for long-horizon agents with context compaction, jointly optimizing task and summary generation.
Cortex: hierarchical embodied agent framework bridging VLM planning to VLA execution for long-horizon robotic manipulation.
FORE: fitted fixed-point method for occupancy-ratio estimation in offline RL via adjoint Bellman recursion.
GaP combines interpretable robot programming with model-free policies for variational automation tasks, bridging TAMP and agentic systems for industrial reliability.
SPEARBench evaluates naturalness in streaming speech-to-speech models across timing, prosody, and turn-taking using controlled dialogue prompts.
REDDIT addresses timestamp drift in autoregressive ASR systems across non-speech spans using replay-based distribution editing without performance degradation.
SovereignPA-Bench evaluates user-owned personal agents on privacy, consent, and user sovereignty across evolving intent and platform mediation.
Graph Sparse Sampling reduces exponential sampling complexity in continuous MDP planning by sharing futures across multiple agents and lookahead depths.
Neuron selector attribution methods are audited for causal importance in LLM pruning and safety editing via paired one-shot zeroing interventions.
HeRo enables selective disclosure watermarking in LLMs, allowing verification of watermark subsets without exposing full embedded metadata.
First multiplayer world model for dynamic environments attributes scene changes to multiple agents' actions, trained on 10K hours of Rocket League gameplay.
OptiAgent is a multi-agent framework that converts natural language OR problems to mathematical formulations and executable code via iterative self-correction.
TREK uses distillation for exploration support expansion in GRPO, enabling reasoning on hard prompts via black-box or white-box teacher trajectories.
Siamese networks require threshold selection for embedding-space verification; paper proposes unsupervised threshold definition without labeled data.
PSA: A change to Google's privacy settings let it train its AI on more of your data. Here's how to opt out.
Diffusion autoencoders exhibit distinct optimization regimes during training; study links reconstruction vs. disentanglement trajectories to early dynamics.
Medical imaging CNN struggles with aneurysm-bifurcation confusion; topology-aware framework using Smooth Euler Characteristic Transform reduces false positives.
Linear probes decode remaining output length from LLM hidden states across 7-8B open-weight models, revealing internal response-length estimation.
M3Bench evaluates model editing for medical vision-language models under clinical variation; tests reliability, precision, generalization post-deployment.
Quantum anomaly detection via spectral methods; computes quantum state deviation from normal dataset without explicit eigenvector recovery.
Pathway Activity Autoencoders integrate multi-omics data with biological priors for cancer risk stratification; balances interpretability and capacity.
Anthropic accused of spying on users; engineer says “experiment” is over.
Subspace-constrained LoRA adaptation on Flan-T5 resists poisoning; restricts updates to 128 coordinates in shared adapter subspace, blocking malicious fine-tuning.
MetaSkill-Evolve enables recursive self-improvement of LLM agents; agents rewrite and evolve their own improvement procedures from execution traces.
PM2.5 super-resolution downscaling over Europe using station-guided pseudo-supervision; bridges coarse atmospheric model fields to local ground observations.
Wavelet scattering transform applied to EEG for schizophrenia biomarker discovery; medical imaging application, not frontier AI.
Quantum hardware fingerprinting via noisy output distributions raises privacy concerns for cloud quantum services.
Neural Controlled Differential Equations for continuous-time modeling of irregular time series; incremental method advance.
Defense mechanism for web agents against prompt injection via content masking with formal security guarantees.
ProPS framework generates speaker embedding distributions conditioned on natural language descriptions; specialized speech task.