Research Briefing — June 9, 2026

This work formalizes the audit gap between behavioral safety evaluations and representation-level robustness, constructing dissociated models that appear safe but remain vulnerable in latent space. It introduces intervention-based evaluation via harmful fine-tuning and latent perturbations.

This paper introduces an information-theoretic definition of open-ended learning based on the bit-equivalent, the information required to attain each reward level, defining open-endedness as linear growth in bit-equivalent. It shows classical bandits are not open-ended, constructs one that is, and provides an algorithm achieving open-ended learning.

SWE-Marathon is a benchmark of 20 ultra-long-horizon software engineering tasks, each with executable environments, reference solutions, and multi-layer verification, where logged agent attempts average over 27 million tokens. It targets measuring agent planning, long-context, and memory capabilities far beyond typical short-task benchmarks.

This paper provides a theoretical framework explaining empirical data mixing scaling laws, extending Kaplan/Chinchilla perspectives to multi-domain settings. It identifies capacity competition and skill overlap as key factors governing domain losses under different data mixtures.

This work demonstrates an indirect prompt-injection attack where payloads caught in plain text evade detectors when transported as structured float parameters and reconstructed as fragmented telemetry. Across 14,400 trials on three commercial LLM APIs, the float-array carrier preserves 94.3% leakage attack success under strong dual-layer text-classifier defenses.

This paper presents the first large-scale multilingual study of LLM sycophancy, benchmarking six instruction-tuned models across 1.1 million instances spanning 38 languages and 33 topics. It finds a consistent resource-tier effect where sycophancy spikes in low-resource and zero-shot language settings, topic-agnostically.

This work proposes scaling participation, a paradigm where modular AI systems are built bottom-up from many small stakeholder-trained models that collaborate compositionally, reportedly outperforming monolithic LLMs by up to 15.4% across 15 tasks. It challenges centralized model development with a participatory alternative.

Ego-Pi investigates fine-tuning VLA models (built on pi_0.5) across human and humanoid embodiments with dexterous hands, showing egocentric human data enables robots to learn new task semantics and compose skills without corresponding robot data. It addresses robotics data scarcity by leveraging easily-collected human demonstrations.

This paper proves identifiability for causal discovery in location-scale noise models with hidden variables, where causes modulate both mean and variance of effects. It establishes the first identifiability result for causally insufficient models beyond additive noise and provides a sound two-stage algorithm.

Sparrow accelerates long-context RL with verifiable rewards by using sparse attention during rollout generation, navigating the stability-efficiency tradeoff via analysis of sparse-to-dense actor-policy mismatch. It observes that most sparse tokens align with dense even under aggressive sparsity, enabling selective sparsification.

RLDT is an online RL algorithm for fine-tuning flow-matching policies by viewing policy improvement as transport of action densities toward high-reward regions, constructing a transport field from a maximum-entropy objective using Stein Variational Gradient Descent. It avoids biased gradients and preserves multimodal modeling capacity unlike distillation approaches.