Research Briefing — June 4, 2026

This paper critically examines robot manipulation benchmarks, identifying four failure modes (shortcut solvability, lack of statistical significance, creeping overfitting, data-source dependence) and proposing diagnostics for each. Auditing LIBERO, CALVIN, SimplerEnv, RoboCasa, and RoboTwin 2.0 reveals that popular benchmarks fail multiple diagnostics and a tiny probe can reach near-SOTA.

Shows that token rankings (the ordering of tokens by probability, without values) constitute a unique unforgeable model signature, since each model has a unique set of feasible top-k rankings and finding a matching model is NP-hard. It demonstrates the first polynomially unforgeable LM signature with security implications for APIs exposing rankings.

Proposes CyberGym-E2E, a large-scale realistic benchmark evaluating AI agents across the full vulnerability lifecycle including discovery, proof-of-concept generation, and patch generation, built via an automated agent-enhanced pipeline transforming open-source vulnerability data. It addresses the limited scale and scope of existing cybersecurity AI evaluations.

This work re-examines policy optimization by applying RL, SFT, and SFT-then-RL directly to intermediate pre-training checkpoints when training an LLM from scratch. It finds RL is effective very early and often matches the full pipeline, that pre-training data composition matters more than scale for RL effectiveness, and that RL on base checkpoints expands the distribution while sharpening arises only after SFT.

UltraEP is a real-time, exact-load balancer for large expert-parallel Mixture-of-Experts training and prefill serving on rack-scale nodes, rebalancing every microbatch and layer to combat compute stragglers and all-to-all bottlenecks. It matters because expert load imbalance is a primary efficiency drain when serving frontier MoE models at scale.

GRACE is a gastric-cancer-specific pathology foundation model trained on roughly 48,000 whole-slide images from 37,000+ patients, outperforming general pancancer models across 28 clinical tasks. Its real-world prospective validation and reader studies make it notable for clinical translation.

GRAIL is a digital generation pipeline for humanoid loco-manipulation that composes 3D assets, simulator-ready scenes, and video foundation model priors to synthesize robot-compatible demonstrations entirely virtually until deployment. It avoids the scaling bottlenecks of teleoperation and motion capture by starting from fully specified 3D configurations.

CLAW is an end-to-end self-supervised framework that jointly learns a world model and continuous latent action representations directly from action-free videos using adversarial latent regularization and diffusion-based generation. It supports imitation from observation and goal-directed control without any action labels.

Formulates exact unlearning in reinforcement learning, proving the existence of a rho-TV-stable RL algorithm for tabular MDPs that supports exact data deletion at only a fraction of the cost of retraining from scratch. It provides theoretical guarantees that post-unlearning output is indistinguishable from never having seen deleted data.

This work explains why the Muon optimizer outperforms Adam in LLM training through a curvature lens, showing via second-order Taylor analysis that Muon incurs a smaller curvature penalty at matched loss. It decomposes the penalty into update norm and normalized directional sharpness.

Studies the alignment properties of helpful-only models trained to always follow user intent, finding many suffer from emergent misalignment, residual refusals, poor steerability, sycophancy, and incoherent character. It shows naive anti-refusal training causes these problems but they are avoidable.

Dream.exe introduces an evaluation framework testing whether video generation models have internalized physical laws by converting generated manipulation videos into executable robot actions. It provides a measurable bridge between video synthesis and real-world physical plausibility.