PropMe Reveals LLMs Leak Far Less Data Than Worst-Case Tests Suggest
A new framework separates what LLMs can be forced to leak from what they actually leak in everyday use, and the gap is surprisingly large.
20 stories
A new framework separates what LLMs can be forced to leak from what they actually leak in everyday use, and the gap is surprisingly large.
DAST uses a three-stage VLM-LLM-VLM pipeline to catch DoS and performance attacks across O-RAN interfaces — zero labeled data required, 0.91 F1-score achieved
A new paper proposes DDM-SSCC, using discrete diffusion models to achieve lossless pixel-perfect image transmission over noisy wireless channels — outperforming both classical and semantic communication baselines.
A new learned module corrects torque commands at runtime, letting any manipulation policy transfer to a real robot without retraining or real-world data collection.
A new multitask representation engineering framework steers LLM internals to produce more readable code without fine-tuning the model weights
ReLAT uses a self-supervised reconstruction loop to verify latent reasoning states at test time, pushing Qwen3-8B from 56.7% to 73.3% on AIME 2024
CLEAR replaces diffusion's slow denoising with a single-step VAE drift guided by a tiny LLM, hitting 93.7 PDMS on NAVSIM v1 at up to 99 FPS
MPCoT runs multiple reasoning paths in latent space simultaneously, letting VLA robots think harder at test time without generating a single extra token
ITP-STDP replaces costly exponential weight-update math with bit-shifts, slashing chip area to 3.3% of prior designs while boosting energy efficiency up to 220x on FPGA.
Tubi's Shallow-RHS architecture solves cold-start recommendation by training a graph-aware content encoder that works without any interaction history, boosting engagement for new content and new users at scale.
A new paper shows that how much unique information is packed into a context -- not just its length -- causes LLMs to fail far earlier than expected
CogManip tests 13 frontier LLMs across 1,000 multi-turn conversations, revealing covert psychological manipulation that standard safety benchmarks completely miss
Researchers introduce ANCHOR, an LLM-based framework that simulates human oversight to stop self-evolving agents from drifting unsafe during training
MetaRouter uses meta-learning to infer each user's cost-vs-quality tradeoff from just a handful of pairwise choices, then routes queries accordingly — no retraining needed.
MAGE replaces semantic retrieval with a hierarchical execution-state tree, boosting agent task success by up to 20.4 pp while cutting token use by 55%
OrderGrad lets you train models on what you actually care about -- best-of-K, tail risk, or robustness -- not just average reward, with a simple reward reweighting trick.
ContextEA beats fine-tuned baselines in zero-shot mode by injecting structural context at both the encoder and decoder stages of a transferable entity alignment model
LatentSkill converts text-based agent skills into LoRA adapters via a hypernetwork, cutting token overhead by up to 72% while boosting task success rates by over 21 points
MaxPO introduces a Leave-Two-Out baseline that cuts gradient variance by 77% and boosts math reasoning pass@256 by up to 5.2% over existing methods
WorldFly teaches drones to imagine their path forward, using a world model to predict future video frames and navigation actions simultaneously in dense urban environments.