A new method exposes cheating in reinforcement learning by asking the model to argue with itself, catching contamination that likelihood-based detectors miss entirely.
Dual Adversarial Training replaces the unstable Langevin sampler behind Joint Energy-Based Models, scaling to ImageNet 256 while unifying robust classification, generation, and counterfactuals.
A new benchmark tests whether nine leading models can trace the hidden mechanisms between cause and effect, and the results expose reasoning shortcuts.
A dual-loop agent architecture separates turn-by-turn dialogue tactics from cross-session strategy, letting an LLM run coherent, adaptive counseling arcs.
A new prequential evaluation protocol reveals that standard train-test carbon audits systematically misestimate the real long-term emissions of ML models in production.
A new benchmark tests 10 routing methods across 17 vision-language models, revealing routers can slash costs by 5x while nearly matching the strongest single model.
A production-grade benchmark of speculative decoding in vLLM reveals that verification cost dominates, batch size erodes gains, and the gap to theoretical maximum is huge.
A new framework from Georgia Tech dynamically prunes 3D convolutional networks per input, cutting energy use on edge devices by up to 1.47x with no accuracy loss.
NYU and partners release an open-source repository of modular Colab-based teaching units for LLM-assisted RTL design, verification, and hardware security.
A five-agent framework that decomposes long-video reasoning like human cognition beats prior systems by 5 points on LVBench while cutting inference cost.
An offline reinforcement learning framework called PIER cuts shipping CO2 by 10% and eliminates catastrophic fuel waste without needing weather forecasts.
Peking University researchers built a data-driven badminton simulator that skips physics engines, uses two probabilistic models, and trains PPO agents to 98% win rates.
A new ICML 2026 paper reframes permutation generation by diffusing continuous soft ranks instead of shuffling cards, with sharp gains on long sorting and combinatorial tasks.
A new alignment method called PreSafe locks in a refusal decision before chain-of-thought starts, cutting jailbreak success rates to near zero without hurting reasoning.
Google DeepMind shows an online RLHF algorithm with uncertainty-guided exploration hits offline RLHF's quality using 10x less human feedback, projected to reach 1,000x.
A new diffusion-based codec handles any bitrate below 0.1 bpp with a single model by combining arbitrary-scale downsampling and degradation-aware reconstruction.
A new virtual cell foundation model recasts perturbation prediction as latent transport, beating STATE by 12% on Tahoe-100M with 12.5x faster training.
A production security architecture for nine autonomous LLM agents in healthcare combines kernel sandboxing, credential proxies, egress allowlists, and prompt integrity controls.