DDOR Cuts AI Overrefusals by 86% Without Touching Model Weights
DDOR borrows a classic software debugging algorithm to automatically find and fix the exact phrases that make LLMs refuse perfectly safe prompts
20 stories
DDOR borrows a classic software debugging algorithm to automatically find and fix the exact phrases that make LLMs refuse perfectly safe prompts
PGPO borrows viscous-fluid physics to dynamically scale gradient steps in self-distillation training, fixing the instability that causes SDPO to collapse

A new two-stage pruning framework fixes the core flaw in attention-based token selection, boosting both spatial coverage and task-relevant precision in VLMs.
A new method adds adversarial robustness to any pre-trained RL agent at inference time, no retraining required, using model-predictive control and worst-case rollouts.

A new paper reveals that when you refresh the teacher in self-distillation training matters more than how old it is, and proposes a smarter gating mechanism that eliminates training collapse entirely.
A new benchmark reveals that AI agents learn more from peer histories than solo practice — but only when they can abstract, not just copy.

A new formal framework treats AI agent delegation as a contractual primitive, solving the recursive permission chain problem that OAuth and IAM systems can't handle.
A new ranked binary classifier approach lets APT malware attribution systems say 'I don't know' — rejecting 94% of unknown-group samples while keeping 92% precision on what they do classify.
Researchers from CNRS, Université Paris Cité, IonQ, and Quantum Signals cut QNN training circuit evaluations from O(n²) to O(log n), enabling gradient-based training on real quantum hardware at 16 qubits.
StepFinder replaces slow LLM-as-judge debugging with a lightweight temporal model that cuts inference time by 79% while outperforming GPT-4-class methods on multi-agent failure diagnosis.
NeuroArmor uses prompt-specific safe variants to check hidden-state consistency at runtime, cutting jailbreak success rates from 41% to under 2% while reducing over-blocking of legitimate requests.
A new paper reveals that weaker teacher models can produce better-trained agents, and that how a trajectory is structured matters more than who generated it.
HARC fixes a critical but overlooked failure mode in MoE model merging, using second-order math to recalibrate broken routers without any retraining
CP-Agent pairs a CLIP-style image-metadata encoder with an agentic LLM to explain why a drug changed cell shape -- hitting F1 of 0.896 on mechanism-of-action classification
PRISM uses a dual-stream Mixture-of-Experts architecture to merge CLIP, DINOv2, and SAM into one model without the gradient conflicts that plague standard distillation

Researchers from NYU Abu Dhabi's eBRAIN Lab introduce PrimeSVT, an automated structured pruning framework that cuts Spiking Vision Transformer memory by 26.68% while keeping accuracy within 3% of the original model.
FORGE links exploit generation, vulnerability prioritization, and detection rule creation in one automated pipeline, hitting 67.8% success across 603 CVEs at $1.50 each
GTR replaces PPO's rigid trust region with a Gaussian-shaped constraint that lets agents break out of stale behaviors in continual and non-stationary RL settings
A new paper exposes a critical blind spot in AI model theft defenses: they all assume a single attacker, and coordinated groups can exploit that to steal models undetected.
FlowGuard uses generative flow models to catch model-stealing attacks on energy grid AI, surviving distributed Sybil attacks where existing defenses collapse to 0% detection